University of Southern California Dissertations and Theses

Zero-power sensing and processing with piezoelectric resonators

(USC Thesis Other)

Zero-power sensing and processing with piezoelectric resonators

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content ZERO-POWER SENSING AND PROCESSING WITH
PIEZOELECTRIC RESONATORS
by
Anton A. Shkel
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulllment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ELECTRICAL ENGINEERING)
December 2018
Copyright 2018 Anton A. Shkel
Acknowledgments
I would like to express my sincere appreciation for and deepest gratitude to my advisor,
Professor Eun Sok Kim. His patience, guidance, encouragement, advice, and insight has
been essential in my pursuit of a doctorate degree and my development as a researcher.
Professor Kim has provided me with many life-changing opportunities to present at inter-
national conferences, serve on administrative committees, and present at meetings with
funding agencies. I would not have been able to complete this work without his support.
I would like to thank Professor Shri Narayanan and Professor Kirk Shung who have
been valuable members of my dissertation committee and who have provided insightful
discussions and advice toward improving this thesis.
I am extremely grateful to the former members of the USC MEMS group, including
Dr. Lukas Baumgartel, Dr. Lingtao Wang, Dr. Qian Zhang, Dr. Arash Vafanejad, and
Dr. Yufeng Wang, who have provided the foundation for my research, lab training, and
mentorship. I would like to thank my current group members, Lurui Zhao, Yongkui Tang,
Hai Liu, and Matin Barekatain, for their valuable advice, discussion, and friendship. I
am also very thankful to all of my friends, Dr. Gene B. Kim, Dr. Nirakar Poudel, and
all others at USC and elsewhere for their valuable support.
ii
My sincerest thanks goes to Dr. Donghai Zhu and Alfonso Jimenez for their contin-
ued support at the state-of-the-art clean-room facilities at USC. Their eorts have been
essential in completing this work.
Finally, I am eternally grateful to my family, whose advice, encouragement, and sup-
port in providing me with the best opportunities made everything I have done possible.
I want to express the utmost love and gratitude to my wife, Jane Shkel, who has been an
incredible source of happiness, support, stability, encouragement, and inspiration to me
in all areas of my life.
iii
Table of Contents
Acknowledgments ii
List of Tables vii
List of Figures viii
Abstract xvi
Chapter 1: Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Scope of Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Overview of Chapters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Chapter 2: Piezoelectric Resonant Microphone 9
2.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 Piezoelectric Thin Films . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.2 MEMS Microphones . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2 Modeling and Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Fabrication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4.2 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.5 Low Frequency MEMS Microphone . . . . . . . . . . . . . . . . . . . . . . 29
2.5.1 Series-Spring Microphone . . . . . . . . . . . . . . . . . . . . . . . 30
2.5.2 Silicon Proof-Mass Microphone . . . . . . . . . . . . . . . . . . . . 31
2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Chapter 3: Acoustic Signal Classification with Pre-filtered Microphone 36
3.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Cepstral Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2.1 Conventional Cepstral Feature Extraction . . . . . . . . . . . . . . 39
iv
3.2.2 Modied Array-Based Feature Extraction . . . . . . . . . . . . . . 43
3.3 Classication Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Chapter 4: Acoustic Signal Classification Experiments 51
4.1 Resonant Microphone Digital Communication Link Experiments . . . . . 51
4.1.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.1.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2 Speech Recognition Experiments . . . . . . . . . . . . . . . . . . . . . . . 55
4.2.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.3 Lung Sound Recognition Experiments . . . . . . . . . . . . . . . . . . . . 61
4.3.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.4 Neural Network Recognition Experiments . . . . . . . . . . . . . . . . . . 75
4.4.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Chapter 5: System Integration and Low Power Implementation 91
5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.2 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.3 Hardware Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.3.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.4 Results - Power Consumption . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Chapter 6: Piezoelectric Vibration-Energy Harvesting 107
6.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.2 PZT Bimorph Energy Harvester . . . . . . . . . . . . . . . . . . . . . . . 108
6.2.1 Modeling and Design . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.3 Macro-Fiber Composite Flexible Energy Harvester . . . . . . . . . . . . . 112
6.3.1 Macro-Fiber Composite . . . . . . . . . . . . . . . . . . . . . . . . 113
6.3.2 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.4 Bistable Energy Harvester . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.4.1 Bistable Energy Harvester Utilizing Magnetic Mass . . . . . . . . . 116
6.4.2 Bistable Energy Harvester Utilizing Flexible Substrate . . . . . . . 119
6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
v
Chapter 7: Zero-Power Amplification with Helmholtz Resonators 123
7.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.2 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
7.3 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.4 Multiphysics Modeling and Simulation . . . . . . . . . . . . . . . . . . . . 133
7.5 Fabrication and Implementation . . . . . . . . . . . . . . . . . . . . . . . 134
7.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
7.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Chapter 8: Zero-Power Wireless Authentication with Film-Bulk Acoustic
Resonators 143
8.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
8.2 Device . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
8.2.1 Film Bulk Acoustic Resonator (FBAR) . . . . . . . . . . . . . . . 145
8.2.2 RFID Antenna . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
8.2.3 Integrating with Energy Harvester . . . . . . . . . . . . . . . . . . 155
8.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Chapter 9: Conclusion and Future Direction 159
9.1 Microphone Array for Signal Processing . . . . . . . . . . . . . . . . . . . 159
9.2 Piezoelectric Energy Harvesting . . . . . . . . . . . . . . . . . . . . . . . . 161
9.3 Helmholtz Resonator for Passive Amplication . . . . . . . . . . . . . . . 162
9.4 FBAR-Based Tampering Detection . . . . . . . . . . . . . . . . . . . . . . 163
Bibliography 165
Appendix A
Digital Signal Processing Methods . . . . . . . . . . . . . . . . . . . . . . . . . 174
vi
List of Tables
2.1 Summarized piezoelectric parameters for ZnO, AlN, and PZT thin-lms [9]. 14
2.2 Summarized coecients of mechanical properties of materials for ZnO [12],
AlN, PZT, Si, and SiO
2
[10] . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Summary of the selected geometric design parameters and the predicted
resonant frequency of a selected subset of array elements. . . . . . . . . . 22
2.4 Summary of the measured parameters of a selected subset of array elements. 28
2.5 Summary of the parameters of a resonant microphone element in compar-
ison to commercial MEMS microphones. . . . . . . . . . . . . . . . . . . . 29
4.1 Summary of center frequencies of selected lters for classication experi-
ments, and associated average Kolmogorov-Smirnov distance of the feature
set and GMM classication accuracy with a digital implementation. . . . 67
5.1 Summary of the system resource requirements for testing the classication
accuracy with a PC, and implementing on low-specication system, with
estimated memory usage of each. . . . . . . . . . . . . . . . . . . . . . . . 100
5.2 Estimated power consumption of Cypress PSoC 4200 BLE in active oper-
ation, as specied by the datasheet. . . . . . . . . . . . . . . . . . . . . . 100
6.1 Table summarizing PZT bimorph cantilever design parameters. . . . . . . 110
6.2 Summary of the selected geometric design parameters. . . . . . . . . . . . 115
7.1 Summary of the selected geometric design parameters (in mm) for Helmholtz-
resonator-enhanced MEMS microphone. . . . . . . . . . . . . . . . . . . . 133
vii
List of Figures
1.1 Diagram showing the proposed fully-integrated system. . . . . . . . . . . 5
2.1 Diagram showing (a) sensing if using the longitudinal piezoelectric coef-
cient
33
and (b) sensing if using the transverse piezoelectric coecient

31
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Scanning electron microscope (SEM) images of a single transducer showing
(a) design elements, and (b) denitions of geometric dimensions. . . . . . 20
2.3 Bode plot of the predicted microphone array sensitivity. . . . . . . . . . . 22
2.4 Diagram showing fabrication processes of resonant MEMS microphone. . . 23
2.5 Photograph of fabricated microphone array with center frequencies indicated. 25
2.6 (top) Photograph of experimental measurement and characterization setup
in anechoic chamber and (bottom) diagram showing schematic of measure-
ment setup with LabVIEW control. . . . . . . . . . . . . . . . . . . . . . . 26
2.7 Electrically measured sensitivity of the 13 array channels . . . . . . . . . 28
2.8 An example of proposed series-spring microphone design, with simulated
reduction in resonance frequency from 2.3 kHz to 1.8 kHz in the same
footprint. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.9 Image of series-spring microphone design FEM simulation of (a) displace-
ment of the rst fundamental mode, and (b) strain distribution of the rst
fundamental mode. (c) The mask layout for patterning of etched cavities
and electrodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
viii
2.10 Diagrams showing the proposed microphone with silicon proof-mass de-
sign, and the necessary KOH mask modication for (a) proof-mass etching
utilizing corner-compensation, and (b) proof-mass formation with through-
wafer DRIE etching. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1 Block diagram showing the necessary steps for the calculation of Cep-
stral features with (a) a standard approach using a signal
at-band mi-
crophone with digital ltering, and (b) an n-channel resonant microphone
array with a pre-ltered input. By using a pre-ltered microphone array,
both discrete Fourier transform (DFT) and digital ltering can be avoided
to obtain an equivalent energy vector prior to taking the discrete cosine
transform (DCT). Analog-to-Digital Conversion (ADC) with n-channels
can be achieved with parallel ADCs or time-multiplexing. . . . . . . . . . 40
3.2 Plots showing the (a) average spectral signature of frames containing wheez-
ing and containing normal breathing. Shaded regions represent the rst
and the third quartiles of the datasets, (b) the measured sensitivities of
an optimized MEMS resonant microphones array with 7 elements, and (c)
the 7 overlapping digital triangular lters with center frequencies matching
the resonant microphone array. . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3 Plot showing the in
uence of sample rate on computational complexity for
cepstral feature extraction with resonant array ltering, digital ltering
with the use of an FFT, and digital ltering with the use of a DFT. . . . 44
3.4 Plots showing the memory requirement relationship of digital ltering and
resonant array ltering in cepstral feature extraction with (a) variation in
number of lters, and (b) variation in sample rate. . . . . . . . . . . . . . 46
4.1 (a) Waveform graph of pulse-width-modulated input signal. (b) Signal with
-15dB SNR additive white Gaussian noise. (c) Measured MEMS resonant
microphone electrical output. (d) Measured reference microphone output
with digital band-pass post-ltering. . . . . . . . . . . . . . . . . . . . . . 53
4.2 (a) Measured bit error rate for 860 Hz signal with 2 kHz sinusoidal noise.
(b) Measured bit-error rate for 6.3 kHz signal with 300 Hz sinusoidal noise. 55
4.3 Experimental measurements of minimum SNR such that BER < 25% for
a microphone element with 3 kHz resonant frequency. The x-axis plots
(a) the ratio of the electrical response of the microphone at the noise fre-
quency to that at the resonant frequency and (b) the absolute value of the
dierence between noise frequency and signal frequency . . . . . . . . . . 56
ix
4.4 Plot of measured word-error rates (WERs) of the acoustic resonator array
and the reference microphone used to process automatic speech recognition
with various levels of 400 Hz sinusoidal noise. . . . . . . . . . . . . . . . . 59
4.5 Spectrogram of the utterance "fty-two nineteen" in the case of (a) a clean
audio le without any ltering, (b) the reference microphone for -10 dB
applied noise, and (c) a summation of the outputs from 13 microphone
array elements in -10 dB noise. . . . . . . . . . . . . . . . . . . . . . . . . 60
4.6 Block diagram showing the experimental testing process for evaluation of
lung sound classication, comparing resonant microphone array processing
and
at-band processing. . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.7 Plots of average spectral distribution of frames containing wheezing (shown
in red) and normal breathing (shown in blue). The averages are plotted for
(a) all recordings, (b) only recordings of infant breathing, (c) only record-
ings of children's breathing, and (c) only recordings of adult breathing.
Dotted lines show spectral distribution for an arbitrarily selected single
frame from each category, and shaded regions are bounded by the 25th
and 75th percentiles of the data distribution. . . . . . . . . . . . . . . . . 64
4.8 Plots showing the selection of lter-banks used in these studies, including
(a) the measured sensitivities of 7 MEMS resonant microphones, which
are scaled articially to occupy the optimized lter positions; (b) the over-
lapping digital triangular lters designed to match the optimized center
frequencies of the resonant microphone array; and (c) the 13 mel-Spaced
triangular lters which give maximum classication accuracy for this ap-
plication using a digital processing approach. . . . . . . . . . . . . . . . . 66
4.9 Spectrograms of recorded breathing, obtained with the summed 7 channels
of the resonant microphone array (left column) and with the single-channel

at-band microphone (right column). From top to bottom the noise sources
are 60 Hz monotonic noise, 1200 Hz monotonic noise, and heartbeat noise,
all with noise intensity of -26 dB SNR. Larger intensity of the recording
appears with a brighter color. . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.10 Measured classication accuracy (F1 score) of wheezing detection for vary-
ing levels of background noise intensity with a 60 Hz noise source. Cepstral
features used in classication are computed with 7 resonant array lters
(red), 7 digital array lter (black, dotted), and 13 mel-spaced digital lters
(black, solid). Features are classied using a Gaussian Mixture Models,
k-Nearest Neighbors, Support Vector Machines, and Naive Bayesian clas-
sication. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
x
4.11 Measured classication accuracy (F1 score) of wheezing detection for vary-
ing levels of background noise intensity with a 1200 Hz noise source. Cep-
stral features used in classication are computed with 7 resonant array
lters (red), 7 digital array lter (black, dotted), and 13 mel-spaced digital
lters (black, solid). Features are classied using a Gaussian Mixture Mod-
els, k-Nearest Neighbors, Support Vector Machines, and Naive Bayesian
classication. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.12 Measured classication accuracy (F1 score) of wheezing detection for vary-
ing levels of background noise intensity with a heart noise source. Cepstral
features used in classication are computed with 7 resonant array lters
(red), 7 digital array lter (black, dotted), and 13 mel-spaced digital lters
(black, solid). Features are classied using a Gaussian Mixture Models,
k-Nearest Neighbors, Support Vector Machines, and Naive Bayesian clas-
sication. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.13 Diagram showing experimental conguration for construction and testing
of a fully-connected time-delay neural network, with a variable number of
features and frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.14 Plot showing classication accuracy (F1 score) of neural network as a func-
tion of the number of consecutive 50% overlapping 40 ms frames used in
the input. Classication of wheezing is performed on a test data set of
respiratory sounds containing 35 dB SNR heart noise. . . . . . . . . . . . 80
4.15 Plot showing classication accuracy (F1 score) of neural network as a func-
tion of the number of consecutive 50% overlapping 40 ms frames used in
the input. Classication of wheezing is performed on a test data set of
respiratory sounds containing -26 dB SNR heart noise. . . . . . . . . . . . 81
4.16 Plot showing classication accuracy (F1 score) of neural network as a func-
tion of the number of training epochs with a xed 25 input frames. Clas-
sication of wheezing is performed on a test data set of respiratory sounds
containing 35 dB SNR heart noise. . . . . . . . . . . . . . . . . . . . . . 83
4.17 Plot showing classication accuracy (F1 score) of neural network as a func-
tion of the number of training epochs with a xed 25 input frames. Clas-
sication of wheezing is performed on a test data set of respiratory sounds
containing -10 dB SNR heart noise. . . . . . . . . . . . . . . . . . . . . . . 84
4.18 Plot showing classication accuracy (F1 score) of neural network as a func-
tion of heart noise intensity, with a model congured for 25 input frames
(320 ms) and training for 50 epochs. . . . . . . . . . . . . . . . . . . . . . 85
xi
4.19 Plot showing classication accuracy (F1 score) of neural network as a func-
tion of heart noise intensity, with a model congured for 50 input frames
(1.02 s) and training for 50 epochs. . . . . . . . . . . . . . . . . . . . . . . 86
4.20 Plot showing classication accuracy (F1 score) of neural network as a func-
tion of heart noise intensity, with a model congured for 100 input frames
(2.02 s) and training for 50 epochs. . . . . . . . . . . . . . . . . . . . . . . 87
5.1 Diagram showing the fully-integrated proposed system for zero-power acous-
tic signature monitoring, with illustrated components. . . . . . . . . . . . 92
5.2 Conceptual illustration of system architecture for low-power continuous
breathing-monitoring system. Key components are the resonant micro-
phone array (consisting of 7 elements), pre-amplier and signal condition-
ing electronics, analog-to-digital converter (ADC), microcontroller, and
Bluetooth low energy (BLE) radio for wireless data transmission. The
microcontroller implements continuous wheeze detection. . . . . . . . . . . 93
5.3 Spectrogram of recorded lung sounds from an infant with wheezing signa-
ture present during the inspiration phase of respiration. . . . . . . . . . . 94
5.4 Block diagram showing the experimental testing process for evaluation of
power consumption of lung sound classication on Cypress PSoC 4200
BLE chip with both resonant microphone array processing and
at-band
processing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.5 Measurements of power consumption in Cypress PSoC 4200 for the case
of (a) processing with a standard digital signal classication method, and
(b) processing with the pre-ltered microphone array method. The clas-
sication period is held constant, with one frame of digitization, feature
extraction, classication, and wireless transmission completing every 5.11
seconds. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.1 Schematic of the PZT bimorph structure, composed of two PZT layers
separated by a brass layer. . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.2 Schematic of PZT bimorph cantilever with attached proof mass. . . . . . 110
6.3 (a) Photograph of fabricated PZT bimorph energy harvester, and (b)
schematic of measurement setup using a variable impedance board, with
buered oscilloscope input. . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.4 Plots of (b) the power delivered to a matched load (32.3 k
) by the PZT
bimorph energy harvester, and (a) the RMS power harvested at varying
vibration levels at the resonant frequency. . . . . . . . . . . . . . . . . . . 112
xii
6.5 Illustration showing the micro-structure of a Macro Fiber Composite (MFC)
energy harvester operating in a bending mode which utilizes the d
3
1 piezo-
electric coecient. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.6 (a) Schematic of multilayer MFC cantilever with proof mass, and (b) Pho-
tograph of multilayer MFC cantilever with proof mass and specied di-
mensions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.7 (a) Plot of RMS power delivered into matched load (459 k
) by series-
connected two-layer MFC cantilever energy harvester with 1.0g vibration
source, and (b) plot of RMS power delivered into matched load at varying
vibration intensities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.8 (a) An illustration of the proposed bistable magnetic cantilever energy
harvester with indicated magnet polarities, and (b) photograph of the fab-
ricated device used in preliminary testing. . . . . . . . . . . . . . . . . . . 117
6.9 Characterization of the response of a bistable magnetic cantilever energy
harvester, with plot of RMS power delivered to a 1 M
load with 0.5 g
vibration strength for three magnet separation distances. . . . . . . . . . 118
6.10 Conceptual diagram of a proposed low-frequency vibration energy harvest-
ing system with an array of bistable magnetically-coupled cantilevers. . . 119
6.11 (a) Diagram of an MFC beam energy harvester with both sides clamped
and bistable states, and (b) Photograph of the fabricated bi-stable MFC
energy harvester with dimensions indicated. . . . . . . . . . . . . . . . . . 120
7.1 (a) Diagram showing basic structure of acoustic transducer using Helmholtz
resonance for sensitivity enhancement. (b) Diagram showing COMSOL
simulated pressured distribution demonstrating the amplication of dier-
ential pressure across the cantilever. . . . . . . . . . . . . . . . . . . . . . 124
7.2 Diagram showing (a) top view and, (b) cross-sectional view of two-wafer
Helmholtz resonator design, with indicated geometric parameters. . . . . . 125
7.3 Circuit diagram showing the electroacoustic modeling of a Helmholtz resonance-
enhanced microphone, with coupling between acoustic domain models and
mechanical domain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
7.4 Graph showing simulated dierence in microphone sensitivity with and
without the use of Helmohltz resonator for passive amplication. . . . . . 134
7.5 Diagram showing the fabrication process of Helmholtz resonator cavity
using KOH as a silicon etchant. . . . . . . . . . . . . . . . . . . . . . . . . 135
xiii
7.6 Diagram showing fabrication process of hemispherical cavity using HNA
as a silicon etchant. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
7.7 SEM image of the cross section of an adhesively bonded device with Helmholtz
resonant cavity and neck. . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
7.8 Graph showing measured dierence in microphone sensitivity with and
without the use of Helmohltz resonator for passive amplication. . . . . . 138
7.9 Photograph the fabricated low-frequency Helmholtz resonators, with(left)
1.57 kHz resonant frequency, and (right) 520 Hz Helmholtz resonant fre-
quency. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
7.10 Measured tip displacement of 1.57 kHz resonance device (left) and 520 Hz
device (right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
8.1 System diagram of the wireless tamper detection scheme. An external
RFID reader transmits a sinusoidal signal and measures the backscat-
tered signal strength from the tamper detection chip. The strength of
the backscattered signal depends on the relationship of impedances be-
tween the RFID dipole antenna and series-connected FBAR. A piezoelec-
tric energy harvester, connected to the FBAR via a rectier and capacitor,
generates a large voltage spike when a sucient mechanical impulse is
applied during the detaching process of the chip. This voltage spike will
irreversibly break down the FBAR and permanently alter the characteristic
of the backscattered signal. . . . . . . . . . . . . . . . . . . . . . . . . . . 144
8.2 (a) Cross-section diagram of the FBAR device, which is composed of silicon
nitride (SiNX), zinc oxide, and aluminum layers on a silicon wafer. (b)
Photograph of the FBAR prior to tampering, and (c) photograph of the
same FBAR after permanent dielectric breakdown induced by tampering. 146
8.3 Diagram showing the charging and discharging paths of the energy har-
vester. An additional capacitor is required to ensure that a sucient cur-
rent is developed to permanently breakdown the FBAR and to leverage
multiple impacts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
8.4 Measured characterization of FBAR breakdown as a function of applied
voltage for a Zinc Oxide thickness of 0.54 m. Breakdown occurs in two
phases, with dielectric breakdown occurring at 5 V, and delamination of
the metal layer occurring at 13 V. Time-dependent dielectric breakdown
occurs in the region between 5 V and 13 V, with delamination occurring
given sucient exposure time. . . . . . . . . . . . . . . . . . . . . . . . . 149
xiv
8.5 Plot of measured current through FBAR under a step function with a 20 V
DC voltage applied. The FBAR draws a current with a peak of 0.141 mA
for a duration of 25 ms before permanent breakdown (delamination). . . . 150
8.6 Plots of measured (a) FBAR resistance, (b) FBAR reactance, and (c)
estimated quality factor before and after the breakdown of the FBAR,
showing a destruction of resonance eects due to tampering. . . . . . . . . 151
8.7 (a) Diagram illustrating the meandering dipole design of the RFID antenna
with key dimensions. (b) Plot of simulated S11 parameter of the RFID
antenna. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
8.8 Circuit diagram showing the method of estimating the backscattered signal
power. The re-radiated power is a function of the backscatter coecient
K and the antenna gain G. K is the relative strength of the backscattered
power as a function of the antenna impedance and FBAR impedance. . . 153
8.9 Estimated backscatter coecient K based on the simulated RFID antenna
impedance and measured FBAR impedances before and after the break-
down of the FBAR. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
xv
Abstract
This dissertation presents several micro-electromechanical (MEMS) sensors and devices
based on thin-lm piezoelectric materials to enable zero-power and ultra-low-power intel-
ligent systems in power-constrained scenarios.
A MEMS resonant microphone array has been developed and evaluated as a mechan-
ical lter-bank front end for speech recognition and respiratory monitoring experiments.
These experiments consistently demonstrate robustness to ambient noise relative to tra-
ditional digital signal processing methods. With cepstral features computed from 40 ms
frames, we measured up to a 57.8% increase in F1 score by using a resonant-array process-
ing method for a signal-to-noise ratio of -26 dB. Using spectro-temporal cepstral features
classied with a dense neural network, improvements of up to 58.8% were measured for
signal-to-noise ratios of -26 dB over an equivalent digital lter implementation.
A complete sensing and low-power signal processing system was developed and eval-
uated to integrate array-based respiratory sound sensing, vibration energy harvesting,
and wireless transmission of data upon detection of wheezing. Using an ultra-low power
processor with 16 kB SRAM and 24 MHz processing speed, a resonant-array based res-
piratory classication system was implemented with classication cycles completing in a
0.46 second period with an average power consumption of 0.596 mW, a factor of 11.1
xvi
improvement over a typical implementation.
A method for passively amplifying the sensitivity of the developed microphone arrays
was hypothesized, modeled, fabricated, and experimentally validated. The method, based
on a micro-fabricated Helmholtz resonator cavity, was shown to improve peak sensitivity
and quality factor of resonant microphones by up to 13.9 in centimeter-scale devices, and
by up to 2.16 in micro-scale devices.
Piezoelectric energy harvesters have been developed on both bulk ceramic and
exible
substrates, and were characterized for harvesting energy from mechanical vibrations.
Bimorph PZT cantilevers were experimentally measured to show an RMS power density of
104.4//
2
. Two bistable nonlinear energy harvesters were also modeled, fabricated,
and characterized, and show a gure of merit up to 35.3 //
2
.
A zero-power wireless authentication system based on FBARs was fabricated, simu-
lated, and experimentally evaluated as a unique method for wireless and passive detection
of tampering activity within integrated circuits. This proof-of-concept system has a RFID
interrogation frequency of 2.6 GHz, and an energy harvester generating a 5 V pulse was
demonstrated to permanently alter the RFID spectral characteristics.
These demonstrations of low-power systems based on MEMS resonators and thin-lm
piezoelectrics provide several creative solutions to emerging power-constrained applica-
tions, including wearable health monitoring, distributed sensor nodes, and internet-of-
things.
xvii
Chapter 1
Introduction
The work presented in this thesis presents several MEMS devices, based on piezoelectric
thin-lms, which target zero-power applications. Although the term "zero-power" is not
physically accurate, as power is always required and expended in electrical systems, it
expresses the goal of developing devices that have no dependency on external power
supplies or batteries, and function as electrically-passive systems.
This thesis outlines the development of a wireless sensing-node-system that is fully
powered by ambient vibration energy harvesters. A MEMS piezoelectric resonant mi-
crophone array, inspired by the human auditory system, was designed, fabricated, and
implemented for low-power and noise-robust acoustic signature detection. The array's
performance in signature detection applications was evaluated, and an integrated system
was developed.
1
1.1 Motivation
Zero-power sensors and systems have attracted research interest in recent years due to the
emergence of the Internet-of-Things (IoT), Wireless Sensor Networks (WSN), wearable
and implantable health monitoring, and many other applications that require robust
performance with little-to-no maintenance by the user or provider.
There is a wide need for wireless sensor networks with the capability of detecting
acoustic signatures and wirelessly signaling events of interest. Four examples of applica-
tions that were specically considered in the design of the systems that are studied in
this thesis are:
Oil pipeline monitoring is a critical but currently inecient task [1]. Leaks in these
pipelines can take weeks to detect manually, and the damage to the environment
and loss of resources for the responsible company can be very signicant. Pipeline
leaks and other structural defects can be detected by the acoustic signature of the
sound emitted by water
ow. A self-powered and reliable WSN can potentially use
this signature to monitor pipeline integrity and to notify a maintenance crew in the
case of a potential leak.
Military reconissance would greatly benet from deployable wireless sensor nodes
that monitor vehicle trac within a particular distance [2]. The unique signatures
of vehicles can be detected and classied, and the seismic vibration created by the
passing vehicle could be harvested. For example, a typical truck has two sharp
2
peaks in the vibration spectrum at 55 Hz and 79 Hz, which can be both classied
and harvested for energy within the same system.
Early detection of asthma attacks can most eectively be achieved by continous lis-
tening of lung sounds with a stethoscope [3]. However, there is currently no wear-
able, wirelessly-connected, and always-on lung-sound tracker. A ultra low-power
device with continuous monitoring of the acoustic signature of sensed lung sounds
can identify precursors of asthma attacks, and automatically notify for assistance.
Vibration from human walking motion can potentially be implemented to realize a
self-powered continuous breath monitoring system.
Naval vessels have a need for distributed wireless sensors for detection and identi-
cation of mechanical problems, and monitoring of heating, ventilation, and air con-
ditioning (HVAC) systems [4]. However, these sensors must be wireless and battery-
less to prevent clutter and unnecessary maintenance in limited space aboard naval
vessels. The engine room and HVAC-related machinery can provide large levels of
vibration energy to power the sensors, signal processing electronics, and wireless
data communication.
These applications all require an ultra-low power or a self-powered detection scheme.
Additionally, they require very accurate signature recognition and detection schemes that
are immune to external noise.
3
1.2 Problem statement
This thesis presents several novel devices and methods based on piezoelectric thin lms
to address the challenges posed by sensing in power-limited, remote, wearable, or dis-
tributed devices. While each technique and device presented here implements a dierent
function within the system architecture, the unifying theme of these implementations
is the utilization of the passive signal-generating property of piezoelectric materials for
use in microfabricated systems in technical areas that have typically been dominated by
active externally-powered devices.
Specically, we outline the results of our research into the use of microfabricated res-
onant microphone arrays for noise-robust acoustic communication and signature recog-
nition and classication for speech recognition and wearable lung monitoring. The use
of micro-fabricated acoustic resonant cavities to enable zero-power signal amplication
in MEMS microphones is proposed, modeled, and demonstrated. Several designs for
piezoelectric vibration energy harvesters are investigated and evaluated for zero-power
applications where active processing is required but access to reliable power sources is
limited or impractical. Finally, the implementation of an acoustic resonator-based se-
curity and authentication system is presented as a zero-power system to be integrated
within integrated circuit (IC) packages to allow wireless assessment of tampering activity
within sensor nodes or other sensitive ICs.
These applications and devices have individual scientic merit, but can also be in-
tegrated within a single wireless sensing node system. Modern wireless sensing node
4
systems rely on purely digital processing of sensed signals, which can be computation-
ally intensive and ineective in the presence of high-intensity background noise. These
sensing systems are typically powered by battery or wired connections. However, many
applications require a vast array of sensors in remote or hard to access locations, where
battery replacement is costly and inecient.
This thesis proposes a complete design, device fabrication, and experimental eval-
uation of the proposed sensor node system, shown in Figure 1.1. The system utilizes
piezoelectric energy harvesting, which is discussed in detail in a Chapter 5, to power the
system through energy harvested from ambient vibrations. A MEMS resonant micro-
phone records audio signals of interest, lters noisy acoustic signals, and uses the array's
digitized outputs for feature extraction and signature recognition within an ultra-low
power microprocessor. The results of the recognition are transmitted with an onboard
antenna and radio.
Figure 1.1: Diagram showing the proposed fully-integrated system.
5
1.3 Scope of Work
A MEMS resonant microphone array has been developed, characterized, and tested as
a lter-bank for simple speech and health-signal processing experiments. These experi-
ments indicated that for certain noise conditions, the use of a resonant microphone array
becomes relatively immune to acoustic interferences while traditional digital signal pro-
cessing methods degrade with increased noise intensity.
A complete system was developed and evaluated to integrate array-based lung sound
sensing with the power harvesting unit, and wireless transmission of data upon detection
of wheezing. The system was evaluated in it's eectiveness and accuracy of classication
for the stated applications and compared to traditional processing methods. Additionally,
the implementation was evaluated against supply power and memory requirements.
A method for passively amplifying the sensitivity of the developed microphone ar-
rays was hypothesized, modeled both analytically and using nite-element simulations,
fabricated, and then experimentally validated.
Piezoelectric energy harvesters have been developed on both bulk ceramic and
ex-
ible substrates. The devices were characterized for harvesting energy from mechanical
vibrations. Two new energy harvesting designs with bistable nonlinear operation have
also been modeled, fabricated, and characterized. Experimental evaluations have been
completed for the persistent powering of a integrated sensing system. We also integrated
energy harvesters with a power management circuit to provide stable power to the pro-
cessing system.
A zero-power wireless authentication system was built, simulated, and experimentally
6
evaluated as a novel and robust method for wireless and passive detection of tampering
and authentication of integrated circuits and other electrical systems.
1.4 Overview of Chapters
In Chapter 1, motivation for the work is introduced, and a high level background of the
thesis is presented. Key concepts are introduced and the overall scope is presented with
reference to existing publications in the eld.
Chapter 2 is a detailed introduction to the concept of resonant microphones, which are
extensively used in the research projects presented in this thesis. This chapter discusses
the modeling, fabrication, and measurement methodology of the devices.
Chapter 3 introduces the signal processing methods with the use of resonant micro-
phones presented in Chapter 2. These methods are compared theoretically with existing
methods.
Chapter 4 presents the experimental methods and results of signal classication ex-
periments with the novel resonant microphone-based processing methods. These methods
are evaluated for three applications by assessing classication accuracy in the presence of
background interference.
Chapter 5 provides an overview of the system integration of the processing methods
in hardware for the application of wheeze detection in lung sounds. A fully integrated
system is evaluated for power-eciency with the resonant microphone-based processing
method.
Chapter 6 highlights the use of macro-scale piezoelectric resonators for use as vibration
7
energy harvesters. The energy harvesters presented in this chapter can be integrated with
the system outlined in Chapter 5 to form a self-powered wireless sensing platform.
Chapter 7 describes a zero-power amplication method of microphone sensitivity
based on Helmholtz resonance. The phenomenon of acoustic Helmholtz resonance is
introduced and modeled, and integration with resonant microphones is experimentally
demonstrated utilizing a micro-fabricated acoustic resonant cavity.
Chapter 8 describes a zero-power authentication methods for wirelessly assessing the
authenticity of integrated circuits. The method, which is based on Film Bulk Acoustic
Resonators (FBAR), is modeled and experimentally demonstrated.
Finally, in Chapter 9, the conclusions of this work are discussed, and future research
directions are proposed.
References
[1] J. Zhang, \Designing a cost-eective and reliable pipeline leak-detection system,"
Pipes and Pipelines International, vol. 42, no. 1, pp. 20{26, 1997.
[2] J. Altmann, \Acoustic and seismic signals of heavy military vehicles for co-operative
verication," Journal of Sound and Vibration, vol. 273, no. 4, pp. 713{740, 2004.
[3] A. Bohadana, G. Izbicki, and S. S. Kraman, \Fundamentals of lung auscultation,"
New England Journal of Medicine, vol. 370, no. 8, pp. 744{751, 2014.
[4] K. L. Gerhard, Non-intrusive vibration monitoring in US Naval and US Coast Guard
ships. PhD thesis, Massachusetts Institute of Technology, 2013.
8
Chapter 2
Piezoelectric Resonant Microphone
A microfabricated resonant microphone array was modeled, designed, fabricated, tested,
and utilized in acoustic signature detection experiments. The array was composed of
13 resonant channels with center frequencies spanning from 860 Hz to 5997 Hz. Each
channel consisted of a silicon cantilever paddle utilizing thin-lm zinc oxide (ZnO) as a
sensing mechanism.
The preliminary design and fabrication for this microphone array was completed by
Lukas Baumgartel [1], with additional characterization and fabrication improvements
made by the author of this thesis. This array was further utilized in the acoustic pattern
recognition experiments outlined in Chapter 4 and the system presented in Chapter 5.
2.1 Background
The resonant microphone array presented in this chapter was inspired by the human
audiotory system, which also features a distributed resonant transduction mechanism.
The cochlea, an organ within the inner ear which houses the hair cells serving as the ear's
9
mechanism to transduce pressure waves into neural signals, has the ability to dynamically
adjust frequency response such that signals of interest are separated from background
noise [2]. This is commonly known as the 'cocktail party eect', and has thus far been
dicult to mimic using purely digital methods [3, 4]. Hair cells within the human cochlea
are sensitive to dierent frequencies within the audible range, and each hair cell transmits
a pre-ltered acoustic signal to the nervous system. The MEMS resonant microphone
array presented in this chapter attempts to mimic this behavior by pre-ltering signals
before processing to improve signature recognition performance in the presence of acoustic
noise.
Resonant acoustic transducers are commonly employed in applications such as acous-
tic ranging [5] and ultrasonic imaging [6]. However, this thesis present the rst application
of resonant microphones for audio-spectrum signal processing tasks.
2.1.1 Piezoelectric Thin Films
Piezoelectric thin lm transducers have several key advantages over electromagnetic and
thermal sensing and actuation. Most notably, piezoelectric thin lms oer minimal power
consumption, quick response, insensitivity to electromagnetic interference, good scaling
properties, and extensive process compatibility [7]. Generally, piezoelectric transduction
is attractive for applications that demand low power, low noise, high frequency, or large
sensitivity.
The most common piezoelectric thin lms used in today's MEMS fabrication pro-
cesses are lead zirconate titante (Pb(Zr,Ti)O
3
called PZT), aluminum nitride (AlN), and
zinc oxide (ZnO). Many other materials exist, but are normally reserved for specialized
10
applications. Some of these include quartz, barium titanate (BaTiO
3
), polyvinylidene

