Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Design of cost-efficient multi-sensor collaboration in wireless sensor networks
(USC Thesis Other)
Design of cost-efficient multi-sensor collaboration in wireless sensor networks
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
DESIGN OF COST-EFFICIENT MULTI-SENSOR COLLABORATION IN WIRELESS SENSOR NETWORKS by Chengjie Zhang A Dissertation Presented to the FACULTY OF THE USC GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (COMPUTER SCIENCE) December 2013 Copyright 2013 Chengjie Zhang Dedication Dedicated to my wife, Yifei, whose unconditional support has accom- panied me through this long and lonely journey, and my parents, who gave me the freedom to choose my path in life. ii Acknowledgments I am grateful to many, many people who helped me, more or less, in my long PhD journey. To all those who worked or even just chatted with me on my research, thank you! First of all, I would like to thank my adviser, John Heidemann for his guidance in my PhD endeavor. His mentorship helped me develop essential research skills, which will surely be precious to my future work and research. His challenges honedmycriticalthinkingability, andhetaughtmehowtobeanethicalscientist. His enthusiasm for research and life encouraged me to never give up on my PhD endeavor. Further, his passionate working style ignited my passion for my own research. In addition, he showed me how to balance between work and family. I would like to give special thanks to Prof. Ramesh Govindan, Prof. Bhaskar Krishnamachari, Prof. Clifford Neuman, Dr. Andrei Popa and Prof. Cauligi S. Raghavendra for their service on my qualifying exam and dissertation committee. They gave me valuable feedback and challenges, making me confident to be an independent researcher. I had wonderful collaboration with a few colleagues in ISI, without whom a great part of my research progress would be impossible. Dr. Unkyu Park and Dr. Affan Syed are the first two friends I made in ISI. Their help was consistent across my entire PhD endeavor, and I wish at some point in life, we may collaborate iii again. Dr. Wei Ye helped me ramp up quickly at the beginning of my PhD study. Andrew Goodney, with his background in Electrical Engineering, gave me countless tips and insights to my sensornet research. Prof. Young Cho shared his inspiring academic experience with me. I also have built rapport with many fellow students in ISI, including Xue Cai andLinQuan. Theymademyroutineresearchlifecolorful. Coffee-timewiththem sparked some research ideas, mutually. AchunkofmyPhDresearchwasfundedbyChevron. Duringourcollaboration withChevron, CharlieWebbandGregLaframboiseprovidedmewiththeirselfless help in my (serial) field tests. In addition, thanks for their passion in my work. Last but not the least, I am heartily thankful to my family in China, for their support across the ocean. They never lose faith in me. iv Table of Contents Dedication ii Acknowledgments iii List of Tables ix List of Figures x List of Algorithms xiii Abstract xiv Chapter 1: Introduction 1 1.1 Thesis Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.2 Proving the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 Chapter 2: Multi-Sensor Vehicle Classification 12 2.1 Problem Statement of Multi-Sensor Vehicle Classification . . . . . . 12 2.1.1 Relation to Thesis . . . . . . . . . . . . . . . . . . . . . . . 13 2.1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.2 Matching Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.2.1 Problem Formalization . . . . . . . . . . . . . . . . . . . . . 16 2.2.2 Numbering Based . . . . . . . . . . . . . . . . . . . . . . . . 16 2.2.3 Time-Stamp Based . . . . . . . . . . . . . . . . . . . . . . . 18 2.2.4 Full Raw Signature Comparison . . . . . . . . . . . . . . . . 26 2.2.5 Algorithm Discussions . . . . . . . . . . . . . . . . . . . . . 29 2.3 Multi-Sensor Vehicle Classification Evaluation . . . . . . . . . . . . 30 2.3.1 Data Collection Experiment . . . . . . . . . . . . . . . . . . 30 2.3.2 Matching Algorithm Correctness . . . . . . . . . . . . . . . 32 2.3.3 Parameter Sensitivity . . . . . . . . . . . . . . . . . . . . . . 35 2.3.4 Impact of Matching on Classification . . . . . . . . . . . . . 39 2.4 Conclusions on Multi-Sensor Vehicle Classification . . . . . . . . . . 48 v Chapter 3: Steam-Choke Blockage Detection 49 3.1 Motivations of Choke Blockage Detection . . . . . . . . . . . . . . . 49 3.1.1 Sensing needs in an Oilfield with Secondary Production . . . 50 3.1.2 Relation to Thesis . . . . . . . . . . . . . . . . . . . . . . . 52 3.1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 53 3.2 Target Problem: Blockage at the Steam Injection Choke . . . . . . 54 3.3 System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 3.4 Steam-Power: Harvesting Thermal Energy . . . . . . . . . . . . . . 58 3.4.1 The Opportunity . . . . . . . . . . . . . . . . . . . . . . . . 58 3.4.2 Thermo-electric harvester Design . . . . . . . . . . . . . . . 60 3.4.3 Power Conditioning the TEG Output . . . . . . . . . . . . . 61 3.4.4 Mounting Design . . . . . . . . . . . . . . . . . . . . . . . . 62 3.5 Non-Invasive Sensing of Pipeline Blockages . . . . . . . . . . . . . . 63 3.5.1 Background: pipeline physics . . . . . . . . . . . . . . . . . 64 3.5.2 Design of the Base Algorithm . . . . . . . . . . . . . . . . . 66 3.5.3 Avoiding False Positives . . . . . . . . . . . . . . . . . . . . 69 3.5.4 Tuning for different environments . . . . . . . . . . . . . . . 72 3.6 Long-term energy harvesting and consumption evaluation . . . . . . 73 3.6.1 Energy Production . . . . . . . . . . . . . . . . . . . . . . . 74 3.6.2 Energy Consumption of Sensing . . . . . . . . . . . . . . . . 75 3.6.3 Batteryless operation? . . . . . . . . . . . . . . . . . . . . . 75 3.7 Steam-Choke Blockage Detection Evaluation . . . . . . . . . . . . . 78 3.7.1 Does Our Detection Algorithm Work? . . . . . . . . . . . . 78 3.7.2 Evaluating Avoidance of False Positives . . . . . . . . . . . . 82 3.7.3 Parameter sensitivity of the basic sensing algorithm . . . . . 83 3.7.4 Parameter sensitivity of the extended algorithm . . . . . . . 86 3.7.5 Generalizing to Water Pipelines . . . . . . . . . . . . . . . . 88 3.8 System evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 3.8.1 Sensor and system calibration . . . . . . . . . . . . . . . . . 92 3.8.2 System cost . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 3.8.3 System robustness . . . . . . . . . . . . . . . . . . . . . . . 98 3.9 Conclusions on Choke Blockage Detection . . . . . . . . . . . . . . 100 Chapter 4: Cold Oil Blockage Detection 102 4.1 Motivation for Cold-Oil Blockage Detection . . . . . . . . . . . . . 103 4.1.1 Relation to Thesis . . . . . . . . . . . . . . . . . . . . . . . 105 4.1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 106 4.2 Cold-Oil Blockage Problem Overview . . . . . . . . . . . . . . . . . 107 4.3 Design of Cold-Oil Blockage Detection Algorithm . . . . . . . . . . 110 4.3.1 Overview of approach . . . . . . . . . . . . . . . . . . . . . . 111 4.3.2 Temperature Sensing for Flow Presence . . . . . . . . . . . . 113 4.3.3 Acoustic Sensing to Avoid False Alarms . . . . . . . . . . . . 115 vi 4.3.4 Sensor Fusion for Blockage Detection . . . . . . . . . . . . . 116 4.4 System Implementation. . . . . . . . . . . . . . . . . . . . . . . . . 117 4.4.1 System Hardware . . . . . . . . . . . . . . . . . . . . . . . . 117 4.4.2 Hierarchical Sampling and Aggregation in Acoustic Mote . . 120 4.4.3 Maximizing Acoustic Gain . . . . . . . . . . . . . . . . . . . 121 4.5 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 4.5.1 Calibrating Individual Sensors . . . . . . . . . . . . . . . . . 123 4.5.2 Field Experiment Approach . . . . . . . . . . . . . . . . . . 126 4.5.3 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . 129 4.5.4 Accuracy of Flow Presence Detection . . . . . . . . . . . . . 131 4.5.5 Auto-Configuration of Temperature Measurement . . . . . . 134 4.5.6 Detecting Pumpjack Operation . . . . . . . . . . . . . . . . 136 4.5.7 Blockage Detection Accuracy . . . . . . . . . . . . . . . . . 139 4.5.8 Robustness of Blockage Detection . . . . . . . . . . . . . . . 142 4.5.9 In-lab Near-Full Blockage Detection . . . . . . . . . . . . . . 148 4.5.10 Evaluation Summary and Algorithm Generalization . . . . . 154 4.6 Preliminary Experiments . . . . . . . . . . . . . . . . . . . . . . . . 154 4.6.1 Short-term Initial Data Collection . . . . . . . . . . . . . . . 155 4.6.2 Medium-term Data Collection . . . . . . . . . . . . . . . . . 155 4.6.3 Acoustic Sensor and Thermal Insulation Tests . . . . . . . . 157 4.6.4 First Prototype Test . . . . . . . . . . . . . . . . . . . . . . 160 4.6.5 Second Prototype Test . . . . . . . . . . . . . . . . . . . . . 164 4.6.6 Preliminary Experiment Summary . . . . . . . . . . . . . . 165 4.7 Conclusions on Cold-Oil Blockage Detection . . . . . . . . . . . . . 167 Chapter 5: Related Work 169 5.1 Target Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 5.1.1 Unconstrained Environment . . . . . . . . . . . . . . . . . . 170 5.1.2 Constrained Environment . . . . . . . . . . . . . . . . . . . 171 5.2 Change-Point Detection Algorithms . . . . . . . . . . . . . . . . . . 173 5.3 Pipeline Monitoring Systems . . . . . . . . . . . . . . . . . . . . . . 174 5.4 Oil Line Blockage Detection Applications . . . . . . . . . . . . . . . 176 5.5 Multi-Modal Sensing Applications . . . . . . . . . . . . . . . . . . . 177 5.5.1 Multi-Modal Sensing in Industrial Monitoring . . . . . . . . 177 5.5.2 Multi-Modal Sensing in Academic Projects . . . . . . . . . . 178 5.6 Energy Harvesting Systems . . . . . . . . . . . . . . . . . . . . . . 180 Chapter 6: Future Work and Conclusions 183 6.1 Short-Term Future Directions . . . . . . . . . . . . . . . . . . . . . 183 6.2 Long-Term Future Directions . . . . . . . . . . . . . . . . . . . . . 185 6.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 vii Bibliography 190 viii List of Tables 2.1 Category-level matching results by full raw signature comparison . . 28 2.2 Event types for matching . . . . . . . . . . . . . . . . . . . . . . . . 34 2.3 Matching correctness of algorithms . . . . . . . . . . . . . . . . . . 34 2.4 Classification accuracy, multi vs. single . . . . . . . . . . . . . . . . 40 2.5 Reported once and correctly classified once . . . . . . . . . . . . . . 42 2.6 Multi-sensor classification accuracy . . . . . . . . . . . . . . . . . . 43 2.7 The correlation between matching and classification . . . . . . . . . 45 2.8 Matching vs. classification in STW (dissection) . . . . . . . . . . . 46 2.9 Matching vs. classification in STW . . . . . . . . . . . . . . . . . . 47 3.1 Notations in blockage detection algorithms . . . . . . . . . . . . . . 69 3.2 Energy buffering test at TEG ∆ HC = 83.1℃ . . . . . . . . . . . . . 77 3.3 Pipe temperature variation along time . . . . . . . . . . . . . . . . 99 4.1 Mote/ground-truth correlation coefficients on temperature . . . . . 124 4.2 Experiment schedule and scheme . . . . . . . . . . . . . . . . . . . 129 4.3 The accuracies of blockage detection . . . . . . . . . . . . . . . . . 140 4.4 In-lab experiment schedule . . . . . . . . . . . . . . . . . . . . . . . 151 4.5 First prototype test schedule . . . . . . . . . . . . . . . . . . . . . . 161 4.6 Second prototype test schedule . . . . . . . . . . . . . . . . . . . . 164 ix List of Figures 2.1 Avalanche problem in NN but solved by NwR. . . . . . . . . . . . . 18 2.2 Avalanche problem in STW. . . . . . . . . . . . . . . . . . . . . . . 22 2.3 Four pre-scaled raw signature pattern comparisons. . . . . . . . . . 25 2.4 Real deployment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.5 Travel time distribution of the 65 Normal vehicles. . . . . . . . . . 36 2.6 Accuracy % of STW. . . . . . . . . . . . . . . . . . . . . . . . . . . 37 2.7 Best accuracy % of DTW with different Shift Values. . . . . . . . . 38 2.8 Best accuracy % of WETW with different Wheelbase Windows. . . 38 3.1 Steam injection (right) and oil production (left) in oilfield. . . . . . 51 3.2 March 2010 field deployment of our sensing system. . . . . . . . . . 56 3.3 Mote system hardware. . . . . . . . . . . . . . . . . . . . . . . . . . 57 3.4 Mounting design for TEG. . . . . . . . . . . . . . . . . . . . . . . . 62 3.5 Up- and downstream temperatures and ∆ ud in controlled blockages. 68 3.6 A cause of false positives and our solution. . . . . . . . . . . . . . . 71 3.7 Power measurement of the TEG. . . . . . . . . . . . . . . . . . . . 73 3.8 Instantaneous load can cause failure. . . . . . . . . . . . . . . . . . 76 3.9 Algorithm results on the thermocouple-pair straddling the valve. . . 80 3.10 Extended algorithm result on the sensor pair straddling the choke. . 83 3.11 Accuracy and false alarm at different th block (α l = 1 / 16 ). . . . . . . 84 x 3.12 δ fluctuation influenced by different α l . . . . . . . . . . . . . . . . . 85 3.13 The extended algorithm with a varying th block. . . . . . . . . . . . 86 3.14 the extended algorithm with a varying th maint. . . . . . . . . . . 88 3.15 In-lab, water-based pipeline prototype. . . . . . . . . . . . . . . . . 89 3.16 Blockage detection in a water pipeline. . . . . . . . . . . . . . . . . 90 3.17 Amplifier board repeatability test result. . . . . . . . . . . . . . . . 94 3.18 Temperature measured by mote and CR1k. . . . . . . . . . . . . . . 95 3.19 Invasive flow sensing with solar panel for power. . . . . . . . . . . . 96 3.20 Aggregation on all three downstream temperatures. . . . . . . . . . 100 3.21 Mote radio packet loss distribution. . . . . . . . . . . . . . . . . . . 100 4.1 Seasonality analysis shows winter production loss. . . . . . . . . . . 108 4.2 A diagram for cold-oil blockage problem statement. . . . . . . . . . 110 4.3 Our temperature and acoustic sensor hardware and deployment. . . 119 4.4 Temperature measured by mote and USB data logger at T 2 d . . . . . 124 4.5 Acoustic measured by mote and PC microphone.. . . . . . . . . . . 126 4.6 November 2012 field deployment. . . . . . . . . . . . . . . . . . . . 128 4.7 Flow presence detection results. . . . . . . . . . . . . . . . . . . . . 133 4.8 Twobox-plotsforthedifferenceintemperatureamongthreelocations.135 4.9 A box-plot for the fluctuation of temperature when flow is normal. . 136 4.10 Acoustic pumpjack status detection result by mote. . . . . . . . . . 138 4.11 Acoustic pumpjack status detection result by PC. . . . . . . . . . 139 4.12 Blockage detection results. . . . . . . . . . . . . . . . . . . . . . . . 143 4.13 A monotonically decreasing example for robustness analysis. . . . . 146 4.14 A stable with fluctuation example for robustness analysis.. . . . . . 147 4.15 A monotonically increasing example for robustness analysis. . . . . 148 4.16 In-lab, near-full blockage test on water. . . . . . . . . . . . . . . . . 149 xi 4.17 Temperature flow presence detection in water pipeline. . . . . . . . 152 4.18 Acoustic pump status detection in water pipeline. . . . . . . . . . . 153 4.19 Lavalier Microphone mounted on pipe without direct contact. . . . 156 4.20 Temperature data from our second data collection. . . . . . . . . . 157 4.21 New lavalier microphone mounting design. . . . . . . . . . . . . . . 158 4.22 FFT results on three pump cycles. . . . . . . . . . . . . . . . . . . 159 4.23 Temperature data shows pipe skin temperature fluctuation. . . . . . 160 4.24 Mote microphone mounting without decoupling. . . . . . . . . . . . 161 4.25 Comparing mote and USB data logger temperature data. . . . . . . 163 4.26 Basic statistics of acoustic data. . . . . . . . . . . . . . . . . . . . . 163 4.27 Mote radio packet loss in prototype test. . . . . . . . . . . . . . . . 164 4.28 Temperature data at mote after circulation valve. . . . . . . . . . . 165 4.29 Acoustic data from second prototype test. . . . . . . . . . . . . . . 166 xii List of Algorithms 2.1 Numbering with Resynchronization algorithm . . . . . . . . . . . . . 19 2.2 Static (Dynamic) Time Window algorithm . . . . . . . . . . . . . . 20 2.3 Wheelbase Enhanced Time Window algorithm . . . . . . . . . . . . 24 3.1 Blockage detection algorithm . . . . . . . . . . . . . . . . . . . . . 70 xiii Abstract Recent years have witnessed an increasing demand for monitoring tasks in indus- trial applications. Although effective in some applications, sensornet uptake has been slow in this emerging area, as shown by slow sensornet deployment in indus- try. Industries prefer well-tested but often expensive or complex sensing solutions because they provide known levels of accuracy, but this choice makes many prob- lemsuneconomic. Thesensornetresearchcommunityhasmadesignificantprogress towards real-world applications with some pilot deployments. While sensornet research is promising, current prototypes are often not cost-effective or still in the early stages. Industry remains slow to adopt these approaches, because they are not seen to address concerns about accuracy for industry-relevant problems. In this thesis, we propose Multi-Sensor Collaboration to achieve low-cost yet accurate event detection and enable sensornet usage in cost-sensitive industrial applications. Wefindthatfouradvantagesofcollaborationimprovecost-effectiveness. First, collaborative sensing can improve accuracy by suppressing false alarms with redundant or heterogeneous sensors. Second, collaborative sensing reduces capi- tal cost by enabling low-cost sensors to be accurate enough to reach actionable results. Third, collaboration can reduce deployment cost, because it enables non- invasivesensingtoprovidesufficientaccuracywithmuchlessexpensiveinstallation. Finally, auto-tuning is important to reduce deployment costs, and collaborative xiv sensing can assist auto-tuning by allowing sensors in different modalities to tune each other with their unique information. The thesis of this proposed dissertation is that Multi-sensor collaboration enables sensor networks to accomplish real-world event detection tasks that are impractical for single-sensor systems. To frame the application domain, we categorize sensing applications by their two orthogonal properties— collaboration scheme and sensing modality. Collaboration scheme, namely the relationship between sensors, is usually in two distinctive forms—competitive and complementary. Competitive collaboration means redundantly suppressing less accurate sensors’ results; complementary collaboration combines all relevant sen- sors’ partial results to form the final result. Modality means the type of the raw input of a sensor. To explore the application domain, we study both collabora- tions with either modality choice in three example applications. First, we eval- uate signature matching in a vehicle classification context to study single-modal competitive collaboration. Second, we present a design of steam-choke blockage- detection system to study single-modal complementary collaboration. Finally, we implementandevaluateadetectionsystemforoil-retrieval-lineblockagetoextend our complementary type study to multi-modality. We prove sensor collaboration can achieve cost-efficiency in the forgoing three specific applications. These appli- cations are useful and allow deployment of sensing where not economically viable before. Further, the commonality between these applications and a large group of relevant applications strongly suggest that collaborative sensing help a larger application domain. xv Chapter 1 Introduction Recent years have witnessed an increasing demand for monitoring tasks in indus- trial applications. Although effective in some applications, sensornet uptake has been slow in this emerging area, as shown by slow sensornet deployment in indus- try. Industries prefer well-tested but often expensive or complex sensing solu- tions because they provide known levels of accuracy, but this choice makes many problems uneconomic [DS99,HTTO02]. The sensornet research community has made significant progress towards real-world applications with some pilot deploy- ment [SNMT07,ZTLZ06,YSM + 06,KPC + 07]. Sensornet applications have many merits, including low cost, automated, ubiquitous and portable. By forming a network, multiple sensors located at different places can communicate and col- laborate. While sensornet research is promising, current prototypes are often not cost-effective or still in the early stages. Industry remains slow to adopt these approaches, because they are not seen to address concerns about accuracy for industry-relevant problems. Event detection and data collection are the two major classes of applications that are studied by the wireless sensor network community. For example, typical eventdetectionapplicationsincludewildlifemonitoring[TGC + 07],structuraldam- agedetection[CJG05],humanactivitymonitoring[AMM + 08],machinewear-down prediction[ZTLZ06]andmilitaryshooterlocalization[SLV11]. Ontheotherhand, 1 data collection applications are good for large area temperature profile establish- ment [MCP + 02], 3D tomography [YSM + 06], remote meteorological and ecological data [BB02,ALE13], camera imaging [RBI + 05], and agricultural data [RSS + 09]. Ofthesetwoclasses,wefocusoneventdetection,becauseweseeitsthreeunique properties distinguish itself from data collection. First, because of the uncertainty of event occurrence, sensors have to be in passive but vigilant state as often as possible in order to capture bursty event signals [DGA + 05]. On the other hand, sensors solely collecting data do not have to be always on. Users may specify the dutycycleaccordingtotheirapplications’temporal(orspatial)resolutionrequire- ment and power supply constraint. Second, in event detection, it is most likely unnecessary to report all raw data back to the backend system. Usually we only need to identify signal presence under certain criteria. Therefore the problem of eventdetectionisfundamentallyafeaturedetectionproblem, anditprovideslarge opportunity for in-node or in-network processing to reduce data transmission cost. Third,theperformancemetricsforeventdetectionaredifferentfrom,orevenmore complex than data collection. Many researchers evaluate in their detection appli- cationcorrectdetectionprobabilities, falsealarmrate, true/falsepositive/negative rate, (mis-)classificationaccuracy, detectionperiodoreventpredictionaccuracyas a derivative of detection [Coi98a,PLH + 11,MVW08]. In this thesis, we study how to leverage multi-sensor collaboration to achieve low-costyetaccurateevent detection (adetailedthesisstatementisinSection1.1). We find that four advantages of collaboration improve cost-effectiveness. First, collaborative sensing can improve accuracy by suppressing false alarms with redundant or heterogeneous sensors. Second, collaborative sensing reduces cap- ital cost by enabling low-cost sensors to be accurate enough to reach actionable 2 results. Third, collaboration can reduce deployment cost, because it enables non- invasivesensingtoprovidesufficientaccuracywithmuchlessexpensiveinstallation. Finally, auto-tuning is important to reduce deployment costs, and collaborative sensing can assist auto-tuning by allowing sensors in different modalities to tune each other with their unique information. To frame the application domain, we categorize sensing applications by their twoorthogonalproperties—collaborationschemeandthenumberofsensingmodal- ities. Collaboration scheme, the relationship between sensors, is usually in two distinctive forms—competitive and complementary. Sensing modality is the type of the raw input of a sensor. Sensing can be single- or multi-modal, depending on the number of input types (modalities). As one of the two collaboration schemes, competitive collaboration means each sensor has its own separate detection and interpretation of the object or the phenomenon and the end result is chosen among them under certain metrics, including signal quality, sensor reliability, timestamp (Chapter 2). One example is information-driven sensor querying in vehicle tracking [ZSR02]. Zhao et al. optimize sensor selection based on sensors’ communication cost and information usefulness, because not all the sensors in the field should be turned on to do track- ing at the same time. Another example is that Gupta et al. fuse both microwave and eddy current image and choose the better one to evaluate corrosion under aircraft paint and in lap joints [GGK + 07]. A third example is Wang et al. com- petitively using acoustic and seismic data in different multi-modal classifiers for military vehicle classification [WQI02]. 3 Complementary collaboration is the other type we study (Chapter 3). In this type of collaboration, the detection of each sensor only forms partial interpreta- tion of the object or the phenomenon. Therefore, we have to combine all of them to obtain a meaningful result. Many sensornet researchers have exploited com- plementary collaboration in their own projects. Kim et al. collect readings on all sensors at every branch in household water pipeline network to decide personal water usage [KSC + 08]. Zhu uses the differential between upstream and down- stream temperature on branch pipe to detect blockage in pulverized coal injection system [Zhu05], as another example. Sensing modality is the other property than collaboration scheme we used to categorize event detection applications. Applications with single-modal sens- ing only utilize one type of raw input [TGC + 07,KSC + 08,ZSR02]. Trifa et al. use acoustic sensing to monitor wild yellow-bellied marmots [TGC + 07]. Kim et al. deploy vibration sensors on household water pipe to monitor personal water usage [KSC + 08]. Likewise, multi-modal sensing (Chapter 4) takes plural raw input types [ZTLZ06, SRT + 08, SB97, AMM + 08]. Zeng et al. use vibration, force and acoustic emission sensors to monitor health of high-speed milling machine [ZTLZ06]. Stiefmeier et al. integrate several sensors into one wearable sensors to monitor worker’s activity [SRT + 08]. Singhal and Brown join audio and video data to predict obstacle in navigation [SB97]. Toexploretheapplicationdomaintosupportourthesis,westudybothcollabo- ration schemes with either modality choice in three example applications—single- modal competitive and complementary collaboration, and multi-modal comple- mentarycollaboration. Multi-modalcompetitivecollaborationispartofourfuture work. 4 To study single-modal, competitive collaboration, we propose signature matching algorithms in a multi-sensor vehicle classification context (Chapter 2 and [ZH11]). We study vehicle classification because it is a typical application in urban surveillance. In addition, signature matching is an critical part of com- petitive collaboration. Spatially separated sensors generate detection to vehicles passing by and best detection is picked to represent the vehicle for future classifi- cation. An essential component of competitive collaboration is how the sensornet relatesoneormoresensordetectionstooneormoreactualtargets. However, most current sensornet work pays little attention to this sub-problem. For example in object tracking, tests often assume as single target, sufficient spatial separation that target-to-sensor mapping is clear [OHRS05,ZSR02,SGZ03], or that some out- of-bandsourceprovidesdetection-to-targetmapping(forexample,[ZCP07]). More specifically, there has been little study of algorithms to associate detections with targets, and the effect of their accuracy on detection results, particularly in envi- ronments with dense targets. Weassesshowsingle-modal,complementarycollaborationimprovesoversingle- sensors in the context of an oil field monitoring application. We propose an inex- pensive multi-sensor system to monitor steam injection, especially steam-choke blockage (Chapter 3 and [ZSCH11a,ZSCH11b]). We pick this problem to show complementary collaboration too improves detection accuracy while lowering the cost in industrial sensing domain. Automation of monitoring and control is essen- tial in today’s industrial processes. Since the 1960s, supervisory control and data acquisition (SCADA) systems have automated monitoring and control of indus- trial processes in applications including water management, power grids, chemical processing, and oil production [HTTO02]. One effective use of SCADA system is 5 inoilindustry. However, itscapitalandinstallationcostlimitstheSCADAsystem coverageoverarangeofapplications,saysteaminjectionmonitoring. Wefindthat by combining readings from sensors located at different spots on steam pipe, we may improve steam-choke blockage detection accuracy and suppress false alarms that are impractical to single-sensor monitoring. We study multi-modal, complementary collaboration in a context of a more complex oil field monitoring application. Long oil lines are prone to be blocked during winter time at their sags or fitting joints. Since the monitoring requires large area coverage, traditional SCADA solution is impractical to deploy here. We propose a new non-invasive multi-sensor algorithm with multi-modal sensing to monitor and detect cold-oil blockages (Chapter 4 and [ZH13a,ZH13b]). We con- clude that multi-sensor collaboration with multi-modality sensing solves oilfield monitoring problem that is prior too complex for multi-sensor single-modality col- laboration. Wefurtherassertthatmulti-modalityofferssimilarhelpinotherevent detection applications. In this thesis, we prove sensor collaboration can achieve cost-efficiency in the forgoingthreespecificapplications. Further,thecommonalitybetweentheseappli- cations and a large group of relevant applications strongly suggest that collabora- tivesensinghelpalargerdomainofeventdetectionapplications. Whetherandhow multi-sensor collaboration can help the other class of applications, data collection is beyond the scope of this thesis. Multi-sensor collaboration is not new to the sensornet community. However we are the first to study its applicability in oil field with real deployment. We are also the first to study the detection-to-target mapping and its effect on end result. We 6 next provide the thesis statement, followed by a list of contributions towards the thesis proposal. 1.1 Thesis Statement The thesis of this proposed dissertation is that Multi-sensor collaboration enables sensor networks to accomplish real-world event detection tasks that are impractical for single-sensor systems. We provide three studies to substantiate the thesis. First, we evaluate signature matching in a vehicle classifi- cation context. Second, we present a low-cost design of multi-sensor, steam-choke blockage-detection system. Finally, we implement and evaluate an oil retrieval line blockage detection system. These three studies evaluate different types of multi-sensor collaboration— competitive, complementary and multi-modal, in different categories of event detection—urban surveillance and industrial monitoring, which as a whole sup- port our thesis. The first part of our research, multi-sensor vehicle classification, shows we can improve vehicle classification by choosing the best signature from different sensors for each vehicle. We particularly focus on signature matching, an essential part of competitive collaboration, and how it affects the end-to-end detection. The second part, multi-sensor steam-choke blockage detection, studies how complementary multi-sensor collaboration suppress false alarms in detection. In the final part, oil line blockage detection, we study the feasibility of multi- modality in solving complex sensing problems. We design a new sensing algorithm and demonstrate that multi-sensor, multi-modal sensing has good accuracy with 7 low-cost. Inall,weprovethatinvehicleclassification,steam-chokeblockagedetec- tion,multi-sensorcollaborationimprovestheendresultoversingle-sensor. Further, multi-sensor, multi-modal collaboration enables cold-oil blockage detection. Our work suggests multi-sensor collaboration outperforms single-sensor solu- tion in many other event detection problems. In particular, our vehicle classifica- tion work suggests competitivecollaboration shouldhelp inapplications wherethe same event under detection occurs repeatedly at different place or time, say object tracking or environment sensing. In addition, our steam-choke blockage and cold- oil blockage detection both suggest complementary collaboration should help in applications where the event causes simultaneous signal changes at different places or in different forms, for example human activity detection or target ranging. 1.2 Proving the Thesis To prove our thesis, we study different types of multi-sensor collaboration with either modality choice in different problems and verify if collaboration improves eventdetectionresultorenablesdetectionatall. Wefirststudysingle-modal,com- petitive collaboration in multi-sensor vehicle classification (Chapter 2). We next study single-modal, complementary collaboration in steam-choke blockage detec- tion (Chapter 3). Finally, we explore multi-modal, complementary collaboration in cold-oil blockage detection (Chapter 4). 1.3 Contributions Ourthesisisthatmulti-sensorcollaborationenablessensornetworkstoaccomplish real-world event detection tasks that are impractical for single-sensor systems. 8 Our first contribution is therefore we prove the thesis. We separately prove both competitiveandcomplementarycollaborationcanimprovesensingaccuracyinthis class of applications. We study the competitive collaboration in a multi-sensor vehicle classification, quantifying the relationship between signature matching and end-to-end classification accuracy. We study how single-modal or multi-modal complementary collaboration suppresses false alarms in steam-choke and cold-oil blockage applications by our non-invasive algorithm. The successful employment of multi-sensor collaboration in these three applications suggests that it could increase detection accuracy in other event detection applications. Second, we demonstrate that competitive collaboration can improve event detection accuracy. We use a multi-sensor vehicle classification as an example (Chapter 2). We design signature matching algorithms, integrate them into a multi-sensor vehicle classification system and carry out field test to study com- petitive collaboration. We propose several classes of algorithms (Section 2.2), evaluating approaches to matching using different features: ordering, timing, tar- get features, and raw target observations. We evaluate signature matching as part of a full system on an active roadway (Section 2.3). We first evaluate matching by itself comparing five algorithms. Unsurprisingly, algorithms using little infor- mation (such as ordering), are easily misled by missing readings. Surprisingly, full signature comparisons are also easily mislead by overly specific details. We conclude that a fairly simple static time window (STW) as the best considering both correctness and simplicity, correctly matching 73% of the time. A complete system must classify vehicles into different categories (passenger car, SUV, truck, etc.), and so complete system performance depends on both signature matching and sensor fusion. Since those algorithms can have correlated errors, evaluation of 9 a full system on real data is essential to confirm our algorithm choice. We show that the cost of imperfect matching (with STW) on overall classification accuracy is only 7–11% compared to perfect (oracle) matching, while a poor matching algo- rithm can reduce end-to-end classification accuracy by 21%. By testing signature matchinginareal-worldsystem, weexplorethehighdegreeofcorrelationbetween matchingandclassificationerrorsinreal-worlddata,andwefindthattheaccuracy ofend-to-end,multi-sensorclassificationaccuracywithouralgorithmsisconsistent with theoretical predictions of partial correlation. Third, we show that complementary collaboration with single-modal sensing can suppress false alarms in event detection applications. We use a steam-choke blockage detection as a case study (Chapter 3). We design multi-sensor self- adaptive non-invasive sensing techniques to detect problems in steam distribution (Section 3.5). Current approaches to detect blocking usually measure differential steam pressure with sensors that must pierce the pipeline, incurring a high cost of labor, equipment, and stopped production. Instead, we observe that external temperature observation is sufficient to detect problems such as blockage or flow constriction, provided we observe at multiple locations. We argue that pervasive industrial sensing requires this sort of non-invasive sensing to reduce deployment cost. We demonstrate that our approach of non-invasive, “steam-powered” sens- ing works as a complete system, through both laboratory experiments and field tests (Section 3.7). Our system employs a custom thermoelectric energy harvest- ing/conditioning unit, and a custom amplification board with calibrated thermo- couples, controlled by a standard Mica-2 mote running new detection algorithms. We show the importance of relatively simple hardware and sensing to solve real- world problems: simple hardware makes operation on harvested energy feasible, 10 and simple, non-invasive sensing, when taken at multiple locations can provide actionable decisions. We also show that our approach generalize to other pipeline networks, say hot water network (Section 3.7.5). Finally, we demonstrate multi-modal, complementary collaboration can sup- pressfalsealarmsincomplexproblems. Wechooseanotherindustrialapplication— cold-oilblockagedetectionasanexample(Chapter4). Weidentifytheopportunity for multi-modal non-invasive sensing to reduce error rates with low-cost sensors, and design corresponding algorithms to detect cold-oil blockage Section 4.3. Sim- ilar to the steam-choke problem above, the high cost of current approaches limits a pervasive deployment. Although capable of detecting flow presence or absence, temperature sensing alone may be confused between real blockage and regular pump-off. Therefore, we need multi-modal sensing and adding acoustic sensing on pumpjackstatuscansuppressfalsealarmstriggeredbypump-off. Wedemonstrate that our multi-modal approach with both temperature and acoustic successfully solves cold-oil blockage in field tests (Section 4.5). We embed parameter auto- configuration in the algorithm to generalize our detection to different wells. While some prior research have explored multi-modal sensing with expensive sensors (for example,cameras)andPC-levelcomputation(includingmobilephonesorlaptops), we believe we are the first to show these approaches apply to low-cost embedded sensors. 11 Chapter 2 Multi-Sensor Vehicle Classification Our thesis is to study how to use multi-sensor collaboration to improve event detection. In this chapter, we study single-modal, competitive multi-sensor col- laboration. Particularly, we study the problem of detection-to-target mapping in the context of a vehicle classification system for urban roadways. Competitive collaboration suppress less accurate results with those of other sensors, in order to improvetheoverallaccuracyandrobustnessagainstsingle-pointfailure. Aversion of the contents of this chapter appears in [ZH11]. 2.1 Problem Statement of Multi-Sensor Vehicle Classification Object tracking is one of the canonical problems for sensor networks. A fixed field of autonomous, inexpensive sensors observes their environment, identifies objects, thencomparesobservationstotrackobjectsthatmovethroughthefield[SMK + 07, LLR + 03]. Objecttrackinghasapplicationsinmilitarysecurity,biologyandanimal detection and counting, and in workplace sensornet deployments [CHK + 05]. An essential component of an object tracking algorithm is how the sensornet relates one or more sensor detections to one or more actual targets. However, 12 most current sensornet work pays little attention to this sub-problem. Tests often assume as single target, sufficient spatial separation that target-to-sensor map- ping is clear [OHRS05,ZSR02,SGZ03], or that some out-of-band source provides detection-to-target mapping (for example, [ZCP07]). More specifically, there has been little study of algorithms to associate detections with targets, and the effect of their accuracy on detection results, particularly in environments with dense targets. In this chapter, we explore the question of detection-to-target mapping with competitive multi-sensor collaboration in the context of a vehicle classification system for urban roadways, where vehicles pass fixed sensors at varying rates. We tie detections at different sensors to individual vehicles by signature matching algorithms using features of the signatures or signature timing. Our goal is to understand how imperfect detection mapping affects end-to-end accuracy. Although our results are evaluated in the context of this particular applica- tion, the observation that matching affects sensor fusion accuracy applies to other applications where multiple sensors observe multiple targets. 2.1.1 Relation to Thesis This case study supports our thesis that multi-sensor collaboration improve event detection by studying competitive collaboration, one of the two major types of multi-sensor collaboration. Results show vehicle classification after collaboration by signature matching can have better accuracy than single-sensor classification. In addition, we find collaboration (signature matching) and end classification have a partial correlation. 13 Since vehicle classification is a typical kind of event detection, we believe it shows that competitive collaboration can help other applications where the same event repeats at different place. In this case, a vehicle passes both sensor stations and generates two detections, and both detections can be classified into vehicle separately. 2.1.2 Contributions The main contribution of this chapter is the design of signature matching algo- rithms (Section 2.2). We propose several classes of algorithms, evaluating approaches to matching using different features: ordering, timing, target features, and raw target observations. Unsurprisingly, algorithms using little information (such as ordering), are easily misled by missing readings. Surprisingly, full signa- ture comparisons are also easily mislead by overly specific details (Section 2.2.4). We conclude that a fairly simple static time window (STW) algorithm is the best choice, even over algorithms that are more complex or use more informa- tion (Section 2.3.2). We have integrated of signature matching into a multi-sensor vehicle classification system. While prior researchers have considered vehicle re- identification and vehicle-specific applications, to our knowledge we are the first to explore the specific effects of signature matching in a multi-sensor system. The second contribution is evaluating signature matching as part of a full sys- tem on an active roadway. We first evaluate matching by itself (Section 2.3.2), comparing five algorithms and confirming STW as best considering both correct- ness and simplicity, correctly matching 73% of the time. A complete system must classify vehicles into different categories (passenger car, SUV, truck, etc.), and so complete system performance depends on both signature matching and sensor 14 fusion. Since those algorithms can have correlated errors, evaluation of a full sys- tem on real data is essential to confirm our algorithm choice (Chapter 2). We show that the cost of imperfect matching (with STW) on overall classification accuracy is only 7–11% compared to perfect (oracle) matching (Chapter 2), while a poor matching algorithm can reduce end-to-end classification accuracy by 21%. By testing signature matching in a real-world system, we explore the high degree of correlation between matching and classification errors in real-world data, and we find that the accuracy of end-to-end, multi-sensor classification accuracy with ouralgorithmsisconsistentwiththeoreticalpredictionsofpartialcorrelation(Sec- tion 2.3.4). Finally, our third contribution is to explore how these results to general multi- sensor and multi-target tracking algorithms. We find that defining metrics to compare algorithms is surprisingly difficult (Section 2.3.4), because end-to-end classification results are not just right or wrong, but also duplicated or omitted. Although similar problems occur in pattern recognition, this analysis has been little explored in sensornets and we expect our metrics are useful to characterize other matching problems. Our numeric results are specific to our case study, but we show the importance of good matching algorithms through comparison of several algorithms against a perfect (oracle). This result suggests future multi- sensor tracking must consider error due to individual sensors, multi-sensor fusion, and detection-to-target mapping. 2.2 Matching Algorithms We want to understand how imperfect signature matching affects multi-sensor classification. To do that, we design our new signature matching algorithms in 15 this section. We start with formalizing the signature matching problem to make clear the base of our algorithm design. And then we show our numbering based, timestamp based and raw signature matching algorithms. 2.2.1 Problem Formalization The goal of signature matching is to determine when observations at two sensors observe the same of different actual targets. Given two sensors, a match is when they observe the same target and a non-match as when only one sensor observes a target,perhapsbecausethetargetdoesnotintersecttheviewofoneofthesensors. Weexplorematchinginthecontextofvehiclere-identificationonaroadway,where non-matches indicate vehicles that park or turn between sensors. Signaturematchingistrivialifidealsensorsgeneratetwoperfecteventstreams andalltargetspassbothsensors. However, arealworldapplicationitcanbequite challenging: signatures are missed because vehicles straddle lanes; they can be mergedbytailgatingvehicles;theremaybenomatchingsignatureifavehicleturns between sensors; or signature timing may vary greatly if vehicles change speeds or park. A reliable, real-world matching algorithm must therefore detect matches and also report non-match signatures when no counterpart can be found. Vehicle classification algorithms can then build on it to do multi-sensor fusion [PSH + 06]). 2.2.2 Numbering Based We start with two variants of a simple, order-based algorithm: Na¨ ıve Numbering (NN), and Numbering with Resynchronization (NwR) handling missing signatures. With perfect sensors, i th signature detected upstream should match the i th downstream. Therefore in Na¨ ıve Numbering, each sensor numbers its signatures, 16 and then we merge the detections sequentially. NN suffers from the problem that any missing signatures throw off the stream alignment and result in mis-matches. We call this problem an avalanche, since one observation error causes many incor- rect matches (the left case in Figure 2.1). In the real world, signatures often are missing, either because of sensor error (perhaps a transient fault), sensor-target interactions (for example, a vehicle partially missing a sensor because it changes lanes), ortargetsnotpassingbothsensors(forexample, vehiclesturningorpulling over between two stations). Since such errors are common in the real world, we present NN as the base to build our next algorithm. Numbering with Resynchronization (Algorithm 2.1) adapts NN to real-world noise by periodically resynchronizing streams. Like NN, each station numbers each signature it detects, but with NwR, all stations reset their numbering at a coordinated regular interval. This reset solves the avalanche problem as shown in Figure 2.1. Suppose vehicle 1, 2, 3 and 4 pass both stations and are detected as U i (upstream) and D i (downstream). The mis-detection of D 2 causes following signatures to match incorrectly. NwR can reset to a new numbering span after D 3 and before U 4 ; U 4 and D 4 could still be correctly matched. NwRaddsoneparametertoNN,thedurationbetweenresets. Wewanttoreset frequently enough to prevent avalanches, but resetting makes it difficult to match vehicles in transit between stations, so we do not want to reset too frequently. Reset frequency depends on travel time between stations and how many vehicles are detected by only one sensor. The more frequent singletons occur, the shorter the reset interval should be. The reset interval should be in proportion to vehicle travel time. For our deployment, a large number—about one-third—of vehicles are seen by only one sensor, and travel time is about 30s, so we anticipate a reset 17 Figure 2.1: Avalanche problem in NN but solved by NwR. interval of 3× 30 = 90s. We verified this intuition with exhaustive analysis of possible reset intervals, where we found 83s was our optimal reset interval. From othersimulationresults,weconcludethatthisparameterconfigurationgeneralizes. Further discussion about parameters is in Section 2.3.3. We evaluate NwR in Section 2.3.2 and find that it provides reasonable correct- ness (we define recall in Section 2.3.2; its recall is 64%, Table 2.3). However, it is fundamentally difficult to handle vehicles that leave the roadway (for example, by parking)withnumbering-basedalgorithms. Wethereforenextconsidertime-based algorithms to handle this case. 2.2.3 Time-Stamp Based The next group of algorithms match using the actual detection times of vehicles rather than detection order. Stations record a timestamp with each signature, and if we assume travel time between signatures is predictable, we can use differences in these timestamps to match the signatures. 18 Algorithm 2.1 Numbering with Resynchronization algorithm Input: Dataset D up and D down (signatures are sorted in their time- stamp ascending order) and numbering span length resync ≈ Over- all Detection Time/Numbering Span# Output: M (matching set), N up and N down (non-match set on upstream and downstream) 1: start up ←Min(time-stampinD up )andstart down ←Min(time-stampinD down ) 2: numbering span← 1 and index← 0 3: for sig up in D up do 4: if sig up .timestamp≤start up +numbering span×resync then 5: index←index+1 6: else 7: numbering span←numbering span+1 8: index← 1 9: end if 10: sig up .serialNo←index 11: sig up .ns←numbering span 12: end for 13: numbering signatures in D down likewise 14: for sig up in D up and sig down in D down do 15: if sig up .serialNo =sig down .serialNo and sig up .ns =sig down .ns then 16: M←M∪{(sig up ,sig down )} 17: D up ←D up \{sig up } and D down ←D down \{sig down } 18: end if 19: end for 20: Put remaining signatures from either D up or D down into N up or N down corre- spondingly 21: return M,N up and N down Static Time Window assumes travel time between sensors is relatively con- sistent (say, around δ), and so it predicts that an upstream signature at time t corresponds to a downstream signature in t+δ±v, where v is the range of vari- ation allowed in travel time. This assumption is true provided vehicles typically travel at consistent average speeds between two sensors [MM05]. Depending on the window is set, this assumption holds for 90% of vehicles or more as described in Section 2.3.3. 19 This algorithm takes advantage of the reality that vehicles in a normal traffic flow tend to maintain a constant speed. Drivers usually observe a 35mph speed limit in commercial district roadways and 25mph in local residences in California. Hence a coarse speed range can be easily determined. Algorithm 2.2 Static (Dynamic) Time Window algorithm Input: D up and D down and time window [tw lo ,tw hi ] Input: *** a shift value sv Output: M, N up and N down {Note: activate lines with “***” mark in dynamic time window algorithm; ignore them in static.} 1: for sig up in D up do 2: for sig down in D down do 3: if tw lo ≤sig down .timestamp−sig up .timestamp≤tw hi then 4: M←M∪{(sig up ,sig down )} 5: D up ←D up \{sig up } and D down ←D down \{sig down } 6: *** Reset tw hi and tw lo 7: break 8: else if sig down .timestamp−sig up .timestamp>tw hi then 9: N up ←N up ∪{sig up } and D up ←D up \{sig up } 10: *** both tw hi and tw lo decreased by sv 11: break 12: else 13: N down ←N down ∪{sig down } and D down ←D down \{sig down } 14: *** both tw hi and tw lo increased by sv 15: end if 16: end for 17: end for 18: Put remaining signatures from either D up or D down into N up or N down corre- spondingly 19: return M,N up and N down Our STW implementation works as follows (Algorithm 2.2). The down- stream station is responsible for matching; it holds all pending signatures (not yet matched) reported by upstream sensor. When the downstream station detects 20 a new signature, it examines upstream signatures in sequential order. If the times- tamp difference between downstream signature and the first buffered upstream signature falls in the time window (δ±v), a match is declared and the upstream signature is removed from consideration. If the time difference is less than the smallest possible value, we declare the downstream signature a non-match. If more than the largest, we declare the upstream a non-match. STW is a simple algorithm, well suited to on-line processing. We employ STW in our experimental system (Section 2.3.1). However, similar to NN, an incorrect match can throw off future matches if alignment between signatures becomes skewed, as shown in Figure 2.2. Suppose a vehicle arrives every 2s, travel time δ = 5s, and v = 2s. If each vehicle generates a signature at both sensors, all will be correct matched. But if one signature is not recorded (perhaps that vehicle is out of its lane and misses the sensor), all subsequent signatures will be mis-matched, since vehicle separation is within the v window of variation. However, if two vehicles are spaced slightly further apart, STW automatically uses the gap to reset itself and so it is less susceptible than NN to this problem. Asecondpotentialweaknessofthealgorithmisthatthetimewindowδ isfixed. In practice we set the algorithm based on the upstream/downstream sensor dis- tance and typical vehicle travel times; this configuration can easily be automated. However, we next describe two additional algorithms that adapt δ dynamically to account for changing conditions, and we later evaluate choice of parameters for all algorithms in Section 2.3.3. Our next algorithm adapt δ dynamically to account for changing conditions. 21 Figure 2.2: Avalanche problem in STW. WefindFigure2.2suggestsadjustingthetimewindowhelps. Afterwedeclared a non-match on D 1 , we decrease δ by 1s. The reason is that we believe most downstream non-match is mainly caused by vehicles travel too fast, and likewise, over-slow vehicles results in upstream non-match. And speeds of vehicles in a platoon is not independent, meaning follow-up ones are likely to have a shorter travel time than δ−v. We shift the time window back to suppress the transient fast traffic flow. While the next signature U 1 will also be incorrect (because the 7s travel time is outside the window [2,6]s), all further signatures will correctly match. To control how much the widow moves after a non-match, Dynamic Time Window uses Shift Value (sv). This new parameter controls the amount of change to the travel-time estimate each non-match. After a successful match we reset δ to the original value. The effect of sv is discussed in Section 2.3.3. 22 Table2.3showsthatDTWimprovesrecallby2%overSTW. However,vehicles that “legitimately” pass a single sensor (for example, parking between the sensors) trigger DTW incorrectly. We therefore next consider the use of additional infor- mation to determine accidental missed sensor readings from vehicles that truly trigger only a single sensor. Prioralgorithmsconsideronlysignaturetiming. Wenextusesignaturefeatures, such as number of wheels or wheelbase length, to evaluate match correctness. If features are reliable, they can select between multiple potential matches, or rule out incorrect matches. We choose wheelbase (distance from the front to rear vehicle axle) as the fea- ture for Wheelbase-Enhanced Time Window. We already extract wheelbase for classification. WETW starts with the STW algorithm to find a tentative up- stream/downstream signature match. However, it then builds on this match by considering the signatures immediately before and after the upstream one. Each signature is tested to see if it falls within the STW time constraints, and also if it approximately matches the downstream wheelbase (plus or minus a wheelbase window factor to account for observation error). We then take the first signature thatmatchesbothconstraints, evenifthismeansundoingapriormatch. Ourgoal here is throw out obviously poor matches. WETW therefore also delays decisions by one signature to allow this wheelbase-triggered re-matching. The details are in Algorithm 2.3. Feature-based methods like WETW add an additional dependency: feature extraction from signatures is not perfect, so incorrect feature extraction actually degrade matching. WETW works best when most vehicles have different wheel- bases (say, a mix of cars and trucks). Finally, it considers one possible feature 23 Algorithm 2.3 Wheelbase Enhanced Time Window algorithm Input: Pre-matched signatures via STW, M, N up and N down . Wheelbase window and time window. Output: improved Matching results, M, N up and N down 1: repeat 2: for sig i up in N up do 3: if sig i−1 up is matched to sig down then 4: if the wheelbase difference between sig i−1 up against sig down is outside wheel window and the differences between sig i up and sig down fall in both wheelbase and time window then 5: M←M∪{(sig i up ,sig down )} and M←M\{(sig i−1 up ,sig down )} 6: N up ←N up ∪{sig i−1 up } 7: end if 8: else if sig i+1 up has a match then 9: similar process as forgoing 10: end if 11: end for 12: until no new matching declared 13: for all signatures in N up and N down do 14: if thedifferencesoftwosignaturesN up andN down fallinbothwheelbaseand time window then 15: transfer these two signatures from N up and N down to M 16: end if 17: end for 18: return M, N up and N down (wheelbase length) in addition to timing, although in principle one could use other or multiple features. We next consider full signatures as a richer feature. ThebetterrecallofWETW,comparedtoSTW(Table2.3),suggeststhatmore information helps matching. To determine if more information always improves results, we replace the matching function in WETW, changing it from wheelbase length to full comparison of raw signatures. We call the new algorithm Raw- Enhanced Time Window. Raw signatures record the change of loop inductance as the vehicle crosses, sampling at 300Hz; thus they represent the most complete 24 Figure 2.3: Four pre-scaled raw signature pattern comparisons. The horizontal axis is time and the vertical is energy. information about what vehicles pass a sensor. Figure 2.3 shows four example signature pairs. Red and blue (darker) lines represent different signatures. One cannot directly compare raw signatures, because slight differences in vehi- clespeedorsignaturesegmentationresultindifferentlengthsignatures. Tocorrect for this distortion, we use Dynamic Time Warp [SC90] to compute the similarity between raw signatures. This approach has two steps. First, it warps the time axis of one signature iteratively until each data point in this sequence is optimally aligned to a point in the other signature. Second, it evaluates the similarity of the signatures by summing the Euclidean distance of between all point-pairs in the warped signatures. (This evaluation measure is also used to evaluate the quality of warping.) Other than this change of comparison function from wheelbase to time-warped signature, RETW is similar to WETW, using the same constraints on signature sequencing and timing. Surprisingly, Table 2.3 shows RETW has slightly poorer recall than WETW, with one fewer correct non-matches of the total 107 correct matches and non- matches. 25 To evaluate if combining WETW and RETW would yield better results, we compare vehicle-level matching results between these two algorithm. We find that theresultaresimilar;onlytwomatchesoutoftotal138eventsaredifferent. There- fore, an oracle algorithm producing a union of the correct matches of these two algorithms would improve by at most 2% in this study. We next evaluate if com- paring full, raw signatures helps. 2.2.4 Full Raw Signature Comparison Ourtime-windowfamilyofalgorithmsusesomefeaturetoshiftsignaturematches. However, potentially one could compare all signatures against all other signatures, as implemented by Cheung et al. [CCD + 05]. In Section 2.2.3 we modified the time window algorithm to consider full signature comparisons. Here we remove the temporal constraints of RETW to see if full matching freedom can improve results. However, it turns out that more information does not help RETW do matching. To test if more information helps, Full Raw Signature Comparison compares all signatures at the two sensors. This algorithm works by testing each match against all signatures at the other sensor side. We use the dynamic time warp as the comparison function. This comparison declares a tentative match as the closest possible score. We then test this tentative match a threshold to see if it is a good match, or instead to declare that signature as a non-match. We determine the threshold by training on known ground truth and using the median distance score of true matching. Thus the main difference against RETW is that full matching always compares all signatures (using maximal information), while RETW compares a time-constrained subset. 26 We find full matching is much worse than RETW: none of the 65 matchable vehicles find its true counterpart. This somewhat surprising result is mainly due to two problems. First, without temporal constraints (like RETW), raw matching relates many signatures that are unreasonably earlier or later. Second, full match- ing requires a threshold to determine match/non-match, but there is no single fixed threshold that identifies correct matches. When we train with known cor- rect matches, the median distance score among the 65 true matches is four times higher than that of all tentative matches. In other words, for real data, incorrectly matched vehicles always look more similar to each other than true matches—too much information can mislead. We see three causes raw signatures often fail to match. First, environmental noise and measurement error may distort signatures, sometimes causing wheels to be mis-detected. Since distortion is usually independent at the two sites, signa- tures of some true matches are inherently different from each other. For example, whilethetoprowofFigure2.3showstwocompletesignatures, inthethirdrowthe darker (blue) signature recorded only one wheel, likely because the car was strad- dling lanes. However, we can sometimes compute a correct wheelbase for partial signatures, even if one wheel is missed. Less information (a partial signature) can still result in a correctly matching feature (wheelbase). Second, a large number of signatures make accidental mis-identification easy, particularly with many vehicles of similar general type (passenger car, truck, etc.) ormake. Table2.1showsthatrawsignaturematchingoften(about30%ofthesig- natures)findsthebestmatchasavehicleofanothercategory,eventhoughahuman would detect that clear mistake. With more than 100 potential signatures, acci- dental matches are increasingly likely. Thus time constraints (in the time-window 27 Table 2.1: Category-level matching results by full raw signature comparison 105 vehicle passed # of whose best match in downstream is a upstream site passenger car SUV truck 24 passenger cars 15 8 1 64 SUVs 4 57 3 17 trucks 0 15 2 algorithms)helpfocusonplausiblecandidates. Finally, DTWcanarbitrarilywarp signatures, and perhaps too much freedom makes mis-identification easier. Our time-window algorithm corrects for speed differences, but assumes acceleration and higher order derivatives are zero, perhaps a more reasonable assumption than allowing on-zero higher-order derivatives. We examine raw signature comparison to get a best possible result by using all available information. In doing so, we ignore the network bandwidth and energy requirements of sending around full signatures. We conclude that, for our sen- sors, careful chosen features (such as wheelbase) represent vehicles better than full information, because feature detection filters out noise. Other researchers have reported better results comparing full signatures (Che- ung et al. report 100% re-identification [CCD + 05,CEV05], details in Chapter 5). We believe they succeed because their test set is smaller (seven vehicles), their sensor spacing closer (several meters away), and their sensor provide informative (three-axis magneto-meter). While their results suggest the need for more work, to see if their results generalize to larger datasets, and more distant or different sensors. Because of these challenges, we conclude that both full signature compari- son, both with all signatures and time-limited signatures, is not desirable—too 28 much information hurts more than it helps. Instead, wheelbase or other extracted features can provide better results by effectively filtering out noise, and time con- straints help avoid improbable matches. 2.2.5 Algorithm Discussions Theadvantagesofalgorithmsdivergeuponmatchingcorrectness, parametersensi- tivity, complexity, real-time-nessandapplicabilitytodifferentmonitoringsettings. DTW, WETW and RETW should have higher recall than STW. STW might not beabletodowellunderheavytraffic,becausevehiclesarelackoftemporalsepara- tion, while DTW could handle the problem. If the traffic is promiscuous, WETW is sure to utilize wheelbase difference to make better decision. If no intersection amidtheroadandvehicleskeepdrivingorder, NwRcouldyieldsatisfactoryresult. Several parameters have to be embedded into each algorithm, but we want to keep the sensitivity minimized. Section 2.3.3 briefly concludes the relation between parameter and performance. One merit of all these algorithms, except RETW is low complexity, comparing to raw signature comparison. Another advantage of these algorithms is that it is easy to scale them up to more than two stations. The match relation is transitive, meaning that if A and B is a match, B and C match, then A, B and C will match. Hence, if three (or more) stations deployed, we can always make pairwise match between immediate neighbor sensors before reach a final match result among all of them. In all, from the simulation result in Table 2.3 as well as our description above, we draw the conclusion that STW is the most appropriate algorithm for our short term experiment, while others could be analyzed in post-facto processing. STW 29 yieldsahigh-enoughrecall(about73%)withoutlosingapplicabilityandsimplicity. A more comprehensive performance analysis is in Section 2.3.2. 2.3 Multi-Sensor Vehicle Classification Evalua- tion To understand and quantify how multi-sensor collaboration improves event detec- tion,weincorporateouralgorithmsdesignedinforgoingsectionintoamulti-sensor vehicle classification system. Each sensor can do detection and classification by its own, but our hypothesis is that we can improve the classification accuracy by deploy multiple sensors on the road and pick the best detection for each vehi- cle. We collected a 3-hour traffic dataset with our prototype system supplemented by human observers and videotape ground truth data. This section describes the detailsofthatfieldtest,comparematchingresultamongandwithinouralgorithms, examine algorithm parameter sensitivities and finally the effect of matching over vehicle classification. 2.3.1 Data Collection Experiment From 8 a.m. to noon, February 19, 2009, we carried out a field test and traffic data are collected at USC campus. During the nearly 3-hour long field test, we collectedabout300detectionsofvehiclesatupstreamanddownstreamlocationson a public road on our campus. We also took videotape of traffic and later manually examined this record to generate ground truth. The collection stations were on one of USC campus internal streets, with two stations 90m distant, each with 30 Figure 2.4: Real deployment. two adjacent inductive loop “Blade” sensors. Figure 2.4 shows real deployment— each station has a laptop, an IST-222 detector and two loop tapes. Stations were connected by a wireless router, 8dB wireless dish adapter and 15dB high-gain antenna. Although our campus has campus-wide wireless coverage, we deployed our own LAN to mimic the same kind of deployment that would be used on a city street. Thedownstreamsitebothdetectedvehiclesanddidsignaturematchingand sensor fusion from upstream signatures. We used sensor calibration as described previously [PSH + 06], and each station ran local single-sensor classification, while the master (the downstream node) performed on-line signature matching using STW (Section 2.2.3) and then sensor fusion. Before we report our results, we describe the traffic and on-line processing. First, we observed a mix of traffic including general automobiles, campus busses, shuttle vans, construction vehicles, delivery and semi-trailer trucks. We observed 33passengercars, 86SUVsand19trucks, andanumberofcarts, motorcycles, and 31 bicycles. Our system automatically discards the signatures of carts, motorcycles, andbicyclesfromourdatasetbecauseourgoalistoclassifycars. Second,although we did on-line processing in the field, the results reported here have been re- evaluated post-facto. This re-evaluation is necessary because our field experiment was mis-calibrated with incorrect typical vehicle speeds. 2.3.2 Matching Algorithm Correctness Although we evaluated STW on-line, to compare all of our signature matching algorithms (Section 2.2), we replayed the data off-line through each one of the algorithms. Our goal is to maximize matching correctness in the face of real- world noise, and to compare matching algorithm performance and overhead. Our expectation is that exchange of more information (up to full signatures) would enable better matching, but instead we find that real-world noise fundamentally limits the correctness of matching. Before looking at numerical comparisons we must first define our measure of correctness. Within the final output of matching system, there are four major situations: (i) Both sites detect a signature of a vehicle respectively, and the system declare a match on these two signatures, a True Match, or M; (ii) The systemincorrectlydeclareamatchontwosignaturescorrespondingtotwodifferent vehicles, a False Match; (iii) One signature of a vehicle is missing at either site, and the system declared a non-match on the detected one, a True Non-match, or N; (iv) The system incorrectly declared a non-match for a signature which does have a counterpart detected by the other node, a False Non-match. Case (iii), where signatures are missing from one site, is important because it shows how real-world conditions can violate the assumption that every signature 32 mustbematched. Inpractice, not everysignatureshouldbematched,foravariety ofreasons. Signaturescanbemissingfromeithersitebecauseofsensororalgorithm error, undesirable vehicle/sensor interaction (for example, if the vehicle is half in the lane), or driver choices that violate our assumptions (for example, a vehicle that stops and parks between our sites). While sensor or algorithm errors can perhaps be corrected with better software or hardware, matching is impossible if vehicles never pass both sites. Webreakvehicledetectionsintofivegroups(Table2.2showshowmanyofeach we see): Normal: vehicles drive continuously across two sites at a reasonable speed (say 10 to 40mph) Singleton: vehicles only pass one site Over-Segmented: vehicleshavemorethantwosignaturesgeneratedononenode at the same time Pull-Over: vehiclespulloverinbetweenthetwositesandareovertakenbyothers. But they pass both sites. T travel > 200s PONO: (Pull-Over, Non-Overtaken) pull-over vehicles where no other vehicles overtake them (the relative order of vehicles maintained) To evaluate correctness, we must normalize our results by number of true events. An event is an oracle-defined true match or true non-match (M + N). We determine oracle events by manual analysis of videotape to get accurate oracle results representing ground truth. To evaluate our correctness, we draw terms from information retrieval [Rij79]. IR defines recall as tp/(tp + fn), characterizing how much of the true result is found. In our case, tp + fn is the number of events, as defined above, since for 33 Table 2.2: Event types for matching Expected Events Occur- Types per Occurrences ences Events Normal 1 match 65 65 Singleton 1 non-match 46 46 Over-Segmented 1 match 13 13 Pull-Over 2 non-matches 7 14 PONO 1 match 0 0 total events 131 138 Table 2.3: Matching correctness of algorithms Correct Incorrect Reported Alg. Recall matches non-mat. mat. non-mat. signature# Precision STW 101 (73%) 46 55 23 28 152 66% DTW 103 (75%) 41 62 17 43 163 63% WETW 108 (78%) 51 57 19 24 151 72% RETW 107 (78%) 51 56 19 25 151 71% NwR 88 (64%) 31 57 18 65 171 51% oracle 138 (100%) 78 60 0 0 138 100% oracle matching, the number of incorrect non-matches is always zero, while true positivesrepresentcorrectmatchesandnon-matches. Theoutputofouralgorithm isthereforeevaluatedby: recall = ( d CM+ d CN)/(M+N),where d CM representsthe number of correct matches output by a matching algorithm, and d CN the output of correct non-matches. We also report precision, to characterize how often a matching algorithm’s output is incorrect: precision = ( d CM+ d CN)/( d CM+ d CN+ d IM+ d IN) where d IM are the number of incorrect matches (and d IN are incorrect non-matches). In general, we focus on recall to evaluate our correctness, but we also report precision. 34 Table 2.3 shows our evaluation of matching for this experiment. We draw several conclusions from the comparison among all of the algorithms. First, all algorithms are generally good—the poorest algorithm has matching recall above 60%. Second, time-stamp based algorithms generally yield better correctness than others, with recall above 73%, 10% better than the 64% or lower rates of alter- natives. Considering the different time-stamp based algorithms, we observe that DTW, WETW and RETW provide only slight improvements over STW (a 2% or 5% improvement over STW’s 73% recall). We therefore recommend STW as the preferred algorithm overall, because it is much simpler to implement and configure than DTW, WETW and RETW and nearly as accurate. The three derivatives appear to have no significant improvement over the base time-stamp algorithm (only 3%–4%). If we examine more carefully, the change of actual matching cor- rectness numbers indicates we have achieved our designing goal in Section 2.2.3 and 2.2.3. DTW has better recall and fewer incorrect matches but more incorrect non-matches than STW. WETW successes in correcting its base version’s a few incorrect matches into correct matches. Surprisingly, we find more information does not always help. RETW has slightly lower recall than WETW, meaning comparison of full signatures do not improve on evaluation of signature similarity by extracted wheelbase length. 2.3.3 Parameter Sensitivity In this section, we present a detailed evaluation on how sensitive the accuracy of ouralgorithmsaretosettingsofvariousparameters. Wefocusontime-stampbased algorithms because they are the most effective. An ideal algorithm is insensitive to parameter settings, so even if mis-configured it will perform reasonably. 35 0 25 50 75 100 125 150 175 200 0 5 10 15 "Normal" vehicle count travel time (s) Figure 2.5: Travel time distribution of the 65 Normal vehicles. We begin by considering STW, where the time window size and location are the only parameters. Figure 2.6 shows how STW’s matching accuracy varies for all possible window configurations. The x-axis and y-axis are the lower bound and upper bound of time window, while the gray-scale value indicates STW’s recall for our experimental dataset. The best accuracy (73%) occurs for range [14,44]s (δ = 29s, v = 15s). As can be seen in Figure 2.5, this time window acceptsmostvehicles,sincethewindowcenter(δ =29s)isnearthemedianvehicle travel time (32s). However, other outliers, which have a really long travel time is fundamentally difficult for any time-stamp based algorithm. STW is fairly robust to an imperfectly set window. However if our parameters are off by 50%, we lose about 20% accuracy. And a 20% change of window only causesabout10%loss(accuracy62%). Inall,wedrawthreeconclusionsfromabove observations. First, as we expected, we have to locate the time window center (δ) close to median travel time to get optimal parameters. Second, the time window size should be adjusted according to vehicle speed variance. The reason is that the 36 0 20 40 60 80 100 0 10 20 30 40 50 time window upper bound (s) time window lower bound (s) 0 20 40 60 Figure 2.6: Accuracy % of STW. broader the window (large v), the more incorrect matches and less correct non- matches since it is easier to mistakenly match a singleton vehicle to another one. And the narrower the window (small v), the more incorrect non-matches and less correct matches because a small travel time range of Normal vehicles is accepted. Finally, STW works surprisingly well (above 60% accuracy) over a wide parameter ranges—upper/lowerboundcouldvariousinranges[30,50]/[10,20]sapproximately. DTWaddsanadditionalparameter, theShift Value (sv). Toevaluatesensitiv- ity to different shift values, we search the entire parameter space (shift value and window bounds). For each sv, we pick the best accuracy over 2-D TW space and show them in Figure 2.7. We see that DTW is quite insensitive to sv, although accuracydropswhensv≥8sandeventuallylowerthanSTW. Thereasonexplains this is that traffic is sparse and hence our premise in Section 2.2.3 does not always hold. We assert that in a short platoon, vehicle speeds are not independent but in the experiment, most vehicle travels alone. Shifting time window too much makes it harder to match follow-up normal vehicle. 37 0 5 10 15 20 0 20 40 60 80 best accuracy (%) shift value (s) Figure 2.7: Best accuracy % of DTW with different Shift Values. 0 10 20 30 40 50 0 20 40 60 80 best accuracy (%) wheelbase window (%) Figure 2.8: Best accuracy % of WETW with different Wheelbase Windows. We also evaluated WETW using the same approach as DTW (with optimal other parameter) and found it is quite insensitive to the wheelbase window. We omit RETW discussion since it performs similarly to WETW. Reviewing how parameters affects the matching accuracy shows us that our time-stamp based algorithms, in general, are insensitive to internal parameters. STWyieldsreasonableaccuracyoverawiderangeoftimewindows. Besides,those parameters are easy to configure under certain rules. For example, time window should center around the estimated common vehicle travel time. The upper/lower 38 bound of time window should be chosen close to (1±50%) of the travel time. We leave the parameter auto-configuration as an open issue. 2.3.4 Impact of Matching on Classification The goal of signature matching in our system is to support multi-sensor fusion, or more generally, to synthesize conclusions from detections from multiple sensors. In this section, we study how the matching correctness affects multi-sensor fusion. Our hypothesis is that better matching algorithms result in better multi-sensor classification. However, classification and matching have a non-linear interaction sinceerrorsthatmakematchingdifficultalsomakeclassificationdifficult,sostudy- ing real data is important. Park et al. previously showed that classification can benefit from multi-sensor fusion [PSH + 06]. Combining readings from multiple sensors can correct some, but not all, classes of errors. For example, although a vehicle may temporarily leave a lane and so be mis-detected by one sensor, it likely returns to its lane later. Park et al. examine the accuracy of several sensor fusion algorithms relative to human observation and show that sensor fusion can allow automatic classification ratesexceedingthatofhumanobservers. Accuracydependedonhowmanygroups were classified (2or 3, with more categories having lower accuracy because thereis more opportunity to error), and the sensor fusion algorithm. They found the best accuracy was for their quality-best fusion algorithm, giving 97% for 2-category and 74%for3-category. Bycomparison,humanobservationhad87%for2-categoryand 83% for 3-category, and 100% accuracy is actually impossible because of overlap in the categories themselves. 39 Table 2.4: Classification accuracy, multi vs. single Categories Classification Veh. two three single sensor: upstream alone 105 80 (76%) 68 (65%) downstream alone 99 79 (80%) 43 (43%) oracle matching oracle fusion: 138 117 (85%) 90 (65%) quality-best fusion: 138 111 (80%) 78 (57%) This prior work, however, assumed a perfect (oracle-based) signature matching algorithm. We next evaluate classification accuracy with a realistic and therefore imperfect matching algorithm. Building on prior work [PSH + 06], we consider two classification tasks: three- category of passenger cars, light trucks (SUVs or pickups trucks), large trucks (FHWA classes 2, 3, and 4–13) and two-category of trucks and non-trucks (FHWA classes 2–3 and 4–13) [Fed03]. Three-category is inherently harder because many light trucks can easily be confused with cars and even humans have difficulty to make a perfect judging (when small SUVs blur into cars) [PSH + 06]. Our goal here is to evaluate how matching effects results of realistic classification, so we set as our baseline oracle matching, and then compare to realistic matching algorithms. We use quality-best fusion, the best choice from [PSH + 06]. Each sensor assigns a quality value for each signature it extracted, based on wheel detected, signal strength and other factors. Hence when fused, a matched vehicle could have two candidate classifications and we choose the one with higher quality value. Table 2.4 summarizes our baseline. For two-category and three-category, the baseline accuracies are 80% and 57% respectively. With this baseline we now 40 must define how to quantify multi-sensor classification accuracy. Single-sensor classificationaccuracyiseasilydefinedasthefractionofcorrectly-classifiedvehicle number by the total. Multi-sensor fusion with perfect (oracle) matching can also be defined similarly. However, just as matching correctness is complicated by duplicate or under- counts,thosecasesmakeitdifficulttoprovideasimpleaccuracymetricforclassifi- cation with imperfect matching. For example, if two detections of one true vehicle arenotmatched, andoneiscorrectlyclassifiedandtheotherisnot, doesthesetwo reportsrepresent(i)twoerrors(sinceitwasmis-matchedandwecannotdetermine which is correct), (ii) one error and one correct result, or (iii) one correct result (taking the correct classification as overriding the incorrect duplicate)? Or what if a single vehicle was reported twice and classified correctly both times, is this (iv) incorrect,sinceitisover-reported,or(v)correct,sincebothreportsareconsistent? We can define the number of vehicles in each case as V c m , where c indicates how manytimesatruevehiclewascorrectlyclassified,andmindicateshowmanytimes it was reported. We therefore define two levels of accuracy: strict and relaxed. Strict Accuracy is the most demanding: we require that each vehicle be correctly classified exactly once—conclusions (i) and (iv) above. If we define V 1 1 as the number of vehicles seenandclassifiedexactlyonce, andV asthesetofalltruevehicles(events), strict accuracy is: Accu strict = V 1 1 /|V|. Relaxedaccuracyisrelevantif,insteadofdemandingperfectcounts,ourgoalis to approximate the percentages of each vehicle class. Here we consider overcounts due to incorrect matching to be correct provided both signatures are classified correctly, thus taking cases (ii) and (v) in the examples. If a vehicle is seen twice 41 Table 2.5: Reported once and correctly classified once signatures matching classification ups. downs. correct? result U i D i correct matches C(i)=F(U i ,D i )=G(i) U i X incorrect C(i)=F(U i ,X)=G(i) U x D i C(x)=F(U x ,D i ) U i D x incorrect C(x)=F(U i ,D x ) X D i C(i)=F(X,D i )=G(i) U i – correct non-matches C(i)=F(U i , –)=G(i) U i D x incorrect matches C(i)=F(U i ,D x )=G(i) – D i correct non-matches C(i)=F(–, D i )=G(i) U x D i incorrect matches C(i)=F(U x ,D i )=G(i) U i ,D i : upstream and downstream signatures generated by vehicle i X: ”don’t care”, i.e., non-matches or matches against a fake signature (noise) or a signature of some other vehicle –: means non-match G(i): the ground truth category of vehicle i is x C(i): multi-sensor classification result of vehicle i F(x,y): fusion of one or two signatures and classified correctly for twice, we define it as one V 2 2 . We further define relaxed accuracy: Accu relaxed = (V 1 1 +V 2 2 )/|V|. AlthoughwetalkaboutV 1 1 andV 2 2 here, thereareactuallyanumberofspecific cases. We enumerate how we handle each case of V 1 1 in Table 2.5. We next consider the effects of matching accuracy on end-to-end classification accuracy. For this evaluation, we compare against the baseline of perfect (ora- cle) matching, shown in Table 2.6. Each algorithm uses optimal parameters (as defined in Table 2.3)). The resulting signatures use quality-best fusion [PSH + 06] to generate the final classification result. 