uoride (PVDF), and lithium tantalate (LiTaO
3
). This thesis focuses on devices based
on PZT and ZnO thin lms.
Fundamentals
Piezoelectrics are materials with crystalline structures that develop electric dipoles in
response to an applied mechanical strain. The reverse eect is also true, allowing piezo-
electrics to be used for mechanical actuation. PZT is a ferroelectric ceramic compound
with a perovskite crystalline structure, while ZnO and AlN are both dielectrics with
wurtzite crystalline structures. Because ZnO and AlN share an identical structure, they
exhibit similar characteristics and properties.
The piezoelectric eect is described analytically by the following governing equations:
=

+

(2.1)
= +
(2.2)
Where strain and stress are second-rank tensors related to electric eld and
electric
ux density vector . The tensor is the mechanical compliance, is the piezo-
electric coecient, and is the dielectric permittivity. In the case of AlN, ZnO, and
PZT, crystal symmetry allows the complexity of the piezoelectric coecient tensor, elas-
tic constant tensor, and permittivity tensors to be reduced to simplied matrices. This
simplied form, called reduced matrix notation, is shown in equations (2.3) and (2.4). It
11
has the same form for AlN, ZnO, and PZT - though it is important to note that this
simplication is not necessarily equivalent for all piezoelectric crystals [8].
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣

1

2

3

4

5

6
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
=
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣

11

12

13
0 0 0

12

11

13
0 0 0

13

13

33
0 0 0
0 0 0
44
0 0
0 0 0 0
44
0
0 0 0 0 0 2(
11
−
12
)
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣

1

2

3

4

5

6
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
+
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
0 0
31
0 0
31
0 0
33
0
15
0

15
0 0
0 0 0
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎣

1

2

3
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎦
(2.3)
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎣

1

2

3
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎦
=
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎣
0 0 0 0
15
0
0 0 0
15
0 0

31

31

33
0 0 0
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎦
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣

1

2

3

4

5

6
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
+
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎣
11
0 0
0 11
0
0 0 33
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎦
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎣

1

2

3
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎦
(2.4)
Typically for MEMS applications, we are most interested in piezoelectric coecients

33
and
31
. The longitudinal piezoelectric coecient,
33
, describes the electrical re-
sponse on face 3 to stresses along the direction 3 acting on face 3, with charges developed
on face 3. The transverse piezoelectric coecient,
31
, describes the charges developed
on face 3 by stresses acting on face 1 in direction 1. These two modes are visualized in
Figure 2.1.
12
Figure 2.1: Diagram showing (a) sensing if using the longitudinal piezoelectric coecient

33
and (b) sensing if using the transverse piezoelectric coecient
31
.
Another commonly cited material property is the piezoelectric coupling coecient k,
which is dened as:

2
=
(Piezoelectric energy density stored in the material)
2
Electrical energy density× Mechanical energy density
(2.5)
The coupling coecient k can be interpreted as the ratio of energy used to the total
energy delivered to the material [8].
Material Properties
The values of important piezoelectric parameters for PZT, ZnO, and AlN are shown
in Table 2.1. These parameters can be highly dependent on composition, orientation,
crystalline quality, and grain size [9].
While ZnO and AlN have similar piezoelectric constants, we see that PZT has a value
of
33
that is roughly an order of magnitude larger. This is the preeminent motivation for
using ferroelectric ceramics such as PZT. It has been found experimentally that ceramic
PZT has maximum piezoelectric sensitivity with a solid solution of 55% lead zirconate
13
Parameter ZnO AlN PZT

33
(pm/V) 12.4 5.1 360

31
(pm/V) 5.0 2.6 180

31
(C/
2
) -1.0 -1.05 -8..-12
33
10.9 10.5 300-1,300

33
- - 0.69

31
- - 0.35
Table 2.1: Summarized piezoelectric parameters for ZnO, AlN, and PZT thin-lms [9].
and 45% lead titanate. In sensing systems, the larger piezoelectric constant allows for
higher sensitivities, and consequently PZT thin lms have become the preferred material
for medical ultrasound and ultrasonic ranging applications.
Nonetheless, PZT suers from several drawbacks that make it unsuitable for many
applications. PZT has poor temperature stability, with its performance as a piezoelectric
being a strong function of temperature. Additionally, under elevated temperatures, PZT
undergoes aging which results in gradual depolarization.
Another dierence is that PZT can sustain a maximum voltage of about 1 MV/cm
prior to breakdown, while wurtzite structures like AlN support a breakdown eld of about
5 MV/cm [10]. Piezoceramics such as PZT also exhibit pyroelectric eects, meaning that
they produce an electrical response to temperature changes. This is a disadvantage in
strain sensing applications, as temperature
uctuations cannot be dierentiated from the
desired signal.
AlN has a relatively large band gap of 6 eV, and thus acts as a good dielectric insulator
with high resistivity [11]. ZnO is essentially a semiconductor, with a small band gap of
about 3.37 eV. As a result, it suers from large dielectric losses at low frequencies. This
is one of the reasons why an insulation layer is typically used with ZnO. The resistivity of
14
ceramics such as PZT is in general lower than that of single crystals, like ZnO and AlN.
Parameter ZnO AlN PZT Si SiO
2
Young's Modulus (GPa) 123 300 63 220 70
Coecient of Thermal
Expansion (ppm/
∘ C)
4 5 9.2 2.1 0.4
Residual Stresses (MPa) 100 -450/0 200 400-1000 -280
Table 2.2: Summarized coecients of mechanical properties of materials for ZnO [12],
AlN, PZT, Si, and SiO
2
[10]
AlN is a very sturdy material with large mechanical stability. On the other hand,
PZT is mechanically weak and brittle; it is not suitable for use in structures that are
exposed to signicant stresses. To compensate for this weakness, there have been some
attempts to create sturdier composites by integrating PZT with a polymer material. This
increases the mechanical stability at the expense of piezoelectric sensitivity. One example
is porous PZT, a composite of PZT with air, which has advantages in ultrasonic appli-
cations because of its low density, low stiness, and good acoustic impedance matching
with human tissue [13].
Fabrication
To maximize the piezoelectric coecient of these studied materials for MEMS applica-
tions, the crystal planes of the piezoelectric must be aligned with the electrodes and
actuation axis of the device. This is accomplished using two dierent mechanisms for
perovskite ceramics and wurtzite crystalline structures.
AlN and ZnO have a single axis of polarity, referred to as the c-axis. The alignment of
this axis is achieved by using a bottom electrode layer that closely matches in structure
with the desired crystal plane of growth, a process that is called lattice-matching epitaxy.
15
ZnO and AlN has been reliably deposited on Al, Au, Si, SiO
2
and Sapphire substrates,
with varying degrees of quality [14].
AlN and ZnO are typically deposited using RF sputtering. Since ZnO suers from
large dielectric loss, it is typically deposited at low temperatures to maximize the material
resistivity [11]. This is not as serious of a concern for AlN, which has a very high resistivity,
so it can be deposited at higher temperatures. Optimal temperatures for AlN sputter
deposition are between 200
∘ C to 450
∘ C [10]. Alternative deposition methods of AlN and
ZnO include sol-gel spin coating, chemical vapor deposition, and pulsed laser deposition.
The dipoles of PZT are oriented post-deposition by applying an electric eld at an ele-
vated temperature. This process is called poling, and it consists of the following steps [8]:
1. PZT is heated to a temperature between 600
∘ C and 700
∘ C, which is slightly lower
than the transition temperature.
2. A large DC voltage is applied across the thin lm for 2-3 hours to ensure that the
dipoles are oriented in the direction of the electric eld.
3. Without removing the supply, the PZT is cooled back to room temperature. After-
words the eld is removed, and the material is polarized as desired.
This polarization procedure can be dicult to control, and process variations can lead to
structural defects that degrade piezoelectric performance.
There are several techniques used to deposit thin-lm PZT. Most commonly, a method
called metalorganic chemical vapor deposition (MOCVD) is used [15]. Metalorganic refers
to the precursor compounds used in the deposition process. For PZT, these are typically
16
tetraethyl lead, zirconium tetrabutoxide, and titanium isopropoxide. Films of thickness
in the range of 50-1400 nm can be grown using this method [16].
Alternatively, RF sputtering deposition can be employed. Sputtering of PZT has
several disadvantages compared to MOCVD. The deposition rate of sputtering is several
times slower in comparison. There are also issues related to the generation of surface
defects, and diculties controlling the PZT lm composition. Other deposition meth-
ods that are commonly used include sol-gel processing, electron beam evaporation, spin
coating, and ion beam deposition. Many of these methods have have diculties with
consistency and control over physical parameters.
AlN and ZnO are typically etched using wet chemistry. ZnO lms are highly reactive,
with signicant etching taking place in most acids and bases [17]. Additionally, ZnO is
etched by common solutions such as photoresist developer. To minimize the exposure
and damaging of ZnO lms, a passivation layer such as silicon nitride is typically used.
PZT can be patterned using several wet and dry etching techniques. Wet etching of
PZT can be achieved by using buered HF, followed by a 2HCl:H
2
O solution to remove
the remaining residues and byproducts [18]. Dry etching has been demonstrated using
reactive ion etching (RIE), deep reactive ion etching (DRIE), ion-beam etching (IBE),
electron-cyclotron resonance (ECR) etching, and inductively coupled plasma (ICP) etch-
ing [19]. IBE and ICP have shown the best results with minimal undercut, however it
has been found that exposure of PZT to ion bombardment results in a degradation of
electrical properties.
AlN is completely compatible with the CMOS fabrication process, with robustness to
temperature and most common wet chemistry. For this reason it it has become widely
17
commercialized in MEMS RF lters for cell phones. Because of the reliability and versatil-
ity of AlN, the MEMSCAP company was able to introduce the rst piezoelectric standard
MUMPS (multi-user MEMS) process in 2013 [20]. This provides users with cost eective
access to MEMS prototyping and fabrication services for piezoelectric micro-systems.
Due to the reactive nature of zinc oxide, modications to typical IC processing must
be made for reliable integration. Another major diculty in dealing with ZnO is the fast
diusion rate of Zinc into silicon [9]. Both AlN and ZnO can be sputtered at low tem-
peratures, allowing compatibility with temperature-sensitive materials, such as parylene.
PZT has several compatibility shortcomings. Since dipole-alignment requires large
temperatures, temperature-sensitive materials cannot be used in conjunction with PZT.
Additionally, PZT is not compatible with high temperature processes because polarization
is lost at the Curie temperature, which occurs around 300
∘ C.
2.1.2 MEMS Microphones
Micro-fabricated microphones and acoustic transducers are ubiquitous in commercial
products ranging from hearing aids, automotive sensors, mobile phones, and other con-
sumer electronics. In contrast to the resonant microphone described in this research,
commercial microphones typically have a
at-band response in the audible spectrum (20
to 20,000 Hz), with resonant frequencies between 20 to 50 kHz. The most common
microphone design types are condenser, electret, and piezoelectric microphones.
Condenser microphones, which operate on a capacitive sensing principle, are the
most commercially used microphone design due to their high sensitivity [21]. Con-
denser microphones require a DC bias such that the changes in capacitance can be
18
sensed when the diaphragm vibrates relative to the backing plate. The microphone
requires a separation between the backing plate and the microphone diaphragm by
a narrow air gap, which adds to the fabrication complexity. The largest developers
of condenser microphones are InvenSense, Knowles, and STMicroelectronics, with
typical sensitivities of 13 mV/Pa and signal-to-noise ratios (SNR) of 60-70 dBA.
Electret microphones (ECM) are a variation of capacitor-based microphones, which
eliminate the need for an external supply by utilizing materials with implanted
static charge [22]. These microphones have high sensitivity, but suer from lower
performance-density, are vulnerable to high-temperatures, and degrade in perfor-
mance over time. Both piezoelectric and ECM microphones require a low-noise
pre-amplier circuit.
Piezoelectric microphones utilize thin-lm or bulk materials exhibiting the piezo-
electric eect to generate an electrical potential when subject to a mechanical strain.
These microphones oers advantages over other transduction mechanisms in their
simplicity of fabrication, linearity, and low-power operation. A commonly cited dis-
advantage is low sensitivity, which is addressed in this research by operating at the
resonance frequency and integration with passive acoustic ampliers. A commercial
piezoelectric microphone was introduced by Vesper in late 2015 claiming a
at-band
sensitivity of about 14 mV/Pa and the highest SNR (up to 80 dB) of a commercial
microphone, with preliminary design presented in [23].
19
2.2 Modeling and Design
The intended purpose of microphone arrays developed in this thesis is resonant ltering
of the sensed acoustic signals; thus precise modeling of the relationship between design
geometries and resulting resonance frequency is critical.
(a) (b)
Figure 2.2: Scanning electron microscope (SEM) images of a single transducer showing
(a) design elements, and (b) denitions of geometric dimensions.
An analytical model was derived using the Rayleigh method in [1] for a paddle-shaped
cantilever microphone with the nal design shown in Figure 2.2. As the structure bends,
it bears both kinetic and potential (bending strain) energies. By equating the two, the
th mode resonance frequency

is obtained with value:

2

=

(2 )
2
∫︀

0
()(
2
()

2
)
2
∫︀

0
()
2

()
(2.6)
Where() is the area moment of inertia,

() is a shape function that approximates
the transverse displacement of the cantilever-beam diaphragm, () is the cross-sectional
area, is the length of the cantilever, and is the distance along that length. and 20
are the material's Young modulus and density, respectively.
The shape function () can be any function that satises the boundary conditions
for the desired clamped cantilever beam geometry. If the shape function is a perfect
match to the actual bending shape, the calculated resonance frequency will be equal
to the physical resonance frequency. Any error in the shape function will result in a
calculated resonance frequency that is higher than the true resonance frequency. Several
shape functions were used to calculate the resonance frequency in Equation 2.6, and the
following polynomial shape function was chosen:
() =

3
4
(
4
− 4
3
+ 6
2

2
) (2.7)
where is the displacement at the cantilever beam's tip. The calculated resonance fre-
quency using this method closely agrees with results given by Finite Element Method (FEM)
modeling using COMSOL's structural mechanics module [24]. Using an analytical model
of the cantilever is preferable to relying on FEM simulations as it allows quicker compu-
tation and allows for backward calculation of parameters when given a desired resonance
frequency.
Using the derived analytical model, a 13-channel microphone array was designed with
parameters given in Table 2.3.
The predicted sensitivities of the microphone array elements are plotted against fre-
quency in Figure 2.3. Due to the overlapping frequency responses of adjacent array ele-
ments, a minimum sensitivity of about 5/ is achieved within the covered frequency
range. An added benet of the usage of a microphone array is this boosted minimum
21
Microphone 1 3 5 7 9 11 13
(mm) 2.5 2 1.6 1.3 1.2 1.1 1
(mm) 0.75 0.27 0.28 0.34 0.27 0.24 0.24
(mm) 0.75 0.4 0.4 0.4 0.35 0.34 0.34

0
(Hz) 875 1726 2585 3448 4291 5144 5997
Table 2.3: Summary of the selected geometric design parameters and the predicted reso-
nant frequency of a selected subset of array elements.
sensitivity if the channels of the array are summed and the output is alternately used as
a broadband microphone.
Figure 2.3: Bode plot of the predicted microphone array sensitivity.
2.3 Fabrication
The array is fabricated on silicon-on-insulator (SOI) wafers with 4.5 thick device
layer using standard microfabrication techniques. The fabrication steps are outlined in
Figure 2.4 and these steps are discussed in detail below:
22
Figure 2.4: Diagram showing fabrication processes of resonant MEMS microphone.
1. The SOI wafer is encapsulated in silicon nitride (SiN

) via low-pressure chemical
vapor deposition (LPCVD). This SiN

layer is used as an etch mask for back cavity
etching.
2. The backing cavity window is dened in a standard photolithography step and
the SiN

is patterned by CF
4
reactive-ion etching (RIE). Anisotropic potassium
hydroxide (KOH) bulk silicon micromachining is then used to etch away the backing
cavity. The SOI's buried silicon oxide (SiO
2
) layer serves as an etch stop, due to
23
the large etch sensitivity of Si to SiO
2
in KOH. This allows us to very accurately
control the cantilever thickness.
3. The SiN

on the surface of the device is removed with RIE. Hot phosphoric acid
wet etching can be employed as an alternative method, with the added benet of
reduced damage to the silicon.
4. Deep reactive-ion etching is used to partially dene the paddle shape and to create
alignment marks on top of the wafer (with a depth of about 0.5 ). Alignment
to the back cavity is achieved using an infrared (IR) bulb, which is visible through
the thin device and oxide layers of the SOI wafer.
5. The surface of the wafer is cleaned by etching in oxygen plasma RIE. The buried
oxide layer is removed by wet etching in buered hydro
uoric acid, with the added
benet of removing any native oxide from the silicon surface.
6. An aluminum layer with a thickness of 0.20 is deposited by thermal evaporation
or RF sputtering. This layer is patterned and etched for use as the bottom electrode.
7. Piezoelectric ZnO, which serves as the sensing layer, is deposited with a thickness
of 0.59 by RF sputtering. Plasma-enhanced CVD is used to deposit a 0.1
layer of SiN

, which serves as an electrical isolation layer to prevent charge leakage
between the two electrodes. An aluminum layer of thickness 0.22 is deposited
and patterned to be the top electrode. This same mask is used to pattern the
SiN

and ZnO layers. The necessary thicknesses are determined such that the net
residual strain in the completed cantilever is 0.
24
8. The wafer is diced prior to the release of the cantilevers, due to the fragility of the
diaphragms. As additional protection against water
ow and debris from the dicing
saw, a thick layer of photoresist is deposited before dicing. This layer is afterwords
removed by soaking in acetone and ashing with oxide plasma. After dicing, the nal
release of the cantilevers is done by DRIE etching using the mask from Step (4).
This fabrication process is CMOS compatible, and necessary electronics such as pre-
ampliers, analog-to-digital converters (ADC), and digital signal processing (DSP) cir-
cuits can be integrated on-chip. A photograph of a completed microphone array is shown
in Figure 2.5.
Figure 2.5: Photograph of fabricated microphone array with center frequencies indicated.
2.4 Characterization
The microphone array was characterized to determine the sensitivity and noise
oor, and
to compare these characteristics to those of commercial microphones.
2.4.1 Experimental Setup
The fabricated microphone array is wirebonded to a custom PCB containing a simple
pre-amplier circuit for each of the 13 array elements. The circuit utilizes a LTC6244
25
low-noise CMOS op-amp, connected in a non-inverting amplier conguration for a gain of
101 (40.1 dB). The input to the amplier is connected to a corresponding array electrode
and the cathode of a diode, which prevents DC charge buildup in the top electrode by
allowing discharging through the reverse bias leakage current of the diode.
Figure 2.6: (top) Photograph of experimental measurement and characterization setup in
anechoic chamber and (bottom) diagram showing schematic of measurement setup with
LabVIEW control.
26
The microphone and PCB are enclosed in an aluminum enclosure with a mesh opening
which electronically shields the array and PCB. All 13 output channels are digitized with
a ROGA data acquisition module (DAQ) at a rate of 44 kSa/s and a 16-bit resolution. A
calibrated reference microphone (GRAS 40A0) is connected to the second input channel
of the DAQ. The reference microphone exhibits a
at response up to 20 kHz, and pro-
vides a
atband reference sensitivity of 12.5 mV/Pa by which sensitivity in the MEMS
microphones can be accurately determined in units of mV/Pa.
A PC running a custom LabVIEW measurement control routine saves and analyzes
the MEMS microphone and reference signals through a USB connection to the DAQ.
The LabVIEW routine controls a function generator to sweep the frequency of an input
acoustic signal, which is generated by a ceramic card speaker. The entire setup is placed
within an anechoic chamber walled with foam to acoustically isolate the test setup from
the lab environment and to prevent any scattering eects. The complete characterization
setup is shown in Figure 2.6.
2.4.2 Experimental Results
The measured sensitivity is plotted in Figure 2.7 for the 13 channels of the microphone.
Table 2.4 summarizes the measured center frequency, the error between the designed and
measured center frequencies, the peak sensitivity of each channel, and the quality factor.
The average deviation from the designed resonant frequency of the microphones is 1.81%,
and the average quality factor is 43.1.
For the respiratory sound classication experiments performed in Chapter 4, a single
resonant microphone was used, and lower resonant frequencies were eectively obtained
27
Figure 2.7: Electrically measured sensitivity of the 13 array channels
Microphone 1 3 5 7 9 11 13

0
(Hz) 860 1750 2591 3454 4341 5300 6263

0
(%) 1.71 -1.39 -0.23 -0.17 -1.17 -3.03 -4.44
(mV/Pa) 202.6 81.85 24.62 21.12 17.03 12.65 10.8
− 44 43.1 21.6 50.6 59.3 51.4 51.5
Table 2.4: Summary of the measured parameters of a selected subset of array elements.
by increasing the playback speed of the audio signal under test. This is a work-around
that was necessary to test resonant microphone array for low-frequency signal processing
applications, while development of low-frequency resonant microphones was still under
development. The parameters of this microphone element are summarized in Table 2.5,
and compared to the state-of-the-art commercial microphones. It is evident that by using
a resonant microphone, we are able to obtain greater sensitivities and lower noise
oors
than commercial
atband microphones can oer.
28
Microphone f