42 Table 2.6: Multi-sensor classification accuracy Matching 2-cat. accu. (%) 3-cat. accu. (%) Algor. Recall (%) strict relaxed strict relaxed STW 73 (-27) 69 (-11) 73 (-7) 49 (-8) 50 (-7) DTW 75 (-25) 68 (-12) 76 (-4) 47 (-10) 50 (-7) WETW 78 (-22) 70 (-10) 75 (-5) 50 (-7) 51 (-6) RETW 78 (-22) 70 (-10) 75 (-5) 50 (-7) 51 (-6) NwR 64 (-36) 59 (-21) 74 (-6) 39 (-18) 46 (-11) oracle 100 (0) 80 (0) – 57 (0) – Thecomparisonprovesourhypothesis,confirmingthatcorrectnessinsignature matching has a large, but correlated effect on end-to-end classification. (Here we refer to three-category classification; two-category classification is similar.) First of all, with our algorithms, the strict accuracy can approaches that of the baseline (50% for WETW vs. 57% oracle, with a 7% penalty due to incorrect matching). Categorization with oracle matching is not perfect because the underlying single and multi-sensor classification methods are imperfect. The addition of realistic matchingfurtherlowersclassificationaccuracybecausemis-matchedsignaturescan results in signature duplication, omission, or incorrect multi-sensor classification. We further study this correlation in Section 2.3.4. Second, as expected, accurate matching helps improve classification while poor matchinghurts. WefindWETWmatchessignaturesmostaccurately(78%against 64%, the worst case with NwR). Higher matching recall here shows a correspond- ing improvement in end-to-end classification accuracy, with WETW allowing 50% classificationaccuracy(vs.39%worst-case,NwR).WETW’simprovementisdueto morecorrectmatches(51of78cases,Table2.3),whilepoorermatchingalgorithms cause duplicated or omitted signatures. 43 Finally, our multi-sensor classification has moderate improvement over single- sensor in terms of both accuracy and robustness, even when coupled with realistic (imperfect) matching algorithms. We see that the downstream sensor is less accu- ratethanupstream,possiblyduetodifferencesinvehiclespeedsandchannelization at the two sites. However, multi-sensor fusion improves downstream classification accuracyresultby7%, evenwithimperfectsignaturematching(WETW),showing that multi-sensor fusion can be more robust to deployment or sensor error. While Table 2.6 shows how end-to-end classification accuracy changes due to matching, it doesn’t show why the results differ. We look at that question next. Wenextlookmoredeeplyatwhy signaturematchingandclassificationaccuracy affect each other. Both matching and classification use the same sensor data, so inaccurate data at one sensor (perhaps due to target or environment noise, or deployment differences) can both make matching difficult and affect multi-sensor classification accuracy. If the algorithms were completely correlated, then end-to- end accuracy should be the minimum of either algorithm’s correctness. If they were strictly uncorrelated, then end-to-end accuracy should be their product. Table 2.7 shows there is partial correlation between matching correctness and classification accuracy. We report matching recall for each matching algorithm, and three-category vehicle classification accuracy (with oracle matching), then compare expected accuracies with no and full correlation to experimental results. Wefindthattheend-to-endmulti-sensorclassificationaccuracywithourmatching algorithms is always between what would be predicted by no or full correlation. Theseexperimentalresultssuggestthataccuracyofthetwoalgorithmsispartially correlated. This correlation shows up in end-to-end accuracy, where uncorrelated 44 Table 2.7: The correlation between matching and classification Match. Class. Correlation recall accu. none full Algor. (m) (c) m·c min(m,c) experiments STW 73% 57% 42% 57% 49% DTW 75% 57% 43% 57% 47% WETW 78% 57% 44% 57% 50% RETW 78% 57% 44% 57% 50% NwR 64% 57% 36% 57% 39% WETW would predict a 13% classification penalty (from incorrect matching), but correlation means that experimentally the penalty is only 7%. Tounderstandwhatcausesthesecorrelations, wenextreanalyzetheSTWcase from Table 2.6: Table 2.9 shows STW accuracy grouped by correctness in either or both matching and classification. The first case (yes, yes) is both matching and classification are correct, our goal. The second (yes, no) includes vehicles that are correctly matched or non-matched, but where multi-sensor classification gives an incorrect result. Presence of the third category where matching fails but classification succeeds (no, yes) is unexpected, but in these cases multi-sensor fusionselectsthecorrectsignatureandresulttorecover. Vehiclesinthefourthcase (no, yes/double) are correctly classified, but because matching fails there appear to be two vehicles (one at each sensor), so we over-count (theV 2 2 case from defined earlier). Infinalcase(no,no)bothmatchingandclassificationfail. Table2.8shows the full details of all cases we considered, and how we classified them. This table supports our prior results and illustrates the complexity in evaluating correctness when faced with two sensors, each of which may omit or duplicate or incorrectly detect a vehicle. 45 Table 2.8: Matching vs. classification in STW (dissection) up- down- matching reported # in STW class. sig. sig. correctness veh. class. correct. 2-cat. 3-cat. V 1 1 U i D i correct mat. C(i)=F(U i ,D i )=G(i) strict; relaxed 37 29 V 1 1 U i X incorrect C(i)=F(U i ,X)=G(i) 3 2 Ux D i C(x)=F(Ux,D i ) V 1 1 U i Dx incorrect C(x)=F(U i ,Dx)=G(i) 0 0 X D i C(i)=F(X,D i )=G(i) V 1 1 U i – correct non-mat. F(U i )=G(i) 22 18 V 1 1 U i Dx incorrect mat. F(U i ,Dx)=G(i) 6 5 V 1 1 – D i correct non-mat. F(–, D i )=G(i) 24 12 V 1 1 Ux D i incorrect mat. F(Ux,D i )=G(i) 3 1 V 2 2 U i X incorrect F(U i ,X)=G(i) relaxed 6 2 X D i F(X,D i )=G(i) V 1 2 U i X incorrect F(U i ,X)=G(i) none 3 5 X D i F(X,D i )6=G(i) V 1 2 U i X incorrect F(U i ,X)6=G(i) 0 0 X D i F(X,D i )=G(i) V 0 1 U i D i correct mat. C(i)=F(U i ,D i )6=G(i) 9 17 V 0 1 U i X incorrect C(i)=F(U i ,X)6=G(i) 4 5 Ux D i C(x)=F(Ux,D i ) V 0 1 U i Dx incorrect C(x)=F(U i ,Dx) 0 0 X D i C(i)=F(X,D i )=G(i) V 0 1 U i – correct non-mat. F(U i ,–)6=G(i) 6 10 V 0 1 U i Dx incorrect mat. F(U i ,Dx)6=G(i) 1 2 V 0 1 – D i correct non-mat. F(–,D i )6=G(i) 3 15 V 0 1 Ux D i incorrect mat. F(Ux,D i )6=G(i) 0 2 V 0 2 U i X incorrect F(U i ,X)6=G(i) 1 3 X D i F(X,D i )6=G(i) V 0 0 – – – – 10 10 total vehicles 138 138 Wedrawthreeconclusionsaftercomparingmatchingagainstend-to-endclassi- fication result. First, incorrect matches do not always result in incorrect classifica- tions. In8outof138(6%)(no,yes)cases,matchingfailsbutclassificationiscorrect in three-category classification. Should matching and classification are completely correlated, these incorrectly matched signatures would never be correctly classi- fied. Second, in the (no, yes/double) case matching fails and we overcount one vehicle twice, at each sensor. This case prompted us to consider strict and relaxed 46 Table 2.9: Matching vs. classification in STW correct? categories matching classification two three yes yes 83 (60%) 59 (43%) yes no 18 (13%) 42 (30%) no yes 12 (9%) 8 (6%) no yes/double 6 (4%) 2 (1%) no no 19 (14%) 27 (20%) 138 events (100%) accuracy, although with only 2 cases of 138 (1%), this event is rare. The only exception is with NwR matching, where 65 incorrect non-match (more than other algorithms, Table 2.3) results in more of these V 2 2 events (7% vs. others 1–3%, Table 2.6). Finally, we find correct matches do not always result in correct clas- sification, either. Unfortunately, although our STW algorithm did well on 42 out of 138 (30%) (yes, no), our imperfect single-sensor classification and fusion fail to turn them into correct classification. We see opportunity that with better vehicle classification and sensor fusion might help us to achieve a 30% improvement. Overall, these results demonstrate that correlation between these algorithms has significant, quantifiable effects on end-to-end performance. While both algo- rithms can be studied and improved independently, we conclude that a full eval- uation must consider both in the context of real data, and good overall accuracy requires a balance of good algorithms for matching, classification and multi-sensor fusion. 47 2.4 Conclusions on Multi-Sensor Vehicle Classi- fication This section supported our thesis by showing competitive multi-sensor collabora- tion could improve vehicle classification, a typical application of event detection. More specifically, we explored how signature matching affected end-to-end multi- sensor vehicle classification accuracy. We further infer that multi-sensor collabo- ration helps event detection especially other similar classification applications. In this chapter, we first proposed designs of a suit of signature matching algo- rithms and evaluation results showed a simple static-time window algorithm was bothefficientandthemostaccurateofarangeofalgorithms,includingfullrawsig- naturecomparison. Wethenshowedthatsignaturematchinghadsignificanteffect on the end-to-end accuracy of a multi-sensor classification system. An important effect was the correlation between matching and classification; we quantified that the algorithms were partially correlated. Ultimately, we showed that real-world matching algorithms could reach end-to-end classification accuracies within 90% of perfect matching when evaluated in real-world field tests. 48 Chapter 3 Steam-Choke Blockage Detection Our thesis is to show multi-sensor collaboration enables sensornets to accomplish real-world event detection tasks beyond the reach of single-sensor systems. The previous section supported this thesis by studying single-modal, competitive col- laboration in a typical class of event detection applications—urban surveillance, with vehicle classification as a case study. We next study the other type of multi- sensor collaboration, complementary collaboration. Complementary collaboration combines all relavent sensors’ partial interpretations on interested phenomenon to form the final result. Without complementary collaboration, the partial result of sensors may be misleading and cause false alarms. We study complementary collaboration in another domain—industrial sensing, in order to support our the- sis from a larger prospect. A version of the contents of this chapter has been published in the 9 th ACM Conference on Embedded Networked Sensor Systems (SenSys ’11) [ZSCH11a]. 3.1 Motivations of Choke Blockage Detection Automated monitoring and control of industrial processes are becoming increas- ingly important as industrial operations grow in complexity and size. Since 1960s, supervisory control and data acquisition (SCADA) systems have been used to automate monitoring and control of industrial processes in applications ranging from water management, power grids, chemical refining and processing, to oil 49 production [HTTO02]. Today, SCADA systems are a multi-billion-dollar-per-year industry,andtheneedforwirelessanddistributedsensornettechniquesisgrowing. One effective use of SCADA systems can be seen in oil industry. While the fabled “gusher” produces oil from internal pressure, in most cases this kind of pri- mary production can only extract a fraction (5–10%) of oil in the ground. Today manyolderfieldsdependonsecondary productiontechniques, wherewater, steam, or CO 2 is injected to force out oil, allowing extraction to approach 30–60% of reserves. While such techniques are essential to meet energy demands, the key limiting factor is cost, not technology. The cost of automation needed for effec- tive secondary production guide SCADA and sensornet deployments. Although oil companies have great technical sophistication, solutions as simple as monthly human monitoring of a service are often seen as sufficient and more cost effective than expensive automation. Even with relatively inexpensive hardware, the cost to install the sensor to monitor a pipe can easily top US$10k to power it and tap the pipe. 3.1.1 Sensing needs in an Oilfield with Secondary Produc- tion Most modern oilfields employ secondary production, where water, steam, or CO 2 is injected into the ground to release otherwise difficult to extract oil. In addition to helpingreleasetrappedoil, injectionhelpsmaintainundergroundpressuretoavoid ground subsidence, a potential environmental problem and a source of damage to wells. Whilesecondaryproductionisessentialtoextractingoilinolderfieldswhere thenaturalpressureisinsufficientforprimary(unaided)production,itgreatlyadds to the complexity of the field. 50 Figure 3.1: Steam injection (right) and oil production (left) in oilfield. Figure 3.1 depicts a simplified oil production scenario with steamflood-based secondary production. Steam is produced at a central site, (often a co-generation facility that also provides electrical power) and is distributed throughput the field at high temperature and pressure (250℃ and 5000kPa or more) [But97]. Steam in the distribution network is actually an approximate mix of 70% steam and 30% water. Maintaining this ratio (called steam quality) is important to control injec- tion characteristics. Special devices called splitigator are necessary to maintain steam quality at any pipeline branch. In addition, steam pressure is regulated by a choke, a small, controlled-size hole (about 1cm or more in diameter) just before an injection well. In the ground, steam helps heat oil and bitumen, and provides fluid pressure to release and drive oil to a nearby production well. On the production side, an oil well extracts oil from underground and sends it downstreamtowardstherefinery. Itisimportanttomonitoreachwell’sproduction, both to understand the field’s behavior, and because individual wells are often leasedfromdifferentowners,eachofwhommustbepaid. Insomefields,eachwell’s production is individually instrumented. However, in most production fields, each productionlineisdirectedtoanautomatic well testing (AWT)facilitybeforebeing 51 merged in order to reduce cost. The AWT allows multiple wells to share common monitoring hardware for periodic production audits. Wells also occasionally need to be flushed with steam to remove blockages, so the steam distribution system connects to the AWT, and a production well can be isolated and its flow reversed to inject steam when needed. This brief description highlights the essential role instrumentation plays in an oilfield. Steam quality must be monitored in the steam distribution network; flow rates at injection wells and chokes must be observed; well monitoring is essential attheproductionside; theabilitytoinjectsteaminproductionsystemsmeansthe injection and production sides are cross-linked and must be monitored for leaks. Yetallthisworkmustbeaccomplishedcost-effectively, evenforwellsthatproduce only a few barrels of oil per day, and in fields that have hundreds or thousands of production and injection wells! 3.1.2 Relation to Thesis We studied competitive collaboration in the prior chapter. This chapter supports out thesis by studying the other type, complementary collaboration, in a different problem domain—industrial sensing. Here we find complementary collaboration can reduce sensing cost with cheap sensors but reach high detection accuracy by suppressing false alarms. We believe the high cost is one of the reasons why the oil industry gives up pervasive sensing on steam-choke blockage. Therefore by lowering the cost, multi-sensor collaboration can potentially enable pervasive deployment. 52 We believe complementary collaboration applies to applications beyond steam- choke blockage detection, and even industrial sensing. In particular, complemen- tary collaboration potentially helps those applications where one event may cause simultaneous but spatially-distributed signal change. In this case, a blocked choke causes downstream temperature to drop but upstream relatively unchanged. 3.1.3 Contributions In this chapter we propose an inexpensive sensornet system to monitor steam injection in oilfields. We directly address the cost of current approaches through two contributions. First, we employ non-invasive sensing techniques to detect problems in steam distribution. Current approach to determine steam flow rate typically requires direct measurement of differential pressure within the pipeline. Installation of pressure sensing devices require production halt and piercing of the pipelines; costing thousands of dollars. We, instead, observe that external temperature observation is sufficient to detect problems such as blockage or flow constriction, and with care, even to infer flow rates, provided we can observe at multiple locations. We argue that pervasive industrial sensing requires this sort of non-invasive sensing to reduce deployment cost. Second, we demonstrate a new approach to harvest energy from temperature differential inherent in the phenomena we are studying. We exploit the Seebeck effect induced by heat that is naturally present in the steam injection system to generate enough power for sensor nodes, eliminating any need for external power or batteries. Although many prior systems have demonstrated energy harvest- ing, they have typically been from solar power or vibration. We are the first to 53 exploit heat for large-scale industrial sensing and to show operation solely on har- vested power without batteries operating as buffer. Our approach is important for extended operation where solar is ineffective (for example, the north shore of Alaska) vibration is insufficient, and batteries are not easily replaceable. Our final contribution is to demonstrate that our approach to non-invasive, steam-poweredsensingworksasacompletesystem,throughbothlaboratoryexper- iments and field tests. Although low-power sensing and energy harvesting have been demonstrated before, we are the first to demonstrate an integrated targeted at a new application. To provide this system, we added a custom thermoelectric energy harvesting/conditioning unit and a custom amplification board with cali- brated thermocouples to sense temperature using a standard Mica-2 motes, and developed new detection algorithms that run on this platform. Although we validate our approach with a very specific oilfield deployment, the approaches are applicable to a wide range of industrial sensing. Many indus- trial processes have moderate or large temperature differentials that could support energy harvesting, and non-invasive sensing is important to bring the cost of sens- ing in line with inexpensive communications and computation. 3.2 Target Problem: Blockage at the Steam Injection Choke Toevaluationcomplementarycollaboration, wefirstfocusontheproblemof block- age at the injection-well choke in a steamflood field. We define blockage as the decreasing of the choke’s cross section due to obstructions. Field engineers report 54 that steam-choke blockage is a serious problem in field operation. Chokes are eas- ilycloggedbysmallobjectsbecausetheirsmallboreisanaturalpointofblockage. Sources of blockage occur naturally in a steam distribution system due to scaling andcorrosioninthepipe,buildupofanyimpuritiesormineralcontentinthewater, and aging of the network and choke. Partially or totally blocked choke is a serious problem because it alters the steam injection rate, throwing off field management, reducing production, and potentially eventually causing ground subsidence. Although our current work focuses of blockage at the choke for steamflood fields, we expect that the work also applies to several related problems as well. Other points of operational concern include splitigator operation and AWT mon- itoring; both could use systems similar to ours. We focus on steamflood systems, but we also show that our sensing algorithm applies to waterflood networks (Sec- tion 3.7.5), and we believe that our energy harvesting system could be adapted for different thermal conditions. 3.3 System Overview The goal of our sensing system is to show that multi-sensor collaboration can detect blockages at the choke of steam injection wells, and at a cost much lower than current invasive sensing. We next briefly review the hardware and software we have taken into the field to evaluation solutions to this problem. Figure 3.2 shows our system in the field deployed in March 2010, with a logical diagram on the left and a photograph of the deployment on the right. Each sensor nodeinoursystemincludesamotewithtwotemperaturesensors,athermalenergy harvester, and a wireless network connection. Figure 3.3 shows a deployment with two such sensor nodes and a base station that is connected to the field SCADA 55 (a) Logical view of deployment. (b) Physical view of deployment. Figure 3.2: March 2010 field deployment of our sensing system. system. We review the hardware and software below, and discuss details of this field experiment in Section 3.7.1. By comparison, Figure 3.19 shows as a current invasive pressure sensor; we compare deployment costs of our approach to current approaches in Section 3.8.2. Each sensor node consists of a compute platform based on a Mica-2 running TinyOS-1. Figure 3.3(b) shows a mote packaged for field deployment, and 3.3(a) individual system components and sensors. The sensors themselves are NANMAC 56 (a) A Mica-2, a custom amplifier board, a helio-mote and a hose- clamp thermocouple. (b) Deployedmount,inapelican box with lid open. Figure 3.3: Mote system hardware. J-type thermocouples with hose clamps to attach to the pipe; Section 3.8.1 dis- cussesthecarethatmustbetakentogetaccurate,calibratedtemperaturereadings. Because the voltage output by thermocouples is quite small (less than 15mV), we addacustomamplificationboardtoboostthissignal100-fold. Thewholepackage is powered by a custom-built thermo-electric generator described in Section 3.4. Software on our sensor node includes our new problem detection algorithm (described in Section 3.5). We run the sensing algorithm locally on the mote and report alerts as they occur to the field SCADA system via the base station. In addition, welogtemperatureovertheradiotothebasestation, andlocallytoflash memory for debugging and long-term analysis. In our field experiments we disable logging to flash as described in Section 3.6.3, but in operation, we would expect local logging to serve as backup in case of temporary network outages. The base station should be a devices with wireless communication with the sensor nodes that bridges data into the field network and SCADA system. In principle, a mote with a wired network connection, or a multi-hop mote network 57 could service this purpose. We do not currently have permission to integrate with the field SCADA system, so for our experiments our base station is a mote that connects directly to a laptop that logs data to disk. Although we have built and tested our prototype system, work remains before thesystemcanbefieldedforlongtermoperation. Openissuesincludefullyweath- erizing the packaging and fully integrating it with field-wide SCADA systems. In addition, additional packaging work would be required to insure the system is explosion proof (Class I, Division 1) to be safe for use near production wells. 3.4 Steam-Power: Harvesting Thermal Energy 1 We now describe our motivation and design choices to build an energy harvesting system exploiting the thermal energy present in the oil field’s steam distribution network. This “steam-powered” system will then provide the energy to power the blockage-sensing system we describe in Section 3.3. 3.4.1 The Opportunity The main reason for injecting steam into certain secondary production oil field is so that heat from steam can cause the crude oil to drain into the underground reservoir. The collected oil in the reservoir is then pumped out from the well. The goal of our project is to persistently monitor the flow of steam into the injec- tion sites and report any problems when they occur. While framing our research problem, we observed that heat from steam can be harvested to sufficiently oper- ate our sensors without any external power source. In our system steam-power is 1 Section 3.4 and the work on the thermal harvester is done by Affan Syed. We include this section for the completeness of steam-choke blockage detection. 58 obtained by exploiting the Seebeck effect to generate electricity from a tempera- ture differential. Steamflood pipes operate at around 260℃, while the ambient temperature averages from 0–38℃ over the course of the year at our location, providing a significant temperature differential. An electric circuit converts this temperaturedifferentialintoaflowofchargewithasmallvoltagedifferential;many such circuits in parallel form a thermoelectric generator (TEG) to provide usable voltage and current [Cus,JG06]. While prior work has looked at TEG for auto- motive applications or to augment other power sources, we are one of the first to look at powering sensors in production oil field application. Also, due to efficient TEG, high temperature of the steam, and a large thermal mass of the pipelines, our prototype sensor node is the first of its kind operated directly off the energy harvester without a use of energy buffers such as batteries. We will discuss other TEG related work in more detail on Section 5.6. Although we explore TEGs to monitor a steam distribution network, many industrialapplicationshavetemperaturedifferentialsthatcanbetapped,including many pipeline systems, many systems with engines or motors (even the ones that do not directly produce electricity) and exothermic chemical processes. Although the high temperature differential in our scenario provides significant energy with a relatively small TEG, other scenarios can use larger surface areas coupled with energy buffers. Our system demonstrates the concept of energy sufficiency, beyond energy- efficiency, for ambient-powered sensor networks. Energy-sufficiency argues design- ing ambient-energy harvesting systems that trade-off a more efficient design for lower cost but that generate energy sufficient to sustain the monitoring system. Thus, while thermal-to-electric power generation is inefficient, our cost-conscious 59 design (Section 3.4.2) sufficiently powers our sensing system (Section 3.6). In fact, Section 3.6.3 shows that we can go further to batteryless operation. Before those experimental results, we next describe our basic TE harvester (Section 3.4.2), power conditioning (Section 3.4.3), and the physical mount (Section 3.4.4). 3.4.2 Thermo-electric harvester Design We had several system requirements for our TEG-based power harvesting system to satisfy. First we need a thermo-electric module that works at the 250℃ plus temperature typical of the steam injection pipes. Secondly we want the TEG module to harvest sufficient energy to directly power mote-class devices. Lastly, wewantthepower-harvestingmoduletobelow-costtoengenderdensedeployment. To satisfy the first two requirement we choose the 1261G-7L31-04CQ thermal power generation module from Custom Thermoelectric [Cus]. This module has a maximum temperature rating of 260℃ and, under ideal conditions, is rated to generate up to 5.9W. With transmit power at highest gain (+10dBm) reported to be 82mW [CK06] we should have sufficient energy a mote-class device. We do not use active cooling, or any buffer energy in batteries, to minimize the cost of deployment and maintenance. The cold side of our TE-harvester mates to analuminumheatsink,measuring5 3 / 8 ×5 3 / 8 ×1 3 / 8 inches. Whileweapplyarelatively cheap thermal paste on the cold side, for heat transfer to the heatsink, we do not use any thermal paste on the hot side of the TE-harvester due to its high cost. In Sections 3.6.1 and 3.6.1 we show that our efficiency-cost trade-off in TEG design is sufficient to meet our application needs. 60 3.4.3 Power Conditioning the TEG Output Weaddpowerconditioningcircuitrytoadaptelectricityfromtheenergyharvester to voltage and current usable by our sensor node. For this purpose, we chose the heliomote board designed to power motes using solar power at UCLA [KHZS07] for an off-the-shelf solution. Under normal oper- ation, the heliomote board trickle charges two NiMH AA batteries from its solar panel input and regulates power to a sensor node. Given the Heliomote as the power conditioning unit of our system, our first challenge was to adapt it to work with the thermo-electric power. Since electrical property of a typical TEG drastically differs solar panels, simply replacing it with TEG did not work. The heliomote needs an input voltage larger than 2.4V to charge the batteries that then supply the energy to be conditioned to Mica-2. While a solar panel can produce this voltage, a typical TEG generates much lower voltage. We considered charging a single battery at 1.2V, but some IC’s in the heliomote were unable to operate at that voltage. Finally, we configured one of our heliomote unit such that it directly regulates the output voltage of the TEG to 3.0V. With this modification, the TEG directly powers the Mica-2 and thus needs to provide all instantaneous power requirements. We verify this capability in Section 3.6.3. A final concern is the system behavior during the times when the thermo- electric power is insufficient for our prototype system run properly. (for example, see Section 3.6.3). To gracefully handle brownout, we enable the brown-out detec- tion on the Atmega128 processor to proactively shutdown if the voltage drops below a user specified threshold. 61 2.375” OD Insulation" Heat Sink Base block 2.75" height Hose clamps Extender block 1.3" height (a) Design of the TEG pipe-mounting appara- tus (b) Deployed mount, without extender block Figure 3.4: Mounting design for TEG. While our application scenario provides continuously available thermal energy and directly powering our system is acceptable, we are currently also looking into using capacitors (super- or otherwise) to ride any load transients (Section 3.6.3). Weplantolookatmodifyingexistingpowerregulationcircuitdesignedspecifically for thermal power [MCL + 07] and incorporate design into our thermocouple board. 3.4.4 Mounting Design A final step for the deployment of the thermal energy harvesting system is the pipe mounting design. We want a robust and stable design harvesting sufficient energy to power motes. However there were several constraints regarding the pipe mounting. First, the outside pipe diameters in many oil fields vary between 2 3 / 8 and 3 1 / 2 inches. Second, most steam pipes are insulated to prevent heat loss. While we have permission to remove the insulation at locations of interest, the insulation thicknessvariesbetween2to4inchesatdifferentlocations. Sincewewanttheheat 62 sink to be exposed to airflow, we prefer it is located above the surrounding insu- lation. Third, we want a mechanism to securely clamp our system to the heated pipe. Our solution to the above constraints is shown in Figure 3.4. The base block of our TEG mounting system is curved to match the exact pipe diameter at our target site (2 3 / 8 inches). This base block also has a height of 2 3 / 4 inches, to allow it to extend past an expected 2inches of insulation. A separate additional block can be added to accommodate deployment to locations with thicker insulation. Finally, the base block also has curved grooves going through its one side allowing us to pass two hose-clamps on either side and clamp the entire apparatus securely to the pipe. 3.5 Non-Invasive Sensing of Pipeline Blockages Wewanttoproveoneofthetwocollaborationtypes—complementarymulti-sensor collaboration helpssteam-choke blockage detection, as part of ourthesis. Westart withthephysicsbackgroundaboutwhyablockageeventuallyreducesdownstream temperature while upstream remains. We next describe the design of a base algo- rithm and show how multi-sensor non-invasive, temperature-based sensing can detect blockages with pairwise upstream and downstream sensors. After that, we identifyproblemsinourbasealgorithmandfurtherdesignextensionstoavoidfalse positives in detection. 63 3.5.1 Background: pipeline physics Our hypothesis is that pipe surface temperature can indicate internal steam-choke blockages. In this section we summarize the physics of fluid flow in the pipe to show how a blockage decreases downstream pressure, which in turn decreasing surface temperature, a phenomena we can detect. To understand what happens in the pipe, we must understand what happens when supersaturated steam passes through the choke (see [PG84] and [Cra88] for general background). The choke is an intentionally narrow opening in the pipe (a choke bean) designed to keep steam at critical flow, where the fluid reaches sonic velocity, effectively isolating pressure upstream and downstream of the choke [Cra88]. This isolation is essential for oilfield operation, since down- stream and downhole conditions may vary, and also our algorithm since we can observe temperature differences on upstream and downstream of the choke (T U and T D ) to detect blockage. As we described in Section 3.2, scaling inside the pipe, steam impurity, and device wear can all cause blockages, which change the cross-sectional size (A) of the choke. We write this change as: A ′ <A where A ′ indicates the value after a blockage occurs. The volume of steam passing inaunittime( ˙ m,themass flow rate)isdeterminedbythechokeaperturesize,soa partial blockage reduces steam volume. The Thornhill-Craver choke rate equation shows mass flow for straight-bore chokes [CD46]: ˙ m = 73YA 1− 0.00625L √ A ! q ρP U 64 Theflowratedependsongasexpansionfactor(Y), aperturesize(A), chokelength (L), upstream pressure (P U ) and steam density (ρ), calculated by vapor-phase and liquid-phase specific volumes. Field operations keep ρ constant during normal operation. By the definition of choked flow, P U is constant as well. From the last flow rate equation, we see that a partial blockage (d ′ <d) reduces flow rate: ˙ m ′ < ˙ m Since steam is compressible, a decrease in mass flow decreases pressure [PG84]: P ′ D <P D Alowerdownstreampressurereducesinternalsteamtemperature(T D,i )andthere- fore pipe surface temperature. Experimental data shows this relationship with internal temperature [GA96], as shown by the the following empirical equation provided by field engineering: T D,i = 7006.3 9.48654−ln P D 144.9 −382.55 (3.1) Sinceweknowthatsurfacetemperaturesfollowinternaltemperatures(T D ∝T D,i ), drop in pressure implies drop in temperature: T ′ D <T D 65 and upstream temperature and pressure are not changed (T ′ U ≈ T U , because it is choked flow), so we can therefore detect blockage by looking for relative tempera- ture differences: T ′ U −T ′ D >T U −T D The above reasoning suggests why choke blockage is visible in our system. However,oilfieldsarecomplex,andchokeblockageisnottheonly possiblecauseof pipetemperaturechanges. Weatherchangesonthesurface,anddownholepressure changes are both potential sources noise. Our detection algorithm (Section 3.5.2) triggers on sudden and relative temperature differences, so it should not trigger on surface changes that affect both sensors (such as weather, since the relative differencesbetweenthesensorsareunchanged),orgradualdownholechanges(such as reservoir changes, since they take place over days or weeks). In this section we summarized how blockage eventually reduces downstream pipe skin temperatureand we provide theoretical and empirical equations to prove that. Griston et. al. observes similar phenomenon that smaller choke bean size doesnotaffectupstreamtemperaturemuchwhilesignificantlyreducesdownstream temperature in their experiments [GA96]. These results are consistent to our hypothesis that we can use temperature to detect choke blockage remotely. This background is used in our algorithm design to provide good detection accuracy (Section 3.7.2). 3.5.2 Design of the Base Algorithm Wenextapplyourobservationofpipetemperaturetodetectchokeblockage. Since blockages represent changes in flow behavior, the principle of our algorithm is to look for changes in temperature upstream and downstream of a possible point 66 of blockage, then look for short-term changes in temperature. Our algorithm adaptively learns typical pipe temperatures from long-term averages. We study two kinds of blockages, each with its own inference derived from choke physics. For a partial blockage, upstream pressure is unchanged while downstream pressure drops as the orifice size is reduced. For total blockage, upstream pressure drops as flow stagnates, and downstream pressure drops to ambient, in-pipeline pressure. In both cases, we measure temperature to infer pressure, detecting blockage because of a sudden drop in downstream temperature relativetoupstream. FromourexperimentalresultinthebottomplotofFigure3.9, we see that whenever blockage happens, upstream temperature (the dotted line) is relatively stable, although it dips slightly upon full and near-full blockage. The downstreamtemperature(thesolidblueline)ismuchmoresensitivetopipestatus, and the temperature differential (∆ ud , the wide red line) shows ten distinct peaks. We next show experimental results to illustrate this hypothesis. We present a detaileddescriptionanddiscussionofourfieldexperimentsinSection3.7.1. Letters (F, NF, O, P) in Figure 3.5 indicate the degree of emulated blockage, and vertical dotted lines in the figure show points when we change blockage level. We see that upstream temperature (the dotted line) is relatively stable, although it dips slightly upon full and near-full blockage. The downstream temperature (the solid blue line) is much more sensitive to pipe status, and the temperature differential (∆ ud , the wide red line) shows ten distinct peaks corresponding to ten controlled full or partial blockages. Full analysis of this experiment is Section 3.7.1. Ouralgorithm(pseudocodeinAlgorithm3.1,withnotationsinTable3.1)works in two steps. We first compute the short- (s(∆ ud )) and long-term (l(∆ ud )) history of the difference between upstream (T u ) and downstream (T d ) temperature via 67 180 200 220 240 260 280 . temperature ( o C) 0 3000 6000 9000 12000 15000 18000 −20 0 20 40 60 80 time (seconds) difference ( o C) . O F O NF O P O F O NF O P O F O NF O P O grad. O T u T d Δ ud Figure 3.5: Upstream and downstream temperatures (top lines and left scale) and ∆ ud inseveralcontrolledblockages. Avalveemulatesthefollowingblockagelevels: F: full, NF: for near-full, P: partial, and O: open (no blockage). exponentialweightedmovingaverage. WechooseEWMAbecauseitislightweight and easy to implement on our 8-bit mote platform; short- and long-term EWMAs may use separate gains (α s and α ℓ ). Whenever the difference (δ sℓ ) between two history exceed pre-defined threshold (th block), the system declares pipe blockage. After a blockage is detected, we expect responders to investigate the problem and reset the algorithm after it is corrected. By far our system does not transmit alarms back to field surveillance room, which is out of scope of this work. Since in our field tests (Section 3.7.1) we artificially induce blockages rapidly (in tens of minutes), we employ two addition rules specifically to aid testing. First, after blockage detected, when two history series converge (δ sℓ = 0), we automatically reset pipe state in order to precede to follow-up tests. Second, we stop updating the long-term history when the pipe is in any non-normal state. While these rules were designed to allow our short-term tests to mimic long-term operation, they do not negatively affect normal operation. 68 Table 3.1: Notations used in the description and analysis of blockage detection algorithms T u ,T d Temperature at Up- and downstream ∆ ud Up- and downstream temperature difference s(∆ ud ),l(∆ ud ) Short- and long-term ∆ ud history th norm normal state threshold th block blockage state threshold th maint upstream maintenance state threshold α s ,α ℓ Short- and long-term EWMA gain δ sℓ Short- and long-term history disparity s pipe status Our algorithm successfully detects pipe blockage, as shown experimentally in Section 3.7.1. Our basic algorithm detects problems using two sensors on either side of the blockage site. These sensors could be network-connected, or directly connected to a common controller (as in our implementation). Although this algorithm is correct, upstream maintenance can cause false alarms when pressure for the entire steam system drops. The next section shows how we can employ networked sensors to avoid these false alarms. 