Q
Sensitivity
(dBV)
SNR
(dBA)
Resonant Microphone 1.3 kHz 45.8 -17.2 85.8
Vesper VM1000 - - -38 64
Knowles SiSonic - - -38 65
STMicroelectronics MP34* - - -26 64
Table 2.5: Summary of the parameters of a resonant microphone element in comparison
to commercial MEMS microphones.
2.5 Low Frequency MEMS Microphone
A major diculty of implementing low-frequency mechanical resonators on a micro-scale
is that scaling a mechanical design to lower frequencies becomes dicult to fabricate for
high-aspect ratio structures. The yield quickly becomes unmanageable as devices become
thinner than 2 m or larger than about 2 mm on each side.
Another problem is that with reduced thickness comes reduced accuracy. A typical
SOI wafer with device layer of 2 m has an manufacturing error threshold of± 5 m.
With a 25% swing in cantilever thickness we can observe up to a 40% swing in resonance
frequency.
Two alternative designs are proposed as future work for integration into arrays for
low-frequency signal recognition, while maintaining an acceptable yield. Both of these
proposed methods are to be experimentally evaluated for fabrication consistency, yield,
and lowest achievable resonance frequency as future work. Although the proposed system
design requires an array of resonant microphones, it is also possible to implement a hybrid
approach in which low frequency lter banks are applied digitally, and high frequency lter
banks are applied in the acoustic domain.
29
2.5.1 Series-Spring Microphone
One potential method explored was based on the use of a springs-in-series-like design of
a microphone (if considering a mass-spring-damper model for the cantilever). Figure 2.8
demonstrates a reduction in resonance frequency of 56% in a xed 1
2
area.
Figure 2.8: An example of proposed series-spring microphone design, with simulated
reduction in resonance frequency from 2.3 kHz to 1.8 kHz in the same footprint.
A benet of using such a method is the improved stability due to the higher-frequency
'twisting' mode. Twisting in the original device has a central clamping point which is
easier to damage, and leads to asymmetry in the device. It was hypothesized that by using
this modied approach, the fabrication yield of low-frequency devices can be improved.
The proposed device was simulated by FEM and the fundamental mode is plotted in
Figure 2.9, along with the mask layout of such a proposed device. A preliminary fab-
rication iteration has been completed to evaluate yield of large devices, with resonance
frequencies ranging from 400 Hz to 1000 Hz. Although most devices survived fabri-
cation (94.7% yield), a misalignment between the electrode and DRIE release pattern
signicantly damped the response and increased the resonant frequency. This defect is
30
Figure 2.9: Image of series-spring microphone design FEM simulation of (a) displacement
of the rst fundamental mode, and (b) strain distribution of the rst fundamental mode.
(c) The mask layout for patterning of etched cavities and electrodes.
suggested to be corrected in a future iteration by providing sucient tolerance for align-
ment errors. An iterative approach to fabrication and measurement is would be necessary
for future work.
2.5.2 Silicon Proof-Mass Microphone
A potential design modication which can be used to reduce resonant frequency is to
use a silicon proof-mass, shifting from a continuous cantilever-based system to a discrete
mass-on-spring system. This technique could potentially increase yield in the fabrication
processes by retaining structural integrity, while helping to reduce the resonance frequency
via mass loading.
Low-frequency resonant MEMS cantilevers with silicon proof-masses have been demon-
strated for accelerometers [25] and energy harvesters [26]. Adoption for microphone ap-
plications presents additional challenges in maximizing sensitivity, which demands large
surface areas, to maximize incident pressure, and narrow etch gaps, to maintain a pressure
dierential across the cantilever.
31
Figure 2.10: Diagrams showing the proposed microphone with silicon proof-mass design,
and the necessary KOH mask modication for (a) proof-mass etching utilizing corner-
compensation, and (b) proof-mass formation with through-wafer DRIE etching.
Anisotropic KOH wet etching of silicon cannot be applied with a direct pattern out-
lining the cantilever, due to the well-documented convex corner problem [27] in which
convex-corners of<111> crystalline planes are etched at a much faster rate. This problem
can be corrected using a technique called corner-compensation, which is implemented in
Figure 2.10a. Although this technique enables the use of bulk micromachining to fabricate
the proof mass, it can only be achieved with a carefully timed process.
Another potential solution is to partially release the silicon cantilever with KOH, but
complete the release with a through-wafer DRIE step. This step avoids the timing chal-
lenges of corner-compensation and has the benet of having fully-supported cantilevers
until the nal step. A major diculty with this method is the long etch time required for
through-wafer DRIE, which will require innovation on the masking technique to execute
it successfully.
Simulations with COMSOL models for the mass-on-spring design have demonstrated
32
the eectiveness of using such a design. Fabrication and experimental validation of a
silicon-mass microphone are suggested as a future research direction in wearable health
monitoring with MEMS resonators.
2.6 Summary
The design, modeling and fabrication of a MEMS resonant microphone has been reviewed
in this chapter. We proposed and implemented some modications to the existing fabri-
cation processes to improve yield of low-frequency microphones with high aspect ratios,
which has been a bottleneck in reducing the resonant frequency of such devices. The
design of this resonant microphone was also compared with existing commercial micro-
phones, and it was demonstrated that an improvement in peak sensitivity and noise
oor
can be achieved. We proposed a future direction for the fabrication of low frequency
microphones through the addition of a silicon proof mass and through modifying the
anchoring structure of the cantilever microphones.
The resonant microphone design introduced in this chapter is the foundation of the
ultra-low-power signal processing and classication techniques discussed in the following
chapters.
33
References
[1] L. Baumgartel, A. Vafanejad, S. J. Chen, and E. S. Kim, \Resonance-enhanced piezo-
electric microphone array for broadband or preltered acoustic sensing," Journal of
Microelectromechanical Systems, vol. 22, no. 1, pp. 107{114, 2013.
[2] R. F. Lyon and C. Mead, \An analog electronic cochlea," IEEE Transactions on
Acoustics, Speech, and Signal Processing, vol. 36, no. 7, pp. 1119{1134, 1988.
[3] M. Namiki, T. Hamamoto, and S. Hangai, \Spoken word recognition with digital
cochlea using 32 DSP-boards," in Acoustics, Speech, and Signal Processing, 2001.
Proceedings.(ICASSP’01). 2001 IEEE International Conference on, vol. 2, pp. 969{
972, IEEE, 2001.
[4] B. W. Edwards, C. J. Struck, P. Dharan, and Z. Hou, \New digital processor for
hearing loss compensation is based on auditory system," Hearing Journal, vol. 51,
pp. 38{50, 1998.
[5] M. I. Haller and B. T. Khuri-yakub, \Ultrasonic Air Transducer," October, vol. 43,
no. 1, 1996.
[6] T. L. Szabo, Diagnostic ultrasound imaging: inside out. Academic Press, 2004.
[7] C.-B. Eom and S. Trolier-McKinstry, \Thin-lm piezoelectric MEMS," MRS Bul-
letin, vol. 37, no. 11, pp. 1007{1017, 2012.
[8] M. Vijaya, Piezoelectric materials and devices: Applications in engineering and med-
ical sciences. CRC Press, 2016.
[9] A. Safari and E. K. Akdoan, Piezoelectric and acoustic materials for transducer
applications. 2008.
[10] E. Defa y, Integration of Ferroelectric and Piezoelectric Thin Films: Concepts and
Applications for Microsystems. 2013.
[11] S. Trolier-Mckinstry and P. Muralt, \Thin lm piezoelectrics for MEMS," 2004.
[12] M. J. Madou, Fundamentals of Microfabrication: The Science of Miniaturization.
2002.
[13] H. Kara, R. Ramesh, R. Stevens, and C. R. Bowen, \Porous PZT ceramics for receiv-
ing transducers," IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency
Control, vol. 50, no. 3, pp. 289{296, 2003.
[14] W. Wang, W. Yang, Z. Liu, Y. Lin, S. Zhou, H. Qian, F. Gao, and G. Li, \Epitaxial
growth of high quality AlN lms on metallic aluminum substrates," CrystEngComm,
vol. 16, no. 20, pp. 4100{4107, 2014.
[15] Y. Sakashita, T. Ono, H. Segawa, K. Tominaga, and M. Okada, \Preparation
and electrical properties of MOCVD-deposited PZT thin lms," Journal of Applied
Physics, vol. 69, no. 12, pp. 8352{8357, 1991.
34
[16] G. Gautschi, \Piezoelectric Sensors," in Piezoelectric Sensorics, pp. 73{91, 2002.
[17] M. J. Vellekoop, C. C. O. Visser, P. M. Sarro, and A. Venema, \Compatibility of
zinc oxide with silicon IC processing," Sensors and Actuators: A. Physical, vol. 23,
no. 1-3, pp. 1027{1030, 1990.
[18] L. P. Wang, R. Wolf, Q. Zhou, S. Trolier-McKinstry, and R. J. Davis, \Wet-etch Pat-
terning of Lead Zirconate Titanate (PZT) Thick Films for Microelectromechanical
Systems (MEMS) Applications," Materials Research, vol. 657, pp. 1{6, 2001.
[19] C. Soyer, E. Cattan, and D. Rmiens, \Electrical damage induced by reactive ion-
beam etching of lead-zirconate-titanate thin lms," Journal of Applied Physics,
vol. 97, no. 11, 2005.
[20] A. Cowen, PiezoMUMPs Design Handbook. MEMSCAP Inc., 2013.
[21] J. Weigold, T. Brosnihan, J. Bergeron, and X. Zhang, \A MEMS Condenser Micro-
phone for Consumer Applications," 19th IEEE International Conference on Micro
Electro Mechanical Systems, no. January, pp. 86{89, 2006.
[22] M. Goel, \Electret sensors, lters and MEMS devices: New challenges in materials
research," Current science, vol. 85, no. 4, pp. 443{453, 2003.
[23] R. J. Littrell, \High Performance Piezoelectric MEMS Microphones," PhD Thesis,
pp. 1{111, 2010.
[24] C. Multiphysics, \Structural Mechanics Module - User's Guide," tech. rep., 2013.
[25] P. L. Chen, R. S. Muller, and A. P. Andrews, \Integrated silicon pi-fet accelerometer
with proof mass," Sensors and Actuators, vol. 5, no. 2, pp. 119{126, 1984.
[26] H. Liu, C. J. Tay, C. Quan, T. Kobayashi, and C. Lee, \Piezoelectric MEMS energy
harvester for low-frequency vibrations with wideband operation range and steadily
increased output power," Journal of Microelectromechanical Systems, vol. 20, no. 5,
pp. 1131{1142, 2011.
[27] M. J. Madou, \Fundamentals of microfabrication and nanotechnology. Volume two:
Manufacturing techniques for microfabrication and nanotechnology," 2011.
35
Chapter 3
Acoustic Signal Classification with Pre-filtered Microphone
With the emergence of next-generation smart sensory systems, the integration of sensors,
hardware and software-enabled intelligence is becoming a focal point in the eort to in-
crease battery life while improving processing capabilities. Devices addressing important
applications such as remote sensing, internet-of-things, and wearable health monitoring,
have stringent requirements for long battery life, small form factors, always-on capabil-
ity, and robust performance. These are often con
icting objectives that are dicult to
achieve with reliance on purely digital processing methods. Algorithms delivering robust
performance and always-on capability are often computationally expensive and require
frequent recharging as a result. The use of bulky batteries to prevent frequent recharging
is not a solution for small form factor devices such as those required for wearable sensors.
In this chapter, we report the application of the fabricated MEMS resonant micro-
phone array for implementing and developing more ecient and noise-robust ltering,
machine learning, and signal classication tasks, especially for improving performance
in environments with high levels of acoustic interference. The experiments performed
in Chapter 4 validate these signal processing methods and compare performance to the
36
state-of-the-art through the applications of single-frequency communication channels with
single resonant microphone elements, automatic speech recognition (ASR) with resonant
microphone arrays, and the implementation of a wearable continuous lung-sound classi-
cation system with use of a low-frequency microphone array.
3.1 Background
Resonant acoustic transducers are commonly employed in applications such as acoustic
ranging [1] and ultrasonic imaging [2]. For pattern recognition applications, however, em-
phasis has typically been placed on sensors with a
at pressure-frequency response used in
conjunction with digital signal processing. These applications include speech processing,
acoustic leak detection, electronic health monitoring, and vehicle classication.
While digital ltering has traditionally been seen as superior to a analog signal pro-
cessing due to its versatility and cost eectiveness, analog ltering and signal conditioning
has been shown to have a number of advantages over a purely digital approach [3]. Dig-
ital signal processing is power hungry, making sensors relying only on digital processing
dicult to implement for low-power and self-powered applications. Additionally, digital
ltering suers from limitations in sampling signals with a large dynamic range in am-
plitude. In very low-SNR conditions, analog-to-digital converters may become saturated
and the features of interest may be lost or distorted before processing.
These limitations serve as motivation for the development of high quality factor acous-
tic sensors for signal recognition applications in noisy environments. In this chapter we
37
report experimental evidence for the eective use of resonant acoustic sensors as a me-
chanical method for improved noise ltering.
Signal processing systems similar to the one proposed in this chapter have been devel-
oped using MEMS resonators, analog resonant circuits, and analog active circuits. Arrays
of MEMS resonators have widely been adopted for RF wireless frontends, to be employed
as lter banks for RF ltering applications [4]. The natural signal processing capabilities
of the human cochlea have also served as an inspiration for distributed-frequency arrays
of mechanical sensors [5] and analog implementations of auditory models [6, 7, 8]. The
research presented in this chapter is similarly inspired by the human cochlea, and is the
rst demonstration of resonant mechanical sensor arrays as a hybrid signal processing
element in classication applications.
3.2 Cepstral Feature Extraction
In this research we focus on a dimensionality reduction technique called cepstral feature
extraction, which is commonly employed for spectral classication and detection tasks,
such as speech recognition. Automatic detection of health signals, such as the detection
of wheezing in lung sounds, which is emphasized in this thesis, has been shown to be
classied automatically with cepstral features with a high degree of accuracy [9, 10].
Cepstral features have also been demonstrated as eective features for detection of other
adventitious respiratory and pulmonary sounds, as well as classication of many other
non-medical audio signals.
The use of this feature extraction algorithm is evaluated in this study through the
38
classication of features and assessment for accuracy and computational complexity of
the implementation. A simplied comparison of the sequence of steps necessary for ex-
tracting the cepstral feature vector is shown for both
at-band and resonant microphone
approaches in Figure 3.1.
3.2.1 Conventional Cepstral Feature Extraction
Cepstral features are computed from a digitally sampled signal using the series of functions
summarized in Figure 3.1a. In the typical implementation, lung sounds are recorded with
a
at-band transducer and immediately digitized with an ADC. Signal conditioning is
then applied, typically in the form of a pre-emphasis lter to emphasize higher frequencies.
De-noising and removal of DC osets is also commonly applied at this stage. The audio
is then segmented into shorter frames and windowed using a Hamming window. Window
lengths are typically 50 ms in length, but for classication of lung sounds frames of
length 40 ms with a 50% overlap between consecutive frames have been determined to be
optimal [11].
The spectrum of the framed audio segment is then computed with a Discrete Fourier
Transform (DFT), which is the most computationally expensive operation in the feature
extraction sequence, with a time-complexity on the order of (
2
), where n is the number
of samples in each frame. The Fast Fourier Transform (FFT) is commonly implemented
as an alternative to reduce the complexity to (log
2
).
A dierence in the average audio spectra between frames containing wheezing and
normal breathing is illustrated in Figure 3.2a, showing the average spectral distribution
of 13 audio les obtained from the R.A.L.E. database with patients ranging from infants
39
Figure 3.1: Block diagram showing the necessary steps for the calculation of Cepstral
features with (a) a standard approach using a signal
at-band microphone with digital
ltering, and (b) an n-channel resonant microphone array with a pre-ltered input. By
using a pre-ltered microphone array, both discrete Fourier transform (DFT) and digital
ltering can be avoided to obtain an equivalent energy vector prior to taking the discrete
cosine transform (DCT). Analog-to-Digital Conversion (ADC) with n-channels can be
achieved with parallel ADCs or time-multiplexing.
40
Figure 3.2: Plots showing the (a) average spectral signature of frames containing wheez-
ing and containing normal breathing. Shaded regions represent the rst and the third
quartiles of the datasets, (b) the measured sensitivities of an optimized MEMS resonant
microphones array with 7 elements, and (c) the 7 overlapping digital triangular lters
with center frequencies matching the resonant microphone array.
41
to adolecsents. To estimate the energy distribution of the audio spectrum, a triangular
lter bank is applied in the frequency domain. The triangular lters are typically spaced
logarithmically according to the mel scale, which is in approximation of the linearity of
human perception of sound [12]. A common conversion formula to transform from Hertz
on a linear frequency scale to mels on the mel-scale is:
= 2595· log
10
(1 +

700
) (3.1)
When using a mel-spaced lter bank, the resulting features are referred to as mel-
Frequency Cepstral Coecients (MFCC). As perceptual linearity is not critical in the
analysis of lung sounds, Linear Frequency Cepstral Coecients (LFCC) can be imple-
mented with similar results (Figure 3.2c). The energy of each lter in the lter bank is
then computed through summation of the squares of each bin of the ltered spectrum.
The logarithm of the resulting energy vector is then passed through a Discrete Cosine
Transform (DCT) to convert the frequency-domain signal into the so-called cepstral do-
main. The rst element of the resulting cepstrum is typically discarded, and the remainder
of the elements form the cepstral feature vector, on which classication is performed.
In a typical speech recognition implementation, 20 to 30 lters are used, and only
the rst 13 cepstral coecients are retained for classication, as the cepstral features
become less statistically signicant with higher orders. The rst cepstral coecient is
also discarded in typical implementations, as it contains information about the loudness
of the audio signal, and classication is expected to be robust to variations in audio
intensity.
42
3.2.2 Modified Array-Based Feature Extraction
The standard algorithm must be modied to accommodate a pre-ltered array input,
with the modied sequence of feature extraction steps illustrated in Figure 3.1b. The
new algorithm necessitates a multi-channel frontend, consisting of either multiple parallel
ADCs or a single ADC that is shared by all channels through time multiplexing. The rele-
vant spectral content of low-frequency signals, such as many health signals and breathing
sound recordings, is limited to sub-kHz frequencies. For these applications sample rates
as low as 2 kS/s can be tolerated, and such low sample rates can easily accommodate
time multiplexing of an array of audio input channels with a single ADC.
A disadvantage of the array-based processing approach is that a larger number of
data buers are required to store the digitized audio data prior to feature extraction. For
this reason, array-based feature extraction of cepstral coecients is more demanding on
memory requirements than a standard approach. From a hardware perspective, a larger
memory buer is easier to implement and requires less power than increasing processing
capabilities [13, 14].
The use of a resonant microphone array for pre-ltering allows bypassing the costly
DFT (or FFT) and digital ltering steps in the calculation of cepstral features, since
ltering is already eectively completed in the array (Figure 3.1b). An energy vector
can be calculated from the time-domain signal of the pre-ltered array, that is equivalent
to the quantity calculated in the frequency domain through the standard process. This
43
time-domain energy vector is ideally identical in value to the frequency-domain energy
vector, as supported by Parseval's theorem for the discrete Fourier transform:
− 1
∑︁
=0
|[]|
2
=
1

− 1
∑︁
=0
|[]|
2
(3.2)
where [] is the DFT of [], both of length N. This relationship can be used to
evaluate an equivalent energy vector in the time domain and equate this energy through
a constant factor of N to that computed in the frequency domain, under the condition
that the mechanical resonant ltering eect is of an equivalent spectral shape to that of
the digital lter. By avoiding a DFT, we eectively reduce the computational complexity
of cepstral feature extraction from (
2
) to ().
Figure 3.3: Plot showing the in
uence of sample rate on computational complexity for
cepstral feature extraction with resonant array ltering, digital ltering with the use of
an FFT, and digital ltering with the use of a DFT.
44
Overall computational complexity is most dramatically in
uenced by the choice of
sampling rate and frame duration :

∝·

∝··
2
(·)

∝ (·)
2
(3.3)
Memory requirements are most signicantly in
uenced by the amount of memory
dedicated to the input buers following the ADC:

∝··

∝·
(3.4)
where is the number of lters, is the sample rate in samples per second, and is
the frame length in seconds. The in
uence of sample rate on computational complexity
is shown in Figure 3.3 for the three methods of cepstral feature extraction. The factors
in
uencing memory requirements are plotted in Figure 3.4. It is evident from these plots
that the use of a resonant microphone array oers a signicant improvement in process-
ing eciency, while requiring more memory to buer the additional channels for each
frame. The memory requirements of FFT and DFT implementations are approximately
equivalent.
A benet of exploiting the resonant response of the MEMS microphone design is that
the ltering occurs prior to digital signal processing, which reduces the appearance of
spectral distortion and artifacts during digital processing. In addition to the advantages
45
Figure 3.4: Plots showing the memory requirement relationship of digital ltering and
resonant array ltering in cepstral feature extraction with (a) variation in number of
lters, and (b) variation in sample rate.
46
to computation time and power consumption, the high-Q resonance of the microphone
array elements helps eliminate background noises, such as the heart beat, and pick up
weaker signals with a much higher sensitivity than commercial
at-band microphones.
This higher sensitivity also translates to a lower noise
oor for signals at the microphone's
resonant frequency, which improves relative signal quality and robustness to background
noise.
3.3 Classification Methods
Classication of MFCC features can be achieved using many methods. Speech recog-
nition is typically implemented with spectro-temporal features passed through a hidden
Markov model (HMM) [15]. The classication of lung sound recordings has most eec-
tively been implemented with Support Vector Machines (SVM) [16], Gaussian Mixture
Models (GMM) [17], k-Nearest Neighbor (kNN) [18], HMMs, or Articial Neural Net-
works (ANN) [9, 10, 19].
This study will assess the case of classication of speech with a typical HMM. Lung
sounds are primarily classied with features extracted from a single 40 ms frame, without
dependency on neighboring frames. This approach is selected in order to work within the
constraints of low power hardware, and because adventitious signatures in lung sounds
do not typically vary signicantly from frame to frame. For comparison to the spectro-
temporal classication methods implemented in literature, the use of neural networks
using multiple frames of respiratory audio features was also evaluated.
47
To evaluate the performance and provide a point of comparison of the proposed feature
extraction algorithms in ultra low power hardware, we selected to use a computationally
inexpensive Naive Bayes classier. The classier seeks to maximize the conditional prob-
ability for a feature vector

= (
1
,
2
...
3
) as follows:
(|
1
,
2
...

) =
()
∏︀

=1
(

|)
(
1
,
2
...

)
(3.5)
(|
1
,
2
...

) is the posterior probability of the class given a predictor vector

. () is the probability of observance of class , which is xed at 0.5 for the binary
classication implemented in this research, but can be adjusted in practice to reduce false-
positives. (

|) is the likelihood of observing the feature

given class , which is
given by the Gaussian distribution for each combination of features and classes obtained
through the training process. (
1
,
2
...

) is the predictor prior probability, which is
the probability of observing the feature vector independent of class. Since this probability
is identical for all classes, it does not need to be calculated to determine which class
probability is greater.
Although classication with a Naive Bayes model is not optimal for maximizing ac-
curacy, it can be more easily implemented within the memory and processing constraints
of low-power embedded processors.
48
3.4 Summary
A signal processing front-end was developed to take advantage of a resonant microphone
array transducer as both a processing and acquisition element in acoustic signal classi-
cation applications. We demonstrated that the algorithm is computationally ecient,
and in subsequent chapters the classication accuracy and computational eciency is
experimentally validated.
References
[1] M. I. Haller and B. T. Khuri-yakub, \Ultrasonic Air Transducer," October, vol. 43,
no. 1, 1996.
[2] T. L. Szabo, Diagnostic ultrasound imaging: inside out. Academic Press, 2004.
[3] R. Sarpeshkar, C. Salthouse, J. J. Sit, M. W. Baker, S. M. Zhak, T. K. Lu, L. Turic-
chia, and S. Balster, \An ultra-low-power programmable analog bionic ear proces-
sor," IEEE Transactions on Biomedical Engineering, vol. 52, no. 4, pp. 711{727,
2005.
[4] G. Piazza, P. J. Stephanou, and A. P. Pisano, \One and two port piezoelectric higher
order contour-mode MEMS resonators for mechanical signal processing," Solid-State
Electronics, vol. 51, no. 11-12, pp. 1596{1608, 2007.
[5] R. D. White and K. Grosh, \Design and Characterization of a MEMS Piezoresis-
tive Cochlear-like Acoustic Sensor," in ASME International Mechanical Engineering
Congress and Exposition, vol. 2002, pp. 1{10, 2002.
[6] M. Kucic, P. Hasler, J. Dugger, and D. Anderson, \Programmable and adaptive
analog lters using arrays of
oating-gate circuits," in Proceedings - 2001 Conference
on Advanced Research in VLSI, ARVLSI 2001, pp. 148{162, 2001.
[7] R. F. Lyon, \Analog Implementations of Auditory Models," 1991.
[8] R. F. Lyon and C. Mead, \An analog electronic cochlea," IEEE Transactions on
Acoustics, Speech, and Signal Processing, vol. 36, no. 7, pp. 1119{1134, 1988.
[9] R. X. A. Pramono, S. Bowyer, and E. Rodriguez-Villegas, \Automatic adventitious
respiratory sound analysis: A systematic review," PLoS ONE, vol. 12, no. 5, 2017.
49
[10] R. Palaniappan, K. Sundaraj, and N. U. Ahamed, \Machine learning in lung sound
analysis: A systematic review," 2013.
[11] N. Sengupta, M. Sahidullah, and G. Saha, \Lung sound classication using cepstral-
based statistical features," Computers in Biology and Medicine, vol. 75, pp. 118{129,
2016.
[12] S. S. Stevens, J. Volkmann, and E. B. Newman, \A Scale for the Measurement of
the Psychological Magnitu Pitch," The Journal of the Acoustical Society of America,
vol. 8, no. 3, pp. 185{190, 1937.
[13] T. Yemliha, Performance and Memory Space Optimizations for Embedded Systems.
PhD thesis, Syracuse University, 2011.
[14] P. Juang, H. Oki, Y. Wang, M. Martonosi, P. Peh Li-Shiuan, and D. Rubenstein,
\Energy-Ecient Computing for Wildlife Tracking: Design Tradeos and Early Ex-
periences with ZebraNet," Proceedings of the 10th International Conference on Ar-
chitectural Support for Programming Languages and Operating Systems (ASPLOS
2002), pp. 96{107, 2002.
[15] L. Rabiner and B.-H. Juang, \Fundamentals of speech recognition," Signal Process-
ing Series. Prentice Hall, Englewood Cliffs, NJ, 1993.
[16] I. Steinwart, Support Vector Machines, vol. 13. 2010.
[17] M. Bahoura and C. Pelletier, \Respiratory sounds classication using Gaussian mix-
ture models," Electrical and Computer Engineering, 2004. Canadian Conference on,
vol. 3, no. 2, pp. 1309{1312, 2004.
[18] C. H. Chen, W. T. Huang, T. H. Tan, C. C. Chang, and Y. J. Chang, \Using K-
nearest neighbor classication to diagnose abnormal lung sounds," Sensors (Switzer-
land), vol. 15, no. 6, pp. 13132{13158, 2015.
[19] I. Ozer, Z. Ozer, and O. Findik, \Noise robust sound event classication with con-
volutional neural network," Neurocomputing, vol. 272, pp. 505{512, 2018.
50
Chapter 4
Acoustic Signal Classification Experiments
Classication experiments were performed, using the algorithms outlined in Chapter 3, to
assess the eectiveness of resonant-array-based processing in applications where interfer-
ence from background noise is dicult using a traditional approach. The applications that
served as a demonstration platform in this chapter are acoustic digital communication
with a single-frequency, automatic speech recognition, and detection and classication of
respiratory sounds.
4.1 Resonant Microphone Digital Communication Link
Experiments
Preliminary experiments using individual resonant microphone elements for the sensing
of single-tone audio signals were performed. These experiments served as a method of
independently assessing the pattern recognition performance of a resonant microphone,
without the additional variables present in more complex signal recognition applications.
51
Acoustic digital communication links are a relatively new eld of research, with ap-
plications in intra-body communication networks [1], undersea communication [2], and
on-chip short-range sonic communication [3].
4.1.1 Experimental Setup
These experiments were performed in an anechoic chamber containing a speaker, single
resonant MEMS microphone, and LabVIEW control system, with experimental conditions
and setup similar to that shown in Figure 2.6.
The input signal used was a sequence of randomly-generated bits that was amplitude-
modulated with a carrier frequency corresponding to the resonant frequency of the indi-
vidual sensing element (860 Hz and 6.3 kHz). Sinusoidal noise was combined additively
with the signal at a SNR ranging from -72 dB to -40 dB. The microphone output is
digitized and demodulated using LabVIEW's Modulation Toolkit. For each SNR point,
800 bits are transmitted, received, and demodulated to obtain a Bit-Error Rate (BER).
4.1.2 Results
Preliminary experiments with a pulse-width-modulated signal that is buried under noise
(15 dB higher than the signal, i.e., -15 dB SNR) show that both the microphone array
(without any ltering) and the reference microphone (with digital ltering) can recover
the signal (Figure 4.1). However, it was found that the advantages of the resonant
microphone array are limited in white noise conditions, as noise at the resonant frequency
has signicant contributions in both cases.
52
Figure 4.1: (a) Waveform graph of pulse-width-modulated input signal. (b) Signal with
-15dB SNR additive white Gaussian noise. (c) Measured MEMS resonant microphone
electrical output. (d) Measured reference microphone output with digital band-pass post-
ltering.
53
For this reason we investigate the eect of sinusoidal noise at a single frequency.
Although not a generally realistic environmental condition, this noise prole mimics en-
vironments where noise is dominant in a limited bandwidth. In practice, if the typical
spectral conditions of the noise in a given environment are known, microphone resonances
can be designed to capture signals with minimal acoustic interference.
The results are plotted in Figure 4.2, and two results are shown. In the rst case,
detection of a low-frequency signal (860 Hz) with high-frequency noise (2 kHz) was per-
formed using the microphone #1 of the resonant microphone array (of which the resonant
frequency is 860 Hz). The MEMS microphone sensitivity at the noise frequency of 2 kHz
was 0.66 mV/Pa, which is 0.3% of the sensitivity at resonance. In the second case, detec-
tion of high-frequency signals (6.3 kHz) with low-frequency noise (300 Hz) was performed
using microphone #13 (of which the resonant frequency is 6.3 kHz). The MEMS can-
tilever sensitivity at 300 Hz is 0.14 mV/Pa, 1.3% of the sensitivity at resonance. These
results show a signicant decrease in error rate as a result of using a highly resonant
acoustic sensor rather than a wide-band microphone.
We see that in both of these cases, sensitivity is quite low for the noise frequency.
BER is nearly zero down to about -70 dB SNR, where it sharply saturates to nearly 0.5,
which is the probability of random guessing. For situations where noise frequency is much
closer to the resonant frequency of the microphone, we expect to see an increase in the
minimum SNR required for reliable demodulation.
In Figure 4.3, we see that this is indeed the case. As noise frequency approaches the
resonant frequency, the sensitivity to the noise increases and thus the BER performance
54
Figure 4.2: (a) Measured bit error rate for 860 Hz signal with 2 kHz sinusoidal noise. (b)
Measured bit-error rate for 6.3 kHz signal with 300 Hz sinusoidal noise.
degrades. However, we see that even with only about a 2.4% oset between noise fre-
quency and carrier frequency, SNRs of about -15 dB can be achieved before BER reaches
25%. Thus, resonant microphones with high quality factors can eectively recover signals
buried under noise that has spectral components very close to the center frequency of the
microphone.
4.2 Speech Recognition Experiments
The results of the digital acoustic communication experiment were extended to the more
widely-researched and commercially-signicant application of Automatic-Speech Recog-
nition (ASR). Speech recognition in noisy environments has been a major area of research,
with the vast majority of proposed solutions focusing on noise-robust digital models [4]
and digital speech enhancement and de-noising [5]. In this section, we compared a con-
ventional ASR algorithm, with an algorithm integrating a resonant microphone array
frontend.
55
Figure 4.3: Experimental measurements of minimum SNR such that BER < 25% for a
microphone element with 3 kHz resonant frequency. The x-axis plots (a) the ratio of
the electrical response of the microphone at the noise frequency to that at the resonant
frequency and (b) the absolute value of the dierence between noise frequency and signal
frequency
56
4.2.1 Experimental Setup
Figure 2.6 illustrates the experimental setup for testing of the microphone array in acous-
tic signature detection applications, which is identical to the setup used for microphone
characterization. The MEMS microphone array and pre-amplier circuit are contained
within an aluminum electromagnetic shielding box. A commercial speaker was used as
the signal source, while a calibrated precision microphone with
at response served as
a reference. A pre-amplier circuit, based on the LM386, amplied the reference mi-
crophone output to an amplitude comparable to that of the resonant microphone array
in order to ensure equivalent sampling resolution conditions. This entire assembly was
placed within an anechoic chamber for acoustic isolation during testing.
A ROGA 2-channel USB data acquisition system was used with a sampling resolution
of 305 V at a sampling rate of 16 kHz. A PC was used with the NI LabVIEW develop-
ment environment for data acquisition, experiment automation, and signal analysis.
Speech processing was performed using the CMU Sphinx4 toolkit [6] with a custom
front-end based on Linear-Frequency Cepstral Coecients (LFCC). To compute LFCC
features, rst the speech data is divided into 25 ms frames with 10 ms separation between
data blocks. Several lters are applied for pre-emphasis and windowing. This is followed
by a Fast Fourier Transform (FFT) for each data block to obtain a normalized magnitude
frequency response.
57
4.2.2 Results
The subsequent step of applying bandpass ltering was performed dierently for reference
microphone data and microphone array speech output. Thirteen triangular lter banks
were used for the reference microphone. Each of these lters had a quality factor of 43.1
and linearly spaced center frequencies were placed to coincide with microphone array
resonances. Microphone array data was processed with thirteen rectangular lter banks
to isolate resonance peaks from one another. The rectangular lter bandwidth coincided
with the base dimension of the triangular lter banks used for the reference microphone.
In both cases, the energy coecients for each of the thirteen lter banks undergo a
discrete cosine transform to obtain a 13-element feature vector to be directly used for
ASR decoding.
The microphone input consisted of 140 alphanumeric speech recordings from the AN4
database. This database was used due to its reduced vocabulary set, which allowed for
large variations in recognition accuracy over a range of wide range of SNR conditions. A
400 Hz sinusoidal noise was additively combined with the speech data at SNRs ranging
from -15 dB to 25 dB. The speech output at both microphones was processed using
the Sphinx toolkit to obtain decoded transcriptions, which were then compared to the
original speech text to obtain a Word-Error Rate (WER). Figure 4.4 shows the average
WER of the resulting transcriptions as a function of the SNR. These results demonstrate
that speech acquired using the resonant microphone array is relatively unaected by large
levels of out-of-band acoustic noise, while a typical
at-response microphone relying on
digital band-pass ltering has degrading performance under high levels of acoustic noise.
58
Figure 4.4: Plot of measured word-error rates (WERs) of the acoustic resonator array
and the reference microphone used to process automatic speech recognition with various
levels of 400 Hz sinusoidal noise.
59
Figure 4.5: Spectrogram of the utterance "fty-two nineteen" in the case of (a) a clean
audio le without any ltering, (b) the reference microphone for -10 dB applied noise,
and (c) a summation of the outputs from 13 microphone array elements in -10 dB noise.
It is hypothesized that the improvement in recognition accuracy of the microphone
array is related to the limitations in dynamic range of the analog-to-digital conversion
process. As out-of-band noise amplitude increases, the signal becomes distorted at all
frequencies for a wideband microphone (Figure 4.5b). As a result, digital ltering is
unable to accurately recover the original signal. In contrast, the signal output of the high-
quality-factor MEMS microphone array does not experience any signicant distortion
(Figure 4.5c) due to environmental noise.
In addition to these benets, the use of a microphone array oers the potential for
higher Q-factor ltering, which is limited by sampling rate and window-length in the
60
digital domain. Due to the reduced digital processing requirements, it is also expected
that the use of a microphone array will increase computation speed and reduce memory
footprint, both of which are especially evident in embedded processing applications.
4.3 Lung Sound Recognition Experiments
The use of a resonant microphone array is well-suited for the application of wearable
health sensing in the form of a wearable, low-power, and noise-robust stethoscope. For a
robust wearable stethoscope system, it is required to have immunity to ambient external
noise sources as well as noise sources originating within the wearer's body. Additionally,
the use of a resonant microphone addresses some of the diculties with obtaining good
coupling to the skin, since high sensitivities will make losses due to poor coupling more
tolerable.
Other MEMS devices, such as contact accelerometers [7], have been proposed for
wearable lung and heart auscultation sensors. Such devices suer from many of the same
problems that exit in electronic stethoscopes, namely low sensitivity and high process-
ing power. Contact accelerometer devices partially address this limitation through the
improvement in acoustic coupling to the skin, however the sensitivity of the device suf-
fers since it operates at sub-resonant frequencies. Such devices are also susceptible to
interferences that are unrelated to the health signals of interest, such as motion of the
skin.
61
4.3.1 Experimental Setup
In the experimental testing of classication performance (process shown in Figure 4.6a),
a PC control program is implemented with the NI Labview development environment for
data acquisition, experiment automation, and signal analysis. A ROGA 2-channel USB
data acquisition ADC digitizes the input with a xed sampling resolution of 305 V.
Signals were resampled to 5120 S/s such that classication accuracies are representative
of low-power hardware implementations of the algorithm.
Figure 4.6: Block diagram showing the experimental testing process for evaluation of
lung sound classication, comparing resonant microphone array processing and
at-band
processing.
Experiments were performed in an anechoic chamber to limit re
ections and external
interference during measurement. The chamber contains two mid-bass full range speakers,
with each speaker serving as a dedicated channel; one for playing the breathing audio
and the other for background noise. A GRS 4PF-8 4" speaker served as the signal source,
62
with a power capacity of 40 W RMS and sensitivity of 84 dB 1W/1m. A Galaxy Audio
S5N-8 5" speaker served as the noise source, with a power capacity of 100 W RMS and
sensitivity of 91 dB 1W/1m. The use of high-power and high-sensitivity speakers allowed
us to sweep through a wide range of signal-to-noise ratios (SNRs) while ensuring linearity
and good dynamic range.
Breathing audio recordings are taken from the R.A.L.E. database [8], which contains
examples of normal breathing and wheezing for infants, children and adults. In total, 5
audio les of normal breathing and 8 wheezing les are used, with 2,707 frames (108 sec-
onds) labeled normal breathing and 1,057 frames (42 seconds) labeled as wheezing. An
Onkyo A-9050 stereo amplier was the interface between the PC control and speakers to
ensure that minimal unintended noise is introduced to the measurement.
To standardize the in
uence of noise
oor, the breathing signal volume was xed
throughout the experiment while noise intensity was swept from minimum to maximum
volume. Average signal intensity and noise intensity were characterized to obtain an
estimated SNR for each noise volume combination.
Filter Distribution Selection
The location of spectral lters used in this study was optimized through analysis of the
spectral distribution of wheezing within respiratory recordings of the R.A.L.E. reposi-
tory. The average power spectral densities of all frames used in this study are plotted
in Figure 4.7. These plots indicate that the majority of distinguishing characteristics of
wheezing occur between 200 Hz to 400 Hz, and 600 Hz to 900 Hz.
63
Figure 4.7: Plots of average spectral distribution of frames containing wheezing (shown in
red) and normal breathing (shown in blue). The averages are plotted for (a) all recordings,
(b) only recordings of infant breathing, (c) only recordings of children's breathing, and
(c) only recordings of adult breathing. Dotted lines show spectral distribution for an
arbitrarily selected single frame from each category, and shaded regions are bounded by
the 25th and 75th percentiles of the data distribution.
64
An iterative approach was used to optimize the distribution of selected lters. The
optimization sought to minimize the number of lters, while maximizing the average
distance between calculated cepstral features of wheezing and normal breathing. We used
the Kolmogorov-Smirnov distance metric [9] to quantify the dierence in distributions
between calculated cepstral features. The Kolmogorov-Smirnov distance is dened as the
maximum distance between two empirical distribution functions:

,
= sup

|
1,
()−
2,
()| (4.1)
where sup is the supremum function, and
1,
and
2,
are empirical distribution func-
tions dened to be:

() =
number of elements in the sample≤

=
1

∑︁
=1

[−∞ ,]
(

) (4.2)
where

is the set of observations, and
[−∞ ,]
(

) is the indicator function, equal to
1 if

≤ , and equal to 0 otherwise.
Two sets of lters are selected for classication experiments, and summarized in Ta-
ble 4.1. The optimal selection of lter locations with a minimum quantity was determined
to contain 7 lters ranging from 175 Hz to 700 Hz. Additionally, to provide a point of
reference to typical algorithms presented in literature, we implemented a digital lter
bank with 13 overlapping mel-spaced lters, with center frequencies ranging from 100 to
1000. The lter positions and values selected for this study are illustrated in Figure 4.8.
The fabricated microphone array spans the audible spectrum with a higher resonance
65
Figure 4.8: Plots showing the selection of lter-banks used in these studies, including (a)
the measured sensitivities of 7 MEMS resonant microphones, which are scaled articially
to occupy the optimized lter positions; (b) the overlapping digital triangular lters
designed to match the optimized center frequencies of the resonant microphone array;
and (c) the 13 mel-Spaced triangular lters which give maximum classication accuracy
for this application using a digital processing approach.
66
7 Filters 13 Mel-Spaced
Filters
1 175 Hz 100 Hz
2 275 Hz 152 Hz
3 359 Hz 207 Hz
4 450 Hz 266 Hz
5 550 Hz 329 Hz
6 625 Hz 395 Hz
7 700 Hz 466 Hz
8 - 542 Hz
9 - 622 Hz
10 - 708 Hz
11 - 799 Hz
12 - 897 Hz
13 - 1000 Hz
Distance 0.329 0.204
Accuracy 85.5% 88.7%
Table 4.1: Summary of center frequencies of selected lters for classication experiments,
and associated average Kolmogorov-Smirnov distance of the feature set and GMM clas-
sication accuracy with a digital implementation.
frequency than the required range. To evaluate the processing concept and devices despite
this limitation, a tempo-scaling factor in frequency was applied to perform acquisition of
breathing data with the array. By repeating the experiment for each channel and scaling
the playback tempo of the audio, we simulated the eect of recording with multiple
resonant microphones, eectively forming the lter bank in Figure 4.8a.
Cross Validation Methodology
Several cross validation methods were assessed for these studies. All experiments were
performed using a distribution of 70% of frames randomly assigned to the training set,
and the remaining 30% of frames assigned to the testing set. The random seed used for
67
distribution of frames is kept constant such that the same training and testing sets are
chosen between noise cases and processing methods, so that equal comparisons can be
made.
We also studied the eect of Leave-One-Out Cross Validation (LOOCV) as a cross
validation methodology [10], and presented the results. It was observed that the addi-
tional training data available for LOOCV did not signicantly impact the results, and
the observed trends were consistent between methodologies.
4.3.2 Results
We evaluated the classication accuracy of feature extraction with the selected lter con-
gurations under varying degrees of noise and with the selected classication algorithms
(GMM, SVM, kNN, and NB), implemented with the Scikit library [11] in python.
The algorithms were evaluated under 3 noise conditions: single-tone 60 Hz noise,
single-tone 1.2 kHz noise, and heart-beat noise, which is the most common noise source
in clinical cases [12] (Figure 4.9).
68
Figure 4.9: Spectrograms of recorded breathing, obtained with the summed 7 channels of
the resonant microphone array (left column) and with the single-channel
at-band micro-
phone (right column). From top to bottom the noise sources are 60 Hz monotonic noise,
1200 Hz monotonic noise, and heartbeat noise, all with noise intensity of -26 dB SNR.
Larger intensity of the recording appears with a brighter color.
69
The resulting classication accuracies are evaluated on the PC through a leave-one-
out cross-validation (LOOCV) testing technique, and the results are plotted for both
digital ltering approaches with a
at-band microphone and array-based processing with
the resonant microphone array. Although it is the most commonly reported metric, accu-
racy is not suitable for this application and dataset due to the imbalanced classication
problem. Since wheezing is underrepresented in the R.A.L.E. dataset (present in about
29% of the data), we use the F1 Score metric as a more appropriate representations of
true detection accuracy. F1 is dened as:

1
= 2· ·
+
=
2·
2· + +
(4.3)
where TP, FN, and FP represent the number of true positives, false negatives, and false
positives, respectively.
The classication metrics are plotted as a function of SNR for the case of a low-
frequency 60 Hz monotonic background signal in Figure 4.10. Monotone 60 Hz noise
is a common noise prole in everyday life, and more generally gives insight into the
in
uence of noise sources with dominant frequencies lower than that of wheezing. The
gures indicate that although digital ltering can achieve a comparable performance in
high SNR environments (i.e., low levels of background noise), the performance of digital
ltering sharply degrades with increased noise. The reason is particularly evident in the
spectrograms of the measured signal in Figure 4.9, where there is a clear presence of
harmonics and other forms of distortion due to low frequency background noise when
recording with a
at-band microphone, while this is not present in the array signal.
70
Figure 4.10: Measured classication accuracy (F1 score) of wheezing detection for varying
levels of background noise intensity with a 60 Hz noise source. Cepstral features used
in classication are computed with 7 resonant array lters (red), 7 digital array lter
(black, dotted), and 13 mel-spaced digital lters (black, solid). Features are classied
using a Gaussian Mixture Models, k-Nearest Neighbors, Support Vector Machines, and
Naive Bayesian classication.
71
The results also show that to achieve a comparable classication accuracy with digital
ltering, almost twice as many lters are necessary. When only 7 lters are used in the
frequency range where wheezing is present, the performance even without background
noise is reduced relative to array-based signal processing.
Figure 4.11: Measured classication accuracy (F1 score) of wheezing detection for varying
levels of background noise intensity with a 1200 Hz noise source. Cepstral features used
in classication are computed with 7 resonant array lters (red), 7 digital array lter
(black, dotted), and 13 mel-spaced digital lters (black, solid). Features are classied
using a Gaussian Mixture Models, k-Nearest Neighbors, Support Vector Machines, and
Naive Bayesian classication.
The accuracy for the case of high frequency noise (1200 Hz) was also evaluated and
72
plotted in Figure 4.11. For this noise case, we can see that classication accuracy us-
ing the array method is maintained until about an SNR of -10 dB, at which point the
high frequency signal can no longer be eectively removed. In this case, digital ltering
presents an advantage, since a high frequency noise source will not distort the signals of
interest, and digital ltering can completely remove the noise from the spectrum. This
arises because a mechanical resonator will still have a certain sensitivity at high frequen-
cies, which may be greater than desired due to higher order vibration modes. A digital
low pass lter can be applied to the pre-ltered signals to remove high frequency noise,
however this addition will reduce some of the computational performance benets of using
the resonant array approach. Regardless, even with these limitations, the array approach
is still an advantageous design when compared to a low-dimension digital lter array with
7 digital lters.
The most common and dominant type of acoustic noise encountered in lung aus-
cultation is that coming from the heart beat [12]. This interference is more prominent
when monitoring bronchial lung sounds rather than tracheal lung sounds, making chest-
mounted breathing analyzers dicult to implement. These noises are particularly trou-
blesome since the intervals of occurrence, intensity, and spectral location is comparable
to that of lung sounds, and they are particularly dicult to lter out, digitally or other-
wise. However, since a majority of the spectral content of heart noises occurs below the
fundamental frequency of wheezes, this interference can be eectively ltered by using a
resonant microphone array. This is supported in Figure 4.9, where it is evident that the
performance of classication with digital lters degrades at a faster rate.
The results of classication with interference from heart noise is plotted in Figure 4.12.
73
Figure 4.12: Measured classication accuracy (F1 score) of wheezing detection for varying
levels of background noise intensity with a heart noise source. Cepstral features used in
classication are computed with 7 resonant array lters (red), 7 digital array lter (black,
dotted), and 13 mel-spaced digital lters (black, solid). Features are classied using a
Gaussian Mixture Models, k-Nearest Neighbors, Support Vector Machines, and Naive
Bayesian classication.
74
Interestingly, the use of 7 digital lters, although showing a worse performance overall,
degrades with noise intensity slower than the 13 mel-Spaced digital lters. This is likely
due to the placement of the lters only in the portion of the spectrum where wheezing
shows a unique spectral signature, and thus avoiding portions of the spectrum where
heart noise may be more dominant.
Similar experiments were performed with other noise sources, including Gaussian
white noise and realistic "crowd" noise taken from recordings in a busy restaurant. Clas-
sication with the presence of these noise sources performed poorly with all processing
methods, with F1 score dropping below 10% for 0 dB SNR. It is hypothesized that these
noise sources are the most troublesome because they occupy the band of frequencies where
wheezing and breathing sounds are most prevalent, and thus spectral ltering techniques
are not sucient for separating these noise sources.
4.4 Neural Network Recognition Experiments
Thus far we have only discussed the case of classication of single frames independently
of neighboring frames. This is done to work within the constraints of low power hard-
ware, and because wheeze signatures do not typically vary signicantly from frame to
frame. However, it has been shown in literature that the use of spectro-temporal features
classied with neural networks can provide a potentially greater accuracy and robustness
to noise [13, 14, 15].
Analysis of multiple sequential frames could be achieved with classication using
variations of convolutional neural networks (CNN), time-delay neural networks (TDNN),
75
or recurrent neural networks (RNN). In this section we present experiments comparing the
accuracy of the resonant array-based cepstral features with conventional cepstral features
classied with a fully-connected time-delay neural network structure.
4.4.1 Experimental Setup
The neural network classication experiments are implemented in python using the Keras
library [16], with the network structure shown in Figure 4.13. The features used as input
to the neural network are recorded and computed using the same 3 congurations that
were presented in section 4.3, specically a resonant array-based method using 7 opti-
mally placed lters, the conventional digital method using triangular lters at the same
7 locations, and a conventional digital method using 13 overlapping triangular lters
placed linearly on the mel-scale.
The neural network is organized as a fully-connected Multilayer Perceptron (MLP)
feedforward network, consisting of an input layer, hidden layer, and output layer. The
input to the network is a distribution of feature vectors, each of size , with total
frames, giving total input nodes. Each subsequent frame is delayed by a time ,
which is set to 20 ms such that we have a 50% overlap in subsequent frames, each of
which have a duration of 40 ms.
The hidden layer is composed of

2
nodes, and scaled to follow the rule of thumb
of having the number of hidden nodes between the input and output layers to be the
average of the number of input nodes and number of output nodes. This is somewhat
of an arbitrary choice, and the number of hidden nodes and layers can be optimized for
76
Figure 4.13: Diagram showing experimental conguration for construction and testing of
a fully-connected time-delay neural network, with a variable number of features and
frames .
a xed set of input frames, but a standardized scaling of hidden layers is necessary to
study the eect with varying the number of input frames.
In this study, the hidden layer activation function, which is the nonlinear function
applied to the weighted sum of inputs at each node, is chosen to be the Rectier Linear
Unit (ReLU) function, dened as:
() =
+
=(0,) (4.4)
77
where the max function selects the greater of the two values, and x is the weighted linear
summation of inputs to the hidden node:
=

∑︁
=0

(4.5)
As this is a binary classication of breathing into wheezing or non-wheezing classes,
the output takes a single binary value between 1 and 0, respectively. The output activation
function, which is the nal nonlinear function applied to the sum of weighted inputs
to the output node, is implemented with the sigmoid function. The sigmoid function
is commonly used at the output layer of binary classication models which implement
cross-entropy as the loss function [17], as we do. It is dened as:
() =
1
1 +
−
(4.6)
The nal output is determined by rounding the output to the nearest integer. Alter-
natively, two output nodes could have been selected, each representing a class (wheezing
or normal breathing), and with the larger probability at the end of classication chosen to
be the predicted output class. Since the application is a simplied binary class problem
only a single node was necessary and evaluated, but a multiple output node approach
would be necessary with a multi-class extension of this system.
The model is trained using the error backpropagation algorithm [18] with Keras' de-
fault implementation of the Adam optimizer [19], and a binary crossentropy loss function.
Training was completed, as in earlier sections of this chapter, with a random 60%/40%
78
division of training data to testing data. Additional practical implementation details are
covered in Appendix A.
4.4.2 Results
The network is optimized for training duration (number of cycles) and number of consec-
utive input feature frames to use, and classied for accuracy with varying intensities of
heart noise. To characterize the optimal number of consecutive frames, we classied with
a neural network trained for 50 epochs with the number of consecutive frames sweeped
from 1 frame (40 ms) to 100 frames (2.02 s). A single frame's occurrence of wheezing
within the input sequence was classied as wheezing for the entire sequence at the output.
The F1 score of the classied test set was plotted against the number of applied frames
for data with 35 dB SNR (Figure 4.14) and for data with -26 dB SNR (Figure 4.15).
For the low-noise 35 dB SNR test data set, the accuracy (F1 score) consistently
improves for all 3 methods, with some noise in the accuracy which most likely arises due
to the randomized distribution of test and train data. For the large-noise -26 dB SNR test
data set, it can be seen that there is an optimal number of consecutive frames of about 15
for the 7 digital lters conguration. The resonant array feature extractor also performed
better than even the 13 mel-spaced digital lters, until about 90 consecutive frames are
used for classication. It appears that the resonant array classication accuracy reaches a
saturation value with about 50 consecutive frames (1.02 s). A trade-o of increasing the
number of consecutive frames is the much greater hardware requirement for processing,
storage of weights, and storage of input data.
79
Figure 4.14: Plot showing classication accuracy (F1 score) of neural network as a func-
tion of the number of consecutive 50% overlapping 40 ms frames used in the input. Clas-
sication of wheezing is performed on a test data set of respiratory sounds containing
35 dB SNR heart noise.
80
Figure 4.15: Plot showing classication accuracy (F1 score) of neural network as a func-
tion of the number of consecutive 50% overlapping 40 ms frames used in the input. Clas-
sication of wheezing is performed on a test data set of respiratory sounds containing
-26 dB SNR heart noise.
81
A challenge in implementing noise-robust neural networks is overtting the network to
the training data set and hampering the ability of the network to generalize to variations
in patients and environmental noise conditions. Including noisy data in the training set
is a poor solution because the actual characteristics of noise encountered in practice will
dier from the training environment. In this study, the optimal performance was achieved
by training with a clean data set, although methods exist for improving noise robustness
through noise-aware training [20] and recurrent neural networks [21].
Figure 4.16 shows the eect of training cycles on the resulting F1 score, with a xed
25 input frames (320 ms), and indicates that the model performance saturates after
about 40 epochs. Training is completed in batch sizes of 15, and each batch of training is
referred to as an epoch. Our objective is to nd an optimal number of epochs such that
overtting does not impact the performance at high levels of background interference.
Figure 4.17 shows the eect of training epochs on the resulting classication accuracy
with -10 dB SNR heart noise test data. Accuracy of classifying features extracted with
the resonant array method reaches a maximum after about 25 epochs, and begins to
decrease after about 50 epochs. The eect of overtting is less evident with a larger
number of input frames.
The F1 scores of the classication of respiratory sounds are plotted as a function of
heart interference intensity for models composed of 25 input frames (Figure 4.18), 50
input frames (Figure 4.19), and 100 input frames (Figure 4.20). In all cases, the use of
a resonant array to compute cepstral features shows relative robustness to heart noise,
even with many mel-spaced digital lters performing better in low-noise conditions. This
advantage in noise-robustness appeared more pronounced with a greater number of input
82
Figure 4.16: Plot showing classication accuracy (F1 score) of neural network as a func-
tion of the number of training epochs with a xed 25 input frames. Classication of
wheezing is performed on a test data set of respiratory sounds containing 35 dB SNR
heart noise.
83
Figure 4.17: Plot showing classication accuracy (F1 score) of neural network as a func-
tion of the number of training epochs with a xed 25 input frames. Classication of
wheezing is performed on a test data set of respiratory sounds containing -10 dB SNR
heart noise.
84
Figure 4.18: Plot showing classication accuracy (F1 score) of neural network as a func-
tion of heart noise intensity, with a model congured for 25 input frames (320 ms) and
training for 50 epochs.
85
Figure 4.19: Plot showing classication accuracy (F1 score) of neural network as a func-
tion of heart noise intensity, with a model congured for 50 input frames (1.02 s) and
training for 50 epochs.
86
Figure 4.20: Plot showing classication accuracy (F1 score) of neural network as a func-
tion of heart noise intensity, with a model congured for 100 input frames (2.02 s) and
training for 50 epochs.
87
frames. With the use of 100 consecutive input frames, the classication accuracy of
respiratory sound classication with -26 dB SNR heart noise was 58.8% greater than
with the equivalent 7 digital lter implementation, and 12.1% greater than with the 13
mel-spaced digital ltering implementation.
4.5 Summary
Preliminary experiments were performed for single resonant microphone communication
experiments, and it was demonstrated that the use of a resonant transducer was robust
to out-of-band noise up to signal-to-noise ratios of about -70 dB SNR, while
at band
microphones with digital ltering began to degrade in performance with additive out-of-
band noise with lower levels of noise.
Experiments were performed for automatic speech recognition with a resonant micro-
phone array, with 13 resonant elements spaced linearly in the frequency spectrum between
860 Hz and 6000 Hz, and it was similarly observed that the use of resonant array ltering
was relatively immune to out-of-band noise relative to a
at-band microphone using a
standard ASR front-end.
We similarly performed classication of respiratory sound recordings with both reso-
nant array front-ends and
atband front-ends for cepstral feature extraction. It was ob-
served through classication of these features with GMM, k-NN, SVM, Naive Bayes, and
time-delay neural networks that there is a consistent robustness to low-frequency, high-
frequency, and heart beat noises across classication algorithms. The resonant transducer
88
performs best when ltering out-of-band noise, and the recognition accuracy in white-
noise environments could not be improved by using this method.
The performed comparative experiments make a strong case for the use of resonant
acoustic transducer arrays in signature detection applications, especially in noisy envi-
ronments. This is particularly valuable in wearable sensing and monitoring applications
and wireless sensor networks, where environmental conditions cannot be controlled or
accurately predicted.
References
[1] B. Herrera, E. Demirors, G. Chen, R. Guida, F. Pop, N. Dave, C. Cassella, T. Melo-
dia, and M. Rinaldi, \PMUT-Based High Data Rate Ultrasonic Wireless Communi-
cation Link for Intra-Body Networks," in Hilton Head Solid-State Sensors, Actuators
and Microsystems Workshop, (Hilton Head, SC), pp. 105{108, 2018.
[2] M. Stojanovic and J. Preisig, \Underwater acoustic communication channels: Prop-
agation models and statistical characterization," Communications Magazine, IEEE,
vol. 47, no. 1, pp. 84{89, 2009.
[3] J. Hoople, J. Kuo, S. Ardanuc, and A. Lal, \Chip-scale sonic communication using
AlN transducers," in IEEE International Ultrasonics Symposium, IUS, pp. 1934{
1937, 2013.
[4] J. Li, L. Deng, Y. Gong, and R. Haeb-Umbach, \An overview of noise-robust auto-
matic speech recognition," 2014.
[5] P. C. Loizou, Speech enhancement: theory and practice. CRC Press, 2007.
[6] CMU Sphinx, \CMU Sphinx Project," 2009.
[7] P. Gupta, Y. Jeong, J. Choi, M. Faingold, A. Daruwalla, and F. Ayazi, \Preci-
sion High-Bandwidth Out-Of-Plane Accelerometer as Contact Microphone for Body-
Worn Auscultation Devices," in Hilton Head Solid-State Sensors, Actuators and Mi-
crosystems Workshop, (Hilton Head, SC), 2018.
[8] PixSoft, \The RALE repository."
[9] F. Massey, \The Kolmogorov-Smirnov test for goodness of t," Journal of the Amer-
ican Statistical Association, vol. 46, no. 253, pp. 68{78, 1951.
89
[10] J. Shao, \Linear model selection by cross-validation," Journal of the American sta-
tistical Association, vol. 88, no. 422, pp. 496{494, 1993.
[11] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blon-
del, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Courna-
peau, M. Brucher, M. Perrot, and