3.5.3 Avoiding False Positives The above algorithm detects blockages around the target, but regular steam dis- tribution maintenance also changes system pressure and temperature. Our base algorithmisunabletodistinguishmaintenancefromchokeblockage,thusincurring false positives. We avoid false positives by employing networked sensor readings with different threshold. We avoid false positives by employing networked sen- sor readings from more distant parts of the steam distribution system. While we could, in principle, record scheduled maintenance events and explicitly disable 69 Algorithm 3.1 Blockage detection algorithm Input: T u ,T d ,th block,th maint,th norm,α s and α ℓ . Output: Pipe state s. 1: s← NORMAL 2: while system on do 3: ∆ ud ←T u −T d 4: s(∆ ud )←s(∆ ud )+α s ×(s(∆ ud )−∆ ud ) 5: l(∆ ud )←l(∆ ud )+α ℓ ×(l(∆ ud )−∆ ud ) 6: δ sℓ ←s(∆ ud )−l(∆ ud ) 7: if (δ sℓ ≥ th block)∧(s = NORMAL) then 8: s← BLOCKAGE 9: print “Pipe blocked” {*** below are extensions from Section 3.5.3.} 10: else if (δ sℓ ≤ th maint)∧(s = NORMAL) then 11: s← MAINTENANCE 12: else if (δ sℓ ≥ th norm)∧(s = MAINTENANCE) then 13: s← STABILIZATION 14: start timer 15: else if (timer fired)∧(s = STABILIZATION) then 16: s← NORMAL 17: end if 18: end while blockage detection during those times, we strongly prefer a sensor-based solution. If we can infer maintenance at the sensors, we avoid dependencies on manually logged events and the system integration and additional errorconditions that such coupling entails. We can also adapt to environmental changes such as seasonal temperature drift. We next describe how we extend our base detection algorithm to distinguish system-wide changes from local blockages. ThelowerplotofFigure3.6showsthedetailfromthesecondtrialofFigure3.10 to illustrate how our extension solves a potential false alarm. We distinguish upstream maintenance from blockage by detecting both the start and completion of a maintenance period. We decide maintenance starts if δ sℓ < 0 because of the inertiaatthechokemakingtheupstreamtemperaturedropbeforethedownstream 70 −120 −90 −60 −30 0 30 60 temperature ( o C) s b / n n u δ sl th_block th_maint 3500 4000 4500 5000 5500 90 120 150 180 210 240 270 time (seconds) temperature ( o C) O NF O T d 2 T d 3 Figure 3.6: A cause of false positives and our solution. n, u or s means the pipe is in normal, upstream maintenance or stabilization state respectively. A b / shows a false blockage detection suppressed. (solid vertical line in the upper plot at 3800s). Likewise, we detect the stop at the next δ sℓ > 0 since the same inertia causes a reverse process (time≈ 4700s). Our extended algorithm then gives the pipe stabilization time, depicted in the upper plot as dotted shading area, suppressing any following potential δ sℓ peaks that would otherwise be misinterpreted as blockage (for example, the crosses at 4710s). We detect blockage by any δ sℓ that larger than th block but falls out of the stabilization period (i.e. not preceded by a sign of maintenance start). Thiskindoffalsealarmalsoshowswhywerequiremultiplesensors. Ourdown- stream sensor alone can not distinguish the temperature drop incurred by block- age and maintenance. A pair of sensors, on the other hand, can distinguish these events, and also can adapt to environmental changes. In Section 3.7.2 we show that this false-positive elimination algorithm successfully distinguishes upstream maintenance from blockage. 71 3.5.4 Tuning for different environments Our algorithm has several parameters that require tuning. We next discuss the parameters and how we can tune them to support not only steam blockages, but also blockages in hot-water distribution networks. In Section 3.7.5 we show that, with proper parameters, our work generalizes to blockage detection in other types of pipeline networks. Currently we configure our system manually; automatic configuration is an area of future work. The detection thresholds (th block, th maint and th norm) are critical to trade between accuracy, responsiveness and reliability. We assume ∆ ud follows a normal distribution N(μ,σ 2 ). Usually th block is set higher than 3σ, according to 3-sigma rule [Wes56]. The water in PVC pipe has lower temperature and hence we observe less significant ∆ ud variance upon anomaly. To make accurate detection, we set both th block and th maint closer to 0. The th norm parameter should be set to a small enough value to ensure hysteresis in our algorithm; we set it by default to 0. Short and long-term gains (α s and α ℓ ) determine how our algorithm reacts to noise and blockage. Long-term history should be relatively stable while short- term agile. Pipe material (metal or PVC), fluid type (steam or water), and the ambient environment all affect how quick the temperature reacts to changes in pipe status. With PVC and water, the pipe has better heat insulation than metal, and hence we want to keep long-term history more stable because of the sluggish short-term change. We therefore we propose 1 / 2 for short-term EWMA gain, and 1 / 64 for long-term in steam blockage detection, while 1 / 4096 for water pipelines. Finally, the stabilization time period helps avoid false positives. We find that a 360s timer suppresses most noise due to upstream maintenance. The reason is that it takes δ sℓ about 950s to subside, and with α ℓ of 1 / 64 and peak usually 72 (a) TEG power measurement harness 0 0.5 1 1.5 2 2.5 3 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Voltage (V) Power (W) 80 o C 65 o C 46 o C (mote−class requirements) 82mW (b) TEGpowercurvesatdifferenttemperaturedifferentials. Figure 3.7: Power measurement of the TEG: With optimal loads, but depending on temperature differential, our TEG can provide between 0.3 to 0.8W. occurs within the first half of the period. For water pipelines, with a smaller α ℓ and non-metallic pipe, we set a longer timer because it takes accordingly longer to stabilize. 3.6 Long-term energy harvesting and consump- tion evaluation 2 To evaluate self-sufficient energy harvesting, we first evaluate the capability of the TEG to generate energy compared to the power consumption of our sensing system, concluding that batteryless operation is possible. 2 Section 3.6 and evaluation of the thermal harvester is done by Affan Syed. We include this section for the completeness of steam-choke blockage detection. 73 3.6.1 Energy Production In the lab: To measure the energy production of our TEG sandwich, we built a prototype TEG with a flat surface that is placed on a laboratory hotplate (Ther- molyne 900, rated to 260℃). We place the TEG on a small aluminum block about 3.8cm above the hotplate and use a small external fan to maintain a constant airflow over the heatsink. To measure power generation we vary the load offered to TEG using a high- wattageresistornetwork. Wethenmeasuretheoutputvoltageandcurrentsourced by the TEG as we vary both the load and hotplate temperature (Figure 3.7(a)). (Wecannotdirectlymeasurecurrentbecausemeasurementisdifficultwithitslarge dynamic range, from 3 to 800mA.) Figure 3.7(b) shows power characteristics of our prototype TEG in the lab. First, we observe that under most operational conditions (temperature and load), our prototype provides more than the the 82mW needed to power mote-class devices [CK06]. Even after a blockage when pipeline temperature falls, the 180℃ temperature (Figure 3.5) provides sufficient power. Second, we verify that TEG delivers 0.3–0.8W at the optimal load corresponding to 1.2V. However, since our actual system does not necessarily operate at optimal load for a thermal harvesting source, we next confirm system operation with in-field experiments. In the field: We perform in-field tests to confirm that our fielded TEG mount (Section 3.4.4) reproduces laboratory experiments, and to evaluate real oilfield conditions, where ambient temperature can reach 50℃ in summer. Our sensing platform was under development when we carried out our initial harvesting experiments (December 2009), so we log the open-circuit voltage of our TEG and spot-measure temperature. We find a nearly constant temperature 74 differential of 100℃ and open circuit voltage of 3.8V. This observation verifies that our pipe-mount TEG can harvest energy comparable or even more than the lab prototype, prompting full system tests described next. 3.6.2 Energy Consumption of Sensing Withourunderstandingofenergygeneration,wenextturntoenergyconsumption. Our steam-sensing systems employs a Mica-2 mote with an additional hardware for power conditioning (Heliomote) and thermocouple signal amplification. We therefore measure the power draw of our system running the blockage detection algorithm. Our results show that the average power draw is 70mW, similar to previous measurements of motes that include radio transmission (82mW [CK06]). ThesevaluessuggestthatourTEGcaneasilycoverlong-termenergyrequirements and power our system. Given the large headroom of harvested power, one possible future work is to reduce the size of harvesting unit and hence the total cost of our system. We next evaluate the instantaneous power requirements of our system to understand if batteryless operation is feasible. 3.6.3 Batteryless operation? WehaveshownthattheTEGshouldprovidesufficientenergyforlong-termopera- tion(Section3.6.1). Giventhelargeamountofheadroomshownthere,weexpected that batteryless operation would be straightforward. In fact, we have confirmed that our system can successfully sense and communicate batteryless, both in the lab and in the field. 75 (a) Logging to flash (erase-write) causes two load spikes. (b) At temperature differentials below 80℃, spikes cause the mote to fail and reboot. Figure 3.8: Instantaneous load can cause failure. However, we found that some debugging modes of our system require high short-term power that cannot be provided by harvested energy alone. Specifi- cally, writing to flash has peak current draws that starve the CPU (peak power of 260mW, as shown in Figure 3.8(a)), causing the mote to reboot. TEG power is a function of ∆ HC and the absolute temperature; we observe power shortages only for smaller ∆ HC and at lower absolute temperatures. We found brown-outs occur at ∆ HC around 80℃ (dynamic temperatures are difficult to measure, we estimate measurement accuracy at around±5℃). Thus we conclude that short- term, power-intensive operations like flash require significantly additional energy generation headroom for batteryless operation. Adding minimal energy buffering: A small energy reserve can bridge brief peak power requirements. We therefore evaluate traditional capacitors (1000 to 9000μF) to support flash logging, while avoiding the maintenance problems of batteriesandthecostofsupercapacitors. Wecanusethecapacitorbothtotolerate flash logging in the field, and provide sensing-and-transmit operation at lower, in both cases operating at smaller ∆ HC . 76 Table 3.2: Energy buffering test at TEG ∆ HC = 83.1℃ capacitor System operation status none sensing and radio fine, but always reboots upon flash logging 1000μF sensing and radio fine, but reboots after 2 packets flash logging 3300μF sensing and radio fine, but reboots after 12 packets flash logging 4300μF sensing, radio, and flash logging always correct 9900μF sensing, radio, and flash logging always correct Wefirstadda1000μFcapacitorandlower∆ HC untilthemotestops. Without a capacitor our system reboots at around 80℃ ∆ HC , and the mote will not boot at all at 50℃ ∆ HC . With the capacitor we are able to sense-and-transmit as low as 60℃ ∆ HC . Thus a small capacitor supports operation at about a 20% lower temperature differential. We next vary the capacitor size to see how much energy buffering is required at around 83℃. Table 3.2 shows the correlation between capacitor and system robustness. The scan interval is 10s, the same as that of our field deployment and we do one radio transmitting and flash logging at the end of every cycle. We find that larger capacitors support power-intensive flash logging, with a 4300μF capacitor sufficient to support continuous operation at this ∆ HC . (There is ample time for the capacitor recharge between 10s cycles, much larger than the sub- second capacitor recharge time.) We conclude that while batteryless operation is possible, a small energy buffer is important to support peak loads. 77 3.7 Steam-ChokeBlockageDetectionEvaluation This section is to study how complementary multi-sensor collaboration improves event detection. We use steam-choke blockage as a case study, since it is another typical event detection application, other than vehicle classification. We first designed a multi-sensor non-invasive algorithm for choke blockage detection in the previous section. To test whether algorithm works and quantify the improvement in detection accuracy, we implemented it on a sensor platform and deployed our systeminaproducingfieldinMarch2010. Thetwosensorsiteseachconsistedofa Mica-2 mote running TinyOS-1.x with two temperature thermocouples, a thermal energy harvester and a wireless network connection. Our base station potentially can bridge sensor nodes to SCADA system. However, we do not currently have permission to integrate with the field SCADA system and hence for our experi- ment our base station is a mote that connects directly to a laptop that logs data to disk. Hill et al. show an infrastructure of event-detection sensor networks in oil field which might serve as the backend of our system [HCCI08]. One sensor site straddled a steam choke while the other did a valve (Figure 3.2(a)). We next show the evaluation on our base and extended algorithms and how our algorithm generalizes to another water pipeline environment. 3.7.1 Does Our Detection Algorithm Work? We have described our blockage detection algorithm using multi-sensor, non- invasive algorithm. We next review our methodology and present data to show that our algorithm accurately detects full, near-full and partial blockages. Since it is difficult to non-invasively induce controlled blocks in a real choke, and such actions might interfere with production, field engineers manually control 78 a valve to emulate blockages. We believe this emulation models a small orifice restricting steam flow, the key similar physical property, and so we expect our results here to apply to real choke blockages. We emulate four blockages levels: full (flow rate ˙ m = 0), open ( ˙ m = 100%), near-full blocked ( ˙ m≈ 10%), open, partially blocked ( ˙ m≈ 50%) and open again; the blockage approximate are best estimates by the field engineers. We repeat the procedure three times and then slowly but continuously shut off the pipe over 9 minutes to observe a gradual blockage. These events are shown by letters and vertical lines in the lower plot of Figure 3.9. We collect data with one mote controlling two thermocouples straddling the valve to measure T u and T d , as indicated in Figure 3.2(a) and seen in the top left of Figure 3.2(b). We also deploy a second mote and pair of thermocouples around the actual choke for use in our extended algorithm. A nearby laptop with a mote receiver archives transmissions from both motes. For ground truth, we use aCampbellScientificCR1kdataloggerwitheightNANMACD60-60-Jhose-clamp thermocouples, each thermocouple placed adjacent to a mote thermocouple. We operatedoursystemfrom12:30pmMarch4 th to8:30amMarch5 th , 2010, collecting data for the first twelve hours as described next. Our field experiment is not problem-free. Our experiment ran for 20 hours, however we collected data for only the first 12 hours. At this time our sensors did not have capacitor assist, so we disabled on-mote flash logging and depended on transmission and logging at the central laptop. Unfortunately, external power for the laptop failed overnight, so we lack data for the last 8 hours. Our stored dataincludesallcontrolledexperimentsandissufficienttovalidateouralgorithms. Also, analysis of timestamps the next morning confirms our batteryless prototype 79 −40 −20 0 20 40 60 80 temperature ( o C) δ sl th_block th_maint 0 3000 6000 9000 12000 15000 18000 160 180 200 220 240 260 280 time (seconds) temperature ( o C) F NF P F NF P F NF P grad. T u T d 1 Figure 3.9: Similar base and extended algorithm results on the thermocouple- pair straddling the valve. The bottom plot shows the raw up- and downstream temperatures with pipe status mapping. operated continuously. Finally, although we ran our algorithms live, an electrical coupling error between the thermocouples at the valve spoiled our live run of the algorithm. Fortunately,ourground-truthtemperaturesensorsinthesamelocation recorded the complete data. We therefore replay this data post-facto for the anal- ysis presented in this section. Our sensor testing and calibration (Section 3.8.1) suggests that this substitution does not alter our conclusions. Due to the limited fieldexperimenttime,weusethesamesetupfortrainingandevaluation. However, our evaluation is reliable because our water experiment uses a different setup and reports consistent result (Section 3.7.5). Figure 3.9 separates δ sℓ and shows where it crosses the thresholds to indicate detections (th block = 15,th maint =−16, α s = 1 / 2 and α ℓ = 1 / 64 ). It shows that our algorithm correctly detects all three levels of the blockages, capturing all nine δ sℓ peaks. It correctly detects the final gradual blockage as well. 80 We draw four further observations from the bottom plot. First, consistent to our hypothesis, the upstream temperature is relatively constant while the down- stream one is sensitive to pipe status. Second, surprisingly, nearly-full blockages yield the larger downstream temperature drops, even more than full blockage, while partial blockage has the smallest difference. We believe this behavior is because a nearly closed valve starts choking flow, reducing downstream pressure, but a full blockage completely isolate downstream pipe from the upstream steam network, leaving it occupied by back pressure from reservoir through wellbore. Field engineers confirm this intuition. Third, our algorithm is highly responsive. For example, it takes averagely around 6 samples (60s) to correctly detect prob- lems when th block = 15. Finally, carelessly configured parameters would trigger false alarms. An over-aggressive th block value (≤ 5℃) would confuse normal temperature fluctuation for pipe anomaly. In principle, one could measure typical temperature variation, for example by making sure that the threshold well outside typical variance. In all, we conclude that our base algorithm is capable of detecting three differ- ent degrees of blockage, detecting nearly-full and full-blockage most easily, with the same threshold setting. However, when we run our algorithm over the thermo- couple pair straddling the choke (data omitted due to space), our base algorithm triggersfalsealarmsaswediscussedinSection3.5.3. Thesameblockagethreshold captures seven out of ten positive a δ sℓ peaks in Figure 3.10. We next evaluate how our extended algorithm can avoid the choke false positives. 81 3.7.2 Evaluating Avoidance of False Positives In Section 3.7.1, we show that our base algorithm has good accuracy on emulated blockages. However, application of our basic algorithm to readings of our second sensor-pair (T 2 d and T 3 d around the choke in Figure 3.2(a); downstream of the sensor-pair T u and T 1 d around the valve) shows a number of false alarms, even though there were no blockages at that location. In effect, our experiments at the sensor-pair around the valve emulate maintenance on the steamflood network, changing the pressure for all downstream sensors. We next re-analyze the data from both sensor pairs sensor to show that our extended algorithm in Section 3.5.3 prevents these false alarms. We expect the extension works at both locations, successfully detecting target blockage at the valve and suppressing our the maintenance-like effects at the choke. The extended algorithm yields exactly the same result as the base one over the valve-straddling pair with the same configuration. This result proves that the extension does not impair our algorithm performance. Figure 3.10 shows how the extended version avoids false positives at the choke- straddlingsensorpair(T 2 d andT 3 d ). Thebasealgorithmwiththesamethresholdas before (th block = 15) captures seven out of ten positive δ sℓ peaks upon upstream valve operation because of the asynchronous temperature drop on both sides (Sec- tion3.5.3). However,theextendedversiontriggersnofalsealarmsanymore(seven b / tags). With th maint, the system detects all ten upstream maintenance events (tagged as “u”). The stabilization period (dotted stripes shading tagged as “s” in the upper plot) successfully suppresses all positive δ sℓ peaks following upstream maintenance and δ sℓ rises back above 0. 82 −120 −90 −60 −30 0 30 60 temperature ( o C) s b / s b / s s b / s b / s s b / s b / s s b / n u n u n u n u n u n u nu nu n u n u n δ sl th_block th_maint 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 90 120 150 180 210 240 270 time (seconds) temperature ( o C) F NF P F NF P F NF P grad. T d 2 T d 3 Figure 3.10: Extended algorithm result on the sensor pair straddling the choke with th block = 15,th maint =−16, α s = 1 / 2 and α ℓ = 1 / 64 . The bottom plot shows the raw up- and downstream temperature with pipe status mapping. Our field tests carefully evaluate our extended algorithm in two different scenarios—an emulated blockage, and an emulated, upstream maintenance event. These experiments demonstrate that our extended algorithm yields good accuracy and triggers minimum false alarms. 3.7.3 Parameter sensitivity of the basic sensing algorithm Section 3.7.1 and Section 3.7.2 have shown that our algorithm is able to detec- tion steam choke blockage under certain parameter settings. We next study how sensitive those results are to the parameters settings, to learn how tolerant the application is to different situations or potential misconfiguration. Field engineers help us to turn valve to emulate different blockages following this procedure: 1. let pipe fully open for at least 10 minutes. 83 0 0.33 0.66 1 Accuracy 0 5 10 15 20 25 30 35 0 1 2 3 4 5 6 false alarm occurrences th_block ( o C) full nearly full partial false alarm Figure 3.11: Accuracy and false alarm at different th block (α l = 1 / 16 ). 2. completelyshutoffthepipetomockupacompleteclogforabout10minutes. 3. fully open-up the valve for 10 minutes to let pipe temperature subside. 4. turn the valve to nearly-fully shut-off and wait for another 10 minutes. 5. repeat step 3. Open up the valve. 6. turn the valve off moderately, less than step 4, to simulate a partial clog. Again, wait for about 10 minutes. 7. repeat step 3. Open up the valve. Figure 3.11 shows our algorithm performance with different upper threshold (th block). Short- and long-term EWMA gains are fixed to 1 / 2 and 1 / 16 respec- tively for this analysis according to Section 3.5.4. We configure α l < 1 / 64 to reset the long-term history to normal level between any two controlled blockages. We have examined choices both long- and short-term gains (evaluation omitted due to space) and believe that these parameters are suitable for our application. We run respectively three trials for each type of blockages and we observe false alarms under th block≥ 2. 84 −60 −40 −20 0 20 40 60 ¯ α = 1 16 δ ( o C) full block near−full partial open δ pipe−open −60 −40 −20 0 20 40 60 ¯ α = 1 32 δ ( o C) full block near−full partial open δ pipe−open −60 −40 −20 0 20 40 60 ¯ α = 1 64 δ ( o C) 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 x 10 4 full block near−full partial open time (1s) δ pipe−open Figure 3.12: δ fluctuation influenced by different α l . We draw two observations from Figure 3.5. First, carelessly configured param- eter triggers false alarms. An over-aggressive th block value (≤ 5.3) would mistake normal temperature fluctuation for pipe anomaly. However, one can easily avoid theproblembytakingthepipetemperatureintrinsicpropertiesinaccount,namely mean and σ. Finally, our algorithm is not sensitive to long-term ∆ T history gain (α l ),aslongasitisreasonablysmall. Figure3.12showsthatlargerα l yieldshigher δ peak, because long-term history are more sluggish, but also takes more time to subside when pipe back to normal. However, according to those δ peak values, it is easy to optimize parameter set for all three α l ’s. In all, we conclude that our algorithm is sensitive to detection threshold but not to history gains. However, parameter configuration is still easy and we find a range of parameters to yield 100% accuracy and limited detection delay within limits. 85 0 5 10 15 20 25 30 0 0.2 0.4 0.6 0.8 1 Accuracy 0 5 10 15 20 25 30 0 2 4 6 8 10 FP occurrances th_block ( o C) TP: full TP: nearly full TP: partial FP: blockage FP: maint. Figure3.13: Evaluationoftheextendedalgorithmwithavaryingth block running between T u and T 1 d for choke blockage detection. th maint is fixed at 16℃. 3.7.4 Parameter sensitivity of the extended algorithm Section 3.7.2 shows that our extended algorithm successfully avoids false alarms caused by pipe maintenance. Similar to the sensitivity evaluation of the basic algorithm (Section 3.7.3), we next study the extended algorithm’s robustness to parameter variations. We apply our algorithm to two different sensor pairs and evaluate blockage and maintenance threshold separately. WefirstevaluateourimprovedalgorithmparametersensitivitybetweenT u and T 1 d , straddling the valve for blockage detection scenario. Our hypothesis is that the new algorithm should performs similar to the base algorithm, or better. We fix th maint at 16 to simplify the analysis, since this parameter is not critical for blockagedetection(Likewise,wesetth block as16inthenextscenarioevaluation). This number is not randomly chosen, we prove that it works well for maintenance detection and is good for our algorithm robustness test. Figure 3.13 shows that generally, the algorithm tweak does not affect the detection of this scenario com- paring to Figure 3.11, although we have some maintenance false alarms. We still find a range of th block, [8, 18], which yields perfect detection—100% detection accuracy for all three level of blockage and zero false alarm occurrence. Another 86 similarity is that larger threshold makes it harder to trigger detections and the system fails all detections if th block≥ 27. We separate two kinds of false alarms—false blockage and false maintenance into two dotted lines in Figure 3.13 and discuss them here. Similar to the base algorithm, an over-aggressive (small) threshold is likely to mistake signal series jit- ters for blockage anomaly. But a reasonable parameter configuration, th block≥ 8 easily eliminates the problem. The occurrence of maintenance false positive is because of the asynchronous history variation. Upon blockage, both short- and long-term history would rise in different speed but parameter misconfiguration (th block≥ 22) may prevent the system detects at the first time. After a certain time, if the anomaly settles, the drop of ∆ T would make short-term history drops below long-term history. This δ valley would then be captured by the system and report as maintenance. WenextevaluatethealgorithmonT 2 d andT 3 d ,straddlingtheactualchoke. Our hypothesis is that since the valve is far upstream to the both spots and we use it for blockage simulation, our algorithm should be able to distinguish this anomaly from choke blockage. Figure 3.14 shows that the algorithm successfully detects the maintenance scenario at reasonable false alarm rate. The first observation is a smaller perfect detection parameter range—[15, 17] of th maint, comparing to the previous case. However, we believe this problem is fundamentally harder than choke blockage scenario for two reasons. First, for maintenance the inertia of the choke cause more unpredictable ∆ T variations, which is the purpose of the choke—regulating and separate downstream pressure from the upper one. The other is that valve is relatively farther to the two sensors and we believe distance somehow absorbs the valve operation impact and incur more noise to the data 87 0 5 10 15 20 25 30 35 40 0 0.2 0.4 0.6 0.8 1 Accuracy 0 5 10 15 20 25 30 35 40 0 2 4 6 8 10 FP occurrances th_maint ( o C) TP: maint. full TP: maint. nearly full TP: maint. partial FP: blockage FP: maint. Figure3.14: Evaluationoftheextendedalgorithmwithavaryingth maintrunning between T 2 d and T 3 d for maintenance detection. th block is fixed at 15℃. (briefly discussed in Section 3.8.3). Therefore, parameter misconfiguration, say th maint≤ 7 or th maint≥ 24 triggers a certain number of false blockage or maintenance alarms. Another observation is that, contrary to previous scenario, full upstream blockage is the easiest to detect, easier than nearly-full. The reason is similar to that of the last observation, two sensor are far downstream to the blockage spot and the choke effect subsides. The full blockage creates the most ∆ T drop while partial one has only insignificant. We carefully evaluate our improved algorithm in two different scenario— blockage and upstream maintenance with temperature data straddling the valve and the choke. We demonstrate that our algorithm yields good accuracy and reasonable false alarms in a wide range of parameters for steam pipe. 3.7.5 Generalizing to Water Pipelines We have shown our algorithm detects blockages in steamflood pipelines. Our algorithmandtheconceptofdetectingsuddentemperaturedropsisnotspecificto 88 (a) Logical view of the pro- totype. (b) Physical view of the prototype. Figure 3.15: In-lab, water-based pipeline prototype. steamnetworks,andtogeneralizecomplementarycollaborationtoabroaderappli- cation domain, we test our algorithm with other kinds of fluid flow such as water pipelines. However, water has very different physical properties than steam—it is incompressible and at operates at much lower pressure and temperature. These differencesrequirere-tuningouralgorithmandmakedetectionofpartialblockages difficult. We next show that, after adjusting parameters, our algorithm can detect full blockages in a hot-water distribution system, and therefore can generalize to applications in other domains and industries. To evaluate a second network, we constructed a simple hot water network. Shown in Figure 3.15, our small testbed consists of a tankless water heater with a recirculation pump; a plastic, lidless tank; and a small network of PVC pipes and valves. This experiment precedes our mote implementation, so it uses the same algorithmsbutrunningonaPC,withdatafromGo!Temp,USB-basedtemperature sensors [Ver]. As in the steamflood pipeline, we emulate blockages by controlling valves (V a in Figure 3.15). While our steam experiments use fieldable hardware 89 −10 −5 0 5 10 temperature ( o C) δ sl th_block 0 50 100 150 200 250 300 350 400 35 40 45 50 55 temperature ( o C) time (minutes) O 50% 100% O 25% 50% 60% 70% 80% 90% T u T d Figure 3.16: Blockage detection in a water pipeline: raw up- and downstream temperatures (bottom), with emulated blockage shown by annotations (vertical lines): 100%: full blockage, X%: partial blockage by turning V a X% off, and O: open (no blockage). in an industrial pipeline network, our hot water experiments are laboratory-based with experimental hardware and a simple pipeline. However, they demonstrate the generality of our algorithms in a very different medium, showing our approach can apply to other cases where blockage points can be anticipated. v The lower plot of Figure 3.16 shows raw up- and downstream pipe temper- atures change during a full blockage between 82 and 168 minutes. Vertical lines in the figure indicate when we change the emulated blockage to the approximate percentage shown. The upper plot in Figure 3.16 shows that our algorithm, with parameter (th block = 3), captures both δ sℓ peaks (minutes 93 and 400) caused by full and 90% blockages. We show one representative example of three consistent experimental runs. 90 From this experiment, we see that first, our algorithm successfully detects full and near-full blockage in this very different water network. The downstream tem- peraturedropssignificantly,convergingtoambienttemperatureafterfullblockage, from 44℃ to 39℃ in about 25 minutes, while upstream temperature remains con- stant. Second, we observe that it is difficult to detect partial blockage in this water pipeline,unlikethesteamnetwork. Thetemperaturechangeuponpartialblockage (minutes 34–82 and 247–389) is insignificant, except for near-full (90%) blockage attheendoftheexperiment. Themainreasonsmakepartialblockagehardisthat water is incompressible and PVC pipes in our testbed have a much lower thermal conductivity, showing greater hysteresis. We believe partial blockages are difficult to detect in water networks because water is incompressible and hence partial blockage has little affect on fluid pressure. Another reason is that PVC pipes in our testbed have a much lower thermal conductivity (0.19W/mK) than copper pipes do (401W/mK), and therefore this network shows much greater hysteresis. 3.8 System evaluation Many factors determine if it is practical to pervasively deploy our system with multi-sensor collaboration. Evaluations in prior sections only prove one of them, accuracy—multi-sensorcomplementarycollaborationcan reduceerrorandsoreach accurate detection. In this section, we further discuss three other factors: the ease of calibration, overall cost and system robustness. 91 3.8.1 Sensor and system calibration The premise of non-invasive sensing is that pipe surface temperatures are able to predict blocking and constriction inside the pipe, and that we can evaluate pipe surface temperature reasonably accurately. We next evaluate this claim, exam- ining how we calibrate temperature sensing in spite of several potential sources of observation error. We break down the calibration into following steps, from pipe raw temperature to thermocouple output, to mote input value and finally to mote ADC readings. We show that our system is accurate enough in each steps to detect problems without careful calibration, but that with calibration we can predict internal temperature accurately enough to infer internal flow rates given a known fluid. Thefirststepofcalibrationisfrompipetemperaturetothermocouplewithtwo major factors. One is the pipe insulation because it significantly affects measured temperature distribution (μ and σ). To mount thermocouple on pipe, we have to temporarily tear down the insulation material, usually thermal sponge and aluminumwrap. Wetesttemperaturecollectedbeforeandafterinsulationre-apply from the same sensor and we find that mean temperature rise from 244.86℃ to 265.37℃ and σ drops from 3.00℃ to 0.78℃, with all other condition remaining the same. In other words, we observe less environmental noise and closer to steam temperature,ifpipeproperlyinsulated. Theotheristhermocouple-to-pipecontact. We believe hose-clamp style provide firmest contact than other styles, say magnet or general-purpose (stick) probe. In the field, we observe that our hand-held Sper Scientific thermometer with general-purpose probe could not measure repeatable temperature, mainly because of the pipe curved and rusty surface. 92 Another calibration concern is thermocouple voltage-to-temperature (V-T) non-linearity. We want to know whether V-T relation approximates linear in a small temperature range, since in the entire space, usually we have to use high order polynomial conversion [Rec10]. From J type thermocouple calibration refer- ence table [AST93], we pick up all points of which temperature ranges between [0, 310]℃. We then use linear fitting find the gradient for V-T relation is 18.259 and they-interceptis2.852. TheR-squaredvalueofthefitis0.99995andthestandard deviation of error between look-up table and fitting points is 0.887℃. This shows the V-T relation is linear enough for our algorithm. The second step is from thermocouple output to mote input voltage. We find thattorunourdetectionalgorithm,itisunnecessarytodothecalibration. Accord- ing to Seebeck effect, a voltage potential at the two ends of a conductor when a temperature gradient could be established along it [Rec10]. When two differ- ent material soldered at one point, we measure its temperature gradient by the voltage difference between two legs of the custom thermocouple. We need the reference point temperature (T r ) to calibrate each thermocouple because of the thermocouple-to-voltmeter connection. However, we are more interested in the reading difference between two thermocouples than individual accurate reading. Hence, no T r calibration is required here because we find the reference junction errors from two thermocouples cancel each other if both in the same isothermal box. The last step is mote input voltage to mote ADC value. There is one last variable in this step, our custom amplifier board. We first verify the two op-amp channels connected to thermocouples behaves the same. We connect thermocou- ples, amplifier board and Mica-2, feed the same voltages to both op-amp channel 93 Figure 3.17: Amplifier board repeatability test result. and record ADC values returned from Mica-2. We sample different voltages and repeatthetestforfourtimes. Wereporttestresultfromoneoftheamplifierboards since they are fairly close. The mean difference between two channels is 1.17 and σ is 0.68. Next we quantitatively test the board repeatability (Figure 3.17). From the four test, we conclude the two channels are stable enough. T = 18.259× ADC×3×1000 2 10 ×β +2.852 (3.2) Finally, we combine and test all calibration steps. Figure 3.18 shows that if we directly convert ADC to temperature with minimum calibration, according to Equation 3.2, we only have near constant error for each channel. V is voltage in mV and β = 100 for the amplifier board gain. This clearly proves that we can run our detection algorithm directly over ADC value returned by Mica-2. In all, we draw two conclusions. First, our system is linear enough to run our algorithm without calibration. Second, with post-facto calibration we can predict 94 50 100 150 200 250 300 temperature ( o C) before the choke CR1k mote 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 x 10 4 50 100 150 200 250 300 time (1s) temperature ( o C) after the choke Figure 3.18: Temperature measured by mote and CR1k. temperatureaccurately. Inaddition,accordingtoEquation3.1,webelieveaccurate pipeskintemperatureeventuallyhelpsusestimatein-pipepressureandsteamflow rate. 3.8.2 System cost Our goal is to reduce cost, so we next show the savings from small, non-invasive sensors with a thermal energy harvesting. We consider capital and operational costs of sensor system installation. Many industrial sites, including our test site, haveexistingfieldwirelessnetworks, sowedonotconsidercoststodeploythefield network (such network hardware is typically US$200 or less per node). Currently deployed systems: The circular pipe fitting with two large pro- trudingtapsontheleftofFigure3.19isatypicalinvasivepressuresensorcurrently 95 Figure 3.19: Invasive flow sensing with solar panel for power. in use. This system requires installation of pressure sensor taps between adjoining pipelines and a data acquisition unit tethered to a large solar panel and battery. Cost of data acquisition is high enough that measurement units are transported from site to site by field technicians to manually collect limited samples of data. Installation costs of differential pressure sensing taps and power to the sensor can be as high as US$20000 per site, with mechanical costs from $8000 to $16000 and calibration an additional $3000 to $4000. Cost of the hardware itself, including pressure sensors, a data logger, a solar panel, and battery packs can range from $13000 to $20000. Given such high costs, a great deal of oilfield monitoring today is manual. For example, many surface oil fields contain hundreds or thousands of oil producing wells, and steamflood secondary production may employ half as many injection wells. However, not all the steam injection sites are equipped with the monitoring points due to high cost of installation. In addition, only a limited number of 96 sites can be monitored at any time, since expensive instrumentation is shared over many wellheads rather than permanently installed. Today it is typical for a field technician to visit each site monthly, attaching a sensor and data logger to the sensing points to collect data for several minutes. The technician uploads the data on return to the office, or via the field wireless network. With travel time, one data collection requires multiple person-hours per site. Human-driven sensing reduces capital costs, with the sensing points installed when the line is built and the cost of the pressure sensor is amortized over many measurement points. However, human-in-the-loop makes the operational costs quite high: easily hundreds of dollars per measurement. This high cost for each measurement discourages frequent measurements and so prevents easy detection of problems before they occur. In addition, recurring costs will rise with upward trends in future labor costs. Our sensor-network-based system: By comparison, the capital cost of our system is quite modest. Our prototype unit consists of a Mica-2 (US$100) for control, a modified Heliomote ($125) for power conditioning, a custom amplifier board($50)andtwothermocouplesensors($70intotal)forsensing,theTEG($50) and a custom heatsink and mount assembly ($200). Despite the system being a research prototype and so not benefiting from economies of scale, component costs are less than $600. Volume would reduce these costs, although technical support would add to them. Moreimportantly, ourapproachcandramaticallylowerdeploymentandopera- tionalcost. Deploymentcanbedonebyatechnicianinanhourortwo(deployment time for our field experiment was two hours, and we expect future deployments to be half that). Since deployment is non-invasive, steam flow need not to be 97 interrupted and new plumbing is not required; since it is self-powered, electrical expertise is not required. The primary technical skills are SCADA integration and standard oilfield and steamflood safety training. We estimate deployment cost at around $300. Besides, we see no recurring operational costs for sensing. We believe these significant reductions in both acquisition and operation will allow much greater deployment of sensing with systems such as ours than are possible today. 3.8.3 System robustness Although we showed our system works in the lab and for short-term field deploy- ments, a long-term, real-world deployment raises a number of questions about system robustness. We next look at three questions related to system reliability under different conditions. Do environmental changes affect our algorithm accu- racy? Is our algorithm sensitive to sensor location and placement? Is our radio communication reliable enough? We first look at our system to evaluate if environmental temperature changes affect TEG and sensing performance. During our overnight field deployment the TEG-powered motes operated continuously, even while ambient temperature ranged from 9 to 22℃. This observation confirms diurnal temperature changes are small compared the potential energy in the steam network. We also see no diurnal affects on sensed pipe temperature. Table 3.3 records that pipe temperature remains relatively constant in spite of changes in ambient temperature. In fact, upstream co-generation power cycle or distribution branch configuration is likely to have more impact. In all, both our algorithm and hard- ware platform should work independently from diurnal amplitude. 98 Table 3.3: Pipe temperature variation along time T u (℃) T 1 d (℃) ambient (℃) time (s) μ σ μ σ μ σ Noon 258 2.7 261 2.1 19 1.4 Evening 261 0.8 262 1.0 17 0.7 Midnight 258 1.0 263 1.0 10 0.23 Second, the sensor location can effect system operation because of choked fluid flow (Section 3.5.1). Figure 3.20 compares temperature fluctuation at different downstream spots (after the valve) upon blockage. Letter labels and vertical lines aredefinedinFigure3.5. Hereweshowrunning50-samplemeanstosmoothshort- term variation. We see distinct temperature changes at all three locations for all levels of blockage, but the temperature change varies depending on proximity to the choke. However, contrary to T 1 d and T 3 d , T 2 d shows that temperature upon full blockage is lower than that of near-full blockage. The reason is that T 2 d is farther downstream than T 1 d to the valve but still before the actual choke, experiencing a smaller transient due to choked flow. This analysis suggests that sensor placement can affect results. Since our system can adapt to both steam and water pipelines, we are confident parameter adjustments can accommodate such variation, but automating the process is future work. Finally, we want to verify that the communication between our system and the base station is reliable. Since we hide the mote antenna inside the pelican box to prevent potential environmental hazard to our device, we expect packet loss, despite a short base-to-sensor distance of 4m with build-in B-MAC [PHC04]. In general, the total packet loss rate is low, 0.63% (31 losses in 4914 transmissions) at the valve mote and 0.67% (33 losses in 4914) at choke mote. Figure 3.21 shows 99 0 3000 6000 9000 12000 15000 18000 120 140 160 180 200 220 240 260 280 time (seconds) temperature ( o C) F NF P F NF P F NF P grad. O O O O O O O O O O O T d 1 T d 2 T d 3 Figure 3.20: Aggregation on all three downstream temperatures. 0 100 200 300 400 500 600 700 800 mote−2 mote−1 time (minutes) Figure 3.21: Mote radio packet loss distribution. Each marker represents one sample missing from dataset. loss is generally uniform and uncorrelated across motes, if slightly burst in time. If we consider the low data rate of the system—one tens-of-byte data packet per 10s for current deployment and occasional blockage alarm packets for full system, we conclude our wireless communication channel is adequate and robust enough. 3.9 Conclusions on Choke Blockage Detection This section supported our thesis by showing complementary multi-sensor collab- oration could achieve better choke blockage detection accuracy than single-sensor 100 solutions. We explored how multi-sensor non-invasive sensing improved steam- choke blockage detection accuracy and suppressed false alarms. From the success- ful experimental results, we further infer that complementary collaboration can offer improvement in other similar event detection applications where one event causes simultaneous signal change at different places. We developed an algorithm to detect problems in pipelines using non-invasive sensing and extensions to avoid false alarms. We demonstrated the effectiveness of our algorithm and extensions and the whole system through laboratory tests and fieldexperiments. Theseexperimentsdemonstratedthatourextendedmulti-sensor algorithm yielded good accuracy and triggered minimum false alarms. We also showed our multi-sensor algorithm with complementary collaboration generalized to other environment, in our case hot water pipelines. In addition, we showed the potential to power a sensor network by the phenomena being sensed, to operate without any battery, and successfully tested our TEG in field over-night. 101 Chapter 4 Cold Oil Blockage Detection We have shown that multi-sensor collaboration helps vehicle classification and steam-choke blockage-detection applications. More specifically, what we have proven so far is single-modal multi-sensor collaboration improves certain event detection applications. However, in reality, some problems are not solvable by only one sensing modality, no matter how many sensors are deployed. Multi-modal sensing is necessary is that in some cases it can suppresses false alarms while single-modal cannot. Multiple phenomenon may trigger the same signal to confuse single-modal sensing. Although those phenomenon may trigger other kinds of signal, if the sole sensing modality in question is insensitive to them, false alarms will be inevitable. We use cold-oil blockage as an example. Oil extraction in pipe may congeal and block flowline under low temperature in winter, causing equipment damage and production loss. Therefore, we need to deploy sensors on flowline to detect this blockage problem. The temperature on pipeindicatesifthehotfluidinpipestopsflowing,butitcannottellifthestoppage is because of blockage or the stop of flow actuation (pumpjack). On the other hand, acousticsensingcandistinguishpumpjackoperatingornotbycapturingthe noise emitted from pumpjack. Therefore, we need to first detect suggest blockage by temperature sensing and next suppress false blockages by detecting pumpjack status with acoustic sensing. 102 In order to solve this problem and prove our thesis works on a larger domain of event detection problems, we turn to multi-modal complementary collaboration. We next explore this cold-oil blockage detection—a more complex problem than steam-choke blockage in industrial monitoring. Here any single-modal sensing is insufficient,becauseadetectionrequiresacombinationofphenomenawhosesignals are orthogonal to each other. For example, typical orthogonal signal combination may include temperature, vibration, acoustics, image and electromagnetic emis- sion. More generally, we need multiple sensors with different sensing modalities whose union meets the minimum sensing requirement. A technical report based on this chapter appears in [ZH13b]. In addition, a version of the contents of this chapter has been submitted to a sensornet conference [ZH13a]. 4.1 Motivation for Cold-Oil Blockage Detection Sensor networks are used to collect data, detect problems, and take actions in the physicalworld. Smallandinexpensive,sensornetscanbeeasilydeployedtoaddress many real-world problems, from sewage pipe leakage detection [SNMT07], milling machine wear-out prediction [ZTLZ06] and live stock health monitoring [FIIS11]. In spite of their effectiveness in some applications, sensornet uptake has been slow in many industrial applications. SCADA systems today often employ tradi- tional dedicated and often expensive sensors, or fall back on manual observations where automated sensing is not seen as cost effective. A challenge in use of low- cost wireless sensors is that simple sensing methods often create many false alarms when they are confused by noise or changes in regular operation. In this chapter we propose to use different kinds of sensors to distinguish real anomalies from false alarms. We select a main sensor that detects the anomaly 103 butmaybeconfusedbychangesduringregularoperation. Wethenaddadditional sensors that can distinguish actual problems from false positives, although they cannot detect anomalies alone. Our overall goal is to identify classes of industrial applications where multi- modal sensing with complementary collaboration can resolve sensing ambiguities. We prove this claim in the context of a specific example: cold-oil blockages in flow- lines in producing oilfields. A typical oilfield has many kilometers of distribution flowlines that collect crude oil extracted from wellhead pumpjacks, gathers the oil for measurement and accounting, and ultimately sends it to refineries. Distribu- tion systems near the wellhead are often small, particularly in older fields. In cold weather, oilthickensbecauseoilviscosityhasaninverseproportionalrelationwith its temperature. Oil may then interact with sand or other contaminants in the fluid, and with pipe sags or narrow fittings, resulting in blockages in the lines. Blocks cause production loss, and if left unresolved they can result in pipe leaks, damage to the flowline, or even to the pumpjack. After pipe being fully closed, it takes only tens of seconds for pressure to build up before some parts in line rupture. Although the oil industry has explored several stand-alone sensors, current approaches are either unreliable or too expensive to install and maintain (Sec- tion 4.2). Although some fields contain thousands of wells where production lines are vulnerable to blockage, manual inspection is the most commonly used tech- nique today. Our insight is that multi-modal sensing can not only reduce the cost of detec- tion of cold-oil blockages while avoiding false alarms. Automating sensing can provide much more rapid detection than current approaches. Rapid feedback is 104 important because a shorter gap between blockage reaches critical level and alarm is signaled can minimize different losses, including environmental and equipment. We detect blockage by sensing temperature and acoustic signals. We infer flow interruption from pipe skin temperature, but in addition to blockages, many regu- lareventschangetemperature,includingautomaticpumpjackshut-insanddiurnal environmentaleffects. Weavoidfalsepositivesbycomparingmultipletemperature readings and by using acoustic sensing to monitor pumpjack status. We define our sensing problem and summarize our approach in Section 4.2. Our experimental results focus on cold-oil blockage, but the principle of multi- modal sensing to avoid false positives applies to many other sensing problems. For example, Girod and Estrin suggest using video evidence to correct problems from obstacles in acoustic ranging [GE01]. In human motion detection, Stiefmeier et al. cross-segment data stream between different sensors, including inertial, vibration andforcesensitivesensors[SRT + 08]. Wediscussmoreongeneralizingourapproach to other applications in Section 4.5.10. 4.1.1 Relation to Thesis Multi-sensor vehicle classification and steam-choke blockage detection are both single-modal collaboration. We therefore remove the constraint on modality and supports our thesis by exploring multi-modal (complementary) collaboration in a more difficult problem—cold-oil blockage. We use the acoustic channel in our system to suppress false alarms in the temperature channel. More importantly, no single-modalsensorcansolvethisproblemnon-invasivelyatalowcost. Indeed,we show collaboration can enable pervasive deployment on flowline blockage in field. 105 Like our prior studies, multi-modal collaboration generalizes to other problem domains and especially helps applications where themain sensory channelis easily confused by certain events but a secondary channel is capable of detecting the eventsinquestion. Therefore, thoseapplicationsmayusemulti-modalsensingand filterfalsealarmsbetweenmodalities. Forexample,ifaGPSmodemcannottellifa vehicleisonhighwayoranearbylocalpass,aspeedometercancleartheambiguity due to the speed limit difference. In this chapter, pumpjack duty cycling triggers false alarms of blockage on temperature sensors. Our acoustic sensor therefore suppresses these false alarms by detecting pumpjack on/off, although it cannot directly detect cold-oil blockage in flowlines. 4.1.2 Contributions The first contribution of this chapter is to identify the opportunity for multi- modal sensing to reduce error rates with low-cost sensors. While some prior sensors have explored multi-modal sensing with expensive sensors (for example, cameras [GE01]) and PC-level computation (including mobile phones or lap- tops [ZTLZ06,AMM + 08]), we believe we are the first to show these approaches apply to low-cost embedded sensors. Our second contribution is to prove this claim by exploring a specific applica- tion: we design an embedded sensing approach that detects cold-oil line blockages usingacombinationofinexpensivetemperatureandacousticsensors(Section4.3), then test our specific implementation (Section 4.4) in the field (Section 4.5). 106 4.2 Cold-Oil Blockage Problem Overview To understand how multi-modal collaboration by avoiding false positives can solve complex problem in industrial monitoring, we further focus on cold-oil block- age after steam-choke blockage detection. While in the introduction above (Sec- tion 4.1) multi-modal sensing seems straightforward, the key question is under- standing how real-world sources of noise and false detections affect sensing system design. To that end, we focus on cold-oil blockage as a real-world application. TheProblem: Cold-oilblockageoccurswhenthereturnlinefromaproducing oilwellbecomesblocked, typicallyduetochangesinoilviscosityasaresultofcold weather, sometimes compounded by buildup of sand in the pipe. Blockage typically build up gradually over time. Producing wells often operate intermittently with on/off cycles of 5-15 minutes (to allow downhole pressure to build up for suitable operation); when the pump is not operational, oil can transi- tion from flowing slowly to blocked. A blocked pipe can cause equipment damage and oil leaks, since if well production continues with a blocked flow line pressure in the line will cause flow line rupture or pumpjack damage. Recovery from equip- ment damage can easily amount to ten thousand dollars per event, in addition to reducing production. The equipment cost generally consists of the repair fee and the associated oil production loss during repair. Leakage of a stuffing box can cost US$100 to $1000 to do the cleanup, depending on how much oil spills on the ground. If a rod is parted from the pump, it takes $6000 to move in a rig to pull and replace the parted rod. If a motor burns up, it takes $1200 to replace the motor and belts. Finally, like stuffing box leakage, the cost to fix flowline rupture is most determined by the oil spills, ranging from $500 to $10000. In addition to these repair cost, the production stoppage caused by repair, usually 107 JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC 95% 100% 105% normalized product. 0% 50% 100% % below baseline Figure 4.1: Seasonality analysis shows winter production loss. one to five days, can too introduce heavy loss. The revenue of an average well is about $1000 per day, with ten barrels of production. Therefore, the stoppage may further introduce up to $5000 in lost revenue. Cold-oil blockage is a significant problem in some oil fields. Figure 4.1 shows eight consecutive years of production data of a business unit in the California Central Valley, an oil field where cold-oil blockage is a concern. We normalize production values to remove long-term decreasing trend in field production and show seasonal variation in production. The first step of normalization is comput- ing monthly index by applying exponential decay fitting over the whole dataset. The resulting fitting error is 1.3%, low enough to show we have a good fit. The fitting forecast is the monthly index—the baseline of monthly production which is unaffected by the overall trend. Next, we normalize the raw data by its ratio against its index for every month and the result is in the upper plot. The lower plot summarizes, for each month, how often that month’s production is below its index. It further shows that winter months (November through February) witness consistently lower production rates. There are multiple factors that contribute to 108 this trend (scheduled maintenance is often planed to avoid hot summer months), but field engineers confirm that a significant factor to reduced winter production are well problems due to cold-oil blockage. Although we focus on cold-oil blockage so can evaluate real-world sources of error,inSection4.5.10weconsiderhowmulti-modalsensingappliesinotherappli- cations. Current Approaches: Aware of the problem, oil companies have explored current sensing approaches, including flowmeters, pressure sensors, leak detection, and of course manual inspection. Unfortunately, currently techniques for automa- tion have high installation and maintenance costs. For example, pressure sensors cost US$1000 or more to purchase and have annual costs of $300 or more to recal- ibrate; flow sensors are more costly. As a result, these sensors are used only on a few, very productive wells, while manual inspection remains common, in spite of its large delay. Theseapproachessuggesttheimportanceoftheproblem,butsensorcostmeans they are deployed on only a few of the thousands of wells where there is concern. Fieldengineersconfirmthatinsomecases,thealternativeissimplytopreemptively stop production on certain well that have no monitoring. Degree of Blockage: Blockages build up over time, and one would like to detect them before they happen, or very quickly after they happen. Currently pressure sensors are deployed on only a few wells. Instead, manual inspection is done to identify equipment damage that follows a blocking, something that often occurs twelve hours after the fact. Here we focus on rapid detection of full and near-full blockages. Detection of blockage allows well shut-in and recovery before damage; rapid detection after 109 Figure4.2: Adiagramillustratesourproblemstatement. Squaresymbolsaretem- perature sensors; the oval is an acoustic sensor. S i denote different pipe sections. blockage may avoid equipment damage and will minimize leaks. Our evaluation (Section4.5.7)showswecandetectblockagein10to30minutes,muchshorterthan 12 hours by current manual inspection. While not instantaneous, this detection is potentially able to save large production loss. We emulate full blockages in field experiments (Section 4.5). We cannot test near-full blockages in the field due to safety concerns, but we evaluate near-full blockages in laboratory tests (Section 4.5.9). The success of the both field and lab tests shows the generality of our approach on cold-oil blockage and even a broader range of applications. Earlier detection of partial blockages (50% or less) would be helpful and is future work. However, our current temperature method is not enough to detect the subtlety. 4.3 DesignofCold-OilBlockageDetectionAlgo- rithm To prove that multi-modal collaboration enables complex problem detection, we needtoproposeanalgorithmforcold-oilblockagedetection,particularlyexploiting multi-modal sensing. In this section, we explore how low-cost temperature sensors detectblockage,acousticsensorsdetectequipmentoperation,andthetwotogether provide reliable blockage detection with a low false positive rate. 110 4.3.1 Overview of approach To detect blockages we use two sensing methods: acoustic sensing at the wellhead and temperature sensing at locations along the flow line (Figure 4.2). Typical flow lines are much hotter than ambient temperature (100℃ vs. 0–30℃), particularly in fields that use secondary production techniques such as steam injection. We can therefore infer flow blockage by observing pipe skin temperature: the pipe downstream of a blockage will converge to ambient temperature. In operation we expecttoplacemultipletemperaturesensorsalongtheflowline, nearplaceswhere blockages are expected. Unfortunately, pumpjacks often stop production (“shut-in”) periodically to allow downhole pressure to accumulate. Pumpjack shut-in causes drop in pipe temperature the same as a blockage, so temperature sensing alone will result in false alarms. We therefore add a second acoustic sensor to detect pumpjack operation. And one acoustic sensor placed near the wellhead can provide pumpjack status for all temperature sensors on the production line. temperatures on the same line downstream. The acoustic sensor listens to flow in the pipe and the clanging of the pumpjack rods and tubing to detect pumpjack operation. (These sounds propagate well through the pipe, so acoustic sensor placement can be within 20m of the wellhead.) Our hypothesis is that our combination of temperature and acoustic sensing is both necessary and sufficient to detect cold-oil blockage. Sources of noise: Although we focus on pumpjack operation as our main source of error, we must consider many sources of noise, from the environment, field, and measurement system. 111 Environmental noise contains diurnal and seasonal changes in weather and ambient temperature. Pipe skin temperature changes by a few ℃ over the course of a day due to changes in sunlight and wind or other weather. Our algorithm is insensitive to this change because the temperature difference between normal flow and ambient is much larger. Seasonal weather changes have a greater change, withtemperaturesthatvaryby38℃ormorefromthemininwintertothemaxin summer. However,thislong-term changedoesnotaffectouralgorithmbecausethe detection threshold is hourly auto-retrained against recent pipe skin temperature, quickly adapting our algorithm in as short as hours. Second, field conditions change: including downhole conditions, equipment maintenance and main-line back pressure. Downhole temperature and pressure changes as the field produces oil and due to changes in injection. These changes aregenerallyslow(overdaysorweeks);ouralgorithmretrainshourlyandsoadapts to these. Valve close-up caused by maintenance indeed behaves similar to a real, sudden full blockage. We depend on field engineers to identify maintenance a pos- sible source of false blockage detections. Finally, there will be some temperature propagation from the main line back to a blockage. We expect this effect to be minimal. Thelastgroupismeasurementnoise,whichisrelatedtoourdeploymentsetting, including sensor installation and random glitch. If a sensor has a loose contact with the pipe, the readings are always a weighted average between ambient and pipe skin temperature. Poor connection will reduce our algorithm’s sensitivity, but our tuning accounts for variations. We confirm in tests that our algorithm adapts to loose connections that cut the mid-point between ambient and normal 112 pipe operation temperature in half, still finding the correct reference value and triggers on sub-20℃ drop. From the discussion of three categories of noise—environmental, system and measurement,weconcludethatouralgorithmwithparameterauto-tuningisrobust enough. 4.3.2 Temperature Sensing for Flow Presence Section4.3.1showsflowpresencedetectionisthefirstpartofourmulti-modalcold- oil blockage detection. In this section, we talk about how to detect flow presence by temperature and how to automatically tune parameters. Accordingtoourproblemstatementandhypothesisabove, weneedtomeasure pipeskintemperaturetodetectthepresenceofflow,orinanotherwords,suggested blockage. Since the temperature usually drops gradually (about 20℃ in an hour), weneedanalgorithmtoprocessstreamingtemperaturetraceandidentifyitstrend of approximating ambient. Fortheabovereason, ouralgorithmemploysone-sidedCUSUM(orcumulative sumcontrolchart[Pag54]),originallyastatisticaltechnologydevelopedforprocess quality control. The algorithm starts at low-pass filtering raw temperature obser- vation (by EWMA) to filter transient noise. Next, it compares every observation to a reference value to calculate the deviation from it. Meanwhile, it maintains a running statistics, the cumulative sum of all the deviation in history as basic CUSUM does. We call this cumulative sum of deviation certainty of drop (C d ). When observation is lower than k, C d becomes larger and larger before it exceeds a threshold, which suggests a blockage because the temperature is too low for too 113 long. We use one-sided CUSUM, resetting C d when it is less than zero to respond quickly to temperature drops. We must set two algorithm parameters: the threshold for certainty of drop, and the reference value (k). We set the threshold to 15 times normally observed temperature, in this case 3000, to be robust to transient temperature dips. The reference value, k is set as the mid-point between quality level—normal pipe tem- perature, μ 0 and anomaly level—flow stopped, μ 1 (μ 1 <μ 0 ). Since k is important to the accuracy and responsiveness of the algorithm, we auto-tune it instead of hard-coding. However, due to different sources of noise we list in problem statement, we do not think predefined, fixed μ 0 and μ 1 estimation can best reflect an appropriate k. Hence, it is necessary to first auto-tune μ 0 and μ 1 for its dependency and we embed auto-tuning in our algorithm to adjust the estimation ofthetwolevels. Whenpumpjackis operating(determinedbyacoustic node, introduced later in Section 4.3.3), we constantly update the quality level μ 0 by temperature observation. When pumpjack shuts in, we stop updating μ 0 but start anomaly level μ 1 updating as temperature drops. To generalize this, we are using a second sensory channel, to convert a false-alarm hazard into a helper of parameter tuning. By the time of the shut-in is over and pumpjack resumes operating, a new reference value k will be ready, based on auto-tuned μ o and μ 1 . Another k-tuning feature is that we do not update k at shut-in because during pump-off, temperature detection becomes less important. More importantly, we intend to avoid accidentally updating threshold to an inappropriate value. 114 4.3.3 Acoustic Sensing to Avoid False Alarms OurdiscussioninSection4.2showsthattemperaturealoneisnotenough. Acoustic sensing on pumpjack status can avoid thefalse alarms caused by regular pump-off. In this section, we describe our acoustic algorithm design and next discuss how we automatically tune parameters in that algorithm. We need to determine if pumpjack is operating for end pipe blockage detec- tion. Since pumpjack stroke with engine rumbling generates wide band noise and propagates along pipe, we use microphone mounted on pipe surface to measure thesoundpressurelevel(SPL),ahighlevelofwhichsuggestspumpjackoperating. When pumpjack is off, microphones are expected to pick up much lower energy of environmental noise. Our acoustic algorithm works as follows. First, for each stroke cycle C, we detectifpumpjackisonbycomparingsoundamplitudetoapre-configuredthresh- old θ p . If samples in C exceeds the threshold, mostly because of a significantly loud rod-tube clang noise associated with each stroke, we decide the pumpjack is on during the whole cycle (typically 7s). However, simple pumpjack flip detection is not robust against transient error and hence we need to know if the pumpjack is steady on. In order to make that decision, we check a longer history to see if it was being on for a whole warm-up period W long, usually far longer than a single cycle. Hence,tocorrectlydetectpumpjackstatus,weneedtoproperlyconfigurethree parameters: certaintyofdrop(C),warm-upperiod(W),andthreshold(θ p ). Wedo tuning on base station because the training involves certain intensive computation as auto-correlation and memory storage complexity both beyond mote capacity;so we employ a PC in our experiments. (In principle a mobile-phone class processor 115 could easily accommodate this work, although it is beyond 8-bit motes.) The on-site training step makes acoustic sensors robust against environment noise and mechanical difference across pumpjacks. We next describe our training algorithm for these three parameters, started by training data collection. Before deployment, we collect a short period of acoustic training data contain- ing both pump-on and -off. We next compute C by running auto-correlation over the pump-on trace. The lag yielding the largest coefficient represents pumpstroke cycle. To prevent from choosing harmonics, in implementation we search the high- est coefficient in a possible-cycle range, say [5s, 9s]. Further, based on our prior study, W could be set as five times of C. We consider both pump-on and -off to compute θ p , because it needs to be able to properly denote the difference between those two status. We first compute the noise floor by averaging all the samples in pump-off trace. We next throw away all samples below noise floor in pump-on segmentation. θ p equals the 86-percentile of amplitude among all the rest of the pump-on segmentation. The reason we choose this value for θ p is that during a common 7s pump cycle, our threshold should detect the single sample capturing the loudest rod-tube clang noise against other six under 1Hz sampling rate. Therefore, the signature noise sample is likely to have a higher amplitude than the other 86% (six out of seven in one cycle) samples. 4.3.4 Sensor Fusion for Blockage Detection Wetalkaboutthetwoalgorithmsintheabovesectionsandnextwedescribehowto fusethemtodetectendblockage. Ifweinterpretourbasichypothesis(Section4.2) with technical details, we find that blockage could be detected as flow stops but 116 pump is steady on. In another words, if pumpjack is off, our algorithm ignores all suggested blockage detection by temperature sensing, although the certainty of drop builds up due to stagnant flow. Onthecontrary,ifpumpjackison,ouralgorithmcandetectblockage,allinthe following two different situations. If blockage occurs during pumpjack operation (i.e. pumpjack is steady on), we expect to witness a line temperature drop. As soon as line temperature stays below reference value long enough, blockage detec- tion triggers. Besides, if blockage occurs during shut-in, after pipe cools off and pumpjack resumes, line temperature stays close to ambient and does not increase significantly. Hence, the certainty of drop can too build-up, followed by blockage detection. We evaluate the fusion result later in Section 4.5.7. 4.4 System Implementation To evaluate how multi-modal collaboration performs in real world, we implement an fieldable system for later field test. Before we review the details of our field experiment, we briefly talk about our mote sensing platform with low-cost sen- sors. We first summarize the hardware of our multi-modal sensing system. Next, we discuss the two challenges in acoustic node implementation and our software approaches to solve them. 4.4.1 System Hardware Our sensor network consists of three types of nodes: base node for data collec- tion, acoustic mote for pumpjack status detection and temperature mote for flow presence detection. In this section, we introduce the hardware of their parts. 117 OurbasenodeissimplyaMica-2mote[Cro]connectedtoPCthroughMIB520 programming board. It passively listens and logs all the packets transmitted from acoustic or temperature motes in the network. Our acoustic mote is composed of a Mica-2 mote and an MTS310CA, “Mica Sensor Board” with an on-board electret condenser microphone, Panasonic WM- 62A (Figure 4.3(a)). Figure 4.3(b) shows how we tape and clamp the extended microphone on the pipe with thermal insulation and we discuss how decoupling microphone benefits signal gain later. We cannot directly mount Mica-2 microphone on pipe because the high pipe temperature may damage the equipment, or at least result in inaccurate measure- ment. Common electret condenser microphone has a sub-70℃ operation tem- perature, lower than the pipe skin temperature in operation. Although it is not mentioned in its specification [Pan], we believe this model on Mica sensor board, WM-62Aisnotdesignedforahighertemperaturetask. Evenifitsustainstheheat, electretcondensermicrophonehasanunpredictablefrequencyresponseunderhigh temperature (around 80℃ [YO05]). Therefore in deployment, we apply thermal insulationontopoftheover-warmpipetoprotectourmicrophone. Theinsulation is called Fire Blanket and is made of woven fiberglass. We are aware of some side effects of sandwiching insulation between the microphone and pipe, for example, signal attenuation. However, under the design principle of low-cost sensing, we decide to make this trade-off instead of employing expensive specially-customized microphones, say US$5000-priced Br¨ uel & Kjær 4949 automotive surface micro- phone. 118 (a) Acoustic mote with microphone extended (b) Mote mic on a pipe (c) Temperature mote packed in a box (d) Thermocouple on a pipe. Figure 4.3: Our temperature and acoustic sensor hardware and deployment. Finally, the design principle of temperature motes inherits our steam-choke blockage application Section 3.3. They each consists of a Mica-2 for control, a cus- tom amplifier board to optimize thermocouple signal readings and a thermocouple sensor (NANMAC D6-60-J J-type) for pipe line and ambient temperature mea- surements. Figure 4.3(c) and 4.3(d) shows how we deploy them in our experiment. During experiments, we find surprisingly that our custom amplifier boards are sensitive to their operation temperature, although all components are rated at a much higher range. Our initial field trials show if exposed under the sun 119 directly, temperature sensors with the amplifiers sometimes return random read- ings, but sensors without the amplifiers work correctly. Hence in the latest test (Section 4.5.2), we covered the sensor motes in shade, but we are currently exam- ining our design and seeking a more robust solution. 4.4.2 Hierarchical Sampling and Aggregation in Acoustic Mote To obtain the sound pressure level of pipe, our acoustic sensor samples 2000 times a second. This sampling rate is high for a mote, posing two challenges. First, although the sensor generate and transmit one packet per second, we cannot col- lectively stack 2000 (one-second-long) samples in buffer due to the limited Mica-2 RAM size (4kB for both program and data). The other challenge is that because the sensor samples at such short interval as 500μs, hardware interrupts from other components(radio,flashlogger,etc.) arelikelytocauselargevariationinsampling rate [KPC + 07,HBC + 09]. For accurate sampling, we shut down all external com- ponentswhichmightoccupytheCPUfortoolongtoholdupthetimer. Hence, we design our software able to schedule and interleave processing, transmitting and flashloggingamongcontinualsampling. Wedolocalflashloggingbecauseinoper- ation, it could serve as backup in case of temporary network outages, although in a fully integrated system, data is always streamed back to a central server through field network. We design a hierarchical sampling and aggregation scheme to overcome the two challenges above. Overall, we pause the high frequency sampling and schedule other operations, before next sampling cycle. The pause causes gaps in sampling, and in the worst case we may mis-detect interesting phenomenon. To minimize 120 this sampling gap and coordinate data management, we make following design choices. At a high level, our sensor samples and computes the SPL within a one- second-long window (long window) before logging it to flash and transmitting it out. Atanintermediatelevel, wedivideeachlongwindowintoten0.1-second-long short windows. In each short window, sensor samples for 0.06s at 2kHz rate, and uses the remaining 0.04s to do SPL aggregation. The final 10 th short window does further aggregation by choosing the maximum SPL value among the past ten to represent the entire long window, before flash logging and radio transmission. Our labtestingshows60%/40%dutycycleisoptimalbecauseaslightlymoreaggressive setting (i.e. short than 0.04s gap) causes significantly more packet loss. Besides, the 0.04s gap does not cause mis-detection on the 0.2-second-long signature rod- tube clanging noise. 4.4.3 Maximizing Acoustic Gain Tomaximizetheacousticsignalgain,wetakethreestepsonsoftwareandhardware customization. First, we optimally adjust digital current bias through calibration. The 10-bit ADC channel of Mica-2 returns values ranging from 0 to 1023 mapped to 0 to 3V. As a result, it does not return negative voltage. To avoid losing the negative half of the waveform, MTS310CA is designed to elevate the center of the output acoustic waveform from 0V to approximately 1.5V, which corresponds to 512 in ADC value. We test this feature with our equipment and find that the new ADC waveform centers around 501, slightly off by the theoretical value of 512. We thus use our experimental result to offset mote ADC readings, removing DC bias. 121 Second, wedecouplemicrophoneunitfromtheboardforbettermounting. The flat Mica sensor board does not well match the curved pipe surface. Therefore, we desolder the microphone off and extend it out via wire, which enables us to simply tape it down to pipe in deployment for best contact and windscreen. Finally, we use TinyOS to maximize the microphone analog gain. We use the OS service to tune an resistor in the amplification stage to its largest value, which is an on-board, digitally controlled, variable resistor [Cro07]. 4.5 Evaluation This section evaluates if multi-modal complementary collaboration successfully detects complex problem in industrial applications. In the prior chapter, we show single-modal collaboration improves steam-choke blockage detection and here we further choose cold-oil blockage as a case study, which is more complex. We next describe the experiments we carried out to demonstrate we can detect flow block- age, and that multi-modal sensing can avoid false positives. We first evaluate our inexpensive sensors in the laboratory. We next describe our field experimen- tal setup and evaluation metrics test how temperature and acoustic sensors can infer blockages and equipment operation, and finally show how their combination provides a robust system. Before this experiment, we carried out five preliminary experiments and the details are in Section 4.6. 122 4.5.1 Calibrating Individual Sensors Ourpremiseisthatlow-costsensorsaresufficienttodetectflowblockages. Wenext compare inexpensive mote-based temperature and acoustic sensors against high- quality PC-based sensors to confirm that inexpensive sensors are “good enough”. Temperature Sensor Measurement We show our acoustic mote is close enough to ground truth in previous section. Next we compare temperature data by mote against USB data logger to verify if our low cost temperature collection solution performs well enough or not. The major differences between the two systems lies in hardware and calibration. The software processes are likely the same, although only limited information about USB data logger disclosed by its manufacturer. InSection3.8.1,wefindthatinrelativelylowtemperaturerange(0–200℃),itis unnecessary to calibrate J-type thermocouples before we deploy them in detection tasks. Hence, our mote reports raw ADC readings while USB data loggers are pre-calibrated by manufacturer. Table 4.1 proves our mote data is almost equivalent to USB data logger by showing strong correlation exists between all three mote-ground-truth pairs. We further visualize one data set, T 2 d as an example to better demonstrate that this claim. Figure 4.4 clearly shows that the data by mote is merely off from ground truthbyaconstantbutthefluctuationisalmostthesame. Forclearercomparison, wepost-facto’lyconvertrawADCreadingbymotetoCelsiusscaleunderthesame equation (Equation 3.2) in Section 3.8.1: T = 18.259× ADC×3×1000 2 10 ×β +2.852 123 Table 4.1: The three correlation coefficients between temperature data by motes against their ground truth T u T 1 d T 2 d 0.91 0.80 0.83 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 0 20 40 60 80 100 temperature ( o C) time mote USB Figure 4.4: Temperature measured by mote and USB data logger at T 2 d . where β is 367 as the gain of our pre-amplifier board. In our detection task, the algorithmismoresensitivetotemperaturedropsinsteadoftheabsolutevalue,and hence a constant disparity is acceptable. Acoustic Sensor Measurement We first look into our acoustic mote. Before we compare mote and PC acous- tic sensors, we briefly describes their components and the difference in the data collection approaches. Our acoustic mote is consist of Mica-2 mote, an electret condensermicrophoneandaMicasensorboard. Whereas, ourPCacousticsystem is equipped with more powerful hardware—a laptop with sound card complying to Intelhighdefinitionaudioarchitectureandabattery-poweredlavaliermicrophone. The hardware superiority of PC system alone is enough to justify its cleaner data. 124 Other than the hardware difference, the second major differences is sensor installation. Althoughbothsitontopofinsulation,wetapedownmotemicrophone by duct tape to pipe while we clamp on PC microphone by a customized clasp and hose-clamps, likely to produce larger force to press the microphone against the pipe for a better contact. Finally, there is difference in sampling and aggregation mechanism in software after we abstracted out the OS differences, although the final packet rates of both forevaluationarethesame,1Hz. AswedescribedinSection4.4,therawsampling rate on motes is 2kHz and we next use a hierarchical aggregation to assemble one packet every second from ten 0.1-second-long short packets. On the other hand, the raw sampling rate on PC is 16kHz, much high than what we have with motes, mainly because we plan to keep high quality ground truth data in case we need to investigate the frequency domain of acoustic signal. Since the software we choose, Audacity, does not support a sampling rate as low as 1Hz, and more importantly, we prefer to maintain the consistency between both systems, we re-sample our PC data in 2kHz by the same software before we further aggregate it by one-second- long window. Despite the differences we listed above, Figure 4.5 shows that our mote data is close enough to the ground truth. It is difficult to directly convert their units; hence we keep both trace in their raw units and hand-scale them in the plot. The correlation coefficient between the two traces is 0.44, proving a strong positive correlation. TheotherobservationthatPCdatahashigherSNR,whichisdepicted by much higher state transition spikes and near-zero pump-off noise measurement. 