E. Duchesnay, \Scikit-learn: Machine Learning in
Python," Journal of Machine Learning Research, vol. 12, pp. 2825{2830, 2012.
[12] S. I. Khan, \Respiratory Sound Analysis for Identifying Lung Diseases : A Review,"
International Journal of Science and Research (IJSR), vol. 3, no. 11, pp. 566{571,
2014.
[13] K. Kochetov, E. Putin, S. Azizov, I. Skorobogatov, and A. Filchenkov, \Wheeze De-
tection Using Convolutional Neural Networks," in Progress in Artificial Intelligence ,
vol. 8154, pp. 162{173, 2017.
[14] K. Forkheim, D. Scuse, and H. Pasterkamp, \A comparison of neural network models
for wheeze detection," IEEE WESCANEX 95. Communications, Power, and Com-
puting. Conference Proceedings, vol. 1, no. 95, pp. 214{219, 1995.
[15] B. S. Lin, H. D. Wu, S. J. Chen, G. E. Jan, and B. S. Lin, \Using Back-Propagation
Neural Network for Automatic Wheezing Detection," in Proceedings - 2015 Interna-
tional Conference on Intelligent Information Hiding and Multimedia Signal Process-
ing, IIH-MSP 2015, pp. 49{52, 2016.
[16] F. Chollet, \Keras," 2015.
[17] D. M. Kline and V. L. Berardi, \Revisiting squared-error and cross-entropy functions
for training neural network classiers," Neural Computing and Applications, vol. 14,
no. 4, pp. 310{318, 2005.
[18] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, \Learning representations by
back-propagating errors," Nature, vol. 323, no. 6088, pp. 533{536, 1986.
[19] D. P. Kingma and J. L. Ba, \Adam: a Method for Stochastic Optimization," Inter-
national Conference on Learning Representations 2015, pp. 1{15, 2015.
[20] M. L. Seltzer, D. Yu, and Y. Wang, \An investigation of deep neural networks
for noise robust speech recognition," in 2013 IEEE International Conference on
Acoustics, Speech and Signal Processing, pp. 7398{7402, 2013.
[21] A. L. Maas, Q. V. Le, T. M. O'Neil, O. Vinyals, P. Nguyen, and A. Y. Ng, \Recurrent
Neural Networks for Noise Reduction in Robust ASR.," Interspeech, pp. 3{6, 2012.
90
Chapter 5
System Integration and Low Power Implementation
This chapter presents system-level study and implementation of a fully integrated wireless
sensor network (WSN) node for detection and classication of acoustic signatures, as
illustrated in Figure 5.1. The system makes use of a custom-designed MEMS microphone
array for mechanical ltering of acoustic signals, as presented in Chapter 2, which is shown
to reduce the power requirements and power consumption of acoustic signal detection
in micro-processors. A vibration energy harvester served as the power source for such
a system, and a power-management integrated circuit was evaluated for conditioning
and energy storage, which allows for zero-power operation without an external supply
or batter source. A programmable system-on-a-chip (PSoC) was congured for digital
signal processing and system control. In the proposed system design, when an acoustic
signature of interest is detected, an integrated radio and antenna transmits the collected
data via low-power communication to a central node, such as a cell phone or base station.
In this chapter, we emphasize the integration of the proposed wheezing detection al-
gorithm with a low-power processing platform. Although energy harvesters optimized for
human walking motion are bulky and currently impractical for integration with wearable
91
Figure 5.1: Diagram showing the fully-integrated proposed system for zero-power acoustic
signature monitoring, with illustrated components.
systems, we will still investigate the integration of vibration energy harvesting for use
as a generalized WSN testing platform. The discussion on the details of implementa-
tion of the energy harvester geometries and energy harvesting integration is covered in
detail in Chapter 6. The system discussed in this chapter is illustrated in Figure 5.2,
and consists of a MEMS microphone, multi-channel analog-to-digital converter (ADC),
microcontroller-based processing, and Bluetooth radio for wireless transmission of mea-
surements.
92
Figure 5.2: Conceptual illustration of system architecture for low-power continuous
breathing-monitoring system. Key components are the resonant microphone array (con-
sisting of 7 elements), pre-amplier and signal conditioning electronics, analog-to-digital
converter (ADC), microcontroller, and Bluetooth low energy (BLE) radio for wireless
data transmission. The microcontroller implements continuous wheeze detection.
5.1 Motivation
Low-power and self-powered acoustic sensors present multitude of promising applications
including pipeline leak detection [1], industrial reliability and process monitoring [2],
wildlife monitoring [3], and military vehicle classication [4].
The main application for the integrated device discussed in this chapter is detection of
wheezing within the acoustic signatures of breathing, as a potential precursor of asthma
attacks [5, 6]. Wheezing, as well as other adventitious sounds associated with respiratory
diseases, can be identied separately from normal respiratory sounds in patients using a
wearable sensor. Although not demonstrated in this thesis, such a wearable sensor could
93
implement multi-class breathing characterization with the spectro-temporal methods dis-
cussed in this thesis [7].
Figure 5.3: Spectrogram of recorded lung sounds from an infant with wheezing signature
present during the inspiration phase of respiration.
The eectiveness of this wheeze detection system is demonstrated as a smart wear-
able stethoscope, which continuously monitors breathing sounds for indications of health
risks. The device and accompanying algorithm were implemented to detect constriction
of the respiratory airways, which is manifested as high-pitched wheezing in the breathing
cycle (Figure 5.3). Wheezing is a common symptom of asthma, Chronic Obstructive
Pulmonary Disease (COPD), and other respiratory conditions [8, 9]. Accurate detection
and characterization of wheeze sounds is most eectively achieved by continuous moni-
toring of lung sounds with a stethoscope, which could prompt diagnosis and appropriate
treatment [10]. However, there are currently no wearable, wirelessly-connected, always-on
lung-sound trackers that would be as eective as evaluation by a trained clinician.
A number of automatic wheeze detectors, such as Respiris AirSonea devices [11], are
commercially available to allow patients to monitor their health from home without the
94
presence of a physician. However, state-of-the-art wheeze detectors are limited in their
eectiveness because they are not wearable, are sensitive to background interference,
cannot monitor lung sounds continuously, and require frequent recharging.
We addressed these challenges by (1) improving coupling to low-intensity lung sounds
by enhancing stethoscope sensitivity, (2) improving performance by ltering background
noise during acquisition, and (3) eciently processing to extend battery life. By employ-
ing resonant MEMS microphone arrays for signal acquisition, we were able to miniaturize
the necessary electronics, extend battery life, and enhance performance through improved
classication accuracy, greater sensitivity, and reduced noise
oor.
5.2 System Architecture
The integrated acoustic signature detection system features key innovations on the design
and fabrication of the MEMS resonant microphone array (Chapter 2), on the signal clas-
sication algorithm (Chapter 3), and on the integration of these hardware components.
Resonant Microphone Array
The acoustic resonator array presented in this thesis was composed of paddle-shaped
silicon cantilever elements with a piezoelectric ZnO thin lm as the sensing mechanism,
as presented in Chapter 2.
The use of a piezoelectric sensing mechanism, in contrast to capacitance-based micro-
phones, further helps with the reduction of power demands on the system, as piezoelectric
sensing is passive and does not require a voltage bias.
95
The selected individual resonant microphone element that was used in this study for
classication experiments was characterized to have a peak sensitivity of 138 mV/Pa at a
resonant frequency of 1.36 kHz. The noise
oor was measured to be 10.8 dBA (85.8 SNR).
The G.R.A.S. calibrated precision microphone, which also serves as a
atband microphone
reference for subsequent testing, has a uniform sensitivity of 12.6 mV/Pa with a noise

oor of 44.5 dBA (49.5 SNR), values that are consistent with typical
atband microphones
on the market.
Classification Algorithm
Many of the algorithms that have been developed for high-degree-of-accuracy classica-
tion of lung sound features are not suitable for embedded and wearable applications due
to their large memory footprint or demanding computation requirements. Commercial
implementations and research into automatic detection of adventitious respiratory sounds
focuses primarily on digital signal processing of audio recorded using a single microphone,
electronic stethoscope, or contact accelerometer, which ideally exhibit
at-band spectral
responses. Prior to classication, lung recordings are processed with a feature extractor
for dimensionality reduction. Features on which lung sound classication is performed in-
clude Mel-Spaced Cepstral Coecients (MFCC), Wavelet Packet Decomposition (WPD),
spectrogram images, or temporal features [12].
In this chapter we exploit the band-pass ltering characteristic of the MEMS resonant
microphone to reduce the digital processing requirements on the embedded processor,
thereby reducing the computation time and power consumption. We focus on feature
extraction through calculation of cepstral coecients, such as those employed in MFCCs,
96
with a reduced computational complexity aorded to us by performing the bandpass
ltering tasks at the sensor level with a resonant microphone array rather than in the
digital domain, as experimentally demonstrated in Chapter 4.
Integration of Energy Harvester
Power conversion from the energy harvester to the supply of the processing components
was performed with the Cypress MB39C831 Power Management IC (PMIC). The vibra-
tion energy transduced by the piezoelectric or electromagnetic energy harvester will be
directly input to the PMIC, which will regulate the harvested power to a stable 3.3 V
DC supply. The PMIC contains a diode bridge rectier, DC-DC buck converter to stabi-
lize the output voltage, and a capacitor or rechargeable battery for charge storage. For
low-acceleration vibration applications a PMIC circuit based on DC-DC boost conversion
could be also employed.
We have characterized the performance of this PMIC and demonstrated that with a
1.0g vibration source sucient energy can be generated and stored to power an optimized
Cypress BLE PSoC system. Based on the average power consumption of the PSoC, we
estimated that the system, which includes sensing and wireless transmission of data,
can operate with a duty cycle of 13% when powered by the harvested energy from 1.0g
vibration.
Hardware Overview
To demonstrate the system-level integration of low-power signal processing and wire-
less data communication with the resonant array transducer, we used the Cypress 4200
97
Bluetooth LE Programmable-System-on-Chip (PSoC) with built-in ADC, ARM proces-
sor, radio, and antenna. Feature extraction and classication were implemented on the
integrated processor to demonstrate and validate the potential improvement in power
consumption when using pre-ltered resonant microphones for feature extraction. The
BLE PSoC was congured for ultra-low power digitization of the sensor output, signal
processing, and wireless output upon detection of a sensed signature of interest.
5.3 Hardware Implementation
To interface the resonant microphone with the processing software, it aws wire-bonded
to a PCB containing a pre-amplier circuit, designed for a voltage gain of 40 dB using
a simple non-inverting op-amp conguration. Due to the high sensitivity of the resonant
microphone, large gain in the amplier is not necessary, but can instead be adjusted to
serve as a buer and channel equalizer. Pre-amplication was implemented at relatively
low power with a TLV341 low-power OpAmp, which draws a quiescent supply current of
70 A per channel.
5.3.1 Experimental Setup
For power consumption experiments (Figure 5.4b), we used the low-power Cypress PSoC
4200 BLE platform. We used the on-board 8 channel ADC (12 bit, 1 Msps) for digitization
of the resonant microphones' signals. This corresponds to a 0.5 mV sampling resolution
when using a 2V supply voltage. If additional channels are necessary for large sensor
arrays, an analog multiplexer can be introduced to allow multi-channel digitization at
cost of reduced sampling rate. Alternatively, time-multiplexing can be easily implemented
98
Figure 5.4: Block diagram showing the experimental testing process for evaluation of
power consumption of lung sound classication on Cypress PSoC 4200 BLE chip with
both resonant microphone array processing and
at-band processing.
with a single ADC due to the low-frequency characteristic of the monitored health signals
of interest.
Continuous pattern recognition was performed in the integrated ARM Cortex-M0
processor, which has a specied 0.9 DMIPS/MHz eciency and a power requirement of
85 W/MHz. Once a signal of interest is detected, the integrated BLE transmitter and
antenna was congured to send a wireless notication. The power consumption of the
transmitter is dependent on the target communication range. In this experiment, we
limited transmission power to 0 dBm (without the use of an external power amplier),
which corresponds to an estimated 3-5 meter range.
A major challenge with the implementation of feature extraction and classication
algorithms on low-power platforms, such as the Cypress PSoC chip, is the limited avail-
ability of memory (16 kB SRAM) and processing speed (24 MHz). These specications
99
set a constraint to the complexity and type of features and classication algorithm that
can be practically implemented. Steps were taken to optimize memory usage and pro-
cessing speed for the algorithm to function without running out of system resources. An
estimate of the required algorithm parameters and resulting memory usage is summarized
in Table 5.1.
Sample
Rate
Frame
Length
No. of
Filters
No. of
Features
Estimated
Stack Size
Estimated
Heap Size
PC Test 5120 S/s
40 ms
(0.2 kS)
13 12 11.4 kB 5.7 kB
PSoC Test 1600 S/s
40 ms
(64 S)
7 6 2.3 kB 1.1 kB
Table 5.1: Summary of the system resource requirements for testing the classication ac-
curacy with a PC, and implementing on low-specication system, with estimated memory
usage of each.
From the reported specication of the device, we estimated a total 2.3 mW average
power consumption during active cycles, which can be reduced by proper duty cycling
and optimization of power modes. The estimated power usage, based on specications in
the PSoC datasheet, is summarized in Table 5.2.
ADC Processing Wireless Transmission Total
Average
Power
Consumption
2 mW
(8 channels)
85 W
(at 1 MHz)
25 mW (100% duty cycle)
0.25 mW (1% duty cycle)
2.3 mW
Table 5.2: Estimated power consumption of Cypress PSoC 4200 BLE in active operation,
as specied by the datasheet.
Power consumption was evaluated by measuring the supplied current to the Cypress
PSoC device while performing continuous acquisition, feature extraction, and classica-
tion. Each classication cycle was performed for a single 40 ms frame, with multiple
100
frame acquisition and feature extraction duration increasing linearly with the number of
analyzed frames.
5.4 Results - Power Consumption
To characterize the in
uence of the proposed mixed-signal processing scheme on com-
putational eciency and power consumption, both feature extraction algorithms were
implemented in the low-power PSoC chip in C along with a Naive Bayes' classication
model. A 7 lter feature extractor was evaluated to work within the hardware constraints,
and signal acquisition was congured for a sample rate of 1600 Hz.
The measured results in Figure 5.5 show a result consistent with our estimation of
computation time, with the primary advantage of resonant array processing being in the
reduction in the computational complexity by avoiding digital ltering during feature
extraction. Each recognition cycle evaluated on the PSoC chip executed in an average
of 5.11 seconds using a single input, and 0.46 seconds with an array input. This results
in an 11x (91.0%) reduction in computation time, which also equates to an 11x power
reduction with duty cycling if computation time is xed.
Wireless transmission with Bluetooth LE in both cases was measured to consume an
average power of 0.85 mW (when transmitting with a rated output power of 0 dBm).
Transmission of a single packet containing a byte of data completes in 5 ms, with a
9600 bps baud rate.
The standard method of using purely digital processing completed a classication cy-
cle in 5.11 seconds with an average power consumption of 6.6 mW. To compare this to
101
Figure 5.5: Measurements of power consumption in Cypress PSoC 4200 for the case
of (a) processing with a standard digital signal classication method, and (b) processing
with the pre-ltered microphone array method. The classication period is held constant,
with one frame of digitization, feature extraction, classication, and wireless transmission
completing every 5.11 seconds.
102
the power consumption of using the array method, we maintained a constant classica-
tion period of 5.11 seconds and put the microcontroller into a deep-sleep state for the
remainder of the cycle. This resulted in an average power consumption of 0.596 mW, an
11.1 times (91.0%) reduction in average power over the baseline case.
Power consumption for array-based signal processing was increased during the signal-
acquisition stage due the increased quantity of front-end electronics, increased demand on
the ADC, and increased requirement for data buer capacity prior to feature extraction.
The duration of the ADC stage would vary proportionally with the choice of frame length.
We estimated an additional supply current of 390 A for signal conditioning in the array-
processing case, which is included in the total ADC gure and which accounts for the
primary source of dierence in power consumption during this step.
5.5 Discussion
The array-based approach shows a potential for improved performance in noisy environ-
ments, specically those with low-frequency noise sources. It is hypothesized that the
improvement in recognition accuracy results from the limitations to dynamic range of the
analog-to-digital conversion process. As out-of-band noise amplitude increases, the ac-
quired signal becomes distorted at all frequencies for the case of a wideband (or
at-band)
microphone. In contrast, the signal output of the high Q-factor MEMS microphone array
does not experience signicant distortion due to environmental noise.
We highlighted that there is additional potential to improve classication accuracy by
taking advantage of the temporal information present in a consecutive series of frames,
103
rather than performing classication on a single frame at a time. This can be accom-
plished using HMMs, or variations of Neural Networks, as demonstrated in Chapter 4.
For applications where temporal variations in signals are useful information (such as in
speech recognition), a resonant array-based processing scheme would be advantageous,
since low power (with consequent low performance) hardware would be unable to perform
continuous real-time feature extraction at a sucient rate.
As evidenced by our experiments with the Cypress PSoC chip, the array approach can
complete many feature extractions in the time it takes standard processing to complete
one, which allows a smaller data buer to be used and frees up system resources for
other tasks. As seen with the neural network classication experiments in Chapter 4,
the best noise-robust performance for wheeze detection can be achieved with about 40
consecutive 40 ms frames with 50% overlap (0.82 s of data). By using the resonant
array algorithm we can extract features with a period of 0.38 s, allowing data acquisition
and feature extraction to be performed in parallel. The frames in the data buer can
be discarded after calculation of each feature frame, while a standard digital processing
technique will take much longer to calculate the feature vector than to digitize the data.
In total, it was estimated that feature extraction of 40 consecutive frames will take 15 s
with the resonant array method and 3 min and 22 s with the digital processing method,
a signicant dierence that would make-or-break many applications.
104
5.6 Summary
The results of this study provided evidence for an alternative classication technique for
implementation in embedded low-power processing, such as for continuous lung monitor-
ing via a wearable stethoscope. With the use of a resonant microphone array, measure-
ment accuracy is less signicantly in
uenced by interference from the heart and external
noise sources. Perfect acoustic coupling (to human body) also becomes less critical for
reliable signal acquisition, and the higher sensitivity, lower noise
oor, and improved noise
rejection of the array will compensate for other limitations in practical application.
Experiments with the proposed array-input feature extraction algorithm have indi-
cated that the implementation of such a technique has the benets of (1) markedly reduced
computation time and power consumption for classication tasks, (2) improved robust-
ness to some forms of interference, and (3) improved classication accuracy due to larger
microphone sensitivity and consequent lower noise
oor. These benets have signicant
implications for addressing important challenges in emerging wearable technologies and
applications.
105
References
[1] J. Zhang, \Designing a cost-eective and reliable pipeline leak-detection system,"
Pipes and Pipelines International, vol. 42, no. 1, pp. 20{26, 1997.
[2] D. E. Lee, I. Hwang, C. M. O. Valente, J. F. G. Oliveira, and D. A. Dornfeld,
\Precision manufacturing process monitoring with acoustic emission," International
Journal of Machine Tools and Manufacture, vol. 46, no. 2, pp. 176{188, 2006.
[3] P. Juang, H. Oki, Y. Wang, M. Martonosi, P. Peh Li-Shiuan, and D. Rubenstein,
\Energy-Ecient Computing for Wildlife Tracking: Design Tradeos and Early Ex-
periences with ZebraNet," Proceedings of the 10th International Conference on Ar-
chitectural Support for Programming Languages and Operating Systems (ASPLOS
2002), pp. 96{107, 2002.
[4] J. Altmann, \Acoustic and seismic signals of heavy military vehicles for co-operative
verication," Journal of Sound and Vibration, vol. 273, no. 4, pp. 713{740, 2004.
[5] N. M. Wilson, \Virus infections, wheeze and asthma," 2003.
[6] N. P. Ly, D. R. Gold, S. T. Weiss, and J. C. Celedon, \Recurrent wheeze in early
childhood and asthma among children at risk for atopy.," Pediatrics, vol. 117, no. 6,
pp. e1132{8, 2006.
[7] S. K. Chowdhury and A. K. Majumder, \Frequency analysis of adventitious lung
sounds," Journal of Biomedical Engineering, vol. 4, no. 4, pp. 305{312, 1982.
[8] F. D. Martinez, A. L. Wright, L. M. Taussig, C. J. Holberg, M. Halonen, and W. J.
Morgan, \Asthma and Wheezing in the First Six Years of Life," New England Journal
of Medicine, vol. 332, no. 3, pp. 133{138, 1995.
[9] C. E. Kuehni, A. M. Brooke, and M. Silverman, \Prevalence of wheeze during child-
hood: Retrospective and prospective assessment," European Respiratory Journal,
vol. 16, no. 1, pp. 81{85, 2000.
[10] H. Pasterkamp, \The highs and lows of wheezing: A review of the most popular
adventitious lung sound," 2018.
[11] \AirSonea. A revolution in asthma management," 2018.
[12] R. X. A. Pramono, S. Bowyer, and E. Rodriguez-Villegas, \Automatic adventitious
respiratory sound analysis: A systematic review," PLoS ONE, vol. 12, no. 5, 2017.
106
Chapter 6
Piezoelectric Vibration-Energy Harvesting
Piezoelectric vibration energy harvesters were developed for use in self-powered wireless
sensor node systems. The harvesters were designed for transduction of ambient vibration
energy, which is tuned in frequency depending on the particular application, into electrical
energy which can then be stored. The energy harvesters in our study are all based on
piezoceramic PZT substrates and
exible Macro Fiber Composite (MFC) substrates.
We specically investigated applications related to energy harvesting on naval vessels,
which typically have vibration sources with sharp spectral peaks in the range of 30-60
Hz. We investigated the performance of the developed energy harvesters for 0.1g and
1.0g vibration amplitude.
6.1 Background
Ambient vibration energy exists in many industrial, military, and commercial environ-
ments. Several vibration sources mentioned earlier include naval engine vibration, indus-
trial pipeline liquid
ow, and human walking motion. These vibrations in the environment
107
can be harvested and converted to electrical energy for powering sensing, processing, and
communication electronics [1].
This chapter presents a study of the use of piezoelectric materials for vibration energy
harvesting. Although alternative ambient energy sources exist (thermal, RF, solar, etc),
these sources suer from low-energy densities in our targeted applications [2]. Electro-
magnetic vibration energy harvesters have been demonstrated to be able to supply large
power densities [3], but generally generate very small output voltages making DC-DC
boost conversion necessary for practical integration. Piezoelectric energy harvesters can
produce very large voltages, but have very high series resistances relative to electromag-
netic vibration energy harvesters.
6.2 PZT Bimorph Energy Harvester
We explored the use of a cantilever structure for resonant piezoelectric energy harvesting.
A cantilever has a mechanical resonant frequency, which is determined by the dimensions
and attached proof mass. These parameters are designed such that the resonant frequency
matches the vibration source frequency for maximum energy generation. We fabricated a
cantilever energy harvester based on a lead zirconate titanate (PZT) bimorph substrate.
A bimorph cantilever structure, containing two piezoelectric layers, was used to take
advantage of opposite stress polarities at the clamped end of the cantilever. Had we
instead used a unimorph cantilever, a substrate material would have been necessary and
only half of the available energy would be harvested. A commercial PZT bimorph with
dimensions shown in Figure 6.1 served as the cantilever substrate for these studies.
108
Figure 6.1: Schematic of the PZT bimorph structure, composed of two PZT layers sepa-
rated by a brass layer.
6.2.1 Modeling and Design
We can write the resonant frequency of the cantilever as follows:

0
=
1
2 √︃
3

3

(6.1)
where

is the Young's modulus of PZT;

is the eective moment of inertia of the
rectangular beam cross section, which is adjusted for the dierence in Young's Modulus
between PZT and brass;

is the eective mass of the cantilever beam and

is
the length of the cantilever. These parameters are determined as follows:

=

(

+

)
3
12
(6.2)

= 0.238

+

(6.3)
Since the weight of the energy harvester is not a critical parameter for the target
applications, we can increase the weight of the proof mass in order to reduce the total
volume of the energy harvester or to increase the width. In both cases, the power density
of the energy harvester was increased by introducing a larger mass.
109
Figure 6.2: Schematic of PZT bimorph cantilever with attached proof mass.

30.8 g 5.5 cm 4.0 cm 0.15 mm 0.35 mm
Table 6.1: Table summarizing PZT bimorph cantilever design parameters.
For a target resonant frequency of 27 Hz, we selected a cantilever with dimensions
summarized in Table 6.1. The calculated resonant frequency is thus

0
=
1
2 ⎯
⎸
⎸
⎷3

(

+

)
3
12
(0.238

+

)
3

= 26.6 Hz (6.4)
We fabricated the energy harvester (Figure 6.3) by mechanically dicing the commercial
PZT bimorph and attaching it mechanically and electrically to a PCB with conductive
silver epoxy through the bottom electrode. The mass was attached adhesively and the top
electrode was wire-bonded to the PCB. Electrical measurement of the energy harvester
was performed by connecting the energy harvester output to a variable resistance board
to match the load impedance to the impedance of the PZT bimorph, and maximize the
power transfer. An op-amp buer circuit is used to isolate this load resistance from the
input resistance of the oscilloscope (1
). Vibration was induced by a characterized
110
shaker driven by a function generator and power amplier. For low-frequency vibrations
less than 30 Hz, a PC-controlled linear actuator was used as the vibration source.
(a) (b)
Figure 6.3: (a) Photograph of fabricated PZT bimorph energy harvester, and (b)
schematic of measurement setup using a variable impedance board, with buered os-
cilloscope input.
6.2.2 Results
We evaluated the power delivered to a matched load without the in
uence of the impedance
of the measuring setup. Using an LCR meter we found the resistance of the PZT bimorph
to be about 130 k
and the capacitance to be 173 nF. At the resonant frequency of 27 Hz,
this corresponded to an impedance of about 32.3 k
. The measured power delivered to
a matched load is plotted in Figure 6.4a.
The measured resonant frequency of 27.2 Hz is consistent with our analytical modeling
and analysis. The quality factor of the response was measured to be 13.9. Under a 0.1g
vibration at the resonant frequency the harvester delivered an RMS voltage of 1.83 V,
and RMS power of 0.124 mW, to a matched load. Under a 1.0g vibration source at the
resonant frequency we measured an RMS power of 3.0 mW, corresponding to a peak
RMS voltage of 9.06 V.
111
Since the active volume of the energy harvester is 29.1 cc, the power density of the de-
vice under 0.1g vibration is 4.26 W/cc, the measured power density under 1.0g vibration
is 104.4 W/cc.
(a) (b)
Figure 6.4: Plots of (b) the power delivered to a matched load (32.3 k
) by the PZT
bimorph energy harvester, and (a) the RMS power harvested at varying vibration levels
at the resonant frequency.
6.3 Macro-Fiber Composite Flexible Energy Harvester
We investigated the use of a
exible piezoelectric meta-material called Macro Fiber Com-
posite (MFC) as an alternative to bulk ceramic and thin-lm piezoelectrics for vibration
energy harvesting. A challenge with using bulk piezoceramic materials is the large Young's
modulus and relative thickness, which leads to the necessity of using large proof masses
or large harvester dimensions. By using a
exible substrate like MFC instead of a bulk
material like PZT, we are able to fabricate lower-frequency and smaller-dimension energy
harvesters, and can explore creative designs that are impractical with bulk materials.
112
6.3.1 Macro-Fiber Composite
MFC is a meta-material composed of PZT bers that are sandwiched between metal
electrodes [4, 5]. This structure is supported by a structural epoxy layer and encased
in a polyimide lm to increase the durability. The internal structure of a typical MFC
sheet, designed for energy harvesting with the d
3
1 piezoelectric coecient, is shown in
Figure 6.5. The eective d
3
1 coecient of MFC is 170 / [6], which is comparable to
that of bulk PZT. The use of PZT bers improves the damage tolerance and
exibility
of the material as compared to bulk PZT.
Figure 6.5: Illustration showing the micro-structure of a Macro Fiber Composite (MFC)
energy harvester operating in a bending mode which utilizes the d
3
1 piezoelectric coe-
cient.
6.3.2 Design
We investigated the use of the MFC meta-material in the design of resonant energy har-
vesters. MFC has advantageous properties and allows for interesting geometries compared
to bulk PZT. Due to the
exible nature of MFC, we are able to design geometries with
low resonant frequencies at smaller dimensions and with the use of lighter masses. PZT
113
and other piezoceramic materials are very brittle, so the use of MFC has the potential to
improve the durability and lifetime of our energy harvesters.
An energy harvester based on an MFC cantilever geometry was designed and fab-
ricated. Since MFC is a
exible material, we can stack several MFC layers while still
maintaining a resonant frequency within the desired range of 20 Hz to 60 Hz. Figure 6.6
shows the MFC bimorph cantilever that we fabricated with two MFC layers and with
an attached proof mass. The two MFC sheets were adhesively bonded and electrically
connected in series. With two layers, we were able to increase the output power and
gure of merit by nearly two times without increasing the volume.
(a) (b)
Figure 6.6: (a) Schematic of multilayer MFC cantilever with proof mass, and (b) Photo-
graph of multilayer MFC cantilever with proof mass and specied dimensions.
The Young's modulus of an MFC sheet is approximately

= 15.86 GPa. The
thickness of each MFC sheet is 300 . For the parameters summarized in Table 6.2, we
predict the resonant frequency of the two-layer MFC cantilever to be:

0
=
1
2 √︃
3

(2

)
3
12

3

= 9.73 Hz
We measured the power delivered by the energy harvester to a matched load using
114
m

l

w

t

C
3.1 g 7.5 mm 1.2 mm 300 m 45 nF
Table 6.2: Summary of the selected geometric design parameters.
the linear actuator as the vibration source. The curves in Figure 6.7 were obtained from
this measurement.
(a) (b)
Figure 6.7: (a) Plot of RMS power delivered into matched load (459 k
) by series-
connected two-layer MFC cantilever energy harvester with 1.0g vibration source, and (b)
plot of RMS power delivered into matched load at varying vibration intensities.
The measured response shows a resonant frequency of 7.7 Hz, which varied slightly
from the analytical measurement. The quality factor of the response was 7.11. Un-
der a 1.0g vibration source, the harvester will deliver an RMS power of 48.9 W to a
matched load of 459 k
at the resonant frequency. This corresponds to a power density
of 3.26 W/cc and a bandwidth gure of merit of 0.46 W/cc.
The low resonant frequency of MFC cantilevers allowed us to either reduce the di-
mensions or add additional MFC layers, to improve the gure of merit and bring the
resonant frequency up to ranges suitable for naval applications were vibrations are on the
order of 40 to 60 Hz. A low frequency design can be potentially eective for harvesting
115
energy for other applications, such as from walking motion, ocean wave motion, or other
low-frequency vibration sources.
6.4 Bistable Energy Harvester
Although cantilever design geometries are the simplest to implement, there is also a
potential to use non-linear eects and geometries to achieve amplied responses. Two
non-linear energy harvester designs, based on bistable magnetic eects and bistable me-
chanical buckling, were proposed and characterized.
6.4.1 Bistable Energy Harvester Utilizing Magnetic Mass
Non-linear bistable energy harvester designs were investigated. The proposed design is
based on a non-linear bandwidth extension technique that has been characterized and
modeled [7, 8].
The non-linear bistable energy harvester was modeled by solving for the motion of
the cantilever (), through the solution of the nonlinear ordinary dierential equation
for one-dimensional forced harmonic oscillation:
() =

_ () +

() +

(,) = cos() (6.5)
Where

(,) is magnetic force between the cantilever, which can be determined em-
pirically or through FEM simulation [8].
A simple bimorph cantilever was fabricated with a neodymium magnetic mass, as
shown in Figure 6.8. A xed-magnet is introduced at a distance from the cantilever, with
116
Figure 6.8: (a) An illustration of the proposed bistable magnetic cantilever energy har-
vester with indicated magnet polarities, and (b) photograph of the fabricated device used
in preliminary testing.
polarities positioned such that there is a repelling force. The device was characterized
with a xed-magnet distance of 1 cm, 2 cm, and without a secondary magnet as a baseline
comparison case. The power delivered to a matched load was measured and is summarized
in Figure 6.9.
As indicated by the plot, there is an optimal distance, or range of distances, between
the cantilever and the xed magnet for which low-frequency response was greatly in-
creased, consistent with similar reported devices [9]. The maximum power attained by
this device was 44.1 RMS, giving a gure of merit of 35.3 //
2
. Although
the preliminary results were promising, a major diculty is the optimization of magnet
separation distance and ensuring vibration in the bistable mode. This can be achieved
by redeveloping the fabrication process for increased precision in alignment, which could
be the subject of future research.
117
Figure 6.9: Characterization of the response of a bistable magnetic cantilever energy
harvester, with plot of RMS power delivered to a 1 M
load with 0.5 g vibration strength
for three magnet separation distances.
A potential extension of this technique for a large power-density system is conceptual-
ized in Figure 6.10. An array of magnetic PZT cantilevers are optimally positioned within
a spring-mounted magnetic frame. As the frame vibrates relative to the cantilever base,
a low frequency bistable mode of vibration is induced in the cantilevers. Alternatively, a
"plucking" mode can be generated for frequency up-conversion [10]. This technique can
potentially realize low-frequency piezoelectric vibration energy harvesting with a high
energy density and large bandwidth.
118
Figure 6.10: Conceptual diagram of a proposed low-frequency vibration energy harvesting
system with an array of bistable magnetically-coupled cantilevers.
6.4.2 Bistable Energy Harvester Utilizing Flexible Substrate
The
exible properties of MFC allow for additional interesting and creative design pos-
sibilities. We explored the use of a pre-strained double-clamped structure, which is not
possible in bulk PZT due to its rigidity and brittleness. This structure (Figure 6.11)
has two stable states, which we hypothesized would increase the generated power at the
device's resonant frequency. The beam is composed of a single MFC sheet to allow for
switching between the two states at low vibration intensities.
Bistable energy harvesting of this type has been proposed and implemented in other
work, including with PZT in micro-scale devices [11]. An advantage of the
exible sub-
strate bi-stable energy harvester is the ability to further reduce resonant frequency, in-
crease durability, and ability to withstand large displacements to generate larger poten-
tials.
119
(a) (b)
Figure 6.11: (a) Diagram of an MFC beam energy harvester with both sides clamped and
bistable states, and (b) Photograph of the fabricated bi-stable MFC energy harvester
with dimensions indicated.
There were multiple vibration modes present in such a structure, with the largest
response occurring at 50.1 Hz. This corresponds to peak RMS power delivered to matched
load of 85.1 W at 1.0g vibration, and 2.21 W at 0.1g vibration at 50.1 Hz, and a
resulting power density of 0.873 W/cc at 1.0g, and 023 W/cc at 0.1g at 50.1 Hz. The
power generated at 0.1g vibration was observed to be much less than the power generated
at higher acceleration levels, because a certain threshold must be reached to transition
from one bi-stable state to the other. This transition was measured to occur when a
vibration level of 0.8g is exceeded.
It should be noted that the portions of the strain along the MFC beam will have
opposite polarities and cancel each other out to reduce the harvested energy. This partic-
ular design can be optimized to obtain maximum power at multiple modes of vibration
by optimizing the electrode pattern. Also, the use of simple supports, rolling supports,
and sliding supports could potentially increase the strain by allowing a larger vibration
range.
In addition to proof-mass driven designs, MFC can be utilized as an energy harvester
120
in 'hinge' or 'joint' congurations. These can be employed in mechanical machinery,
rudders, airplane wings, and wearable biomechanical harvesters [12, 13]. This approach
benets from being frequency-independent and simple to implement reliably. This is
proposed as a future direction of research into MFC-based energy harvesters.
6.5 Summary
Multiple vibration energy harvester designs utilizing both bulk piezoelectric PZT sub-
strates and
exible MFC substrates have been demonstrated. Bimorph PZT cantilevers
are experimentally measured to exhibit an RMS power density of 104.4 //
2
. Two
new energy harvesting designs with bistable nonlinear operation have also been modeled,
fabricated, and characterized, and have shown a gure of merit up to 35.3 //
2
, and
with eective wide-band low-frequency performance relative to simple cantilevers.
These energy harvesters generate sucient power to completely supply the low-power
sensor systems presented earlier and other low-power active circuits, eectively forming
zero-power systems with no external supply requirements. Additionally, through minia-
turization these energy harvesters can be implemented as zero-power shock and vibration
sensors, which is further discussed in Chapter 8.
References
[1] A. Khaligh, P. Zeng, and C. Zheng, \Kinetic Energy Harvesting Using Piezoelectric
and Electromagnetic Technologies," Industrial Electronics, IEEE Transactions on,
vol. 57, no. 3, pp. 850{860, 2010.
121
[2] S. P. Beeby, M. J. Tudor, and N. M. White, \Energy harvesting vibration sources for
microsystems applications," Measurement Science and Technology, vol. 17, no. 12,
2006.
[3] S. Naifar, S. Bradai, C. Viehweger, and O. Kanoun, \Survey of electromagnetic and
magnetoelectric vibration energy harvesters for low frequency excitation," Measure-
ment: Journal of the International Measurement Confederation, vol. 106, pp. 251{
263, 2017.
[4] J. W. High and W. K. Wilkie, \Method of Fabricating NASA-Standard Macro-Fiber
Composite Piezoelectric Actuators," Technology, no. June, pp. 1{30, 2003.
[5] R. B. Williams, D. J. Inman, M. R. Schultz, M. W. Hyer, and W. K. Wilkie, \Non-
linear Tensile and Shear Behavior of Macro Fiber Composite Actuators," Journal of
Composite Materials, vol. 38, no. 10, pp. 855{869, 2004.
[6] S. Q. Zhang, Y. X. Li, and R. Schmidt, \Modeling and simulation of macro-ber
composite layered smart structures," Composite Structures, vol. 126, pp. 89{100,
2015.
[7] S. C. Stanton, C. C. McGehee, and B. P. Mann, \Reversible hysteresis for broadband
magnetopiezoelastic energy harvesting," Applied Physics Letters, vol. 95, no. 17,
pp. 14{17, 2009.
[8] S. C. Stanton, C. C. McGehee, and B. P. Mann, \Nonlinear dynamics for broad-
band energy harvesting: Investigation of a bistable piezoelectric inertial generator,"
Physica D: Nonlinear Phenomena, vol. 239, no. 10, pp. 640{653, 2010.
[9] C. Wang, Q. Zhang, and W. Wang, \Low-frequency wideband vibration energy har-
vesting by using frequency up-conversion and quin-stable nonlinearity," Journal of
Sound and Vibration, vol. 399, pp. 169{181, 2017.
[10] P. P. Holmes, E. M. Yeatman, and A. S, \Magnetic plucking of piezoelectric beams for
frequency up-converting energy harvesters," Smart Materials and Structures, vol. 23,
no. 2, p. 25009, 2014.
[11] R. Xu, H. Akay, and S.-G. Kim, \Micro Buckled Beam Based Ultra-Low Frequency
Vibration Energy Harvester," in Hilton Head Solid-State Sensors, Actuators and
Microsystems Workshop, (Hilton Head, SC), 2018.
[12] S. R. Anton and D. J. Inman, \Vibration energy harvesting for unmanned aerial
vehicles," in SPIE 6928, Active and Passive Smart Structures and Integrated Systems
2008, vol. 692824, 2008.
[13] J. M. Donelan, Q. Li, V. Naing, J. A. Hoer, D. J. Weber, and A. D. Kuo, \Biome-
chanical energy harvesting: Generating electricity during walking with minimal user
eort," Science, vol. 319, no. 5864, pp. 807{810, 2008.
122
Chapter 7
Zero-Power Amplification with Helmholtz Resonators
Acoustic transducers with high sensitivity are advantageous in a variety of applications
including ranging, ultrasonic imaging, and acoustic signature detection in noisy environ-
ments.
In this chapter we discuss the design for a cantilever-based acoustic transducer which
is coupled to a resonant acoustic cavity exhibiting the Helmholtz resonance eect to
amplify the input pressure wave sensed by the transducer [1]. Resonant frequencies of
the cantilever and the acoustic resonator are matched, such that maximum displacement
of the cantilever can be achieved.
Although similar structures have been proposed and demonstrated [2, 3], this approach
uses standard micromachining techniques and is fully compatible with the CMOS process.
A similar device was proposed by Takahashi [4], in which a sensitivity amplication by a
factor of 14 and a Q-factor increase by a factor of 2.3 were achieved for a device operating
at 22 kHz. The reported method, however, requires three substrates, which adds to the
fabrication complexity and cost.
123
7.1 Background
The Helmholtz resonance eect arises in geometries containing a narrow neck leading to
a larger volume, such as a glass bottle. Due to the relative compliance of the air in the
volume and the incompressibility of the air in the neck, a resonance eect is exhibited
where the air volume in the cavity is analogous to a mechanical spring, while the air in the
neck of the resonator is analogous to a mass. When excited at the resonance frequency,
the pressure amplitude at the neck of the resonator is amplied.
(a) (b)
Figure 7.1: (a) Diagram showing basic structure of acoustic transducer using Helmholtz
resonance for sensitivity enhancement. (b) Diagram showing COMSOL simulated pres-
sured distribution demonstrating the amplication of dierential pressure across the can-
tilever.
The acoustic resonance frequency can be easily tuned by adjusting the dimensions of
the air cavity. In this way, we can create arrays of acoustic transducers with varying reso-
nance frequencies amplied by varying sizes of acoustic cavities. This enables summation
of the outputs for increased overall bandwidth or measurement from individual elements
for use as a bank of notch lters.
124
7.2 Modeling
A schematic representation of the device is shown in Figure 7.2. The transducer and
neck portion is a silicon MEMS microphone identical to that presented in Chapter 2. An
advantage of this method of passive amplication is the relative simplicity with which it
can be integrated with existing fabrication
ows.
Figure 7.2: Diagram showing (a) top view and, (b) cross-sectional view of two-wafer
Helmholtz resonator design, with indicated geometric parameters.
A parametric model was developed to accurately predict the performance and reso-
nance of such a device. This electroacoustic model is represented with the circuit shown
in Figure 7.3.
For the practical design of such a device, parametric modeling was necessary to predict
resonant frequencies and performance trends, as a system with mismatched resonances
will not provide any benet to device sensitivity. An analytical model was constructed for
the acoustic system based on electroacoustic analogies, allowing us to model the device
125
as an electrical lter and use standard circuit analysis techniques. In the reduced-order
model, we disregarded the in
uence of acoustic scattering and the harmonic motion of
the cantilever on the acoustic behavior.
Figure 7.3: Circuit diagram showing the electroacoustic modeling of a Helmholtz
resonance-enhanced microphone, with coupling between acoustic domain models and me-
chanical domain.
Modeling of Acoustic Properties
A schematic of the design is shown in Figure 7.3, where A
0
is the top dimension of the
neck (i.e.,>+2); A

is the bottom dimension of the neck; L is the neck length; t is the
cantilever paddle thickness; R is the radius of the hemispherical chamber; D represents
the length and width of the microphone paddle; and d is the width of the gap.
Standard electroacoustic representations of Helmholtz resonators feature an inductor,
representing the neck, in series with a capacitor, representing the acoustic cavity [5].
126
These approximations are valid for incident waves with wavelengths much larger than
the maximum device dimensions. The cavity volume is represented as a capacitor with
value given by:

=

2
(7.1)
The neck of the Helmholtz resonator is analogous to an inductor with value given by:

=

(7.2)
where S is the cross-sectional area of the neck. However, since the cross section changes
with thickness, we modify the expression to be:

=
∫︁

0
()
=

0

(7.3)
A length-end correction factor is typically included in the expression for the neck,
to account for the additional space in the chamber volume which presents an eective
extension of the neck length. Physically this can be understood as the additional distance
that the column of air present in the neck can occupy during oscillation. This factor is
determined through simulation or experiment, but has a characteristic value of =
0.6
√
, giving a nal expression for the inductance of the neck:

=
( + )

0

=
( + 0.6
√
)

0

(7.4)
The narrow gap between the cantilever edge and the substrate generates turbulence
127
in the system, and in general the eect is nonlinear. The eect of the gap on the model
can be linearly-estimated using the electroacoustic analog for a narrow rectangular slit
proposed by Beranek [6]. By estimating the length of this slit as 3D+2d, we obtain the
equivalent electroacoustic resistance and inductance pair given by

=
12

3
(3 + 2)
(7.5)

=
6
5(3 + 2)
(7.6)
If we ignore the coupled mechanical domain component, we can compute the transfer
function from

to using a voltage division:
() =
()

()
=

+

+(

+

) +
1
/
(7.7)
By dening the variables

and

and solving for the magnitude and phase re-
sponse of this transfer function, we obtain the following expressions for amplitude, phase,
and resonance frequency.

=

(7.8)

=((

+

)− 1
/ (7.9)
|()| =
√︁

4

+
2

(
2

+
2

) +
2

2

+
2

+
2

(7.10)
̸ () = tan
− 1
{

2

+

(

+

)
} (7.11)

0
=
1
√︀

(

+

)
(7.12)
128
By maximizing the value of the resulting transfer function, we can achieve a tuned
mechanical amplication of the input pressure across the cantilever beam at the resonant
frequency.
Coupling of Mechanical Circuit Model
A mechanical model is developed for the cantilever beam to estimate the resulting dis-
placement. A lumped mass-spring model is used, for which equivalent values are deter-
mined as follows

= 0.236·
,

=

,

=

3
4
3
,
0
=
√︃

(7.13)
where

is the equivalent proof-mass;

is the equivalent spring constant; and

is the equivalent damping constant. This corresponds to equivalent mechanical
circuit values in Figure 7.3 to be:

=

=

=
1

(7.14)
In the mechanical circuit equivalent, force is analogous to voltage and velocity is
analogous to current [7]. The quality factor Q is determined experimentally to have a
value of about 40 for such a geometry. If the mass-spring system is driven by a harmonic
excitation () =
0
() at the free-end of the cantilever, we obtain the following
magnitude and phase response [8]:
129
() =

0
√︀
(−
2
) + ()
2
(7.15)
() =
− 1
(

2
−
) (7.16)
Pressure is incident on the cantilever uniformly, so it will need to be converted to a
free-end loading case. The expression for free-end de
ection due to a point load at the
cantilever tip is:

=

3
3
(7.17)
The expression for free-end de
ection due to uniform loading caused by a pressure
dierential is given by:

=

4
8
(7.18)
Equating the two expressions 7.17 and 7.18, we attain a relationship to convert the
pressure dierential obtained in 7.10 to an external force expression, which is substituted
into 7.15. This equation also gives us the turns ratio of the transformer which converts
from the acoustic domain to the mechanical domain in Figure 7.3.

0
() =
3
8
()
=
3
8

(7.19)
Additional Considerations
The coupling model presented above is a simplication in which it is assumed that the
de
ection of the cantilever has no in
uence on the properties of the acoustic resonator.
130
The addition of this factor greatly increases the complexity of the analytical model, and
requires the solution of a second-order ordinary dierential equation (ODE) with initial
conditions. After solving the ODE with a numerical method, it was observed that eect
of cantilever displacement is negligible on the acoustic model, with an error of 0.7% in
the estimation of resonant frequency. Thus, these intricacies can be eliminated for the
sake of simplicity and computation speed.
An iterative feedback approach was proposed as a computationally-ecient method of
compensating for the in
uence of the cantilever air-gap. In this method we ensured that
the xed airgap was adjusted to the Root-Mean-Square (RMS) distance when account-
ing for paddle displacement. The inclusion of iterative displacement feedback did not
measurably in
uence the resulting output for the scale of geometries in our applications.
131
7.3 Design
Using the outlined electroacoustic model, we designed devices such that the resonance
frequency matched in both the acoustic and mechanical domain. The dimensions of the
cantilever paddle is the limiting parameter for the minimum size of the Helmholtz neck
cross section. Additionally, KOH etching was used for the neck of the resonator, so we
were constrained to having a 54.74
∘ slope in the sidewall of the neck. The other constraint
is the minimum cavity dimensions, which must be larger than the KOH opening. These
constraints are summarized as:

≥ + 2 +
2
/tan(54.74) (7.20)
≥
/
√
2 (7.21)
Therefore, the only design
exibility we have is to select a value for D, based on the
desired resonance frequency, and to select the value of the cavity volume, such that the
resonant frequencies of the mechanical and acoustic resonators overlap.
We performed a parametric study to determine optimal conditions for maximizing
peak magnitude response at the resonant frequency. It was concluded that only the
gap parameter d has a signicant in
uence on mechanical amplication. This result is
intuitive from the circuit model, as the resistive element dependent on d is responsible
for damping in the system. This is however surprising from the perspective of pressure
equalization in a typical acoustic transducer.
132
This design requires a fairly large resonance frequency in the ultrasonic range, due to
the constrains of the neck geometry and cavity volume. For a desired resonant frequency
of
0
=26 kHz, we calculate the paddle dimension D to be 363 for t = 2 . We can
now determine the parameters, given our optimization criterion and physical constraints
of the design, as shown in Table 7.1.
Parameter A
0
t L D d C H
Value (mm) 0.363 0.002 0.4 0.363 0.02 2.7 0.3
Table 7.1: Summary of the selected geometric design parameters (in mm) for Helmholtz-
resonator-enhanced MEMS microphone.
7.4 Multiphysics Modeling and Simulation
A multiphysics nite element model (FEM) based on the COMSOL acoustics module
was developed for this geometry. An optimized design was derived for the targeted 26
kHz resonance frequency of the air-acoustic system, with the selected dimensions shown
in Table 7.1.
The analytical model and simulation results closely agreed in their prediction of res-
onance frequency, giving a value of 25.8 kHz. The FEM simulation predicts that a sen-
sitivity improvement by a factor of 3.35 can be achieved with the addition of a resonant
acoustic cavity (Figure 7.4).
Although the analytical model considers volume of the cavity as a lump term and
neglects the actual geometry, simulation results indicated that a shallower volume such
as that of the proposed system produces a larger Q resonance whose center frequency
aligns more accurately with analytical predictions.
133
Figure 7.4: Graph showing simulated dierence in microphone sensitivity with and with-
out the use of Helmohltz resonator for passive amplication.
7.5 Fabrication and Implementation
The Helmholtz-enhanced acoustic transducer was fabricated using a two-wafer approach.
An SOI wafer with 2 thick device layer is used for fabrication of the cantilever-based
piezoelectric sensing element, using the process outlined in Figure 2.4. The backing cavity
of the MEMS piezoelectric microphone, serving as the Helmholtz resonance chamber, was
fabricated on a separate standard silicon wafer, which was then bonded to the SOI wafer to
complete the acoustic system. The chamber can be fabricated using several methods, two
of which are outlined in this section. The critical requirement of the chamber fabrication
process is precise control over the volume of the resulting cavity.
134
KOH-Based Cavity Fabrication
A simple method to microfabricate an acoustic chamber on a silicon wafer is the use of
anisotropic KOH etching. Because etching will stop on the < 111 > sidewalls of silicon,
the lateral dimensions of the resulting cavity can be exactly predicted. Silicon nitride
serves as an etch mask for KOH wet etching of silicon. The main disadvantage of this
method is that depth of the etch is controlled as a timed process, so the etching rate must
be greatly reduced. This is performed at room temperature with a concentration of 35%
to slow down the etching time and etching time is about. The resulting wafer is bonded
to the wafer containing the MEMS microphones.
Figure 7.5: Diagram showing the fabrication process of Helmholtz resonator cavity using
KOH as a silicon etchant.
Hemispherical Cavity Fabrication
An attempt was made to fabricate hemispherical cavities using an isotropic silicon etching
method described in [9]. Thick LPCVD-deposited silicon nitride serves as an etch mask
for a solution with 59% hydro
ouric acid, 70% nitric acid and 99.5% acetic acid (HNA).
135
The solution is mixed with a ratio of 4:7:11 and the wafer is etched for about 90 minutes
at a temperature of 30
∘ .
Figure 7.6: Diagram showing fabrication process of hemispherical cavity using HNA as a
silicon etchant.
The benet of HNA etching is the repeatability and consistency in etched volume
provided by the self-limiting etching behavior. This consistency however was not precise
enough to ensure a that acoustic resonance overlapped with mechanical resonance in
several trials. Due to diculty controlling nal etch volume, anisotropic silicon etching
using KOH was instead used as the preferred method, however HNA etching remains an
attractive option.
Wafer Bonding
Bonding was performed using adhesive SU-8 bonding as described in [10]. This method
uses a third 'transfer' wafer, by which spin-on SU-8 photo-resist is transferred to the
bonding interface but does not ll up the acoustic cavities.The adhesive SU-8 layer has a
thickness of approximately 10 . Then the processed 'microphone' and 'cavity' wafers
136
are compressed together with a uniform force of about 500 N at a temperature of 200
∘ C.
It is worth noting that a number of other low-temperature bonding methods would also
be suitable; this includes eutectic bonding, anodic bonding, thermocompressive bonding
and glass frit bonding. The completed wafers are then diced as shown in Figure 7.7. For
electrical testing the devices are wire-bonded to a pre-amplier PCB and packaged to
protect against electromagnetic interference.
Figure 7.7: SEM image of the cross section of an adhesively bonded device with Helmholtz
resonant cavity and neck.
7.6 Results
Device measurements were performed in an anechoic chamber with a GRAS 40AO cali-
brated microphone serving as a pressure reference. The PMUT output is connected with
an Op-Amp based preamplier circuit and enclosed in an aluminum box for electromag-
netic shielding.
The sensitivity of the fabricated PMUT was measured both before and after bonding
of the Helmholtz air cavity. The unamplied sensitivity of the MEMS transducer alone
was measured to be 0.297 mV/Pa (at 29 kHz with Q-factor of 34.9) and the sensitivity
after bonding of the second wafer was 0.64 mV/Pa (at 29 kHz with Q-factor of 73.0).
137
This corresponds to an improvement in sensitivity by a factor of 2.16, through the use of
a matched resonant air cavity.
Figure 7.8: Graph showing measured dierence in microphone sensitivity with and with-
out the use of Helmohltz resonator for passive amplication.
The implemented device demonstrated the concept of using a Helmholtz resonator
acoustic cavity for sensitivity enhancement with results closely following those predicted
by simulation. Although an amplication factor of 2.16 was the largest measured with
this design, other devices in the same fabrication batch exhibited larger sensitivities,
with the largest showing nearly a 3 times larger microphone sensitivity. Of the measured
devices, the acoustic transducers alone had an average sensitivity of 0.44 mV/Pa, and the
Helmholtz-enhanced devices had an average sensitivity of 1.10 mV/Pa; this corresponds
138
to an average gain of 2.5X. Fabrication variations are also responsible for microphone
resonance frequencies ranging from about 24 kHz to 2.9 kHz.
Future Work
The presented Helmholtz resonance-enhanced acoustic transducer can be improved fur-
ther with additional studies into parameter optimization. Preliminary measurements
indicated the potential for sensitivity enhancement of up to two orders of magnitude [11].
These measurements were performed for lower frequency MEMS devices, which re-
quired the use of a macro-scale Helmholtz resonator. These were fabricated using acrylic
sheets with a laser etcher, with cylindrical cavity sections. Two such devices are shown
in Figure 7.9. The gure on the left has a resonance freuency of 1.57 kHz and approxi-
mate dimensions of 0.7 cm X 0.7 cm X 1.9 cm. The gure on the right has a resonance
frequency of 520 Hz and the device has approximate dimensions of 3.8 cm X 3.8 cm X
1.9 cm.
Figure 7.9: Photograph the fabricated low-frequency Helmholtz resonators, with(left)
1.57 kHz resonant frequency, and (right) 520 Hz Helmholtz resonant frequency.
139
The devices were characterized using a laser Doppler displacement meter (LDDM) to
measure the cantilever tip displacement before and after the addition of the Helmholtz
resonator (Figure 7.10).
Figure 7.10: Measured tip displacement of 1.57 kHz resonance device (left) and 520 Hz
device (right).
The transducer with resonant frequency of 1.57 kHz exhibits an increase in peak
displacement from 0.124 to 1.38 (a factor of 11.14). The transducer with resonant
frequency of 520 Hz exhibits an increase in peak displacement from 0.227 to 3.15
(a factor of 13.9).
These studies also suggest a possibility for much larger sensitivity enhancement for
the microfabricated devices. Preliminary simulations indicated that ratio between neck
and chamber dimensions of Helmholtz resonator will signicantly impact the exhibited
boost to sensitivity. One potential use of this feature is in allowing us to eectively
shift the resonance frequency of microfabricated transducers to sub-kilohertz levels by
coupling high-frequency resonant microphones with low-frequency Helmholtz resonators,
which may improve yield at the cost of sensitivity and miniaturization.
140
7.7 Summary
The use of the Helmholtz resonance eect has been comprehensively modeled, and demon-
strated through simulation and measurement of fabricated devices. The method is shown
to be eective for passively amplifying the sensitivity of resonant acoustic sensors.
High sensitivity resonant microphones have been previously demonstrated for noise-
robust and low-power acoustic signature recognition, and the integration of Helmholtz
acoustic resonators into the resonant microphone fabrication process can further improve
performance with minimal additional processing steps.
141
References
[1] A. A. Shkel and E. S. Kim, \Helmholtz Pmut: Demonstrating Passive Amplica-
tion in Microfabricated Acoustic Transducers," Solid-State Sensors, Actuators and
Microsystems Workshop, pp. 60{61, 2016.
[2] Y. Tomimatsu, H. Takahashi, T. Kobayashi, K. Matsumoto, I. Shimoyama, T. Itoh,
and R. Maeda, \A piezoelectric cantilever with a Helmholtz resonator as a sound
pressure sensor," Journal of Micromechanics and Microengineering, vol. 23, no. 11,
p. 114003, 2013.
[3] S. B. Horowitz, M. Sheplak, L. N. Cattafesta III, and T. Nishida, \A MEMS acoustic
energy harvester," Journal of Micromechanics and Microengineering, vol. 16, no. 9,
p. S174, 2006.
[4] H. Takahashi, A. Suzuki, E. Iwase, K. Matsumoto, and I. Shimoyama, \MEMS
microphone with a micro Helmholtz resonator," Journal of Micromechanics and
Microengineering, vol. 22, no. 8, p. 085019, 2012.
[5] M. Kleiner, Electroacoustics. CRC Press, 2013.
[6] L. L. Beranek, \Handbook of Acoustics," American Journal of Physics, vol. 77,
pp. 1197{1197, dec 2009.
[7] H. A. Tilmans, \Equivalent circuit representation of electromechanical transducers
.1. Lumped-parameter systems," Journal of Micromechanics and Microengineering,
vol. 6, pp. 157{176, 1996.
[8] W. J. Palm, Mechanical vibration. Wiley Online Library, 2006.
[9] H. Zhang and E. S. Kim, \Dome-shaped diaphragm microtransducers," in Micro
Electro Mechanical Systems, 1995, MEMS’95, Proceedings. IEEE, p. 256, IEEE,
1997.
[10] L. Yu, F. E. H. Tay, G. Xu, B. Chen, M. Avram, and C. Iliescu, \Adhesive bonding
with SU-8 at wafer level for micro
uidic devices," Journal of Physics: Conference
Series, vol. 34, no. 1, pp. 776{781, 2006.
[11] A. A. Shkel and E. S. Kim, \Acoustic micro-resonator utilizing hemispherical air cav-
ity for sensitivity enhancement," 2015 IEEE International Ultrasonics Symposium
(IUS), pp. 1{4, 2015.
142
Chapter 8
Zero-Power Wireless Authentication with Film-Bulk
Acoustic Resonators
In this chapter we propose a passive system containing for integrated circuit authentica-
tion, containing an embedded tamper sensor, RFID tag, and antenna, to allow for rapid
wireless scanning of ICs. The system is designed to detect signs of tampering activity and
produce a permanent change to a MEMS-resonator-based RFID tag that can be wireless
interrogated. We demonstrated the wireless tamper detection concept through device
fabrication, simulation, and measurement.
8.1 Background
Counterfeiting of ICs is a growing global problem, costing an estimated annual loss of $7.5
billion for US-based semiconductor companies alone [1]. Sensitive electronics, such as dis-
tributed sensors in military applications, require a tamper-sensitive detection mechanism
to prevent an adversary from modifying, reverse engineering, or reusing the system.
143
Figure 8.1: System diagram of the wireless tamper detection scheme. An external RFID
reader transmits a sinusoidal signal and measures the backscattered signal strength from
the tamper detection chip. The strength of the backscattered signal depends on the
relationship of impedances between the RFID dipole antenna and series-connected FBAR.
A piezoelectric energy harvester, connected to the FBAR via a rectier and capacitor,
generates a large voltage spike when a sucient mechanical impulse is applied during the
detaching process of the chip. This voltage spike will irreversibly break down the FBAR
and permanently alter the characteristic of the backscattered signal.
Existing methods of assessing whether an IC is a counterfeit involve time-consuming
imaging tests, electrical verication, or destructive testing. Additionally, existing authen-
ticity verication methods cannot be applied on a bulk-scale.
IC-level tamper detection sensors have been proposed, for example with integrated
photo-detectors [2], to destroy critical program information through detection of tam-
pering activity, but these rely on power from the IC to function. The use of random
number generators for generating cryptographic keys, such as those based on tamper-
sensitive MEMS arrays [3], is another common counterfeit-prevention technique, however
these are only helpful in preventing reverse-engineering and not sensitive to repackaging
attempts by the counterfeiter.
144
8.2 Device
A common technique employed by counterfeiters to remove an IC from a printed circuit
board (PCB) involves raising the IC to an elevated temperature to melt the solder,
and then striking the PCB with forceful impacts to remove the IC chip. The proposed
system, illustrated in Figure 8.1, converts these impulses to large voltage pulses using
a piezoelectric energy harvester, and irreversibly alters the spectral characteristic of an
RFID tag. Such a system can be embedded within an IC package to passively detect
tampering activity, and can be designed to be extremely dicult for counterfeiters to
reverse the detection mechanism.
The main components of the RFID tag system are the RFID antenna, series-connected
FBAR and energy harvester. The resonant frequency of the FBAR (that depends on the
thicknesses of the layers) serves as the encoded ID of the tag. The RFID tag is designed to
function passively, allowing the integrity of the chip to be assessed without any electrical
connection, special conguration, precise alignment, or other measurement complexities
that existing tamper detection mechanisms rely on.
8.2.1 Film Bulk Acoustic Resonator (FBAR)
The FBAR is designed such that it is permanently destroyed through dielectric breakdown
when exposed to a large voltage. This breakdown will permanently alter the spectral
characteristics of the FBAR, which results in an identiable change in spectral shape of
the interrogated signal.
145
Figure 8.2: (a) Cross-section diagram of the FBAR device, which is composed of silicon
nitride (SiNX), zinc oxide, and aluminum layers on a silicon wafer. (b) Photograph of
the FBAR prior to tampering, and (c) photograph of the same FBAR after permanent
dielectric breakdown induced by tampering.
The FBAR is fabricated on a silicon wafer coated with LPCVD Silicon Nitride (SiN