125 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 0 100 200 300 ADC . time 0 0.05 0.1 . amplitude Figure 4.5: Acoustic measured by mote and PC microphone. In all through the comparison, we show our mote data is close enough to PC databyamoreexpensivehardwaresuite. Thisresultfurthersupportsourhypoth- esis above that low cost sensor is capable of reaching effective yet economical sens- ing. 4.5.2 Field Experiment Approach We next evaluate our system with multi-modal complementary collaboration in field tests. From 9:30am to 5:30pm, November 7 th , 2012, we evaluate our system and a producing oilfield in the California Central Valley, working with field engi- neers from our research partners who operate that field. During the nearly seven- hour-long experiment, our system collected acoustic and temperature traces, did in-node processing and ran the full detection algorithm. We also collect ground truth data, concurrent with operation of our experimental system. Ground truth temperature and acoustic data employ USB thermocouple data loggers (EL-USB- TC[DAT12])andalaptopcomputerwithacommoditymicrophone. Figure4.6(b) shows the test site and the producing wellhead. 126 Our experiment emulates oil blockages by controlling valves. (We are not able to inject actual blockages, nor was it the time of year when they would form naturally.) Figure 4.6(a) illustrates the topology of sensors, pumpjack and valves. Ovalshaperepresentsacousticmoteandsquaresdotemperatureones. T u islocated before the production-circulation branch-out and so upstream to both valves. T 1 d and T 2 d are both on production line and straddle production valve, downstream to T u . To emulate blockages, we activate the production valve to close because it is not practical to create a real blockage in the field. When we close the production valve, oil stops flowing in the pipe and hence we observe a total blockage in line with the valve. In our experiment, we always leave open either the production or circulation valve, since closing both could cause high pressure at the wellhead that would damage the producing well or equipment. We conduct experiments on approximately half-hour intervals to allow the sys- tem time to stabilize between changes. Table 4.2 shows our schedule, with three pump-off periods for all four temperature motes to learn μ 1 and update k, with the last one (28-minute long) ran shorter than the first two (each about 50-minute long)duetotimeconstraint. AccordingtotheblockageintroductioninSection4.2, in reality we may generally categorize blockage in to two types regarding how it is formed. One is caused by a lump of viscous oil or sand clogging narrow fitting duringpumpjackoperation(op),anin-op blockage. Theotheriscausedbyresidue oil in pipe cooling off and turning solid during shut-in before pumpjack resumes operation, a non-op blockage. To better evaluate the generality of our algorithm, we simulate both types in three instances over the course of the day, and each stage runs between 24 to 45 minutes. The simulations are interleaved with other two types of stages. One is valve-open and pipe temperature rebounce, so sensors 127 (a) Logical view of deployment. (b) Physical view of deployment. Figure 4.6: November 2012 field deployment. can learn normal pipe temperature μ 0 during operation. The other is pumpjack shut-in, which configures the sensors’ CUSUM anomaly level μ 1 (i.e. temperature on stagnant flow). In addition to this field test, we carry out two prior field experiments where we evaluate components of our system and collect ground truth data for analysis in the lab. Prior tests were done at a different wellhead. We omit this data here due 128 Table 4.2: Experiment schedule and scheme product. start pump valve purpose 11:01am on open T 1,2 d learn μ 0 11:11am off all learn μ 1 12:01pm on close T 1,2 d non-op 12:25pm open T 1,2 d learn μ 0 1:09pm close T 1,2 d in-op 1:54pm off all learn μ 1 2:35pm on open T 1,2 d learn μ 0 3:05pm close T 1,2 d in-op 3:48pm open T 1,2 d learn μ 0 4:30pm* off all learn μ 1 4:58pm on T 1,2 d learn μ 0 to space, but replay of this ground truth data in the lab shows our system works correctly on another well with different sensor locations. 4.5.3 Evaluation Metrics Section 4.3 shows that we detects cold-oil blockage by filtering out irrelevant flow absence with acoustic pumpjack status detection. Before evaluation, we describe below the temperature and acoustic detection metrics due to their similarity. We evaluate both temperature and acoustic sensing in an event-based man- ner, but with separate event definition. For temperature, one event is one interval betweenchangesofequipmentsetting,becausewecareaboutifflowpresencedetec- tion triggers or does not trigger eventually in certain conditions. Each event starts withsettingchange,includingvalveclose-up/openandpump-on/-offandendswith another change, retaining the same setting across the entire event. According to 129 the schedule in Table 4.2, we divide our experiment after 11:11am into ten stand- aloneevents. Wediscardeventslessthantenminutes(theperiodbetween11:01am and 11:11am) because our algorithm requires 15 minutes to stabilize and our algo- rithm is still learning parameters. Hence, we first define metrics for temperature detectiontodenotethecorrectnessofflowpresenceforeachevent: a True Positive (tp) is when flow stops, due to either pump is off or a valve in-line is closed, and the algorithm triggers during the whole event; a True Negative (tn) is when flow is normal and the algorithm does not triggers anytime during the event; a False Positive (fp) is when flow is normal but the algorithm incorrectly triggers; and a False Negative (fn) is when flow stops but the algorithm incorrectly keeps silent. And we define overall accuracy using terms from information retrieval [Rij79]: Accu all = tp+tn tp+tn+fp+fn Contrarytotemperature,theeventinacousticevaluationisdefinedbyasample (i.e. one-second-long sensor reading), because we care about instantaneous pump- jackdetection. Weusesimilarwaystodefinethefourterms(tp, tn, fp, andfn)out of the pairwise combination between pump-on/-off and algorithm output-on/-off. For example, it is a tp, if pumpjack is on and the algorithm correctly declares it on. We inherit the same equation to compute overall accuracy as in forgoing tem- perature metrics. We can then define accuracy of pump-on and -off events using subsets of these measurements: Accu on = tp tp+fn Accu off = tn tn+fp 130 In our experiment, our acoustic node logged 23129 valid samples. Among them, 15310 are tp, 4893 are tn, 2227 are fp and 699 are fn. After we define metrics, we next evaluate our temperature flow presence detec- tion, followed by acoustic pumpjack status detection. In addition to Accu all , we care about Accu on and Accu off in acoustic sensing for future algorithm improve- ment. 4.5.4 Accuracy of Flow Presence Detection We carry out full system deployment with algorithm on-line in the field for both testing and data collection. In order to incrementally test each component in the system, we ran a manual version of temperature algorithm in parallel with the fullyautomatedversiononeachmote. Thesoledifferencebetweenthetwoversions lies in the process of parameter auto-configuration. we remotely re-program the temperature motes to inform the manual version of the perfect pumpjack status, while the automated obtains an imperfect update from their peer acoustic mote. We use the manual version to evaluate flow presence detection, while the fully- automated is for blockage detection evaluation in Section 4.5.7. The three plots in Figure 4.7 shows that our CUSUM-based, flow presence detection algorithm works perfectly during our field test. We achieve 100% accu- racy for the all ten events without any false positive or negative and we discuss more detailed observation below. Since T u is upstream to both production and circulation valves, oil flows as long as pumpjack is on, regardless the status of the production valve. Figure 4.7(a) shows our algorithm remains silent while pump is on and temperature drops upon the first two pump-off events effectively triggers our algorithm. In addition, our algorithm successfully detects temperature drop 131 causedbyin-linevalveclose-up,showninFigure4.7(b)and4.7(c),whereC d builds up at all “valve:closed” events. Weexpecttoseeatriggerinthethirdpump-offevent(4:30pm–4:58pm,marked “*”), but surprisingly C d does not build up high enough in the algorithm, different from the prior two pump-off events. We still count that a true positive rather than afalsenegative,becauseitisaresultofourcompressedexperimentalschedule—we ran out of the time at the end of our experiment and hence we cut-off the third pump-off event prematurely. The trend shows we require 5 additional minutes to trigger, and since operational conditions do not have a 30-minute time limit, we countthiseventascorrect. Twoevidencesarethatthetemperaturestillmaintains a steep drop trend and C d does start to build up at T 1 d and T 2 d . After comparing the results from the three temperature sensors, we make four observations. First, our approach achieves repeatable detection results, because the results are consistent across all three sensors. Second, one sensor is enough to cover a large pipe segment for detection, because it can detect blockage upstream and downstream to it. Temperature traces before and after the valve, T 1 d and T 2 d is almost equivalent, shown by a high positive correlation (0.98) and low standard deviation in differential(1.5℃).Therefore, in ourcaseeitherofthemisenough for the pipe section after production-circulation branch, at least 20m long to the next junction. Three, our sensor placement shows minimal recirculation from the main line since the temperature falls even downstream of the blockage (both T 1 d and T 2 d showsimilartemperatures). Finally,thedifferentresponsesuponblockagebetween T u and the other two shows multiple sensors can be used to locate blockages by distinguishing pipe locations with and without flow. 132 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 0 100 200 300 400 valve: closed valve: closed valve: closed valve: closed valve: open valve: open valve: open valve: open valve: open* valve: open pump off pump off pump off mote ADC 0 3000 C d time T u k (a) Upstream T u . 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 0 100 200 300 400 valve: closed valve: closed valve: closed valve: closed valve: open valve: open valve: open valve: open valve: open* valve: open pump off pump off pump off mote ADC 0 3000 C d time T d 1 k (b) Downstream before production valve T 1 d . 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 0 100 200 300 400 valve: closed valve: closed valve: closed valve: closed valve: open valve: open valve: open valve: open valve: open valve: open pump off pump off pump off mote ADC 0 3000 C d time T d 2 k (c) Downstream after production valve T 2 d . Figure 4.7: Flow presence detection results. 133 In this section, we conclude that flow presence detection with auto- configuration achieves perfect result in field tests. Therefore we next investigate further the necessity of auto-configuration by cross-comparison between the three datasets. 4.5.5 Auto-Configuration of Temperature Measurement Inthepreviousreview,wedemonstratethatourflowpresencedetectionisaccurate with parameter auto-configuration on quality and anomaly levels. We next show the need to auto-configure the parameters of our temperature algorithm, and that with auto-configuration deployment is robust to different wells and conditions. We first show that baseline temperatures vary at different pipe locations and times, and therefore we require different tuning parameters for different locations. The left graph in Figure 4.8 shows the basic statistics over the entire 6.5-hour-long temperature data, while the right one focuses on the temperature under normal flow, excluding no-flow periods caused by either pump-off or valve-close. We know that temperatures of pipes vary, and they are affected by ambient temperature as well. With only one day for field experiments, we cannot allow the pipe to completely cool. In addition, this statistics is not necessary because the overall and the normal-flow statistics are enough to prove that temperatures vary. We find that the temperature upstream to production-circulation branch-out (T u ) is significantly different from the pair straddling the valve (T 1 d and T 2 d ), showing by the different quartiles (blue boxes) in either plot. Hence, auto-configuration is critical because one parameter setting works on one location does not necessarily work on another. For example, a reference value of 229 (ADC value) gives 100% 134 0 100 200 300 400 T u T 1 d T 2 d all data mote ADC T u T 1 d T 2 d normal flow only Figure4.8: Twobox-plotsillustratethedifferenceintemperatureamongthreeloca- tions. The blue boxes cover both the upper and lower quartile with a red median mark in the center. The whiskers extend to 1.57 interquartile range, excluding outliers (red ’+’ marks). accuracy on T u , but would trigger three false positives on sensor downstream to production valve (T 2 d ). Additional motivation for auto-configuration is that temperature changes con- stantly at the same location. Therefore hand-tuning parameters on each sensor to cope with the location disparity is still insufficient. Figure 4.9 breaks down the T u temperature trace and compares between eight events only when flow is normal. The temperature measurement fluctuation is significant, mostly caused by a noise combination of diurnal amplitude, downhole change and back pressure (Section 4.3.2). The first and last three boxes has no overlap with the other four, whichsuggeststhatmaintainingafixedthresholdduringthewholetimeislikelyto cause mis-detections. Our further study confirm with this observation and hence an adaptive temperature algorithm is necessary. To address the above problems, our flow presence detection auto-configures themostimportantparameter—theCUSUMreferencevaluebasedonthetraining temperature under normal flows and pump shut-ins (details in Section 4.3.2). Our 100% accuracy shows it is effective (Section 4.5.4). More importantly, we find it 135 240 260 280 300 320 340 11:01AM 12:01PM 12:25PM 1:09PM 2:35PM 3:05PM 3:48PM 4:58PM mote ADC Figure 4.9: A box-plot illustrates the fluctuation of temperature when flow is normal at T u during the experiment. The x-axis tick labels are the starting time of each event. is necessary because if a uniform reference value were mis-configured above 229, sensor downstream to production valve (T 2 d ) would start to trigger false positives. 4.5.6 Detecting Pumpjack Operation Previous evaluation shows flow presence detection by temperature is effective. However, inferring blockage solely on flow absence is not enough because regu- lar pumpjack shut-in too stops fluid flow (details in Section 4.2). To avoid false alarms caused by these irrelevant temperature drop, we use acoustic sensing to detectpumpjackstatusandlaterapplytheresultontopoftemperature. Next, we evaluate acoustic algorithm accuracy and draw conclusions based on the results. Wefirstevaluatetheoverall,pump-onand-offdetectionaccuracies. Theoverall accuracy is high, 20203 out of 23129 events (defined by samples in Section 4.5.3) are correct (87%) and the accuracy of detecting pump-on is even higher, 96%. 136 However, the accuracy of detecting pump-off is 69%, which is low compared to the other two metrics. Next, we visualize the algorithm output trace to investigate why pump-off detection only works partially. Pump-off detection is much less accurate than overall and pump-on detection. Tounderstandthedifference,weneedmoreinformationaboutwhymanypump-off samples trigger pump-on detection. Figure 4.10 visualize the details by showing theacousticamplificationtraceandalgorithmoutputsonmote. Aredthicklinein theupperplotindicatesthethreshold(θ p )weusedinourfieldtest. Thehighspikes follow every valve status change is because the relatively loud noise generated by wrench-valve clanging is captured by the acoustic sensor. One reason pump-off detection is not effective is that our acoustic algorithm runs with fixed threshold. For example, the noise floor rises in the third pump-off period (4:30pm–4:58pm), triggering false positive under now-too-low threshold. We are currently working on making our algorithm adaptive to cope with this situation. In addition to an adaptive algorithm, this result suggests that filtering noise during pump-off may too improve pump-off accuracy (or overall). To verify if noise-filtering helps, we next apply our algorithm to a cleaner dataset by PC. We find the easiest way to improve the accuracy of pump-off detection is to upgradethehardwareforacousticmeasurement,becausethesamealgorithmworks perfectly on PC acoustic dataset. Figure 4.11 shows the same experiment as for- going but collected by PC microphones. We take six minutes of the data, cov- ering the first pump-on/-off transition to configure the threshold, with the same auto-configuration algorithm. The end detection result is encouraging, with 100% accuracy. We study the hardware difference helping PC microphone collect better data with higher SNR for future mote improvement guidance. We focus on the 137 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 0 100 200 300 valve: closed valve: closed valve: closed valve: closed valve: open valve: open valve: open valve: open valve: open valve: open pump off pump off pump off mote ADC off on P o * time acoustic θ p Figure 4.10: Acoustic pumpjack status detection result by mote. Pump status is denoted at the top of both plots. Dashed/dotted vertical lines indicate the production valve is open/closed. The tags “valve” means the production valve. three most important stages in line with converting raw vibration to digital signal: microphone, preamplifier and analog/digital converter (ADC). After comparing the specifications, we find that PC beats mote microphone in all the three stages. First, although mote microphones have higher sensitivity (-45dB, [Pan]) than the one with PC (-54dB, [Aud]), but PC microphone has a much smaller resistance (1kΩ against 2.2kΩ, potentially able to produce larger curre nt under the same sound pressure. Second, our PC has better preamplifier embedded in its sound card, SoundMAX AD1988A [Ana06] than mote does [Max12]. For example, the former has higher input impedance but lower total harmonic distortion. Finally, the ADC on the microcontroller of our motes have a much lower resolution, 10-bit comparing to PC sound card’s 24-bit resolution. In short, all these differences results in better quality of PC data with larger SNR than the ones collected by our acoustic mote system. 138 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 0 0.02 0.04 0.06 0.08 0.1 valve: closed valve: closed valve: closed valve: closed valve: open valve: open valve: open valve: open valve: open valve: open pump off pump off pump off PC amplitude off on P o * time acoustic θ p Figure 4.11: Acoustic pumpjack status detection result by PC. In this section, we evaluated our acoustic mote for pumpjack status detection. We further applied the same algorithm to PC acoustic data to provide more com- plete verification of our acoustic detection approach. Based on experiment results, weconcludethatourmoteacousticdetectionsystemperformsreasonablywelland it is easy for our algorithm to achieve perfect accuracy on a cleaner dataset. We next evaluate the end blockage detection on top of both temperature and acoustic sensing. 4.5.7 Blockage Detection Accuracy Inordertoreviewourendblockagealgorithmperformance,wedeployafullsystem in our field test. The blockage detection fuses both pumpjack status and flow presence to determine if pipe is clogged or not. In the section, we evaluate its accuracy after defining the metrics below. 139 Table 4.3: The accuracies of blockage detection correct incorrect location tp tn fp fn Accu all T u – 8 2 0 80% T 1 d 3 6 1 0 90% T 2 d 3 6 1 0 90% total event#: 10 Like temperature and acoustic sensing (Section 4.5.3), we define the metrics of blockage detection in a similar event-based manner. Since blockage detection is mostly based on temperature sensing, event is likewise defined by experiment interval. Thetermstodenotethecorrectnessisasfollows: aTrue Positive iswhen the algorithm correctly declares an emulated blockage; a True Negative is when flowisnormalor thepumpisoff,andthealgorithmremainssilent;aFalse Positive is when flow is normal or the pump is off, but the algorithm incorrectly declares a blockage;aFalseNegativeiswhenthealgorithmmis-detectsanemulatedblockage. We then accordingly defines the overall accuracy, Accu all . We first evaluate Accu all after fusing both temperature and acoustic results. Table 4.3 shows that overall accuracy of our fully-automated system is between 80% and 90%. This result further shows our blockage detection algorithm is very accurate. Thistablesuggeststwofurtherobservations. First,alltemperaturedrops caused by blockages are correctly detected, because no false negative occurs across all three sites. This sensitivity of our blockage detection algorithm to temperature dropisconsistentwiththeresultwehaveinevaluatingourflowpresencedetection in Section 4.5.4. In addition, no false positive further indicates that our algorithm is general to different situations, because we emulate different blockages which forms either during pump shut-in or during pump operation. 140 The second observation is the detection period is short, meaning our system is able to give rapid feedback. In problem statement (Section 4.2), we explain why rapid feedback is important to mitigate the loss. We find it generally takes between 10 to 30 minutes before our algorithm triggers. The third observation is that some false positives are raised. We next evaluate why a perfect flow presence detection does not lead to a perfect blockage detec- tion. To answer this question, we need to investigate the result on each event, particularly on false positives. The three figures in Figure 4.12 visualize our fully- automated system outputs and show why false positives exist. The lower plot in each figure contains both the certainty of drop (C d ) and the ultimate blockage indication. We see that there is transients after the pumpjack resumes operation. A blockage signal raised at 2:35pm in Figure 4.12(a) because the pipe skin tem- perature resumes to normal sightly later than the temperature sensor first receives pump-on signal. We expect our base algorithm to remain silent, although false positive is triggered. However, this can be easily fixed in an extended algorithm which suppresses anomaly outputs a short while after temperature sensor receives pump-onsignal. Therefore,westillcountitatruenegativeinlaterevaluation. One major cause of false positives on all three motes is incorrectly reporting pump-on during the third pump-off period (4:30pm–4:58pm), under effective temperature drop detection. However, the blockage detection algorithm successfully suppress the suggested blockages (i.e. temperature drops) in the first two pump-off period because of a build-in anti-false-alarm feature which ignores sporadic mis-detection (fp) in pumpjack status. The other cause is parameter mis-configuration. Due to the imperfect pumpjack detection (an overall accuracy of 87%) in Section 4.5.6, reported pumpjack status often incorrectly flips in the middle of an event, causing 141 mis-configuration on anomaly and quality levels. However surprisingly, a rela- tive chaotic parameter auto-configuration scheme does not throw off our entire blockage detection. Our algorithm exhibits robustness against the configuration errors, which at best cause only one false positive on T u alone (after 1:09pm in Figure 4.12(a)). In all, our blockage detection algorithm has a high accuracy, 80% in the worst case. A close look on algorithm output plots shows what causes mis-detection. However, the interesting results raises one further question about the robustness of our algorithm against the jittering in pumpjack detection. We next investigate this issue in Section 4.5.8. 4.5.8 Robustness of Blockage Detection Theresultsinevaluatingourblockagedetectionsurpriseusbecauseblockagedetec- tion is often insensitive to the error in pumpjack status detection, after fusing it withflowpresencedetection. Onecommonalityintheseerrorsisthatthethreshold to detect pump-on is such mis-configured that detection frequently (in minutes) flips between pump-on and -off. In order to verify if this feature is systematic or random, we take a three-fold approach in this section. We first formalize three simplified detection cases to theoretically prove our hypothesis. Second, we run simulation over these cases, and finally we further support the simulation result by experimental data. The simplified detection cases consists of both acoustic and temperaturemodalities. Weemploythesameacousticmodelacrossthethreecases 142 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 0 100 200 300 400 valve: closed valve: closed valve: closed valve: closed valve: open valve: open valve: open valve: open valve: open valve: open pump off pump off pump off mote ADC T u k false true block. . time 0 3000 . C d (a) Upstream T u . 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 0 100 200 300 400 valve: closed valve: closed valve: closed valve: closed valve: open valve: open valve: open valve: open valve: open valve: open pump off pump off pump off mote ADC T d 1 k false true block. . time 0 3000 . C d (b) Downstream before production valve T 1 d . 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 0 100 200 300 400 valve: closed valve: closed valve: closed valve: closed valve: open valve: open valve: open valve: open valve: open valve: open pump off pump off pump off mote ADC T d 2 k false true block. . time 0 3000 . C d (c) Downstream after production valve T 2 d . Figure 4.12: Blockage detection results. 143 that pumpjack status detection oscillates continuously at every observation time, starting from t 0 : P o i = on ,i = 2m−1 off ,i = 2m Wenextintroducethethreedifferenttemperaturemodels—monotonicallydecreas- ing, stable with fluctuation, and monotonically increasing. The first case we look into is when the pipe skin temperature monotonically decrease. The physical meaning behind this model is that fluid flow stopped either because of pump-off or blockage. Under this situation, we expected that our flow presence algorithm based on one-sided CUSUM should trigger, because observed temperaturesignalshouldstaybelowthereferencevaluesignificantlylongenough. We denote the temperature readings over time as: s i >s i+1 According to the pumpjack status above and without losing generality, we assume pumpjack switches from on to off at t 2m . At the same time, we update the quality level, μ 0 2m correspondingly: μ 0 i = s i−1 ,i = 2m−1 s i ,i = 2m 144 Please see Section 4.3.2 for more details about parameter tuning. Likewise, at t 2m+1 we update the anomaly level, μ 2m+1 : μ 1 i = s i ,i = 2m−1 s i−1 ,i = 2m Hence, the reference value, k i between [t i , t i+1 ] is: k i = μ 0 i +μ 1 i 2 = s i−1 +s i 2 >s i >s i+1 This result clearly show that the reference is always larger than immediate tem- perature observation and therefore certainty of drop (C d ) builds up, triggering algorithm. The left half of Figure 4.13 illustrates this process. The right half is an excerpt of our experiment data rendering similar behavior. The k (solid thick line) in the top right plot comes from the result of our automated algorithm. On the contrary, the k ′ (dashed line) comes from the hand-set version to show what we expected to see if parameter setting is perfect, which yields C ′ d below. Thesecondcaseiswhenthepipeskintemperaturegenerallystabilized,butwith naturalfluctuation. Thephysicalmeaningisthatfluidflownormally. Weexpected that our algorithm should remain silent because temperature should stay high enough. Although this model can represent the situation when pipe completely cools off, we do not discuss it because of no equivalent data. We denote the temperature readings over time as: s i =s i−2 145 0 20 40 60 80 t 0 t 1 t 2 t 3 t 4 t 5 t 6 t 7 pseudo−ADC signal k μ 0 μ 1 off on pseudo−P o time 11:30AM 12:00PM 0 100 200 300 400 mote ADC time 0 3000 C d time T d 1 k k’ C d C d ’ Figure 4.13: A monotonically decreasing example illustrates the robustness of our blockage detection algorithm. and without losing generality: s i >s i+1 ,i = 2m+1 The quality and anomaly level update schema is the same as the last case, which leads to the reference value, k i between [t i , t i+1 ]: k i = μ 0 i +μ 1 i 2 = s i−1 +s i 2 <s i ,i = 2m+1 >s i ,i = 2m The above relationship between reference value and signal indicates that the tem- perature is likely to envelope the reference value and hence C d cannot easily build up. Our algorithm is capable of suppressing false alarms in this case, shown in Figure 4.14. 146 0 20 40 60 80 t 0 t 1 t 2 t 3 t 4 t 5 t 6 t 7 pseudo−ADC signal k μ 0 μ 1 off on pseudo−P o time 3:00PM 3:30PM 4:00PM 0 100 200 300 400 mote ADC time 0 3000 C d time T u k k’ C d C d ’ Figure 4.14: A stable with fluctuation example for robustness analysis. The final case is when the pipe skin temperature monotonically increase. The physical meaning is that fluid flow resumes after temporary pump shut-in. We expected that our algorithm fall back to silent quickly after temperature becomes high enough. Similar to the stable case, we denote the temperature readings over time as: s i =s i−2 ands i >s i+1 ,i = 2m+1 Likewise the reference value, k i between [t i , t i+1 ]: k i = μ 0 i +μ 1 i 2 = s i−1 +s i 2 <s i <s i+1 Since k is always smaller than the temperature, C d will not build and hence our algorithm achieves true negative. This process reversely mirrors what we observe in the decrease case above, depicted in Figure 4.15. Summarizing how it response to all the three models, we conclude that our blockage algorithm is systematically robust against the negative effect of jittering 147 0 20 40 60 80 t 0 t 1 t 2 t 3 t 4 t 5 t 6 t 7 pseudo−ADC signal k μ 0 μ 1 off on pseudo−P o time 11:00AM 11:10AM 0 100 200 300 400 mote ADC time 0 3000 C d time T u k k’ C d C d ’ Figure 4.15: A monotonically increasing example for robustness analysis. pumpjack status detection on parameter setting. The most significant reason is that under those circumstances, the reference value correctly stay above or below the temperature by tracking it as low-pass-filtered signal. 4.5.9 In-lab Near-Full Blockage Detection In prior sections, we show that our multi-modal detection correctly detects full blockages in the field. However, full blockages can quickly result in damage to equipment, so we would like to detect blockages before they fully close the line. We therefore next extend our work to detect near-full blockage. We verify this extension with laboratory tests. Unlike full blockage, we do not evaluate near- full blockage in oil field because it is not safe to emulate a realistic one—opening production valve slightly but with circulation valve closed would cause over-high pressure at the wellhead. Instead we use a mock-up pipeline system in the labo- ratory. We choose hot water as the fluid because it has some similar properties 148 ! "# # (a) Logical view of lab test. (b) Physical view of lab test. Figure 4.16: In-lab, near-full blockage test on water. as oil—both are incompressible and moderately warmer than ambient. We next introduce our lab experiment and show the results of our multi-modal sensing on near-full blockage. Toevaluatepartial-blockageddetectionwesetupatestbedinourlab. Wecon- structed a recirculating network of hot water similar to that used in Section 3.7.5. It consists of a tankless water heater; a recirculation pump; a plastic, lidless tank; and a small network of PVC pipes and valves (Figure 4.16). Because we are eval- uating multi-modal sensing, we deploy both temperature and acoustic sensors. To detect pipe skin temperature, we tape down USB-based Go!Temp temperature sensor[Ver]onthearterylineafteravalve(Figure4.16(a)). Ouracousticsensoris a lavalier microphone, the same as the one used in our field test to collect ground truth data. For acoustic detection, we must account for differences between the signal of pumpjack operation in the field, and the water recirculation pump in the lab. 149 The major difference between pump-on and -off is average amplitude, and our algorithm is still able to distinguish the two when we adjust parameters. Please see detailed explanation on signal difference and evaluation result in later this section. In deployment, we tape down the microphone on the recirculation pump to detect the pump on/off status by picking up pump operating noise. Similar to the field test (Section 4.5.1), we collect raw pump noise with a sampling rate of 8kHz, and down-sample to 2kHz before aggregating into one-second-long samples measuring average noise amplitude. We later run our detection algorithm over these aggregated samples. Finally, we do not use fieldable hardware in our lab near-full blockage test, contrary to our field test for full blockage. The hardware difference is because the field-test hardware are designed and assembled for oil field environment, and hence it takes extra amount of work on tuning them for the lab environment. Fortunately, the key properties of the two different fluids (oil/gas/water mix in the field network, and water in our laboratory network) are similar enough that the success of this lab test demonstrates that our multi-modal sensing can detect near-full blockages in oil field. In addition, this lab test generalizes our algorithm to applications other than those in oil industry. Ourlaboratorytestsfollowthesameprocedureasourfieldtests(Section4.5.2), however here we emulate near-full blockage by closing about 90% the valve rather thanclosingitcomplete. FollowingFigure4.16(a),weclosevalveV.a ontheartery line in, and open the branch valve to keep water flow and the heater running. Table 4.4 shows our lab test schedule, with two pump-off periods and five near-full blockages. In the six-hour-long test, we collect both temperature and acoustic traces by PC. We run our detection algorithm on a PC and do analysis of the 150 Table 4.4: In-lab experiment schedule Artery start pump valve purpose 1:00pm on open T learns μ 0 1:15pm off T learns μ 1 1:56pm on 90% off non-op 2:29pm open T learns μ 0 2:50pm 90% off in-op 3:31pm off T learns μ 1 4:04pm on non-op 4:35pm open T learns μ 0 5:15pm 90% off in-op 5:55pm open T learns μ 0 6:25pm 90% off in-op data after collection. In principle we can integrate our algorithm with the sensors andrunwithexactlythesamehardwareandsimilarsoftwaretothefield, however, here our goal is to test the generality of the algorithm, so we do analysis off-line to allow us to study a range of algorithm parameters post-facto. We use the same evaluation metrics as in field tests (Section 4.5.3). Figure 4.17 shows that our flow presence detection algorithm, with adjusted parameters, gives rapid and accurate response on abnormal flow (pump-off and near-full blockage). Certainty of drop (C d ) correctly builds up on all seven abnor- mal flow events within 20 minutes (the lower plot of Figure 4.17). In addition, no false positive is raised when we open up the valve to learn μ 0 . Togettheseresults,wemustadapttheparameterstodetectnear-fullblockages of the flow. We found that the base algorithm takes too long (about an hour) to trigger on blockages at 3:00pm, 4:00pm and 6:30pm, when the the reference value (k) is set at the mid point of anomaly (μ 1 ) and quality (μ 0 ) levels. This delay occurs because learning μ 1 out of pump-off temperature underestimates the 151 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 6:00PM 7:00PM 30 40 50 valve: 90% valve: 90% valve: 90% valve: 90% valve: 90% valve: 90% valve: open valve: open valve: open valve: open valve: open pump off pump off temperature ( o C) temp k false true block. . time 0 500 . C d Figure 4.17: Temperature flow presence detection in water pipeline. temperature behavior at near-full blockage. The different temperature behavior is manifested by the two facts—temperature drop at in-op is more gradual than pump-off drop and temperature rebounce a little at non-op. In another words, the suboptimal μ 1 makes the algorithm insensitive to the temperature drops, because it tunes k too low and the lower k, the longer before temperature drops below k and triggers detection. Therefore, to adjust the parameters of our algorithm for near-full blockage detection, we increase μ 1 by 20% every time before computing a new k. We next evaluate our acoustic pump status detection and find theoverall accu- racy is 100%, shown in Figure 4.18. More specifically, all nine pump-on and two pump-off periods are correctly detected. Like temperature detection above, we adjust the parameters in acoustic detec- tion because the signal from the water recirculation pump is different from the pumpjack in the field test. The water pump-on noise is a wide-band signal from the motor and fluid flow, and unlike the oil pumpjack, there is no periodic cycle 152 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 6:00PM 7:00PM 0 0.01 0.02 0.03 valve: 90% valve: 90% valve: 90% valve: 90% valve: 90% valve: 90% valve: open valve: open valve: open valve: open valve: open pump off pump off PC amplitude off on P o * time acoustic θ p Figure 4.18: Acoustic pump status detection in water pipeline. and bursty rod-tubing clang. However, the major difference between pump-on and -off signal lies in the average amplitude—pump-off is much quieter relative to ambient noise, while pumpjack flow is relatively loud. The result shows, with proper parameters, our algorithm is capable of distinguishing average amplitude difference between pump status. In this test, we use a longer cycle (18s) and lower pump-on threshold (60-percentile of amplitude among pump-on training samples), comparing to our field test. On the other hand, the success of our acoustic detec- tion on signal with different properties than pumpjack generalize our approach to a broader range of applications. Combining the perfect temperature and acoustic detections, the end near-full blockagedetectionaccuracyis100%overthetotalofelevenevents. Thelowerplot of Figure 4.17 shows that our algorithm triggers on all five blockages and the two pump-off periods are correctly suppressed. 153 Inall,ourlabtestshowsourmulti-modalsensingworkswithparameterchanges on near-full blockage detection in a hot water network. We believe this result generalizes to real-world near-full blockage in oil flowlines. 4.5.10 EvaluationSummaryandAlgorithmGeneralization We have shown our multi-modal approach works well to detect cold-oil and hot water blockages. We next consider related problems and generalize our multi- modal complementary collaboration to a broader application domain. We believe our general approach—use of two different types of sensors, one accurate but noisy, and the second to filter the noise—generalizes to other sensing problems. Our algorithm may apply to other applications where blockages occur, in addition to cold oil lines. For example, vehicle or building cooling systems may have similar blockage problems in systems that duty cycle, as do hot water distri- bution systems. These systems both have fluid moving through a pipe network, where flow can be detected by temperature variation from ambient, yet one must also coordinate with driving machinery that operates intermittently and would otherwise cause incorrect outage detection. 4.6 Preliminary Experiments To progressively test components of our multi-sensor detection system for cold oil blockage, we carried out a total of five on-site, preliminary experiments prior to the one in Section 4.5. In this section, we briefly talk about their goals and results in the order they were carried out. 154 4.6.1 Short-term Initial Data Collection Toverifyifcertaintemporalorspectralpatternsexistinflowlinedata,wefirstcarry out preliminary data collection. We took this data on February 8 th , 2011. Our initialhypothesisisthatcertainuniquefeaturesmaymanifestinthetemperatureor acoustic data when fluid flowing in pipe. We collect pump-on acoustic samples for about10minutesbyPCandpressingLabtecAM-222handheldmicrophoneagainst pipe. For feature comparison, we collect ambient acoustic data. In parallel, we deploy USB data logger and thermocouple and collect pipe skin temperature data while pumpjack is operating. After analyzing the data, we cannot find significant acoustic signature corresponding to fluid flow, mostly because the acoustic signal is dominated by pumpjack noise. We run FFT on two-second-long sound clip containing swooshing sound of fluid and compare the resulting spectrum against those of the other period. Both spectra look like similar wide-band noise, despite the clear acoustic difference. The temperature data tells us what is the typical pipe skin temperature under normal flow condition. We further conclude that our low-cost mote microphone shall be thermally insulated from the pipe. Because these results are mostly negative, we do not report raw data here. 4.6.2 Medium-term Data Collection On May 9 th , 2011, we carried out our second data collection in field. The purpose of this experiment is two-fold. First, we verify if we can pick up good acoustic samples without direct contact between microphone and pipe to avoid potential heat damage to the microphone. Second, we want to know how internal fluid flow affects pipe skin temperature by collecting long enough data (hours-long) and observing a full cycle of pump shut-in and operation resumption. 155 Figure 4.19: Lavalier Microphone mounted on pipe without direct contact. Our prior experiment in February 2011 shows we have to apply insulation between microphone and pipe, because the sub-100℃ pipe skin temperature may damage our microphone. We decide to hover the microphone above the pipe with- out any contact. Figure 4.19 shows our microphone and the mounting design. The microphone is the same as the one in the PC system deployed in our Novem- ber 2012 experiment (Section 4.5). The microphone is clamped on an “L”-shaped support, which is firmly attached to pipe by a hose-clamp. We apply foam wind- screentoreduceenvironmental(especially,wind)noise. Thereareabout2mmgap between the windscreen tip to pipe. We expect this microphone implementation can pick up equivalent acoustic data to the handheld microphone we used before. However, our new microphone attachment design almost loses all fluid flow signal, but instead picks up large volume of environmental and pumpjack rumbling noise. We conclude that we must directly attach microphone to pipe to get data about internal fluid flow. In parallel, we collected three-hour-long temperature data to see how change on flow condition affects pipe skin temperature. The data from this collection is necessary for identifying any long-term trend of pipe skin temperature, whereas 156 Figure4.20: Temperaturedatafromourseconddatacollection. Theshadedregion in center indicates when pumpjack shuts in. we did not collect long enough data in the prior experiment. We again deploy thermocouples and USB data loggers near the wellhead. Our experiment start from 11:00am to 2:00pm, with pump-on at the beginning. At 12:00pm, we man- ually shut-off the pumpjack for an hour and resume pump operation at 1:00pm. Figure 4.20 shows how temperature reacts to fluid stoppage. Within an hour of stoppage (the shaded region), pipe skin temperature drops from 85℃ down to about 55℃, converging to the ambient temperature (about 20℃ at noon). This result supports our hypothesis that blockage should cause significant temperature drop around it. 4.6.3 Acoustic Sensor and Thermal Insulation Tests To test thermal insulation and collect more data for signal feature analysis, we carry out our third preliminary experiment on August 18 th , 2011. 157 Figure 4.21: New (current) lavalier microphone mounting design (before applying windscreen) for PC acoustic system. Wefirstverifytheeffectivenessofpotentialerrorsourcesonacousticsignal. The potential three error sources we identify include thermal insulation (fire blanket), sensor type and its location. Therefore we take short-term acoustic data with or without applying fire blanket between microphone and pipe to verify if insulation significantlyattenuatessignal. Wecompareacousticdatacollectedbyeitherhand- held microphone or lavalier microphone with new mounting design (Figure 4.21). This design uses a widget to clamp the tip of a lavalier microphone on the curved surface of pipe. Finally, we take acoustic data on pipe at wellhead and far away, about 3m downstream to the wellhead. The final result shows none of these three factors significantly affects the quality of acoustic data. To further investigate if any acoustic feature exists in different components of a pump stroke cycle, we dissect waveforms on cycle-level. Prior two experiments showthatnosignificantfeaturecanbefoundinpump-ondataifweFFTwaveforms holistically. After listening to the sound clips of different cycles, we identify four common components in a waveform: “clang”, loud noise from rod-tubing clang; “gush”, flow signal when oil extraction passing the sensor at maximum volume; 158 0 0.2 0.4 0.6 0.8 1 10 −15 10 −10 10 −5 10 0 Frequency (kHz) Power noise.clang signal.gush signal.torrent noise.ambient (a) Cycle 1. 0 0.2 0.4 0.6 0.8 1 10 −12 10 −10 10 −8 10 −6 10 −4 10 −2 Frequency (kHz) Power noise.clang signal.gush signal.torrent noise.ambient (b) Cycle 2. 0 0.2 0.4 0.6 0.8 1 10 −15 10 −10 10 −5 10 0 Frequency (kHz) Power noise.clang signal.gush signal.torrent noise.ambient (c) Cycle 3. Figure 4.22: FFT results on three sound clips of typical pump cycles dissecting different sound component. “torrent”, flowsignalwhenfluidswooshingdownstream; and“ambient”, minimum audible fluid and pumpjack noise, usually between cycles. We run FFT over each component and compare against each other on a large number of cycles. Because most of power component is in frequency below 1kHz, we only focus on this range. We find other than the absolute power different, no narrow-band feature exists in any component. Figure 4.22 shows FFT results on the acoustic signal of three randomly chosen cycles. To check how temperature reacts to a more realistic pumpjack on/off schedule, we take three-hour-long temperature data in parallel. In prior experiment, we manually shut-in the pumpjack for about an hour. However in real operation, 159 Figure 4.23: Temperature data shows pipe skin temperature fluctuation corre- sponds to the pumpjack status. shut-in are usually shorter. We deploy temperature sensor to a pump which is programmed to do frequent and periodic shut-in. The pumpjack operates for about two minutes and shuts in for about eight minutes, in a total of 10 minutes cycle. Figure 4.23 shows that pipe skin temperature drops about 10℃ in each eight-minute pump-off period. 4.6.4 First Prototype Test Prior data collections and short-term experiments are all component-oriented on PC platform. Based on the data and the analysis, we next implement our cold-oil blockage detection system on mote platform. On April 12 th , 2012, we carry out field test with our integrated prototype. We deploy acoustic sensor at well head and two pairs of temperature sensors straddling production and circulation valves. Figure 4.26 shows our initial design of mote microphone mounting. We do not detach the microphone from the Mica Sensor Board but tape down the whole unit ontopoffireblanket. Ourtemperaturesensoristhesameasalllaterexperiments, 160 Table 4.5: First prototype test schedule start product. valve purpose 10:00am open experiment start 11:08am close production line blockage 12:27pm open normal flow 13:11pm close production line blockage 14:27pm open normal flow 15:11pm close production line blockage 16:17pm open normal flow Figure 4.24: Mote microphone mounting without decoupling. includingtheoneinSection4.5. AllthesefivesensorsareconnectedtothreeMica- 2 motes. We alternate the status of production and circulation valves for multiple times to emulate flowline blockage, the schedule shown in Table 4.5. Due to time limit, we leave the pumpjack on all the time. Therefore, we do not test pump-off detection by acoustic sensor this time. To verify flow presence algorithm, we deploy temperature motes running the algorithm in the field test. Some unknown problem causes one of the two motes returns incorrect readings, which connects to the thermocouple pair straddling the circulation valve. Therefore we omit that dataset and only report the result from theproduction-valvepair. Similartotheacousticalgorithm,suboptimalparameter 161 setting throws off the detection—only one out of three flow stoppages caused by a closed valve is detected. We plot the mote data against data collected by USB data logger in Figure 4.25. Although flow presence algorithm is partially working, we make two observations out of the comparison and they are useful to our future deployment. One is that the temperature motes collect reasonably good data. After we convert the USB data to ADC unit, we see the fluctuation is consistent between mote and USB data. On the other hand, we need to tune up the gain of the amplifier board. In this field test, we use the same amplifier board as the one in steam choke blockage detection in Section 3.8. However, the board is optimized for steam pipes at 100 gain, but not for oil pipe. Here the pipe skin temperature is below 100℃, making average ADC reading around 80, which is too low regarding the ADC range from 0 to 1024. In our later experiments, we modify the amplifier board and increase the gain to 367 for better signal. In parallel to the temperature motes, we run the acoustic algorithm in our field test to verify if pump-on detection works. However, the correctness is only 47% because our predefined threshold is suboptimal. Therefore we look into the basic statistics of this dataset to help us tune the parameter in the future. Figure 4.26 shows the average noise level captured by our acoustic mote when pumpjack is on. In addition to algorithm and mote data quality, we further evaluate the radio connection between sensor to base station. Similar to Section 3.8.3, we expect packet loss because of the frequent transmission from the three sensors. Overall, thepacketlossisacceptable,0.81%atboththeacousticmoteandproduction-valve mote, and 1.51% at circulation-valve mote. Figure 4.27 shows the loss distribution is generally uniform. We further conclude the wireless communication is reliable in this cold-oil blockage system. 162 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 20 40 60 80 100 120 140 160 pv: off cv: on pv: off cv: on pv: off cv: on pv: on cv: off pv: on cv: off pv: on cv: off ADC time mote USB (a) Upstream to the production valve. 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 5:00PM 20 40 60 80 100 120 140 160 pv: off cv: on pv: off cv: on pv: off cv: on pv: on cv: off pv: on cv: off pv: on cv: off ADC time mote USB (b) Downstream to the production valve. Figure 4.25: Comparing mote and USB data logger temperature data. −50 0 50 100 150 200 0 1000 2000 3000 4000 mote ADC samples samples: 20981@1Hz median: 16 mean: 23.3058 stdDev: 26.6031 Figure 4.26: Basic statistics of acoustic data. The red line is normal distribution fitting. 163 Figure 4.27: Mote radio packet loss in prototype test. Each marker represents one sample missing from dataset. Table 4.6: Second prototype test schedule product. circulat. start pump valve valve purpose 10:30am on close open product.: in-op; circulat.: normal 11:22am open close product.: normal; circulat.: in-op 12:13pm off close open pump-off 13:09pm on product.: non-op; circulat.: normal 13:41pm off open close pump-off 14:34pm on product.: normal; circulat.: non-op 14:57pm close open product.: in-op; circulat.: normal 15:57pm open close product.: normal; circulat.: in-op 4.6.5 Second Prototype Test To test our modified amplifier board for temperature sensor and test our acoustic algorithm on pump-off detection, we carry out our second prototype field test on July 16 th , 2012. The deployment plan and sensor placement is similar to our first prototype test and we collect a total of six-hour-long temperature and acoustic data. In this experiment, we manually add two pumpjack shut-in stages, shown in Table 4.6. We deploy temperature motes with modified amplifier boards to see if they improve the temperature data. Figure 4.28 depicts that our motes are almost equivalent to USB data loggers in temperature measurement. The average ADC 164 10:00AM 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 0 100 200 300 400 500 600 pv: off cv: on pv: off cv: on pv: off cv: on pv: off cv: on pv: on cv: off pv: on cv: off pv: on cv: off pv: on cv: off pump OFF pump OFF pump on pump on pump on pump on pump on pump on ADC time mote USB Figure 4.28: Temperature data at mote after circulation valve. reading on normal flow is around 300, much larger than the 100 in the last experi- ment. Temperature fluctuation accurately reflects our pump and valve setting. In the lab we replay our CUSUM-based flow presence algorithm on this dataset. The resultshowsthatwithproperparameters,wecanachieve100%detectionaccuracy. Wealsoverifyifourfullacousticpumpjackstatusdetectionworks. Welearnthe optimalthresholdfromourlastprototypetestdataandhardcodeitintheacoustic sensor. However, this threshold does not reflect the correct signal property in this test and therefore the algorithm accuracy is as low as 30%. Despite the clear visual difference between pump-on/off signal in Figure 4.29, the actual statistical difference is more subtle. We conclude that we must implement parameter auto- configuration, and increase the acoustic signal gain to make pump-on/off signal more distinct to each other. 4.6.6 Preliminary Experiment Summary In this section, we walk through all of our preliminary experiments preceding our November 2012 field test (Section 4.5). Overall, these experiments collectively 165 10:00AM 11:00AM 12:00PM 1:00PM 2:00PM 3:00PM 4:00PM 0 20 40 60 80 pv: off cv: on pv: off cv: on pv: off cv: on pv: off cv: on pv: on cv: off pv: on cv: off pv: on cv: off pv: on cv: off pump OFF pump OFF pump on pump on pump on pump on pump on pump on ADC time Figure 4.29: Acoustic data from second prototype test. and progressively help us eventually achieve our successful November 2012 deploy- ment. In addition, these preliminary experiments show the methodology we took to develop this detection system and solve cold-oil blockage problem. We next highlight what we learn from two aspects—algorithm development and deploy- ment. Theresultsfromthesepreliminaryexperimentshelpusdesignandtestourflow presence and pumpjack status algorithms. After identifying the trend of temper- ature fluctuation in data collected, we choose one-sided CUSUM for flow presence detection (Section 4.3.2). Acoustic samples from different wells and pumpjacks show that it is impractical to extract reliable acoustic features of fluid flow by our low-cost sensors. Therefore we later build our pumpjack status detection based on overall amplitude change (Section 4.3.3). The partially-working algorithm in field facilitates our auto-tuning implementation for both temperature and acoustic sensing. Finally, we carry out our preliminary experiment among three different wells with different pump-off schedules. In result, the consistent temperature and 166 acoustic measurement manifests the generality of our algorithm—it automatically adapts to different wells and pumpjacks. Inadditiontothealgorithms,theexperimentsimproveoursystemdeployment. Thesefieldtestsprovideagreatopportunitytoindividuallyandholisticallytestthe components (hardware and software) in our system, maximizing the accuracy and robustness. We optimize our thermocouple amplifier boards to cold-oil application after the unsatisfactory results in the first prototype test (Section 4.5.1). We witness temperature sensor error in different tests and we later find the culprit is likely to be sun-exposure on amplifier boards (Section 4.4.1). Based on our initial temperature observation, we employ thermal insulation to prevent potential heat damage to microphones. The last example is that the second prototype test shows it is necessary to increase the acoustic signal gain. We later come up with software and hardware modification to obtain the gain-tuning (Section 4.4.3). 4.7 Conclusions on Cold-Oil Blockage Detection Thissectionsupportedourthesisbyshowingmulti-modalcomplementarycollabo- ration enables cold-oil blockage detection. This blockage problem is impractical to single-modal, non-invasivesensing. Thissuccesspromisesthatmulti-modalcollab- oration enables event detection applications where single-modal low-cost sensors is not enough. Inthischapter,weexploredhowmulti-modalnon-invasivesensingwithlow-cost temperature and acoustic sensors reduces error rate in cold-oil blockage detection. We first used temperature sensing to detect suggested blockage and acoustic sens- ing to suppress false alarms from regular pump-off. We next demonstrated the effectiveness of our algorithm on oil-line full blockage with an integrated system in 167 field tests. We further generalize our multi-modal approach to near-full blockage in lab water network. Finally, we discussed our preliminary experiments and their results. 168 Chapter 5 Related Work In this section we outline the research work related to this thesis: multi-sensor vehicle classification (Chapter 2), steam-choke blockage detection (Chapter 3), and cold-oil blockage detection (Chapter 4). We group these related work into six different fields: target tracking (Section 5.1), change point detection algorithms (Section5.2), pipelinemonitoringsystems(Section5.3), oillineblockagedetection (Section 5.4), multi-modal sensing applications (Section 5.5), and energy harvest- ing systems (Section 5.6). 5.1 Target Tracking Target tracking is a typical event detection application where multiple sensors collaborate to decide one or more moving objects’ physical position. Since usually the physical position of sensors is fixed while objects are moving, it is natural to deploymultiplesensorstocoverobjectspossibleroamingspaceasmuchaspossible. In order to save resources, say energy or network bandwidth, or optimize tracking resolutions, tracking systems all leverage multi-sensor collaboration at different levels. Our multi-sensor vehicle classification (Chapter 2), employing two sensors along vehicle travel path, builds on prior tracking works in both unconstrained (Section 5.1.1) and constrained environments (Section 5.1.2). 169 5.1.1 Unconstrained Environment Much early works in sensornets considered target tracking in unconstrained envi- ronments. Early works used dense networks of sensors tracking relatively sparse targets, [SMK + 07,Rei79,OHRS05,SOS08,CWcW06]. Zhao et al. use information theoretic techniques for better vehicle path estimation [ZSR02]. They optimized sensor selection based on cost with an idea that is similar to our competitive collaboration—compare and make use of more appropriate sensors. Sandell et al. and Chen et al. study a specific problem in target tracking—data associa- tion [SOS08,CWcW06]. They focus on how to associate noisy measurements to target tracks, but we instead focus on the correspondence between detections and targets. Shin et al. consider overlapping targets and use information about dis- tinct targets to clarify the status of targets near each other [SGZ03]. As with this prior work, we are concerned with confusion of observations about target near each other (the mis-segmentation in road traffic [PSH + 06]). Competitive collab- oration assumes each individual sensor is locally optimized, and in our case we want to make classification by each sensor as accurate as possible. Although for us the roadway constrains target location (a simplification), greater speed and lower separation complicate our problem, making it harder to do classification on our sensors. We discuss our approaches in Section 2.2. Computer vision provides an alternative approach to tracking. Pahalawatta et al. employ Affine Gaussian Scale Space to match image according to the feature point detected [PDPK03]. The technique was introduced by Baumberg to cope withthesituationthatna¨ ıvedirectcorrelationcoefficientcomparisonisnotenough or even impractical [Bau00]. In Section 2.2.4, we face a similar problem. Besides, 170 they are using “Best-n-match” method, which is infeasible in our scenario because of the uncertainty of incoming vehicle number. 5.1.2 Constrained Environment We next discuss prior works of vehicle tracking or re-identification in constrained environment, which are closer to our vehicle classification application because of the similar roadway constraints. We focus on the effects of signature matching on correctness because they are critical to complementary collaboration in order to make meaningful comparison of results from different sensors. Coifman proposes a system for freeway deployment [Coi98b,Coi98a]. His algo- rithm looks for short sequences of measured vehicle lengths that exhibit a strong correlation between two stations, namely downstream and upstream sensor nodes. The algorithm is similar to our Numbering with Resynchronization (Section 2.2.2), although he employs vehicle length and pattern matching to correct for lane changes. He gets about 65% vehicle matching accuracy at highway speeds. We show better correctness (up to 78%), but at much slower speeds. Cheung et al. consider vehicle classification [CCD + 05,CEV05], and study the matching problem. They use an array of seven, close spaced, three-axis magneto- meters and study seven vehicles on an arterial. They show an impressive 100% re-identification rate. Our results show lower matching accuracy, but over more than 100 real-world vehicles with sensors at 100m separation. Kwong et al. use a statistical model of signatures for vehicle re-identification for travel time estimation [KKRV09]. They claim no ground truth is needed and theirestimatedmatchingrateisabout69%. Weinsteadassumevehicletraveltime is relatively stable and our time-stamp based algorithms have recall above 73%. 171 Oh et al. use heterogeneous detection systems for vehicle re- identification [ORJ07]. They extract rich feature information from each individual vehicle, match them with a lexicographic optimization algorithm, then use the matching for travel-time estimation. Their approach can be computa- tionally expensive; our approaches are quite simple by comparison and generally suitable to run on-line. Our WETW is similar to their prioritized time window algorithm; but we show other algorithms can do almost as well as it, even without extensive features. Sun et al. use a combined inductive loop detector and image process- ing [SARR04]. Their heterogeneous system effectively uses multi-modal multi- sensor collaboration. Their result is encouraging, with up to 91% matching rate in the best case. However, the performance is sensitive to fusion weight and video quality from two stations. We discuss parameter sensitivity of our algorithms in Section 2.3.3. As with Coifman [Coi98b], they consider platoon-comparison and assume highway conditions, while we instead study slower traffic. Three other groups have proposed different method travel time estimation, a part of our time-stamp based signature matching algorithms. Jeong et al. pro- vides a method by the frequency of the vehicle time-stamp difference between sensors [JGHD08]. Dailey [Dai93] and Petty et al. [PBO + 98] have looked at cross- correlation of raw signatures. Because of signature size, they look at options to downsample or aggregate raw signatures. We show that much less information can provide better matching correctness (Section 2.2.4). Park et al. have previously looked at sensor fusion to improve classification accuracy [PSH + 06]. Their work assumed perfect signature matching (an “ora- cle”) and showed that sensor fusion can improve classification accuracy. Here we 172 re-evaluate that work using a realistic signature matching, showing that errors in matching and classification are correlated, so overall accuracy is higher than expected. 5.2 Change-Point Detection Algorithms Change-pointdetectionisacriticalpartinoureventdetectionapplications. Many real-time monitoring systems use abrupt detection [BN93] or change-point detec- tion [TV04] to detect problems in observed data. In our steam-choke, blockage- detection system (Chapter 3) we focus on exponential-weighted moving average (EWMA) for change-point detection because it admits very lightweight implemen- tations, making it well suited to mote-class platforms. In flow presence detection (for cold-oil blockage detection, Chapter 4), we focus on cumulative sum control chart (CUSUM). Our detection algorithms are inspired by several prior systems built on EWMA [Jac88,TGC + 07,KN01] or CUSUM [CGM + 11]. Several sensornets build on the simple EWMA algorithm from TCP [Jac88]. Trifa et al. develop an adaptive alarm call detection system for yellow-bellied mar- mots, using EWMA to estimate environment noise [TGC + 07]. Our steam-choke work uses similar concepts to detect significant change in pipe skin temperature using EWMA. Kim and Noble propose EWMA-based algorithms to optimize streaming estimation of network capacity [KN01]. One of their algorithm, called flip-flop filter, keeps both agile and stable EWMA and switches between the two to find the best baseline. Although our algorithm also maintains two EWMAs, we directly compare these two traces to detect sudden changes in pipeline tempera- ture. 173 Some researchers use CUSUM to analyze time series and we implement it to detect flow presence. Chamber et al. develop a CUSUM-based algorithm to detect vegetationchangesinforests[CGM + 11]. Similartotheirwork, wechooseCUSUM foritscapabilityinidentifyingsmallandgradualchangeandwetoomakeouralgo- rithm adaptive to noise. They post-process year-long time series for small change while we in-node process streaming data. Another difference is their algorithm is designed to identifying multiple decreasing segments in a series, but our algorithm focuses on immediate decision. 5.3 Pipeline Monitoring Systems Wesupportourthesisbystudyingsensorcollaborationintworepresentativeappli- cations of industrial monitoring—steam-choke and cold-oil blockage detections. These two work (Chapter 3 and 4) are inspired by other researchers’ work on pipeline monitoring. In fact, pipeline monitoring is a large class of applications that we believe could benefit from multi-sensor collaboration. SCADA systems have long been used for pipeline monitoring. Tradi- tional SCADA systems use simple in-situ sensors and centralized decision mak- ing [MBL03,DS99], while our approach instead shifts detection algorithms into distributed intelligent, communicating sensor nodes. We find sensor collaboration is much simpler in a centralized architecture than distributed one. Prior sensornet research in pipeline monitoring usually assume low- temperature, single-phasefluid[KSC + 08,JE08,Sin05,SNMT07]. Manyresearchers choose vibration sensing, effectively equivalent to our acoustic sensing, for low- temperaturefluidmonitoring. NAWMSfocusesonpersonalwaterusage[KSC + 08, MCS12]. OurpumpjackstatusdetectionhardwareissimilartotheirsinMicaseries 174 motes and MTS310 sensor boards. However, they use accelerometer while we use the microphone embedded on the same sensor board. They infer flow rate by pipe vibrationfrequencyandlinear-programming-basedalgorithms. Wedonotmeasure vibration in our flow inference because operating pumpjack generates wide-band noise which overwhelms flow vibration signal. Instead, we use vibration as a sec- ondary modality to detect pumpjack operation status. Stoianov et al. prove in PIPENET the feasibility of measuring vibration to detect small leak on water sewage pipe [SNMT07]. However, they do not present a completely integrated system; no detection algorithm is implemented. Their field test only shows that their sensors deployed under urban sewage are capable of collecting certain types of data and relay them back. We instead focus on a complete multi-modal system running oil-line blockage-detection algorithm online. Jin and Eydgahi [JE08] and Sinha [Sin05] utilize acoustic wave propagation for pipe defects detection. Jin and Eydgahi propose a general sensor network frame, while focusing on specific signal processing analytical technique. Sinha’s work is mainly about transducer instru- mentation and calibration for natural gas inspection. Instead of pipe defects, we focus on oil line blockage and real system development and deployment. Zhu’s work is closet to ours, showing the feasibility of temperature monitoring for blockage detection of pulverized coal injection system [Zhu05]. He uses tem- peratureobservation from thermometers mountedon branchpipes in his detection algorithms. Similar to our approach, his algorithm differentiates pipe skin temper- ature and compares the resulting ∆ ud against pre-configured thresholds. Unlike his work, we use EWMA and multiple sensors to adapt to changes, avoiding most hard-coded thresholds. Finally, we use inexpensive and portable hardware (less 175 than US$600), while his system is centralized and likely to be much more expen- sive. Wehavepreviouslyexploredthepotentialofsensornetworksinoilfieldproduc- tion systems [YYH + 11]. While that work suggests the potential, we demonstrates a field-tested system, evaluates specific sensing algorithms, and demonstrates that the whole system can operate on steam-power. 5.4 Oil Line Blockage Detection Applications We study multi-modal complementary collaboration in cold-blockage detection (Chapter 4) to support our thesis. Oil line blockage is not a new event detection problem to oil industry. Other researchers have studied how to detect and localize it with other methods [LLPW10,LS01]. Liu et al. shows it is practical to use decompression wave method to detect blockageinalongoiltransportationline[LLPW10]. Theirsensingpressuremodal- ityisinvasive, differentfromours. Ourmethodcanonlydetectcompleteblockage, the same to theirs. They test their method by emulated blockage by an intermit- tently operating oil line, and we are considering emulated blockage for testing, too. Liu and Scott show a theoretical work to localize blockage in subsea flow- lines [LS01]. They build their work on top of steady-state pressure method for blockage detection. Localization of blockage with pervasive sensing is our future work. Theytoouseinvasivesensingmodality—pressure,whileweusenon-invasive one. Besides, we plan to have a real deployment beyond theoretical models. Inshort,ourmajordifferencefrompriorworkisouruseofnon-invasivesensing. Further, we demonstrate a successful wireless sensornet deployment in the field. 176 5.5 Multi-Modal Sensing Applications Weusecold-oilblockagedetection(Chapter4)asacasestudytolearnmulti-modal sensing in industrial monitoring. Next we review related multi-modality work in the same domain, followed by a broader review in other application domains. 5.5.1 Multi-Modal Sensing in Industrial Monitoring Our cold-oil blockage detection shows complementary collaboration with multi- modal sensing has potentially great feasibility to industrial monitoring applica- tions. A few other multi-modal sensing applications are targeted to the same field [ZTLZ06,FIIS11,GGK + 07]. Zengetal.showusingvibration,forceandacousticemissionsensorstomonitor health of high-speed milling machine and predict wear-out [ZTLZ06]. Our pump- jack status detection is similar to their idea—rumbling machines emit detectable acoustic pattern. However, we are not doing oil pipe blockage prediction and it is partofourfuturework. Futagawaetal.designanintegratedelectricalconductivity temperature sensor for cattle health monitoring [FIIS11]. Their sensing is invasive since they embed sensors into rumens of cattle, while part of the requirement of our oil-line blockage-detection system is non-invasive to lower cost. Gupta et al. use microwave and eddy current image to evaluate corrosion under aircraft paint andinlapjoints[GGK + 07]. Theycollaboratetwomodalitiescompetitively, mean- ing either on can fulfill the task but fusion achieves better results, but our sensor collaboration is complementary. 177 5.5.2 Multi-Modal Sensing in Academic Projects Many academic studies explored multi-modal sensing for different applica- tions, including general sensor fusion [HS07, KSPSV02], target classifica- tion [WQI02], target tracking [GE01,ZSR02,ZB05,AAYA09,KOA + 08,LPR07, KBHK06], human activity or health monitoring [AMM + 08,HZHJ11,SRT + 08, BKTT07,OH05,MVW08,NNT11], Robotic navigation [BL05,SB97], vehicle clas- sification [BRVK11,BGL + 10] and human-computer interaction [QWZ10,TIF05, MFS + 02,Goe06]. Someworkusesmultivariatestatisticsmodelingformultiplesensordatafusion. Goeckeusescoinertiaanalysistofindamathematicalcompromisebetweenthecor- relation of audio and video (3D lip tracking features) [Goe06]. Kushwaha et al. use separate non-parametric model for audio sensors and parametric mixture-of- gaussian model for video sensor in their vehicle tracking application [KOA + 08]. Annavaram et al. use bivariate model for ECG and accelerometer data to monitor sensor bearer’s activity [AMM + 08]. Nguyen et al. correlate different kinds of sen- sors, including acoustic, seismic, infrared and ultrasonic sensors, in a joint sparse modelforhumanfootstepdetection[NNT11]. Ouroil-lineblockage-detectionalgo- rithmhasasimilarwaytodochange-pointdetection. Webuildseparatemodelsfor bothacousticandtemperaturedataandconfigurethresholdsforeach. Inaddition, we plan leverage the correlation between the two modality when pipe is normal. Other work specifically uses Bayesian networks. Tamura et al. use triphone HMM to model audio and video data for speech recognition. They find that a training set of audio-visual data achieves better recognition accuracy than audio- only data [TIF05]. McGuire et al. leverage Bayesian networks to integrate spoken instruction, visual memory, the gesture-based region bias to determine the object 178 tobegraspedbyroboticarms[MFS + 02]. ZouandBhanuevaluatebothtime-delay neuralnetworkmethodandBayesiannetworkmethodforwalkinghumandetection from audio-video data. They conclude that Bayesian network is better because of ease to train, higher accuracy and clearer graphical model [ZB05]. Huang et al. proposeacoupledHMMmethodforaudio-visualjointmodeling,especiallytosolve asynchronization problem in an office activity monitoring application [HZHJ11]. Oliver and Horvitz use layered HMM, a modular and hierarchical HMM method for office activity inference [OH05]. Singhal and Brown uses Bayesian network to joint model audio and video data to predict obstacle in navigation [SB97]. These works model multiple modality jointly but we instead focus on separate modeling, since we do not find significant inter-modality correlation during blockage. Rather than direct fuse multiple sensor channels, a few works utilize a sec- ondary/orthogonal sensory channel to assist the main channel for better percep- tion or sensing. Girod and Estrin suggest using video evidence to solve the obsta- cle problem in their acoustic ranging application [GE01]. Qu et al. add vision (Pan-tilt-zoom camera) and actuator (pan-tilt-unit) to help the LDV automati- cally select the best reflective surfaces, point and focus the laser beam, in order to remotely pickup voice signal [QWZ10]. Stiefmeier et al. integrate several sen- sors into one wearable sensors to monitor worker’s activity. They propose use one kind of sensing result to do automatic data segmentation for other sens- ing streams [SRT + 08]. Bajwa et al. deploy a sensors at highway weight station for truck counting. They use magnetometers to detect vehicle speed and, more importantly, trigger in-pavement accelerometer to count truck axles [BRVK11]. Bischof et al. propose to use acoustic sensor to co-train the classifiers of video sen- sor. The acoustic co-trainer makes their classification system adaptive, requiring 179 less manual labeling of training samples [BGL + 10]. Barakova and Lourens pro- pose event-based data fusion, contrary to fixed time interval based fusion. They use gyroscope data to segment visual data for robotic navigation [BL05]. Our method employs this idea because we are actually using acoustic date to assist temperature date for blockage detection. Our hypothesis indicates if pumpjack is not operating, we have no way to tell if the pipe is blocked. In other words, we use acoustic date to segment temperature date and ignore those when pumpjack is off. 5.6 Energy Harvesting Systems 1 Energy harvesting for sensor network has been an active area of research, which inspire our design of thermal harvester in Chapter 3. Today, electricity can be immediately harvested from several types of energy sources at a relatively low cost. These include light, wind, vibration, heat, magnetic, and radio [CAC + 99]. Low-powersensornodescanbepoweredbyevenasmallfractionofambientenergy. Accordingly, there have been a number of research on scavenging ambient energy to operate wireless sensor nodes, some using traditional sources such as sunlight, vibration, andmechanical[RKH + 05,WY95,PS05,MCS12]whileothersusingmore exotic methods as body heat, radio fields, and multiple energy sources [MCL + 07, SS09,ROC + 03] Oneofthefirstresearchinenergy-harvesterbasedsensornetusedsolarpowerto driveindividualnodes[KS03,VRS03]. Heliomotewasthefirstsystemtointegrated solar-power and power conditioning to drive mote-class hardware [RKH + 05]. We 1 Section 5.6 and literature survey on energy harvesting is done by Affan Syed. We include this section for the completeness of steam-choke blockage detection. 180 leverage this prior work and use a modified version of the heliomote to condition the output of our TEG. Prometheus replaced rechargeable batteries with superca- pacitors to reduce energy loss during the conversion process [JPC05]. AmbiMax platform further increased efficiency by performing maximum power point track- ing (MPPT, essentially matching the source and load impedance) and adds multi- modal (solar and wind) energy harvesting [PC06]. Each of these prior platforms employ battery or supercapacitor to isolate energy harvesting from consumption. We show that this buffer can sometimes be eliminated or reduced. Researchers have looked at optimizing the efficiency of energy harvested with large (25W) waste water systems [RMW + 97]) to lower power generators using automobile waste-heat (4W) [Mat02] and even for micro-generators that use soil- to-air thermal gradient generating a maximum of 0.35W [LS02]. Our work pro- vides a cost-efficient and energy-sufficient solution for powering an embedded sys- tem. Other researchers have investigated harvesting thermal energy for storage in energy buffers. Mateu et al. harvest about 5mW using the thermal gradient between human body and ambient temperature but use NiMH battery to store energy [MCL + 07]. Sodano et al. argues that TEG modules can generate greater power while charging batteries quicker than piezo-electric system under typical conditions [SSDI07]. Our work provides a batteryless solution and focuses on inte- grating thermal harvesting with sensing, optimized for low-power, low-cost, sensor network applications. The Micropelt TE-node is most closely related to our work. The Micropelt platform low-power (sub-10mW) sensor node [Mic10] with an internal 100μF capacitor for energy storage, with harvesting from a custom thermo-electrical gen- erator [BNB + 04,B ¨ 05]. Our work differs in that we use a general purpose sensor 181 platform,evaluatingthepotentialandtrade-offsforbatterylessoperation. Wealso explore a general purpose platform (running TinyOS) which allows us to experi- ment with variety of sensing and sensor fusion algorithms. 182 Chapter 6 Future Work and Conclusions We close our dissertation by listing short-term research suggestions and long-term open research questions. After we address future work, we make some concluding statements for multi-sensor collaboration for event detection. 6.1 Short-Term Future Directions Before open research questions, we talk about the areas of immediate future work for multi-sensor vehicle classification, steam-choke and cold-oil blockage detection. First, some immediate problems remain in multi-sensor vehicle classification (Chapter 2). Currently, some parameters for signature matching are manually input during deployment. Although this is more flexible than hard-wiring param- eters in algorithm, we would like to automate the process. One idea is to use out-of-band information, say sensor geolocation and local speed limit to autotune parameters in time-stamp based algorithm. Another idea is, instead of manual command inputting, we may drive a flag vehicle across the sensors at the begin- ningofdeployment,inordertotriggersensordetectionandoptimizetheparameter with its travel time. Allofoursignaturematchingalgorithmsarebi-sensored-oriented,andwewould like to generalize these algorithms to a network with more than two sensors. The easiest way to port our algorithm a large network is to do hierarchical matching— aggregate matching result after matching between pairwise sensors. The matches 183 are ought to be transitive, meaning if A and B, B and C are matched separately, A/B/C is a match. However, it is interesting to study how non-matches and incorrect matching results affect end matching correctness, because they are likely to complicate the problem. We would like to investigate more on the benefits of raw signature similarity in matching. In Section 2.2.4, we draw a paradoxical conclusion that more infor- mation does not always lead to better decision, mainly because of the poor per- formance of matching by full raw signature comparison. However, we can achieve much better matching if we add temporal coherence to raw signature comparison (RETW). Therefore, we believe we may achieve higher matching correctness by adding other orthogonal information to raw signature comparison, or improving signature similarity computation with more advanced pattern recognition algo- rithms. Second, we demonstrate our temperature based algorithm is able to detect blockages in both steam distribution network and hot water network (Chapter 3), but with different parameters. Therefore, to improve the generality of our algo- rithm, we would like to implement parameter auto-configuration for different envi- ronment. In our current topology, pairwise sensors straddling the blockage point are wired to a mote to do centralized decision. Our next step is make our system distributed by connecting sensors with wireless networks. Although our over-night test of steam-choke blockage detection system is successful, it is necessary to carry out long-term experiments before pervasive deployment practical. To test robust- ness and parameter sensitivity, the system is expected to experience multiple diur- nal cycles, different weather conditions, regular field maintenances and seasonal variations. We need to verify if our thermoelectric generators meet the need of 184 our sensors in harsher condition, for example an unstable steam supply. Andrew Goodney and Young Cho with our research institution are currently working on a generic solution for thermal energy harvesting under low differential temperature. Finally, our multi-modal approach can successfully detect cold-oil blockage (Chapter 4). It is interesting to study how to expand blockage detection to block- age localization. One proposal is to use pairwise sensors to segment a long pipeline and locating blockage by the time difference between alarms from different pairs. Inaddition,wedemonstrateoursystemworksinfieldtests,butmuchworkremains before a pervasive deployment. In short-term, a full integration with field network is a stepping stone to transform this research into a practical field equipment. 