)
(Figure 8.2a). The silicon nitride serves as an etch mask for KOH wet etching, which is
used to open a window on the bottom of the wafer and to form a silicon nitride supporting
diaphragm. After forming the diaphragm, a bottom electrode is deposited with 0.2 m
thick evaporated aluminum, and patterned. A piezoelectric ZnO layer is then deposited
with RF sputtering such that the thickness is half of the wavelength of the target resonant
frequency, for longitudinal waves traveling in bulk ZnO. In this study we target a resonant
frequency of 2.6 GHz, which requires a ZnO layer with a thickness of about 0.54 m.
Finally, an aluminum top electrode is deposited and is patterned along with ZnO. The
active portion of the resonator, which sits on the supporting diaphragm, has a pentagonal
shape in the lateral plane to minimize the lateral resonances and spurious modes. A top
146
view photograph of a completed FBAR is shown in Figure 8.2b.
The breakdown voltage of the FBAR is determined by the thickness of the FBAR's
ZnO layer. The literature on dielectric breakdown voltage of ZnO is limited and con-

icting. However, our measurements agree with with the reported values of breakdown
strength of AlN, which is structuraly similar to ZnO, and who's breakdown strength is
reported to be about 17 kV/mm [4]. For a more sensitive FBAR device that is easier to
force into breakdown, it is necessary to increase the interrogation frequency, which would
limit the interrogation distance and may increase hardware complexity of the interroga-
tor. The top-down photograph of the device after the dielectric breakdown of ZnO is
shown in Figure 8.2c.
Figure 8.3: Diagram showing the charging and discharging paths of the energy harvester.
An additional capacitor is required to ensure that a sucient current is developed to
permanently breakdown the FBAR and to leverage multiple impacts.
The FBAR is connected with the energy harvester through a full-wave rectier to
ensure a xed polarity of the potential developed across the FBAR (Figure 8.3). A
capacitor is necessary to accumulate sucient charge to induce permanent breakdown
in the FBAR. This threshold charge was experimentally determined to be a minimum
of about 12 C for a 2.6 GHz FBAR with a minimum threshold voltage applied (5 V).
During tampering activity, multiple impacts may be used to remove the IC, so such a
147
capacitor can leverage the multiple voltage peaks generated by the energy harvester to
meet the current threshold. The capacitor cannot be too large, however, since the time for
charging to the threshold voltage would be too long and also because a large capacitance
would complicate impedance matching with the antenna.
We characterized the breakdown of an FBAR with a ZnO thin-lm thickness of
0.54 m with an experimental setup consisting of a Rigol DM3058 for current mea-
surement and a Rigol DP831A for voltage sweeping with a current limit of 2 mA. The
breakdown measurements are performed by sweeping the applied voltage from zero to
20 V, with sucient time given at each measurement for the current to stabilize. The
results of sweeping the applied voltage and observing the current through a single FBAR
are plotted in Figure 8.4, with error bars indicating the range of leakage currents observed
over a period of about 30 seconds.
We can see that dielectric breakdown occurs when about 5 V is applied to the FBAR.
Prior to reaching this point, the voltage sweep is repeatable with identical results. Once
the breakdown condition of 5 V is reached, the current drawn through the FBAR is
consistently elevated above previous low voltage measurements, and the DC resistance
drops permanently by about 30 times. But with the current limit from the source,
the breakdown does not cause a thermal run away, and the DC resistance increases as
the applied voltage is increased until 13 V, at which point we observed a permanent
delamination of the electrodes from the FBAR, and the FBAR behaves like an open
circuit. This characteristic is repeatably observed in FBARs on a single wafer, but varies
with ZnO thickness and quality of the material from batch to batch.
The breakdown characteristic of the FBAR is time-dependent, consistent with other
148
Figure 8.4: Measured characterization of FBAR breakdown as a function of applied
voltage for a Zinc Oxide thickness of 0.54 m. Breakdown occurs in two phases, with
dielectric breakdown occurring at 5 V, and delamination of the metal layer occurring at
13 V. Time-dependent dielectric breakdown occurs in the region between 5 V and 13 V,
with delamination occurring given sucient exposure time.
time-dependent dielectric breakdown mechanisms, such as those in the gate oxide of
MOSFETs [5]. This time-dependency can create a breakdown and even delamination at
low voltages when exposed over long periods of time. This phenomenon was observed in
2.6 GHz FBARs when 2 V was applied over a 1 to 2 minute period. In the context of
tamper detection, this is not a dependable breakdown mechanism, since pulses supplied
by a piezoelectric energy harvester will be short in duration.
The output of piezoelectric energy harvesters is primarily characterized by high volt-
age and low current. For this reason, we also characterized the breakdown curve as a
149
Figure 8.5: Plot of measured current through FBAR under a step function with a 20 V
DC voltage applied. The FBAR draws a current with a peak of 0.141 mA for a duration
of 25 ms before permanent breakdown (delamination).
function of time after applying a voltage step from zero to 20 V (Figure 8.5). The peak
current of 0.141 mA
ows through the FBAR, and the current lasts for about 25 ms
before the electrodes are delaminated.
Integrating this pulse indicates that a total charge of about 1.76 C is necessary for
the FBAR breakdown with a 20 V potential. The total charge necessary for breakdown
increases with a reduced voltage peak, and design of a high-voltage output piezoelectric
energy harvester is more practical than optimizing current output at low peak voltages.
The high frequency electrical properties of the FBAR are characterized (Figure 8.6),
and the FBAR is measured to have a resonant frequency of 1.34 GHz with a quality factor
of 29.5 at the rst peak, and a resonant frequency of 2.59 GHz with a quality factor of
91.2 at the secondary peak. After breakdown, the resistance of the FBAR is increased at
150
Figure 8.6: Plots of measured (a) FBAR resistance, (b) FBAR reactance, and (c) esti-
mated quality factor before and after the breakdown of the FBAR, showing a destruction
of resonance eects due to tampering.
151
high frequencies, and the capacitance is reduced.
Figure 8.7: (a) Diagram illustrating the meandering dipole design of the RFID antenna
with key dimensions. (b) Plot of simulated S11 parameter of the RFID antenna.
8.2.2 RFID Antenna
We designed a meandering dipole RFID antenna (Figure 8.7a) and simulated its elec-
tromagnetic characteristics with the SONNET software suite (Figure 8.7b), to ensure a
complementary impedance and a matched antenna resonant frequency of 2.6 GHz. To
further reduce the dimensions of such a dipole antenna to t within an IC package,
additional meandering can be introduced into the design, for example with spiral loop
geometries or fractal geometries [6]. The antenna impedance with this design is primarily
capacitive, but can be made more inductive by using T-matching. An alternative antenna
design is an inductive loop-based tag, which will provide better impedance matching to
the FBAR, but reduce the overall range of the RFID antenna to near-contact.
152
Figure 8.8: Circuit diagram showing the method of estimating the backscattered signal
power. The re-radiated power is a function of the backscatter coecient K and the
antenna gain G. K is the relative strength of the backscattered power as a function of the
antenna impedance and FBAR impedance.
We estimate the backscattered signal strength using the circuit model shown in Fig-
ure 8.8, and the following expressions for re-radiated power and backscatter coecient
K:

−
=

(8.1)
=
4

2
|

+

|
2
(8.2)
From the given expression, we see that the power of the re-radiated signal is dependent
on the gain of the antenna, and the backscatter coecient K. The backscatter coecient
is a function of the antenna impedance and FBAR impedance, with a maximum value
achieved when the two are complex conjugates. The estimated impedance of the an-
tenna is obtained from a software simulation, and is applied to the expression along with
the measured FBAR parameters to estimate the backscatter coecient before and after
breakdown of the FBAR, shown in Figure 8.9.
The tag ID in this system is encoded in the spectral signature of the re-radiated signal.
153
Figure 8.9: Estimated backscatter coecient K based on the simulated RFID antenna
impedance and measured FBAR impedances before and after the breakdown of the
FBAR.
Figure 8.9 shows that the original backscattered signal will have two closely-spaced peaks
at 2.59 GHz and 2.70 GHz, when the FBAR and antenna are designed to have overlapping
frequencies. After the breakdown, the FBAR no longer exhibits high-Q resonance, and
the overall capacitance increases. This increases the impedance mismatch between the
FBAR and RFID antenna, creating a low-Q peak with a center frequency shifted to 3.04
GHz, and an 11 times weaker maximum backscattered power.
This change in spectral signature can be identied with three receiver measurements
of backscattered signal intensity at 2.59 GHz, 2.70 GHz, and the local minimum between
them at 2.61 GHz. Common passive RFID interrogators function with tag IDs embedded
within time-varying backscattered signals, which are typically implemented with surface
acoustic wave (SAW) re
ectors or active circuits. For this application, since the spectral
154
characteristics of the backscattered signal are the indentifying "tag" of the RFID system,
the RFID interrogator must be implemented to read frequency-encoded tags. Such an en-
coding scheme has been demonstrated to be eectively interrogated with ultra-wideband
pulses [7] and frequency-tunable RFID readers [8].
8.2.3 Integrating with Energy Harvester
A prototype tamper detection system was built with an energy harvester based on a
piezoelectric PZT bimorph cantilever, fabricated with two bulk PZT layers separated
with a spun-on adhesive layer. The energy harvester has dimensions of 11 mm x 1 mm x
0.6 mm, which can be further reduced to t within a given IC package. With this device,
an open circuit voltage of about 2.97 V was measured when 2.6 g acceleration is applied.
Since impact pulses of about 50 g are applied during tampering activity, we estimate
that the energy harvester will generate about 56 V during the tampering. Since only 5 V
is necessary to cause irreversible breakdown of the 2.6 GHz FBAR, there is a sucient
margin for the energy harvester with a similar bimorph geometry to be miniaturized to a
such level to be incorporated into IC package. If the generated voltage and current was
suciently large, the circuit shown in Figure 8.3 may not be necessary.
The cantilever has total dimensions of 11 mm x 1 mm x 0.6 mm. Such a cantilever can
generate an open circuit voltage of approximately 56 V due to tampering activity. Other
types of energy harvesters can be easily integrated with this system, for potentially greater
reliability and robustness. For example, to detect the temperature rise of a de-soldering
attempt, a piezoelectric cantilever can double as a pyroelectric voltage generator. PZT is
a reliable pyroelectric material with a pyroelectric coeecient of 300
− 2

− 1
, however
155
greater power density can be acheived with a material such as lead magnesium niobate -
lead titanate (PMN-PT) with a pyroelectric coecient of 1071
− 2

− 1
[9, 10]. In this
application a low Curie temperature, such as that of PMN-PT at 121
∘ C, is not a major
concern since the pyroelectric coecient rises near T

and we expect sucient potential
to be generate for the FBAR to be broken after reaching this target temperature.
8.3 Results and Discussion
our ultimate goal is to integrate the FBAR-based RFID tag system system with a mi-
crofabricated energy harvester as a tamper detector, such that the entire system can be
embedded within a typical IC package. This can be supported by increasing the FBAR
resonant frequency to 5 GHz, which would reduce the ZnO thickness of the FBAR and
lower the necessary breakdown voltage to about 3 V. This would allow us to relax the
constraints on the energy harvester design and further miniaturize the energy harvester.
Operating at a higher frequency would also allow miniaturization of the antenna, with
the disadvantage of reducing the range for wireless interrogation.
This tamper detection scheme has an advantage over other anti-counterfeiting mea-
sures, in that FBAR's exact resonant frequency (which depends on the ZnO thickness)
must be known (down to parts-per-thousand level for Q of 1,000) for a counterfeiter
to roll-back or duplicate that tamper detection system after the permanent breakdown.
Once a tamper detection system is destroyed, the exact resonant frequency cannot be
recovered, since ZnO thickness variation over a wafer is typically more than 1%.
A similar tamper detection system could also be implemented with a SAW-based
156
device. Since the resonance frequency of a SAW lter is determined by the spacing
of interdigitated electrodes, dielectric breakdown will permanently destroy all surface
patterns. Use of SAW lters will also allow integration of re
ectors to enable temporally-
encoded tag IDs [11]. Multi-mode resonators have also been proposed as spectral RFID
tags with distributed resonant frequencies [12], and such devices can similarly be coupled
with RFID antennas and energy harvesters as volatile RFID tags.
8.4 Summary
Our simulations and measurement results are a promising preliminary demonstration of
the approach of using dielectric breakdown of FBAR structures for tamper detection in
the IC supply chain. It has been demonstrated that FBARs coupled with RFID antennas
can be used as a spectral RFID tag, and that such FBARs can be permanently damaged
with piezoelectric energy harvesters.
The proposed detection scheme can be similarly applied for wireless detection of dam-
age during package handling and shipping, and other impact detection applications. The
use of this tamper-detection mechanism can also be easily integrated with existing anti-
counterfeit measures, such as obfuscation, for more complete protection against many
types of attackers.
157
References
[1] F. Koushanfar, S. Fazzari, C. McCants, W. Bryson, M. Sale, P. Song, and M. Potkon-
jak, \Can EDA combat the rise of electronic counterfeiting?," DAC ’12: Proceedings
of the 49th Annual Design Automation Conference, no. November 2011, p. 133, 2012.
[2] D. Shahrjerdi, J. Rajendran, S. Garg, F. Koushanfar, and R. Karri, \Shielding and
securing integrated circuits with sensors," in IEEE/ACM International Conference
on Computer-Aided Design, Digest of Technical Papers, ICCAD, vol. 2015-Janua,
pp. 170{174, 2015.
[3] J. English, D. Coe, R. Gaede, D. Hyde, and J. Kulick, \MEMS-assisted cryptography
for CPI protection," IEEE Security and Privacy, vol. 5, no. 4, pp. 14{21, 2007.
[4] T. Ruemenapp and D. Peier, \Dielectric breakdown in aluminium nitride," High
Voltage Engineering Symposium, pp. 373{376, 1999.
[5] C. H. Ho, S. Y. Kim, and K. Roy, \Ultra-thin dielectric breakdown in devices and
circuits: A brief review," Microelectronics Reliability, vol. 55, no. 2, pp. 308{317,
2015.
[6] G. Marrocco, \The art of UHF RFID antenna design: Impedance-matching and size-
reduction techniques," IEEE Antennas and Propagation Magazine, vol. 50, no. 1,
pp. 66{79, 2008.
[7] P. Kalansuriya, N. C. Karmakar, and E. Viterbo, \On the detection of frequency-
spectra-based chipless RFID using UWB impulsed interrogation," IEEE Transac-
tions on Microwave Theory and Techniques, vol. 60, no. 12, pp. 4187{4197, 2012.
[8] S. Preradovic and N. C. Karmakar, \Multiresonator based chipless RFID tag and
dedicated RFID reader," in IEEE MTT-S International Microwave Symposium Di-
gest, pp. 1520{1523, 2010.
[9] C. R. Bowen, J. Taylor, E. LeBoulbar, D. Zabek, A. Chauhan, and R. Vaish, \Py-
roelectric materials and devices for energy harvesting applications," Energy & Envi-
ronmental Science, vol. 00, no. 12, pp. 1{21, 2014.
[10] S. P. Alpay, J. Mantese, S. Trolier-Mckinstry, Q. Zhang, and R. W. Whatmore,
\Next-generation electrocaloric and pyroelectric materials for solid-state electrother-
mal energy interconversion," MRS Bulletin, vol. 39, no. 12, pp. 1099{1109, 2014.
[11] V. P. Plessky and L. M. Reindl, \Review on SAW RFID tags," in IEEE Transactions
on Ultrasonics, Ferroelectrics, and Frequency Control, vol. 57, pp. 654{668, 2010.
[12] M. Ramezani, A. R. Newsome, and M. Ghatge, \A Nanomechanical Identication
Tag Technology for Traceablility and Authentication Applications," in Hilton Head
Workshop 2018: A Solid-State Sensors, Actuators and Microsystems Workshop,
(Hilton Head, SC), 2018.
158
Chapter 9
Conclusion and Future Direction
In this dissertation, we demonstrated several devices and concepts that advance the state
of zero-power sensing and systems. The demonstrated devices include low power acoustic
signature detection techniques based on resonant MEMS microphone array, piezoelec-
tric vibration energy harvesters, passive acoustic amplication of microphones using an
acousto-mechanical Helmholtz resonance phenomenon, and an integrated circuit authen-
tication system based on lm-bulk acoustic resonators and micro-scale voltage generators.
These systems were conceptualized, designed, modeled, fabricated, and experimentally
veried.
9.1 Microphone Array for Signal Processing
In this thesis, we demonstrated a novel signal processing method based on resonant micro-
phone arrays, and showed that such a method could be used to improve F1 score in binary
classication accuracy in noisy environments over a typical digital lter implementation.
The method was validated for speech recognition as well as for wearable monitoring of
159
respiratory sounds, with classication demonstrated to be eective with Gaussian mixture
models, k-nearest neighbor, support vector machines, and naive Bayesian classication,
and spectro-temporal features classied with a time-delay neural networks.
In all classication tests, the use of a resonant array was shown to be robust to out-
of-band noise that would typically be encountered in these applications. With cepstral
features computed from 40 ms frames, we measured up to a 57.8% increase in F1 score
by using a resonant-array processing method for a signal-to-noise ratio of -26 dB. Using
spectro-temporal cepstral features classied with a dense neural network, improvements
of up to 58.8% were measured for signal-to-noise ratios of -26 dB over an equivalent digital
lter implementation.
A complete sensing and processing system was developed, implemented in hardware,
and evaluated to integrate array-based respiratory sound sensing. The implementation is
evaluated for supply power and memory requirements. With an ultra-low power processor
with 16 kB SRAM and 24 MHz processing speed, a resonant-array based respiratory
classication system is demonstrated with classication cycles completing in a 0.46 second
period with an average power consumption of 0.596 mW, a factor of 11.1 improvement
over a purely digital implementation.
A challenge of developing a wearable health monitoring system, and proposed future
work of this research, is to design and fabricate very low-frequency MEMS resonators. Al-
though we have experimentally demonstrated that resonant frequencies as low as 500 Hz
can be reliably produced with a paddle-like design, it remains a challenge to implement
lower frequency resonators with reliable fabrication yields, due to the large aspect ra-
tios that would be necessary. Several approaches, such as mass weighting, innovative
160
anchor design, and use of low-stiness materials were evaluated as alternative methods
to reduce resonance frequency while maintaining a high sensitivity and quality factor.
Low-frequency MEMS microphones with a resonant frequency of no more than 300 Hz
have been designed, but have not yet been experimentally validated.
In addition to the applications discussed in this thesis, the described microphone array
can be implemented for use in noise cancellation in other real-time auditory systems, such
as hearing aids. Preliminary work has been completed to integrate a neural network-
based control system with a multi-channel resonant array, such that noisy channels are
suppressed and channels containing speech are amplied in real-time. Preliminary eorts
indicated this to be a promising method, however diculty lies in providing accurate
labeling of training data for an eective machine learning system. A simple threshold
detection channel-suppression algorithm can be implemented as a simpler alternative,
with active noise cancellation of noisy microphone channels.
9.2 Piezoelectric Energy Harvesting
Our eorts in the development of piezoelectric energy harvesters have demonstrated the
eectiveness of PZT bimorphs in generating large voltages from vibration sources. Simple
low-frequency PZT bimorph energy harvesters were shown to deliver powers of 48.9 W
to a matched load under 1.0g vibrations.
Two non-linear bistable energy harvesters have been developed. A buckling-type de-
sign on a
exible piezoelectric substrate was developed and demonstrated a 0.873 W/cc
power density at 1.0g. A bistable energy harvester using magnetic coupling to generate
161
large displacements was shown to generate a power density of 35.3 / with a
at
low-frequency response.
We demonstrated the integration of piezoelectric energy harvesting on a system-level
for powering sensing platforms and for enabling on-chip tamper detectors. A suggested
future research direction is the miniaturization of these energy harvester geometries for
on-chip integration and batch fabrication.
9.3 Helmholtz Resonator for Passive Amplification
We demonstrated the design, modeling, and fabrication of zero-power sensitivity ampli-
cation of microphones through the use of a micro-fabricated acousto-mechanical Helmholtz
resonator that can be integrated with existing microphone technology without signi-
cantly increasing fabrication complexity or cost. We demonstrated that such a geometry
can increase the sensitivity and quality factor of MEMS microphones by a factor of 2-14,
depending on the geometry and target frequency of the device.
Future eorts should concentrate on developing an analytical estimation for the gain
that such a passive amplication system can provide. Our existing models can be used
to accurately predict resonant frequency, but our current methods for measuring ampli-
cation have not agreed with experiments.
There is potential to utilize Helmholtz resonators on micro-scales for applications other
than amplication of microphones. Examples of such applications are tuning microphone
resonance frequency, manipulating frequency characteristics, and for large bandwidth
amplication of arrays for system-level objectives such as ltering high-frequency acoustic
162
interference and suppressing harmonics. Evidence for these other potential applications
has been observed in benchtop experiments, but has not been characterized or explored
in further detail, so it is suggested as future work.
9.4 FBAR-Based Tampering Detection
We demonstrated the preliminary design and experimental verication of a zero-power
system for wireless assessment of IC integrity, through the use of a Film-Bulk Acoustic
Resonator as a volatile RFID tag. A full system has been designed, and the functional
mechanisms have been validated through simulation of antenna properties and RFID
behavior, as well as measurement of FBAR behavior when exposed to critical voltage
and current supplies. The critical voltage was determined to be 5 V for a 2.6 GHz device,
and such a generated pulse was demonstrated to permanently alter the FBAR-based
RFID spectral characteristics.
The majority of the eorts in this research are left as future work, including the
bulk of eorts in the development of a miniaturized RFID antenna, miniaturized tamper-
detection energy harvester, and empirical assessment of the wireless system integration.
Additional eorts should be made toward miniaturization of the RFID antenna by ex-
ploring fractal antenna layouts, 3-dimensional antenna geometries, or assessing receive
sensitivity with a miniaturized antenna operating at a sub-harmonic frequency. This
task can be simplied by reducing the operating frequency of the system, at the expense
of interrogation range.
163
The tamper-detector can also be greatly optimized through the use of a batch micro-
fabrication process rather than a pick-and-place assembly process. Studies into adding
mass-loading at a micro-scale, stacking of piezoelectric layers, and through exploring al-
ternative tamper-detection mechanisms can further improve reliability and power density.
A promising potential alternative method for tamper detection is to implement a pyro-
electric voltage generator which responds to large heat gradients rather than mechanical
impact, which would be encountered during a typical de-soldering attempt by a coun-
terfeiter. Such a voltage generator can be implemented with any common piezoelectric
material such at PZT or Lithium Tantalate, or with a high-performance pyroelectric
material such as PMN-PT.
A proposed future branch of this research is to explore the use of a surface-acoustic
wave (SAW) resonator in place of FBARs to reduce the required breakdown eld of the
resonator. A SAW resonator device features interdigitated electrodes separated by / 4,
rather than by the/ 2 spacing in FBARs, which would allow further miniaturization and
relaxation of constraints on the required energy harvesting sensor.
Finally, while this research has served as a preliminary system demonstration, the
electromagnetic behavior of the integrated antenna and FBAR system was estimated
analytically and through numerical simulation. The proposed system requires further
validation through experimental measurement of all integrated system components.
164
Bibliography
\AirSonea. A revolution in asthma management," 2018. [Online]. Available:
http://airsonea.com
S. P. Alpay, J. Mantese, S. Trolier-Mckinstry, Q. Zhang, and R. W. Whatmore, \Next-
generation electrocaloric and pyroelectric materials for solid-state electrothermal energy
interconversion," MRS Bulletin, vol. 39, no. 12, pp. 1099{1109, 2014.
J. Altmann, \Acoustic and seismic signals of heavy military vehicles for co-operative
verication," Journal of Sound and Vibration, vol. 273, no. 4, pp. 713{740, 2004.
S. R. Anton and D. J. Inman, \Vibration energy harvesting for unmanned aerial vehi-
cles," in SPIE 6928, Active and Passive Smart Structures and Integrated Systems 2008,
vol. 692824, 2008.
M. Bahoura and C. Pelletier, \Respiratory sounds classication using Gaussian mixture
models," Electrical and Computer Engineering, 2004. Canadian Conference on, vol. 3,
no. 2, pp. 1309{1312, 2004.
L. Baumgartel, A. Vafanejad, S. J. Chen, and E. S. Kim, \Resonance-enhanced piezo-
electric microphone array for broadband or preltered acoustic sensing," Journal of
Microelectromechanical Systems, vol. 22, no. 1, pp. 107{114, 2013.
S. P. Beeby, M. J. Tudor, and N. M. White, \Energy harvesting vibration sources for
microsystems applications," Measurement Science and Technology, vol. 17, no. 12, 2006.
L. L. Beranek, \Handbook of Acoustics," American Journal of Physics, vol. 77, no. 12,
pp. 1197{1197, dec 2009.
A. Bohadana, G. Izbicki, and S. S. Kraman, \Fundamentals of lung auscultation," New
England Journal of Medicine, vol. 370, no. 8, pp. 744{751, 2014.
C. R. Bowen, J. Taylor, E. LeBoulbar, D. Zabek, A. Chauhan, and R. Vaish, \Pyroelec-
tric materials and devices for energy harvesting applications," Energy & Environmental
Science, vol. 00, no. 12, pp. 1{21, 2014.
C. H. Chen, W. T. Huang, T. H. Tan, C. C. Chang, and Y. J. Chang, \Using K-
nearest neighbor classication to diagnose abnormal lung sounds," Sensors (Switzer-
land), vol. 15, no. 6, pp. 13 132{13 158, 2015.
P. L. Chen, R. S. Muller, and A. P. Andrews, \Integrated silicon pi-fet accelerometer
with proof mass," Sensors and Actuators, vol. 5, no. 2, pp. 119{126, 1984.
165
F. Chollet, \Keras," 2015. [Online]. Available: https://github.com/fchollet/keras
S. K. Chowdhury and A. K. Majumder, \Frequency analysis of adventitious lung
sounds," Journal of Biomedical Engineering, vol. 4, no. 4, pp. 305{312, 1982.
CMU Sphinx, \CMU Sphinx Project," pp. 1{3, 2009.
A. Cowen, PiezoMUMPs Design Handbook. MEMSCAP Inc., 2013.
E. Defa y, Integration of Ferroelectric and Piezoelectric Thin Films: Concepts and Ap-
plications for Microsystems, 2013.
J. M. Donelan, Q. Li, V. Naing, J. A. Hoer, D. J. Weber, and A. D. Kuo, \Biomechan-
ical energy harvesting: Generating electricity during walking with minimal user eort,"
Science, vol. 319, no. 5864, pp. 807{810, 2008.
B. W. Edwards, C. J. Struck, P. Dharan, and Z. Hou, \New digital processor for hearing
loss compensation is based on auditory system," Hearing Journal, vol. 51, pp. 38{50,
1998.
J. English, D. Coe, R. Gaede, D. Hyde, and J. Kulick, \MEMS-assisted cryptography
for CPI protection," IEEE Security and Privacy, vol. 5, no. 4, pp. 14{21, 2007.
C.-B. Eom and S. Trolier-McKinstry, \Thin-lm piezoelectric MEMS," MRS Bulletin,
vol. 37, no. 11, pp. 1007{1017, 2012.
K. Forkheim, D. Scuse, and H. Pasterkamp, \A comparison of neural network models for
wheeze detection," IEEE WESCANEX 95. Communications, Power, and Computing.
Conference Proceedings, vol. 1, no. 95, pp. 214{219, 1995.
G. Gautschi, \Piezoelectric Sensors," in Piezoelectric Sensorics, 2002, pp. 73{91.
K. L. Gerhard, \Non-intrusive vibration monitoring in US Naval and US Coast Guard
ships," Ph.D. dissertation, Massachusetts Institute of Technology, 2013.
M. Goel, \Electret sensors, lters and MEMS devices: New challenges in materials
research," Current science, vol. 85, no. 4, pp. 443{453, 2003.
P. Gupta, Y. Jeong, J. Choi, M. Faingold, A. Daruwalla, and F. Ayazi, \Precision
High-Bandwidth Out-Of-Plane Accelerometer as Contact Microphone for Body-Worn
Auscultation Devices," in Hilton Head Solid-State Sensors, Actuators and Microsystems
Workshop, Hilton Head, SC, 2018.
M. I. Haller and B. T. Khuri-yakub, \Ultrasonic Air Transducer," October, vol. 43, no. 1,
1996.
R. L. Harne and K. W. Wang, \A review of the recent research on vibration energy
harvesting via bistable systems," Smart Materials and Structures, vol. 22, p. 23001,
2013.
166
B. Herrera, E. Demirors, G. Chen, R. Guida, F. Pop, N. Dave, C. Cassella, T. Melodia,
and M. Rinaldi, \PMUT-Based High Data Rate Ultrasonic Wireless Communication
Link for Intra-Body Networks," in Hilton Head Solid-State Sensors, Actuators and Mi-
crosystems Workshop, Hilton Head, SC, 2018, pp. 105{108.
J. W. High and W. K. Wilkie, \Method of Fabricating NASA-Standard Macro-Fiber
Composite Piezoelectric Actuators," Technology, no. June, pp. 1{30, 2003.
C. H. Ho, S. Y. Kim, and K. Roy, \Ultra-thin dielectric breakdown in devices and
circuits: A brief review," Microelectronics Reliability, vol. 55, no. 2, pp. 308{317, 2015.
P. P. Holmes, E. M. Yeatman, and A. S, \Magnetic plucking of piezoelectric beams for
frequency up-converting energy harvesters," Smart Materials and Structures, vol. 23,
no. 2, p. 25009, 2014.
J. Hoople, J. Kuo, S. Ardanuc, and A. Lal, \Chip-scale sonic communication using AlN
transducers," in IEEE International Ultrasonics Symposium, IUS, 2013, pp. 1934{1937.
S. B. Horowitz, M. Sheplak, L. N. Cattafesta III, and T. Nishida, \A MEMS acoustic
energy harvester," Journal of Micromechanics and Microengineering, vol. 16, no. 9, p.
S174, 2006.
P. Juang, H. Oki, Y. Wang, M. Martonosi, P. Peh Li-Shiuan, and D. Rubenstein,
\Energy-Ecient Computing for Wildlife Tracking: Design Tradeos and Early Ex-
periences with ZebraNet," Proceedings of the 10th International Conference on Archi-
tectural Support for Programming Languages and Operating Systems (ASPLOS 2002),
pp. 96{107, 2002.
P. Kalansuriya, N. C. Karmakar, and E. Viterbo, \On the detection of frequency-spectra-
based chipless RFID using UWB impulsed interrogation," IEEE Transactions on Mi-
crowave Theory and Techniques, vol. 60, no. 12, pp. 4187{4197, 2012.
H. Kara, R. Ramesh, R. Stevens, and C. R. Bowen, \Porous PZT ceramics for receiving
transducers," IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control,
vol. 50, no. 3, pp. 289{296, 2003.
A. Khaligh, P. Zeng, and C. Zheng, \Kinetic Energy Harvesting Using Piezoelectric and
Electromagnetic Technologies," Industrial Electronics, IEEE Transactions on, vol. 57,
no. 3, pp. 850{860, 2010.
S. I. Khan, \Respiratory Sound Analysis for Identifying Lung Diseases : A Review,"
International Journal of Science and Research (IJSR), vol. 3, no. 11, pp. 566{571, 2014.
D. P. Kingma and J. L. Ba, \Adam: a Method for Stochastic Optimization," Interna-
tional Conference on Learning Representations 2015, pp. 1{15, 2015.
M. Kleiner, Electroacoustics. CRC Press, 2013.
D. M. Kline and V. L. Berardi, \Revisiting squared-error and cross-entropy functions for
training neural network classiers," Neural Computing and Applications, vol. 14, no. 4,
pp. 310{318, 2005.
167
K. Kochetov, E. Putin, S. Azizov, I. Skorobogatov, and A. Filchenkov, \Wheeze Detec-
tion Using Convolutional Neural Networks," in Progress in Artificial Intelligence , vol.
8154, 2017, pp. 162{173.
F. Koushanfar, S. Fazzari, C. McCants, W. Bryson, M. Sale, P. Song, and M. Potkonjak,
\Can EDA combat the rise of electronic counterfeiting?" DAC ’12: Proceedings of the
49th Annual Design Automation Conference, no. November 2011, p. 133, 2012.
M. Kucic, P. Hasler, J. Dugger, and D. Anderson, \Programmable and adaptive ana-
log lters using arrays of
oating-gate circuits," in Proceedings - 2001 Conference on
Advanced Research in VLSI, ARVLSI 2001, 2001, pp. 148{162.
C. E. Kuehni, A. M. Brooke, and M. Silverman, \Prevalence of wheeze during childhood:
Retrospective and prospective assessment," European Respiratory Journal, vol. 16, no. 1,
pp. 81{85, 2000.
A. Leal, R. Couceiro, I. Chouvarda, N. Maglaveras, J. Henriques, R. Paiva, P. Carvalho,
and C. Teixeira, \Detection of dierent types of noise in lung sounds," IEEE Engineering
in Medicine and Biology Society, pp. 5977{5980, 2016.
D. E. Lee, I. Hwang, C. M. O. Valente, J. F. G. Oliveira, and D. A. Dornfeld, \Preci-
sion manufacturing process monitoring with acoustic emission," International Journal
of Machine Tools and Manufacture, vol. 46, no. 2, pp. 176{188, 2006.
J. Li, L. Deng, Y. Gong, and R. Haeb-Umbach, \An overview of noise-robust automatic
speech recognition," pp. 745{777, 2014.
B. S. Lin, H. D. Wu, S. J. Chen, G. E. Jan, and B. S. Lin, \Using Back-Propagation
Neural Network for Automatic Wheezing Detection," in Proceedings - 2015 International
Conference on Intelligent Information Hiding and Multimedia Signal Processing, IIH-
MSP 2015, 2016, pp. 49{52.
R. J. Littrell, \High Performance Piezoelectric MEMS Microphones," PhD Thesis, pp.
1{111, 2010.
H. Liu, C. J. Tay, C. Quan, T. Kobayashi, and C. Lee, \Piezoelectric MEMS energy
harvester for low-frequency vibrations with wideband operation range and steadily in-
creased output power," Journal of Microelectromechanical Systems, vol. 20, no. 5, pp.
1131{1142, 2011.
P. C. Loizou, Speech enhancement: theory and practice. CRC Press, 2007.
N. P. Ly, D. R. Gold, S. T. Weiss, and J. C. Celedon, \Recurrent wheeze in early
childhood and asthma among children at risk for atopy." Pediatrics, vol. 117, no. 6, pp.
e1132{8, 2006.
R. F. Lyon, \Analog Implementations of Auditory Models," pp. 212{216, 1991.
R. F. Lyon and C. Mead, \An analog electronic cochlea," IEEE Transactions on Acous-
tics, Speech, and Signal Processing, vol. 36, no. 7, pp. 1119{1134, 1988.
168
A. L. Maas, Q. V. Le, T. M. O'Neil, O. Vinyals, P. Nguyen, and A. Y. Ng, \Recurrent
Neural Networks for Noise Reduction in Robust ASR." Interspeech, pp. 3{6, 2012.
M. J. Madou, Fundamentals of Microfabrication: The Science of Miniaturization, 2002.
||, \Fundamentals of microfabrication and nanotechnology. Volume two: Manufac-
turing techniques for microfabrication and nanotechnology," 2011.
G. Marrocco, \The art of UHF RFID antenna design: Impedance-matching and size-
reduction techniques," IEEE Antennas and Propagation Magazine, vol. 50, no. 1, pp.
66{79, 2008.
F. D. Martinez, A. L. Wright, L. M. Taussig, C. J. Holberg, M. Halonen, and W. J.
Morgan, \Asthma and Wheezing in the First Six Years of Life," New England Journal
of Medicine, vol. 332, no. 3, pp. 133{138, 1995.
F. Massey, \The Kolmogorov-Smirnov test for goodness of t," Journal of the American
Statistical Association, vol. 46, no. 253, pp. 68{78, 1951.
C. Multiphysics, \Structural Mechanics Module - User's Guide," Tech. Rep., 2013.
S. Naifar, S. Bradai, C. Viehweger, and O. Kanoun, \Survey of electromagnetic and
magnetoelectric vibration energy harvesters for low frequency excitation," Measurement:
Journal of the International Measurement Confederation, vol. 106, pp. 251{263, 2017.
M. Namiki, T. Hamamoto, and S. Hangai, \Spoken word recognition with digital cochlea
using 32 DSP-boards," in Acoustics, Speech, and Signal Processing, 2001. Proceed-
ings.(ICASSP’01). 2001 IEEE International Conference on, vol. 2. IEEE, 2001, pp.
969{972.
I. Ozer, Z. Ozer, and O. Findik, \Noise robust sound event classication with convolu-
tional neural network," Neurocomputing, vol. 272, pp. 505{512, 2018.
R. Palaniappan, K. Sundaraj, and N. U. Ahamed, \Machine learning in lung sound
analysis: A systematic review," pp. 129{135, 2013.
W. J. Palm, Mechanical vibration. Wiley Online Library, 2006.
H. Pasterkamp, \The highs and lows of wheezing: A review of the most popular adven-
titious lung sound," pp. 243{254, 2018.
F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blon-
del, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau,
M. Brucher, M. Perrot, and