6.2 Long-Term Future Directions Manyopenresearchquestionsremain,althoughmulti-sensorcollaborationachieves good detection in each of our projects, including multi-sensor vehicle classification (Chapter 2), steam-choke blockage detection (Chapter 3), and cold-oil blockage detection (Chapter 4). First, we test our multi-sensor vehicle classification in an urban roadway, a constrained environment. It is interesting to study our approaches in a similar application in unconstrained environment. In addition, we expect our results and analysis to be useful to researchers studying object tracking. We believe it is necessary to study how to distribute this decision process to local sensors, especially after we scale up our system to more than two stations. Currently, all matching and classification decisions are made by one of the sensor 185 station (could be either the upstream or downstream one). A distributed struc- ture has many benefits, including better robustness, lower network bandwidth consumption and flexible process. To make our system more portable and less expensive, we would like to find and test mote-based hardware alternatives. Our sensor stations are all PC-based, mainly due to the constraint from the sensor driver. To our best knowledge, no other researcher has coupled inductive loops to a mote platform. Second, our steam-choke blockage detection is based on temperature sensing and we are investigating other modalities for better cost-effectiveness. We find that acoustic sensing is a plausible candidate. The intuition behind acoustics is that when high pressure steam passes through choke bean, it generates high-pitch sound. The pitch varies according to the cross-sectional size of choke bore, which is qualitatively and experimentally verified in our field tests. Therefore, we may use the unique signature of steam sound to detect if choke is blocked. Finally,ourshort-termfieldexperimentforcoldoilblockagewassuccessful,but additionalworkremains. Wefirstdiscussthegroupofissuesonsystemdeployment. A next step is mid-term or overnight testing to evaluate the robustness of the hardwaresystemandthetoleranceofthedetectionalgorithmagainstenvironment change. Inaddition,tofurtherlowerthecostofoursystem,wewouldliketocouple and test our system with energy harvesting units. One option is TEG, but how to reliably generate energy from a lower difference of pipe-to-ambient temperature is an open question. Our systems works in oil field, and we would like to generalize it to harsher environments, for example, subsea. Many researchers have studied subsea flowline blockage [LS01], but to our best knowledge, none has explore the option of wireless sensor networks. 186 The other group of open issues is on detection techniques. We are using tem- perature sensing to detect flow presence and acoustic for pump status; and we are actively seeking other sensing modalities for better cost-efficiency. For example, fromoursurveywefindsomepotentialalternativesareinfraredimagingandvibra- tion. We have shown full and near-full blockage detection is successful, possible future research is studying partial blockage (around 50%). One challenge is that a partial blockage does not change flow properties perceivably, mainly because oil extraction is incompressible and pumpjack is positive displacement, meaning the volumeliftedisconstant. Sopossibleresearchistostudyhowtoimplementsophis- ticated signal processing on sensor platform, and how to leverage the differential temperature due to pressure change before and after partial blockage. In addition, what we have done so far is blockage detection, but as soon as full or near-full blockage happens, it is likely to be too late to save the equipment from damage by the high pressure at wellhead. Therefore, early blockage warning or prediction may mitigate or even eliminate any possible equipment loss. 6.3 Conclusions Many sensor networks are deployed in real-world monitoring applications. Due to the sensornet’s advantage of low cost, fully automation and flexibility, we believe sensornet will be more popular in daily monitoring tasks among manufacturing, transportation, operation, and other fields. However, currently many applications only achieve mediocre accuracy, based on single-sensor observation. Researchers pay little attention to improve their results by adding and collaborating more sensors. 187 In this thesis, we proposed Multi-SensorCollaboration for cost-efficient sensing and enable sensornet usage in industrial applications. Collaboration achieves cost- efficiency by reducing capital, deployment and operational cost while maintaining actionable accuracy. To prove our thesis, we first framed the application domain by two orthogonal properties—collaboration scheme and modality. We next studied both collabora- tiontypeswitheithermodalitychoiceinthreedifferentproblems. First,westudied competitive collaboration by evaluating signature matching in a multi-sensor vehi- cle classification context. We designed a range of signature matching algorithm andintegratedthemwithavehicleclassificationalgorithm. Toevaluatehowsensor collaboration benefits classification, we carried out field test and find a fairly sim- ple static time window algorithm is the best, in terms of matching correctness and cost-efficiency. We further found a partial correlation between signature matching and end multi-sensor classification. Second, we studied the other type, complementary collaboration with single- modal sensing. We take an industrial application as an example—steam-choke blockage detection. The idea is to use the differential temperature by differ- ent sensors to improve the detection accuracy. We developed an algorithm with multi-sensor non-invasive sensing, and we implemented them in a fieldable system. Through over-night field tests, we demonstrated our algorithm with complemen- tary collaboration has perfect accuracy on detecting steam-choke blockages. We furthergeneralizedourmulti-sensortootherenvironmentthroughlaboratorytests in a hot water network. Finally, we extend our complementary collaboration study to multi-modality domain. We explore this domain by solving another industrial problem, cold-oil 188 blockage. It also shows by collaborative sensing, we can solve problems which are more complex than steam-choke blockage. In this application, we integrated low-cost sensors and carried out field experiments. We demonstrated our flow presence detection by temperature sensors can achieve 100% accuracy and overall pump status detection by acoustic sensors is 87% accurate. Combining these two detection,in-fieldfull-blockagedetectionaccuracyisashighas90%,provingmulti- modal collaboration is effective on real-world problem. We further generalized our multi-modal approach to near-full blockage in lab water network. In all, we studied multi-sensor collaboration in these three representative prob- lems and prove collaboration achieves cost-effectiveness in the forgoing three spe- cific applications. These applications are useful and allow deployment of sensing where not economically viable before. Further, the commonality between these applications and a large group of relevant applications strongly suggest that col- laborative sensing help a larger application domain. Beyond that, multi-modal competitive collaboration and other distantly related applications are open to be verified. We have identified remaining open issues (Section 6.1 and Section 6.2), andweexpectourresearchcontributestomakingindustrialsensornetsubiquitous. 189 Bibliography [AAYA09] Besma R. Abidi, Nash R. Aragam, Yi Yao, and Mongi A. Abidi. Sur- vey and analysis of multimodal sensor planning and integration for wideareasurveillance. ACMComputerSurvey,41:1–36,January2009. [ALE13] ALERT Users Group. Alert. http://www.alertsystems.org/, 2013. [AMM + 08] M.Annavaram,N.Medvidovic,U.Mitra,S.Narayanan,G.Sukhatme, Z. Meng, S. Qiu, R. Kumar, G. Thatte, and D. Spruijt-Metz. Mul- timodal sensing for pediatric obesity applications. In Proceedings of the International Workshop on Urban, Community, and Social Appli- cations of Networked Sensing Systems (UrbanSense), pages 21–25, Raleigh, NC, USA, November 2008. ACM. [Ana06] Analog.com. AD1988A/AD1988B data sheets. http://www.analog. com/static/imported-files/data_sheets/AD1988A_1988B.pdf, 2006. [AST93] ASTM Committee E20 on Temperature Measurement. Manual on the Use of Thermocouples in Temperature Measurement, fourth edition. ASTM, Philadelphia, 1993. [Aud] Audio-Technica. ATR3350 omnidirectional condenser lavalier microphone. http://www.audio-technica.com/cms/wired_mics/ 9c6eca17168eef6f/index.html. [B ¨ 05] H. B¨ ottner. Micropelt miniaturized thermoelectric devices: small size, high cooling power densities, short response time. In Proceedings of the 24th International Conference on Thermoelectrics, ICT ’05, pages 1–8, Clemson, SC, USA, June 2005. [Bau00] Adam Baumberg. Reliable feature matching across widely separated views. In Proceedings of 2000 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, volume 1 of CVPR ’00, pages 774–781, Hilton Head, SC, USA, June 2000. 190 [BB02] Edoardo S. Biagioni and K. W. Bridges. The application of remote sensortechnologytoassisttherecoveryofrareandendangeredspecies. International Journal of High Performance Computing Applications, 16(3):112–121, 2002. [BGL + 10] Horst Bischof, Martin Godec, Christian Leistner, Marcus Hen- necke, Arnold Maier, J¨ urgen Wolf, Bernhard Rinner, and Andreas Starzacher. Autonomous multi-sensor vehicle classification for traffic monitoring. InJuliaD¨ uh,HartwigHufnagl,ErhardJuritsch,Reinhard Pfliegl, Helmut-Klaus Schimany, and Hans Sch¨ onegger, editors, Data and Mobility, volume 81 of Advances in Intelligent and Soft Comput- ing, pages 15–26. Springer Berlin Heidelberg, 2010. [BKTT07] M.S. Benning, A. Kapur, B.C. Till, and G. Tzanetakis. Multimodal sensor analysis of sitar performance: Where is the beat? In Pro- ceedings of IEEE 9th Workshop on Multimedia Signal Processing, MMSP ’07, pages 74–77, Chania, Crete, Greece, October 2007. [BL05] EmiliaIBarakovaandTinoLourens. Eventbasedself-supervisedtem- poral integration for multimodal sensor data. Journal of Integrative Neuroscience, 4(2):265–282, 2005. [BN93] Mich` eleBassevilleandIgorV.Nikiforov. Detection of abrupt changes: theory and application. Prentice-Hall, Inc., 1993. [BNB + 04] H. B¨ ottner, J. Nurnus, M. Braun, J. W¨ ollenstein, F. Volkert, and A. Schubert. MicroPelt: state of the art, road map and applications. Technical Report ETS-2004, Infineon Technologies AG, M¨ unchen, Germany, 2004. [BRVK11] R. Bajwa, R. Rajagopal, P. Varaiya, and R. Kavaler. In-pavement wireless sensor network for vehicle classification. In Proceedings of the 10th International Conference on Information Processing in Sensor Networks, IPSN ’11, pages 85–96, 2011. [But97] Roger M. Butler. Thermal Recovery of Oil and Bitumen. Prentice- Hall, 1997. [CAC + 99] A. Chandrakasan, R. Amirtharajah, Seonghwan Cho, J. Goodman, G. Konduri, J. Kulik, W. Rabiner, and A. Wang. Design consid- erations for distributed microsensor systems. In Proceedings of the 1999 IEEE Custom Integrated Circuits Conference, pages 279–286, San Diego, CA, USA, May 1999. 191 [CCD + 05] SingYiuCheung,SinemColeri,BarisDundar,SumitraGanesh,Chin- Woo Tan, and Pravin Varaiya. Traffic measurement and vehicle clas- sification with a single magnetic sensor. In Proceedings of the 84th Annual Meeting of the Transportation Research Board, pages 173– 181, Washington D.C., USA, January 2005. Transportation Research Board. [CD46] Harold L. Cook and H. F. Dotterweich. Report on calibration of positive flow beans manufactured by Thornhill-Craver Company, Inc. Texas College of Arts and Industries, 1946. [CEV05] SingYiuCheung,SinemC.Ergen,andPravinVaraiya. Trafficsurveil- lance with wireless magnetic sensors. In Proceedings of the 12th World Congress on Intelligent Transport Systems, ITS ’05, pages 173–181, San Francisco, CA, USA, November 2005. [CGM + 11] Y. Chamber, A. Garg, V. Mithal, I. Brugere, M. Lau, V. Krishna, S. Boriah, M. Steinbach, V. Kumar, C. Potter, and S. Klooster. A novel time series based approach to detect gradual vegetation changes in forests. In Proceedings of the 2011 NASA Conference on Intelli- gent Data Understanding, pages 248–262, Mountainview, CA, USA, October 2011. NASA. [CHK + 05] W. Steven Conner, John Heidemann, Lakshman Krishnamurthy, Xi Wang, and Mark Yarvis. Wireless Sensor Networks: A Sys- tems Perspective,chapterWorkplaceApplicationsofSensorNetworks, pages 289–307. Artech House, Inc., 2005. Chapter 19, Nirupama Bulusu and Sanjay Jha, editors. [CJG05] K. Chintalapudi, E.A. Johnson, and R. Govindan. Structural damage detection using wireless sensor-actuator networks. In Proceedings of the 13th Mediterrean Conference on Control and Automation, pages 322–327, Limassol, Cyprus, June 2005. [CK06] M. Calle and J. Kabara. Measuring energy consumption in wireless sensor networks using GSP. In Proceedings of the IEEE 17th Interna- tional Symposium on Personal, Indoor and Mobile Radio Communi- cations, PIMRC’06, pages 1–5, Helsinki, Finland, September 2006. [Coi98a] Benjamin Coifman. A new algorithm for vehicle reidentification and traveltimemeasurementonfreeways. InProceedings of the Fifth Con- ference on Applications of Advanced Technologies in Transportation, pages167–174, NewportBeach, CA,USA,April1998.AmericanSoci- ety of Civil Engineers. 192 [Coi98b] Benjamin Coifman. Vehicle reidentification and travel time measure- ment in real-time on freeways using the existing loop detector infras- tructure. Transportation Research Record No. 1643 Pavement Man- agement and Monitoring of Traffic and Pavements, pages 181–191, 1998. [Cra88] CraneCo. Flow of Fluids through Valves, Fittings, and Pipe. Technical Paper No 410, Crane Company, 1988. [Cro] Crossbow Technology. Xbow mica2 datasheet. http://www. xbow.com/products/Product_pdf_files/Wireless_pdf/MICA2_ Datasheet.pdf. [Cro07] Crossbow Technology. MTS/MDA sensor board user’s man- ual. http://bullseye.xbow.com:81/Support/Support_pdf_files/ MTS-MDA_Series_Users_Manual.pdf, 2007. [Cus] Custom Thermoelectric. Custom Thermoelectric: Your Thermoelec- tric Partner. http://www.customthermoelectric.com/. [CWcW06] Lei Chen, Martin J. Wainwright, M¨ ujdat C ¸etin, and Alan S. Will- sky. Data association based on optimization in graphical models with applicationtosensornetworks. Mathematicaland Computer Modeling, 43(9-10):1114–1135, May 2006. [Dai93] D.J.Dailey. Travel-timeestimationusingcross-correlationtechniques. Transportation Research Part B: Methodological, 27(2):97–107, April 1993. [DAT12] DATAQ Instruments. EL-USB-TC thermocouple data logger datasheet. http://www.dataq.com/support/documentation/pdf/ datasheets/el-usb-tc-data-logger.pdf, 2012. [DGA + 05] Prabal Dutta, Mike Grimmer, Anish Arora, Steven Bibyk, and David Culler. Designofawirelesssensornetworkplatformfordetectingrare, random,andephemeralevents. InProceedings of the 4th International Symposium on Information Processing in Sensor Networks, IPSN ’05, Los Angeles, CA, USA, 2005. IEEE Press. [DS99] A. Daneels and W. Salter. What is SCADA? In Proceedings of the International Conference on Accelerator and Large Experimental Physics Control System,ICALEPCS’99,pages339–343,Trieste,Italy, October 1999. 193 [Fed03] Federal Highway Administration and Office of Highway Policy Infor- mation. FHWA vehicle types. http://www.fhwa.dot.gov/policy/ ohpi/vehclass.htm, October 2003. [FIIS11] Masato Futagawa, Mitsuyoshi Ishida, Makoto Ishida, and Kazuaki Sawada. Study of a wireless multimodal sensing system integrated withanelectricalconductivitysensorandatemperaturesensorforthe healthcontrolofcows. IEEE Transactions on Electrical and Electronic Engineering, 6(2):93–96, 2011. [GA96] Suzanne Griston and Tom Abate. Field test of tapered-bore for steam flow control. In SPE Western Regional Meeting, pages 269– 283,Anchorage,AK,USA,May1996.SocietyofPetroleumEngineers. (SPE 35677). [GE01] Lewis Girod and Deborah Estrin. Robust range estimation using acoustic and multimodal sensing. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Maui, HI, USA, October 2001. IEEE. [GGK + 07] K. Gupta, M.T. Ghasr, S. Kharkovsky, R. Zoughi, R.J. Stanley, A. Padwal, M. O’Keefe, D. Palmer, J. Blackshire, G. Steffes, and N. Wood. Fusion of microwave and eddy current data for a multi- modal approach in evaluating corrosion under paint and in lap joints. Review of Quantitative Nondestructive Evaluation, 26:611–618, 2007. [Goe06] Roland Goecke. Audio-video automatic speech recognition: an exam- ple of improved performance through multimodal sensor input. In Proceedings of the 2005 NICTA-HCSNet Multimodal User Interaction Workshop - Volume 57, MMUI ’05, pages 25–32, Sydney, Australia, 2006. Australian Computer Society, Inc. [HBC + 09] Wen Hu, Nirupama Bulusu, Chun Tung Chou, Sanjay Jha, Andrew Taylor, and Van Nghia Tran. Design and evaluation of a hybrid sen- sor network for cane toad monitoring. ACM Transactions on Sensor Networks, 5(1):4:1–4:28, February 2009. [HCCI08] MatthewHill,MurrayCampbell,Yuan-ChiChang,andVijayIyengar. Event detection in sensor networks for modern oil fields. In Proceed- ingsoftheSecondInternationalConferenceonDistributedEvent-based Systems, DEBS ’08, pages 95–102, Rome, Italy, 2008. ACM. 194 [HS07] KyleHarringtonandHavaSiegelmann. Adaptivemulti-modalsensors. In Max Lungarella, Fumiya Iida, Josh Bongard, and Rolf Pfeifer, edi- tors, 50 Years of Artificial Intelligence, volume 4850 of Lecture Notes in Computer Science, pages 164–173. Springer Berlin / Heidelberg, 2007. [HTTO02] H.Hayashi,Y.Takabayashi,H.Tsuji,andM.Oka. Rapidlyincreasing application of intranet technologies for SCADA (supervisory control anddataacquisitionsystem). In IEEE/PES Transmission and Distri- bution Conference and Exhibition 2002: Asia Pacific, volume 1, 2002. [HZHJ11] Po-Sen Huang, Xiaodan Zhuang, and M. Hasegawa-Johnson. Improv- ing acoustic event detection using generalizable visual features and multi-modality modeling. In Proceedings of the 2011 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing, ICASSP ’11, pages 349–352, May 2011. [Jac88] Van Jacobson. Congestion avoidance and control. In Symposium Proceedings on Communications Architectures and Protocols, SIG- COMM ’88, pages 314–329, Stanford, CA, USA, August 1988. ACM. [JE08] Yuanwei Jin and Ali Eydgahi. Monitoring of distributed pipeline sys- tems by wireless sensor networks. In Proceedings of the 2008 IAJC- IJME International Conference, Nashville, TN, USA,November 2008. [JG06] Velimir Jovanovic and Saeid Ghamaty. Design, fabrication, and testing of energy-harvesting thermoelectric generator. Smart Struc- tures and Materials 2006: Smart Structures and Integrated Systems, 6173(1):142–149, 2006. [JGHD08] Jaehoon Jeong, Shuo Guo, Tian He, and D. Du. APL: Autonomous passive localization for wireless sensors deployed in road networks. In Proceedings of the 27th Conference on Computer Communications, INFOCOM ’08, pages 583–591, Phoenix, AZ, USA, April 2008. [JPC05] Xiaofan Jiang, Joseph Polastre, and David Culler. Perpetual environ- mentally powered sensor networks. In Proceedings of the 4th Inter- national Symposium on Information Processing in Sensor Networks, IPSN ’05, pages 463–468, Los Angeles, CA, USA, April 2005. IEEE Press. 195 [KBHK06] Chris Kreucher, Doron Blatt, Alfred Hero, and Keith Kastella. Adap- tive multi-modality sensor scheduling for detection and tracking of smart targets. Digital Signal Processing, 16(5):546–567, 2006. Special Issue on DASP 2005. [KHZS07] Aman Kansal, Jason Hsu, Sadaf Zahedi, and Mani B. Srivastava. Powermanagementinenergyharvestingsensornetworks. ACMTrans- actions on Embedded Computing Systems, 6(4):1–38, September 2007. [KKRV09] Karric Kwong, Robert Kavaler, Ram Rajagopal, and Pravin Varaiya. A practical scheme for arterial travel time estimation based on vehi- cle reidentification using wireless sensors. In Proceedings of the 88th Annual Meeting of the Transportation Research Board, Washington D.C., USA, January 2009. Transportation Research Board. [KN01] Minkyong Kim and Brian Noble. Mobile network estimation. In Pro- ceedings of the 7th Annual International Conference on Mobile Com- puting and Networking, MobiCom ’01, pages 298–309, Rome, Italy, July 2001. ACM. [KOA + 08] M. Kushwaha, Songhwai Oh, I. Amundson, X. Koutsoukos, and A. Ledeczi. Target tracking in heterogeneous sensor networks using audio and video sensor fusion. In Proceedings of Multisensor Fusion and Integration for Intelligent Systems, MFI ’08, pages 14–19, August 2008. [KPC + 07] Sukun Kim, S. Pakzad, D. Culler, J. Demmel, G. Fenves, S. Glaser, andM.Turon. Healthmonitoringofcivilinfrastructuresusingwireless sensornetworks. InProceedingsof the 6th InternationalSymposium on Information Processing in Sensor Networks, IPSN ’07, pages 254–263, 2007. [KS03] Aman Kansal and Mani B. Srivastava. An environmental energy harvesting framework for sensor networks. In Proceedings of the 2003 International Symposium on Low Power Electronics and Design, ISLPED ’03, pages 481–486, Seoul, Korea, August 2003. ACM. [KSC + 08] Younghun Kim, Thomas Schmid, Zainul M. Charbiwala, Jonathan Friedman, and Mani B. Srivastava. NAWMS: nonintrusive autonomous water monitoring system. In Proceedings of the 6th ACM Conference on Embedded Network Sensor Systems, SenSys ’08, pages 309–322, Raleigh, NC, USA, November 2008. ACM. 196 [KSPSV02] F. Koushanfar, S. Slijepcevic, M. Potkonjak, and A. Sangiovanni- Vincentelli. Error-tolerant multi-modal sensor fusion. In Proceedings of IEEE CAS Workshop on Wireless Communications and Network- ing, pages 1–4, Pasadena, CA, USA, September 2002. [LLPW10] Enbin Liu, Changjun Li, Shanbi Peng, and Xia Wu. Detection tech- nology for oil pipeline plugging based on decompression wave method. InProceedings of the 2010 International Conference on Computational and Information Sciences, ICCIS ’10, pages 820–823, December 2010. [LLR + 03] Juan Liu, Jie Liu, James Reich, Patrick Cheung, and Feng Zhao. Distributed group management for track initiation and maintenance in target localization applications. In Proceedings of the 2nd Inter- national Conference on Information Processing in Sensor Networks, IPSN ’03, pages 113–128, Palo Alto, CA, USA, April 2003. Springer- Verlag. [LPR07] Loukas Lazos, Radha Poovendran, and James A. Ritcey. Probabilistic detection of mobile targets in heterogeneous sensor networks. In Pro- ceedingsofthe6thInternationalConferenceonInformationProcessing in Sensor Networks, IPSN ’07, pages 519–528, Cambridge, MA, USA, 2007. ACM. [LS01] Liangjian Liu and Stuart L. Scott. A new method to locate partial blockages in subsea flowlines. In SPE Annual Technical Conference and Exhibition, pages 1–9, New Orleans, LA, USA, September 2001. Society of Petroleum Engineers. SPE 71548. [LS02] E. E. Lawrence and G. J. Snyder. A study of heat sink performance in air and soil for use in a thermoelectric energy harvesting device. In Proceedings of the 21st International Conference on Thermoelectrics, ICT ’02, pages 446–449, Long Beach, CA, USA, August 2002. [Mat02] K.Matsubara. Developmentofahighefficientthermoelectricstackfor a waste exhaust heat recovery of vehicles. In Proceedings of the 21st International Conference on Thermoelectrics, ICT’02, pages418–423, Long Beach, CA, USA, August 2002. [Max12] Max Integrated. MAX4465–MAX4469 data sheets. http:// datasheets.maximintegrated.com/en/ds/MAX4465-MAX4469.pdf, 2012. 197 [MBL03] Francisco Molina, Julio Barbancho, and Joaquin Luque. Automated meterreadingandSCADAapplicationforwirelesssensornetwork. In Ad-Hoc, Mobile, and Wireless Networks, volume 2865 of LNCS, pages 223–234. Springer Berlin/Heidelberg, 2003. [MCL + 07] LoretoMateu,CosimCodrea,NestorLucas,MarkusPollak,andPeter Spies. Human body energy harvesting thermogenerator for sensing applications. In Proceedings of the 2007 International Conference on Sensor Technologies and Applications, SENSORCOMM ’07, pages 366–372, Valencia, Spain, October 2007. IEEE Computer Society. [MCP + 02] Alan Mainwaring, David Culler, Joseph Polastre, Robert Szewczyk, and John Anderson. Wireless sensor networks for habitat monitoring. In Proceedings of the 1st ACM International Workshop on Wireless Sensor Networks and Applications, WSNA ’02, pages 88–97, Atlanta, GA, USA, 2002. ACM. [MCS12] Paul Martin, Zainul Charbiwala, and Mani Srivastava. Doubledip: leveraging thermoelectric harvesting for low power monitoring of spo- radicwateruse.InProceedingsofthe10thACMConferenceonEmbed- ded Network Sensor Systems, SenSys ’12, pages 225–238, Toronto, Ontario, Canada, 2012. ACM. [MFS + 02] P. McGuire, J. Fritsch, J.J. Steil, F. Rothling, G.A. Fink, S. Wachsmuth, G. Sagerer, and H. Ritter. Multi-modal human- machine communication for instructing robot grasping tasks. In Pro- ceedings of the 2002 IEEE/RSJ International Conference on Intelli- gent Robots and Systems, volume 2, pages 1082–1088, 2002. [Mic10] Micropelt. TE-power node: Self-sufficient wireless sensor sys- tem thermoharvesting explorer. http://www.micropelt.com/down/ datasheet_te_power_node.pdf, March 2010. [MM05] Victor Muchuruza and Renatus Mussa. Traffic operation and safety analysisofminimumspeedlimitsonFloridaruralinterstatehighways. In Proceedings of the 2005 Mid-Continent Transportation Research Symposium, Ames, IA, USA, August 2005. [MVW08] SimonMoncrieff,SvethaVenkatesh,andGeoffWest. Dynamicprivacy assessment in a smart house environment using multimodal sensing. ACM Transactions on Multimedia Computing, Communications, and Applications, 5:10:1–10:29, November 2008. 198 [NNT11] N.H. Nguyen, N.M. Nasrabadi, and T.D. Tran. Robust multi-sensor classificationviajointsparserepresentation. InProceedings of the 14th International Conference on Information Fusion, FUSION ’11, pages 1–8, 2011. [OH05] Nuria Oliver and Eric Horvitz. S-seer: Selective perception in a multimodal office activity recognition system. In Samy Bengio and Herv´ e Bourlard, editors, Machine Learning for Multimodal Interac- tion, volume 3361 of Lecture Notes in Computer Science, pages 122– 135. Springer Berlin / Heidelberg, 2005. [OHRS05] Songhwai Oh, Inseok Hwang, Kaushik Roy, and Shankar Sastry. A fullyautomateddistributedmultiple-targettrackingandidentityman- agement algorithm. In Proceedings of AIAA Guidance, Navigation, and Control Conference, pages1–12, SanFrancisco, CA,USA,August 2005. [ORJ07] Cheol Oh, Stephen G. Ritchie, and Shin-Ting Jeng. Anonymous vehicle reidentification using heterogeneous detection systems. IEEE Transactions on Intelligent Transportation System, 8(3):460–469, September 2007. [Pag54] E.S.Page. ContinuousInspection Schemes. Biometrika, 41(1/2):100– 115, 1954. [Pan] Panasonic. WM-62A datasheet. http://www.panasonic.com/ industrial/components/pdf/em05_wm62_a_c_cc_k_b_dne.pdf. [PBO + 98] Karl F. Petty, Peter Bickel, Michael Ostland, John Rice, Frederic Schoenberg, Jiming Jiang, and Ya’acov Ritov. Accurate estimation of travel times from single-loop detectors. Transportation Research Part A: Policy and Practice, 32(1):1–17, January 1998. [PC06] Chulsung Park and P. H. Chou. AmbiMax: Autonomous energy har- vesting platform for multi-supply wireless sensor nodes. In Proceed- ings of the 3rd Annual IEEE Communications Society on Sensor and Ad Hoc Communications and Networks, SECON ’06, pages 168–177, Reston, VA, USA, September 2006. [PDPK03] P. V. Pahalawatta, D. Depalov, T. N. Pappas, and A. K. Katsaggelos. Detection, classification, andcollaborativetrackingofmultipletargets usingvideosensors.InProceedingsofthe2ndInternationalConference on Information Processing in Sensor Networks, IPSN ’03, pages 529– 544, Palo Alto, CA, USA, April 2003. Springer-Verlag. 199 [PG84] R.H. Perry and D.W. Green. Perry’s Chemical Engineers’ Handbook, Sixth Edition. McGraw-Hill, 1984. [PHC04] Joseph Polastre, Jason Hill, and David Culler. Versatile low power media access for wireless sensor networks. In Proceedings of the 2nd International Conference on Embedded Networked Sensor Systems, SenSys ’04, pages 95–107, Baltimore, MD, USA, 2004. ACM. [PLH + 11] TaiwooPark, JinwonLee,InseokHwang, ChungkukYoo, LamaNach- man, and Junehwa Song. E-Gesture: a collaborative architecture for energy-efficient gesture recognition with hand-worn sensor and mobile devices. In Proceedings of the 9th ACM Conference on Embedded Net- worked Sensor Systems,SenSys’11,pages260–273,Seattle,WA,USA, 2011. ACM. [PS05] J. A. Paradiso and T. Starner. Energy scavenging for mobile and wireless electronics. Pervasive Computing, IEEE, 4(1):18–27, 2005. [PSH + 06] Unkyu Park, Fabio Silva, Mengzhao Hu, John Heidemann, Genevive Guiliano, XiWang, andNikhilPrashar. Single-andmulti-sensortech- niques to improve accuracy of urban vehicle classification. Technical Report ISI-TR-2006-614, USC/Information Sciences Institute, April 2006. [QWZ10] Yufu Qu, Tao Wang, and Zhigang Zhu. An active multimodal sensing platform for remote voice detection. In Proceedings of the 2010 IEEE/ASME International Conference on Advanced Intelligent Mechatronics, AIM ’10, pages 627–632, July 2010. [RBI + 05] MohammadRahimi,RickBaer,ObimdinachiI.Iroezi,JuanC.Garcia, Jay Warrior, Deborah Estrin, and Mani Srivastava. Cyclops: in situ image sensing and interpretation in wireless sensor networks. In Pro- ceedings of the 3rd International Conference on Embedded Networked Sensor Systems, SenSys ’05, pages 192–204, San Diego, CA, USA, 2005. ACM. [Rec10] Gerald Recktenwald. Conversion of thermocouple voltage to tempera- ture. http://web.cecs.pdx.edu/ ˜ gerry/epub/pdf/thermocouple. pdf, July 2010. [Rei79] Donald B. Reid. An algorithm for tracking multiple targets. IEEE Transactions on Automatic Control, 24(6):843–854, December 1979. 200 [Rij79] C.J.VanRijsbergen. InformationRetrieval. Butterworth-Heinemann, Newton, MA, USA, 2nd edition, 1979. [RKH + 05] Vijay Raghunathan, Aman Kansal, Jason Hsu, Jonathan Friedman, andManiSrivastava. Designconsiderationsforsolarenergyharvesting wireless embedded systems. In Proceedings of the 4th International Symposium on Information Processing in Sensor Networks, IPSN ’05, pages 457–462, Los Angeles, CA, USA, 2005. IEEE Press. [RMW + 97] Michael D. Rowe, Gao Min, Simon G. K. Williams, Azzeddine Aoune, Kenji Matsuura, Vladimir L. Kuznetsov, and Wenfu Li. Thermoelec- tric recovery of waste heat-case studies. In Proceedings of the 32nd Intersociety Energy Conversion Engineering Conference, volume 2 of IECEC ’97, pages 1075–1079, Honolulu, HI, USA, July 1997. [ROC + 03] Shad Roundy, Brian P. Otis, Yuenhui Chee, Jan M. Rabaey, and Paul Wright. A 1.9GHz RF transmit beacon using environmentally scav- enged energy. In Digest of IEEE International Symposium on Low Power Electronics and Devices, pages 1–6, Seoul, Korea, 2003. [RSS + 09] J.A. L´ opez Riquelme, F. Soto, J. Suard´ ıaz, P. S´ anchez, A. Iborra, and J.A. Vera. Wireless sensor networks for precision horticulture in southern spain. Computers and Electronics in Agriculture, 68(1):25– 35, 2009. [SARR04] Carlos C. Sun, Glenn S. Arr, Ravi P. Ramachandran, and Stephen G. Ritchie. Vehicle reidentification using multidetector fusion. IEEE Transactions on Intelligent Transportation Systems, 5(3):155–164, September 2004. [SB97] Amit Singhal and Chris Brown. Dynamic bayes net approach to mul- timodal sensor fusion. In Proceedings of the SPIE - The International Society for Optical Engineering, volume 3209, pages 2–10, 1997. [SC90] HiroakiSakoeandSeibiChiba. Dynamicprogrammingalgorithmopti- mization for spoken word recognitioxon. In Alex Waibel and Kai-Fu Lee, editors, Readings in Speech Recognition, pages 159–165. Morgan Kaufmann Publishers Inc., 1990. [SGZ03] Jaewon Shin, Leonidas J. Guibas, and Feng Zhao. A distributed algo- rithm for managing multi-target identities in wireless ad-hoc sensor networks. InProceedingsofthe2ndInternationalConferenceonInfor- mation Processing in Sensor Networks, IPSN ’03, pages 625–641, Palo Alto, CA, USA, April 2003. Springer-Verlag. 201 [Sin05] Dipen Sinha. Acoustic sensor for pipeline monitoring. Technical Report LA-UR-05-6025, Los Alamos National Laboratory, July 2005. [SLV11] J´ anosSallai, ´ AkosL´ edeczi,andP´ eterV¨ olgyesi. Acousticshooterlocal- izationwithaminimalnumberofsingle-channelwirelesssensornodes. In Proceedings of the 9th ACM Conference on Embedded Networked Sensor Systems, SenSys ’11, pages 96–107, Seattle, WA, USA, 2011. ACM. [SMK + 07] Jaspreet Singh, Upamanyu Madhow, Rajesh Kumar, Subhash Suri, and Richard Cagley. Tracking multiple targets using binary proxim- ity sensors. In Proceedings of the 6th International Conference on Information Processing in Sensor Networks, IPSN ’07, pages 529–538, Cambridge, MA, USA, April 2007. ACM. [SNMT07] IvanStoianov,LamaNachman,SamMadden,andTimurTokmouline. PIPENET: a wireless sensor network for pipeline monitoring. In Pro- ceedingsofthe6thInternationalConferenceonInformationProcessing in Sensor Networks, IPSN ’07, pages 264–273, Cambridge, MA, USA, 2007. ACM. [SOS08] N.F. Sandell and R. Olfati-Saber. Distributed data association for multi-target tracking in sensor networks. In Proceedings of the 47th IEEEConferenceonDecisionandControl,CDC’08,pages1085–1090, 2008. [SRT + 08] T. Stiefmeier, D. Roggen, G. Tr¨ oster, G. Ogris, and P. Lukowicz. Wearable activity tracking in car manufacturing. IEEE Pervasive Computing, 7(2):42–50, April 2008. [SS09] Alanson Sample and Joshua R. Smith. Experimental results with two wireless power transfer systems. In Proceedings of Radio and Wireless Symposium, RWS ’09, pages 16–18, San Diego, CA, USA, January 2009. [SSDI07] Henry A. Sodano, Garnett E. Simmers, Remi Dereux, and Daniel J. Inman. Recharging batteries using energy harvested from thermal gradients. Journal of Intelligent Material Systems and Structures, 18(1):3–10, 2007. 202 [TGC + 07] Vladimir Trifa, Lewis Girod, Travis Collier, Daniel T Blumstein, and Charles E. Taylor. Automated wildlife monitoring using self- configuring sensor networks deployed in natural habitats. In Pro- ceedings of the 12th International Symposium on Artificial Life and Robotics, AROB ’07, pages 1–4, Beppu, Japan, January 2007. [TIF05] Satoshi Tamura, Koji Iwano, and Sadaoki Furui. A robust mul- timodal speech recognition method using optical flow analysis. In W. Minker, Dirk B¨ uhler, Laila Dybkjær, and Nancy Ide, editors, Spo- ken Multimodal Human-Computer Dialogue in Mobile Environments, volume 28 of Text, Speech and Language Technology, pages 37–53. Springer Netherlands, 2005. [TV04] Alexander G Tartakovsky and Venugopal V Veeravalli. Change-point detection in multichannel and distributed systems with applications. In Applied Sequential Methodologies: Real-World Examples with Data Analysis, volume 173, pages 337–371. CRC Press, 2004. [Ver] Vernier. Vernier Go!Temp thermometer. http://www.vernier.com/ go/gotemp.html. [VRS03] Thiemo Voigt, Hartmut Ritter, and Jochen Schiller. Utilizing solar power in wireless sensor networks. In Proceedings of the 28th Annual IEEE International Conference on Local Computer Networks, LCN’03,pages416–422,Bonn/K¨ onigswinter,Germany,October2003. IEEE Computer Society. [Wes56] Western Electric. Statistical Quality Control Handbook. Western Elec- tric Corporation, Indianapolis, Inc., 1956. [WQI02] Xiaoling Wang, Hairong Qi, and S. Sitharama Iyengar. Collabora- tivemulti-modalitytargetclassificationindistributedsensornetworks. In Proceedings of the Fifth International Conference on Information Fusion, volume 1, pages 285–290, 2002. [WY95] C. B. Williams and R. B. Yates. Analysis of a micro-electric generator for microsystems. In Proceedings of the 8th International Conference on Solid-State Sensors and Actuators and Eurosensors IX., volume 1 of Transducers ’95, pages 369–372, Stockholm, Sweden, June 1995. [YO05] Y. Yasuno and J. Ohga. Temperature characteristics of electret con- denser microphones. In Proceedings of the 12th International Sympo- sium on Electrets, ISE ’05, pages 412–415, September 2005. 203 [YSM + 06] T.A. York, D.R. Stephenson, S. Murphy, J.L. Davidson, and B.D. Grieve. Ad hoc wireless sensor tomography. In Proceedings of the 5th IEEE Conference on Sensors, pages 493–496, October 2006. [YYH + 11] SunHee Yoon, Wei Ye, John Heidemann, Brian Littlefield, and Cyrus Shahabi. SWATS: Wireless sensor networks for steamflood and water- flood pipeline monitoring. IEEE Network Magazine, pages 50–56, 2011. [ZB05] Xiaotao Zou and B. Bhanu. Tracking humans using multi-modal fusion. InProceedings of IEEE Computer Society Conference on Com- puter Vision and Pattern Recognition, volume 3 of CVPR ’05, pages 4–11. IEEE CS, June 2005. [ZCP07] Wensheng Zhang, Guohong Cao, and Tom La Porta. Data dissemina- tion with ring-based index for wireless sensor networks. IEEE Trans- action on Mobile Computing, 6(7):832–847, July 2007. [ZH11] Chengjie Zhang and John Heidemann. Evaluating signature matching in a multi-sensor vehicle classification system (extended). Technical ReportISI-TR-2011-675,USC/InformationSciencesInstitute,Novem- ber 2011. [ZH13a] Chengjie Zhang and John Heidemann. Reducing false alarms with multi-modal sensing for pipeline blockage. in submission, 2013. [ZH13b] Chengjie Zhang and John Heidemann. Reducing false alarms with multi-modal sensing for pipeline blockage (extended). Technical Report ISI-TR-2013-686, USC/Information Sciences Institute, June 2013. [Zhu05] Jin-ming Zhu. Application of temperature method for jam detection in BF coal injection. Baosteel Technology, 6:15–18, 2005. [ZSCH11a] Chengjie Zhang, Affan Syed, Young H. Cho, and John Heidemann. Steam-powered sensing. In Proceedings of the 9th ACM Conference on Embedded Networked Sensor Systems, SenSys ’11, pages 204–217, Seattle, WA, USA, November 2011. ACM. [ZSCH11b] Chengjie Zhang, Affan Syed, Young H. Cho, and John Heidemann. Steam-powered sensing: Extended design and evaluation. Technical Report ISI-TR-2011-670, USC/Information Sciences Institute, Febru- ary 2011. 204 [ZSR02] Feng Zhao, Jaewon Shin, and James Reich. Information-driven dynamic sensor collaboration for tracking applications. IEEE Signal Processing Magazine, 19(2):61–72, March 2002. [ZTLZ06] HaoZeng,TeckBengThoe,XiangLi,andJunhongZhou.Multi-modal sensing for machine health monitoring in high speed machining. In Proceedings of the 2006 IEEE International Conference on Industrial Informatics, pages 1217–1222, August 2006. 205
Abstract (if available)
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Sensing with sound: acoustic tomography and underwater sensor networks
PDF
Congestion control in multi-hop wireless networks
PDF
Robust routing and energy management in wireless sensor networks
PDF
Multichannel data collection for throughput maximization in wireless sensor networks
PDF
Reconfiguration in sensor networks
PDF
Aging analysis in large-scale wireless sensor networks
PDF
Efficient pipelines for vision-based context sensing
PDF
Transport layer rate control protocols for wireless sensor networks: from theory to practice
PDF
Rate adaptation in networks of wireless sensors
PDF
Improving network security through collaborative sharing
PDF
Distributed wavelet compression algorithms for wireless sensor networks
PDF
Understanding and exploiting the acoustic propagation delay in underwater sensor networks
PDF
Models and algorithms for energy efficient wireless sensor networks
PDF
Techniques for efficient information transfer in sensor networks
PDF
Crowd-sourced collaborative sensing in highly mobile environments
PDF
Application-driven compressed sensing
PDF
A protocol framework for attacker traceback in wireless multi-hop networks
PDF
Efficient and accurate in-network processing for monitoring applications in wireless sensor networks
PDF
Gradient-based active query routing in wireless sensor networks
PDF
IEEE 802.11 is good enough to build wireless multi-hop networks
Asset Metadata
Creator
Zhang, Chengjie
(author)
Core Title
Design of cost-efficient multi-sensor collaboration in wireless sensor networks
School
Viterbi School of Engineering
Degree
Doctor of Philosophy
Degree Program
Computer Science
Publication Date
09/24/2013
Defense Date
08/22/2013
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
blockage detection,multi-modal sensing,OAI-PMH Harvest,sensornets,vehicle classification,wireless sensor networks
Format
application/pdf
(imt)
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Heidemann, John (
committee chair
), Govindan, Ramesh (
committee member
), Krishnamachari, Bhaskar (
committee member
)
Creator Email
chezhang@qti.qualcomm.com,zcjsword@gmail.com
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c3-330649
Unique identifier
UC11296495
Identifier
etd-ZhangCheng-2050.pdf (filename),usctheses-c3-330649 (legacy record id)
Legacy Identifier
etd-ZhangCheng-2050.pdf
Dmrecord
330649
Document Type
Dissertation
Format
application/pdf (imt)
Rights
Zhang, Chengjie
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Access Conditions
The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...
Repository Name
University of Southern California Digital Library
Repository Location
USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA
Tags
blockage detection
multi-modal sensing
sensornets
vehicle classification
wireless sensor networks