E. Duchesnay, \Scikit-learn: Machine Learning in Python,"
Journal of Machine Learning Research, vol. 12, pp. 2825{2830, 2012.
G. Piazza, P. J. Stephanou, and A. P. Pisano, \One and two port piezoelectric higher
order contour-mode MEMS resonators for mechanical signal processing," Solid-State
Electronics, vol. 51, no. 11-12, pp. 1596{1608, 2007.
PixSoft, \The RALE repository." [Online]. Available: http://www.rale.ca
169
V. P. Plessky and L. M. Reindl, \Review on SAW RFID tags," in IEEE Transactions
on Ultrasonics, Ferroelectrics, and Frequency Control, vol. 57, 2010, pp. 654{668.
R. X. A. Pramono, S. Bowyer, and E. Rodriguez-Villegas, \Automatic adventitious
respiratory sound analysis: A systematic review," PLoS ONE, vol. 12, no. 5, 2017.
S. Preradovic and N. C. Karmakar, \Multiresonator based chipless RFID tag and dedi-
cated RFID reader," in IEEE MTT-S International Microwave Symposium Digest, 2010,
pp. 1520{1523.
L. Rabiner and B.-H. Juang, \Fundamentals of speech recognition," Signal Processing
Series. Prentice Hall, Englewood Cliffs, NJ, 1993.
M. Ramezani, A. R. Newsome, and M. Ghatge, \A Nanomechanical Identication Tag
Technology for Traceablility and Authentication Applications," in Hilton Head Work-
shop 2018: A Solid-State Sensors, Actuators and Microsystems Workshop, Hilton Head,
SC, 2018.
T. Ruemenapp and D. Peier, \Dielectric breakdown in aluminium nitride," High Voltage
Engineering Symposium, pp. 373{376, 1999.
D. E. Rumelhart, G. E. Hinton, and R. J. Williams, \Learning representations by back-
propagating errors," Nature, vol. 323, no. 6088, pp. 533{536, 1986.
A. Safari and E. K. Akdoan, Piezoelectric and acoustic materials for transducer appli-
cations, 2008.
Y. Sakashita, T. Ono, H. Segawa, K. Tominaga, and M. Okada, \Preparation and elec-
trical properties of MOCVD-deposited PZT thin lms," Journal of Applied Physics,
vol. 69, no. 12, pp. 8352{8357, 1991.
R. Sarpeshkar, C. Salthouse, J. J. Sit, M. W. Baker, S. M. Zhak, T. K. Lu, L. Turicchia,
and S. Balster, \An ultra-low-power programmable analog bionic ear processor," IEEE
Transactions on Biomedical Engineering, vol. 52, no. 4, pp. 711{727, 2005.
M. L. Seltzer, D. Yu, and Y. Wang, \An investigation of deep neural networks for noise
robust speech recognition," in 2013 IEEE International Conference on Acoustics, Speech
and Signal Processing, 2013, pp. 7398{7402.
N. Sengupta, M. Sahidullah, and G. Saha, \Lung sound classication using cepstral-
based statistical features," Computers in Biology and Medicine, vol. 75, pp. 118{129,
2016.
D. Shahrjerdi, J. Rajendran, S. Garg, F. Koushanfar, and R. Karri, \Shielding and
securing integrated circuits with sensors," in IEEE/ACM International Conference on
Computer-Aided Design, Digest of Technical Papers, ICCAD, no. January, 2015, pp.
170{174.
J. Shao, \Linear model selection by cross-validation," Journal of the American statistical
Association, vol. 88, no. 422, pp. 496{494, 1993.
170
A. A. Shkel, L. Baumgartel, and E. S. Kim, \A resonant piezoelectric microphone array
for detection of acoustic signatures in noisy environments," 2015 28th IEEE Interna-
tional Conference on Micro Electro Mechanical Systems (MEMS), vol. 2015-Febru, no.
February, pp. 917{920, 2015.
A. A. Shkel and E. S. Kim, \Acoustic micro-resonator utilizing hemispherical air cavity
for sensitivity enhancement," 2015 IEEE International Ultrasonics Symposium (IUS),
pp. 1{4, 2015.
||, \Helmholtz Pmut: Demonstrating Passive Amplication in Microfabricated
Acoustic Transducers," Solid-State Sensors, Actuators and Microsystems Workshop, pp.
60{61, 2016.
C. Soyer, E. Cattan, and D. Rmiens, \Electrical damage induced by reactive ion-beam
etching of lead-zirconate-titanate thin lms," Journal of Applied Physics, vol. 97, no. 11,
2005.
S. C. Stanton, C. C. McGehee, and B. P. Mann, \Reversible hysteresis for broadband
magnetopiezoelastic energy harvesting," Applied Physics Letters, vol. 95, no. 17, pp.
14{17, 2009.
||, \Nonlinear dynamics for broadband energy harvesting: Investigation of a bistable
piezoelectric inertial generator," Physica D: Nonlinear Phenomena, vol. 239, no. 10, pp.
640{653, 2010.
I. Steinwart, Support Vector Machines, 2010, vol. 13, no. 4.
S. S. Stevens, J. Volkmann, and E. B. Newman, \A Scale for the Measurement of the
Psychological Magnitu Pitch," The Journal of the Acoustical Society of America, vol. 8,
no. 3, pp. 185{190, 1937.
M. Stojanovic and J. Preisig, \Underwater acoustic communication channels: Propaga-
tion models and statistical characterization," Communications Magazine, IEEE, vol. 47,
no. 1, pp. 84{89, 2009.
T. L. Szabo, Diagnostic ultrasound imaging: inside out. Academic Press, 2004.
H. Takahashi, A. Suzuki, E. Iwase, K. Matsumoto, and I. Shimoyama, \MEMS micro-
phone with a micro Helmholtz resonator," Journal of Micromechanics and Microengi-
neering, vol. 22, no. 8, p. 085019, 2012.
H. A. Tilmans, \Equivalent circuit representation of electromechanical transducers .1.
Lumped-parameter systems," Journal of Micromechanics and Microengineering, vol. 6,
pp. 157{176, 1996.
Y. Tomimatsu, H. Takahashi, T. Kobayashi, K. Matsumoto, I. Shimoyama, T. Itoh, and
R. Maeda, \A piezoelectric cantilever with a Helmholtz resonator as a sound pressure
sensor," Journal of Micromechanics and Microengineering, vol. 23, no. 11, p. 114003,
2013.
171
S. Trolier-Mckinstry and P. Muralt, \Thin lm piezoelectrics for MEMS," pp. 7{17,
2004.
M. J. Vellekoop, C. C. O. Visser, P. M. Sarro, and A. Venema, \Compatibility of zinc
oxide with silicon IC processing," Sensors and Actuators: A. Physical, vol. 23, no. 1-3,
pp. 1027{1030, 1990.
M. Vijaya, Piezoelectric materials and devices: Applications in engineering and medical
sciences. CRC Press, 2016.
C. Wang, Q. Zhang, and W. Wang, \Low-frequency wideband vibration energy harvest-
ing by using frequency up-conversion and quin-stable nonlinearity," Journal of Sound
and Vibration, vol. 399, pp. 169{181, 2017.
L. P. Wang, R. Wolf, Q. Zhou, S. Trolier-McKinstry, and R. J. Davis, \Wet-etch Pattern-
ing of Lead Zirconate Titanate (PZT) Thick Films for Microelectromechanical Systems
(MEMS) Applications," Materials Research, vol. 657, pp. 1{6, 2001.
W. Wang, W. Yang, Z. Liu, Y. Lin, S. Zhou, H. Qian, F. Gao, and G. Li, \Epitaxial
growth of high quality AlN lms on metallic aluminum substrates," CrystEngComm,
vol. 16, no. 20, pp. 4100{4107, 2014.
J. Weigold, T. Brosnihan, J. Bergeron, and X. Zhang, \A MEMS Condenser Microphone
for Consumer Applications," 19th IEEE International Conference on Micro Electro Me-
chanical Systems, no. January, pp. 86{89, 2006.
R. D. White and K. Grosh, \Design and Characterization of a MEMS Piezoresis-
tive Cochlear-like Acoustic Sensor," in ASME International Mechanical Engineering
Congress and Exposition, vol. 2002, 2002, pp. 1{10.
R. B. Williams, D. J. Inman, M. R. Schultz, M. W. Hyer, and W. K. Wilkie, \Nonlinear
Tensile and Shear Behavior of Macro Fiber Composite Actuators," Journal of Composite
Materials, vol. 38, pp. 855{869, 2004.
N. M. Wilson, \Virus infections, wheeze and asthma," pp. 184{192, 2003.
R. Xu, H. Akay, and S.-G. Kim, \Micro Buckled Beam Based Ultra-Low Frequency
Vibration Energy Harvester," in Hilton Head Solid-State Sensors, Actuators and Mi-
crosystems Workshop, Hilton Head, SC, 2018.
T. Yemliha, \Performance and Memory Space Optimizations for Embedded Systems,"
Ph.D. dissertation, Syracuse University, 2011.
L. Yu, F. E. H. Tay, G. Xu, B. Chen, M. Avram, and C. Iliescu, \Adhesive bonding
with SU-8 at wafer level for micro
uidic devices," Journal of Physics: Conference Series,
vol. 34, no. 1, pp. 776{781, 2006.
H. Zhang and E. S. Kim, \Dome-shaped diaphragm microtransducers," in Micro Electro
Mechanical Systems, 1995, MEMS’95, Proceedings. IEEE. IEEE, 1997, p. 256.
172
J. Zhang, \Designing a cost-eective and reliable pipeline leak-detection system," Pipes
and Pipelines International, vol. 42, no. 1, pp. 20{26, 1997.
S. Q. Zhang, Y. X. Li, and R. Schmidt, \Modeling and simulation of macro-ber com-
posite layered smart structures," Composite Structures, vol. 126, pp. 89{100, 2015.
173
Appendix A
Digital Signal Processing Methods
This appendix includes a detailed reference of signal processing techniques and methods
for parties interested in repeating the results presented in this thesis, or continuing this
line of research. As many dierent techniques were used on many dierent platforms, this
appendix serves as a high-level documentation of the tools necessary for repeating the
classication and power-consumption measurement tasks presented in Chapters 4 and 5
respectively.
The source code and test les can be obtained from the USC MEMS Group FTP
server, in ftp: //mems.usc.edu/Archive/Former Members/Anton/dspcode.rar. The
code used in this thesis is a combination of python and C scripts for feature extraction,
classication, data management, and data visualization. Two LabView routines are also
included for automated measurement tasks and microphone classication. Please contact
an active member of the USC MEMS research group for access to the source code.
174
LabVIEW Data Acquisition and Measurement
Custom LabView measurement routines were designed to automate all measurement
tasks, saved in the ftp: //mems.usc.edu/Archive/Former Members/Anton/LabVIEW di-
rectory. LabVIEW is able to interface with the ROGA DAQ as an audio device, which
appears in the device manager as a 2-channel playback and 2-channel recording device
called USB Audio Device. The measurement routine handles output from the ROGA
DAQ with a Play Sound File VI, and input with a Acquire Sound VI.
The input to the automated LabView measurement is a text le, with each line in
the le referencing the audio les to be played-back and recorded in a *.wav format with
the rst channel dedicated to signal data, and the second channel dedicated to noise
data. The RMS intensities of both channels of audio should be equalized such that the
base SNR is 0 dB, but this is not a requirement since the output intensities should be
characterized prior to reporting an SNR value. The same LabVIEW interface provides
an option for a separate characterization mode which translates the volume settings into
signal intensity from each of the channels.
A base directory is specied in the LabVIEW measurement GUI for the input and
output directories. A list of channel audio volumes should be specied, and the routine
will sweep through all specied combinations of signal volume and noise volume. The
input le will be read and played through the speakers from the input base + file
name directory, and the output will be output base + ’/breathing X/’ + file name
where X is the specied volume of the playback signal.
175
Python-Based Classification
After acquiring recorded.wav data, feature extraction and classication is performed with
a combination of C-language-based feature extraction, and python-based data manage-
ment and classication. It is recommended that the Anaconda package (https: //anaconda.org)
of python libraries for version 2.7 is used to run the provided functions, for the best com-
patibility. The provided python packages are designed to be a
exible research platform
and provided as a reference. They are not intended to be used as a black box, so modi-
cation of the existing scripts and ne tuning is necessary for any custom uses.
The process for classication of the acquired data is:
1. Conversion of the .wav recorded audio obtained through LabVIEW into comma-
separated value .csv les, with the proper sample rate and temporal scaling factor.
This is done with the wav to csv function in the data functions.py le. An input
audio file path and output output file path must be specied. Optionally,
a tempo-scaling factor Kspeed, amplitude-scaling factor scale, and new sample
rate resample rate can be specied. Examples of batch le conversion with this
function are provided in the bulk function() function.
2. A python function run feature extractor in data functions.py automates the
process of feature extraction for each of the input les and folders, and gener-
ates .csv output les containing a spreadsheet of the extracted features. An in-
put directory and output directory must be specied, with the variable array set
to True if an array input is used (rather than a single
atband input), and the
array suffix must list the input folder sux to sweep through with the batch
176
process. The function itself must be modied with any changes to the input
le list, le lengths (in number of samples), labeling of classes, and directories.
Additionally, a compiled feature extractor .exe le must be specied within the
function. Pre-compiled feature extractors exist for 7 resonant lter pre-processing
(Test Recognition v2 fbf array7.exe), 7 digital lter processing
(Test Recognition v2 fbf gras7.exe), and 13 mel-spaced digital lter processing
(Test Recognition v2 fbf single13mel.exe), provided in the cbin directory. In
all cases sample rate is 5120 S/s, frame length is 40 ms with 50% overlapping, and
lter distributions are identical to those specied in Chapter 4. Any changes to the
feature extraction parameters, such as sample rate, frame length, and lter distri-
bution, require modication and re-compilation of the C feature extraction code,
which is provided in the csource directory.
3. The classification functions.py le contains functions to perform batch classi-
cation and calculation of performance metrics, utilizing the Scikit-Learn library
for GMM, k-NN, SVM, or Naive Bayes classication, or the Keras library for neural
network classication. The bulk ANN and bulk classification functions perform
a batch classication with a neural network model or other classication model re-
spectively. Thebulk classification function takes as input a string parameter to
specify whether k-NN, GMM, SVM, or Naive Bayesian classication is performed.
Both functions take as input a.csv feature le for training and a list of .csv feature
les for testing. The data is divided according to split ratio, with the specied
proportion of frames assigned to the training set, and the remainder assigned to
177
the test set. This split is randomized, but the randomization seed is held constant
across the test les to ensure that an equal comparison is made.
The source of the feature extraction, and C-language Naive Bayes classication im-
plementation are provided in the csource directory. The feature functions.py le
is necessary to include for .csv le management, splitting of features into a testing and
training set, and implementing neural networking with time-delay-based spectro-temporal
inputs. The feature graphics.py le is not required, but it contains functions that are
useful for visualization of lter distributions, spectrograms, and spectral distributions of
audio les.
PSoC Hardware Implementation
The Cypress PSoC 4200 BLE, that is used for the power consumption experiments in
this thesis, implements a slightly modied variation of the feature-extraction code im-
plemented in C for PC classication. The most signicant dierence is that it does not
have a input le and output le interface, but rather uses a minimum amount of system
memory by creating a data buer which interfaces with the ADC hardware.
A software package provided by Cypress called PSoC Creator is used to program
the hardware frontend, including ADC, multi-channel input, and wireless output, and
connecting these hardware functions to the micro-controller. Two versions are imple-
mented for the array processing method and conventional digital processing method, and
are included in the PSoC directory. Programming of the PSoC hardware should function
without any dependencies, provided that the software and drivers are updated.
178
Communication with Bluetooth LE is implemented for a one-way broadcast of data,
but a receiver application has not been developed. Cypress provides a CySmart applica-
tion for gathering statistics and debugging bluetooth devices and services. Performance
of classication accuracy on the PSoC can be assessed with a wired USB UART or a
custom Bluetooth receiver application, neither of which is implemented in the provided
les.
179

Abstract (if available)

Abstract This dissertation presents several micro-electromechanical (MEMS) sensors and devices based on thin-film piezoelectric materials to enable zero-power and ultra-low-power intelligent systems in power-constrained scenarios. ❧ A MEMS resonant microphone array has been developed and evaluated as a mechanical filter-bank front end for speech recognition and respiratory monitoring experiments. These experiments consistently demonstrate robustness to ambient noise relative to traditional digital signal processing methods. With cepstral features computed from 40 ms frames, we measured up to a 57.8% increase in F1 score by using a resonant-array processing method for a signal-to-noise ratio of -26 ㏈. Using spectro-temporal cepstral features classified with a dense neural network, improvements of up to 58.8% were measured for signal-to-noise ratios of -26 ㏈ over an equivalent digital filter implementation. ❧ A complete sensing and low-power signal processing system was developed and evaluated to integrate array-based respiratory sound sensing, vibration energy harvesting, and wireless transmission of data upon detection of wheezing. Using an ultra-low power processor with 16 kB SRAM and 24 MHz processing speed, a resonant-array based respiratory classification system was implemented with classification cycles completing in a 0.46 second period with an average power consumption of 0.596 mW, a factor of 11.1 improvement over a typical implementation. ❧ A method for passively amplifying the sensitivity of the developed microphone arrays was hypothesized, modeled, fabricated, and experimentally validated. The method, based on a micro-fabricated Helmholtz resonator cavity, was shown to improve peak sensitivity and quality factor of resonant microphones by up to 13.9 in centimeter-scale devices, and by up to 2.16 in micro-scale devices. ❧ A zero-power wireless authentication system based on FBARs was fabricated, simulated, and experimentally evaluated as a unique method for wireless and passive detection of tampering activity within integrated circuits. This proof-of-concept system has a RFID interrogation frequency of 2.6 GHz, and an energy harvester generating a 5 V pulse was demonstrated to permanently alter the RFID spectral characteristics. Piezoelectric energy harvesters were developed on both bulk ceramic and flexible substrates, and were characterized for harvesting energy from mechanical vibrations. ❧ These demonstrations of low-power systems based on MEMS resonators and thin-film piezoelectrics provide several creative solutions to emerging power-constrained applications, including wearable health monitoring, distributed sensor nodes, and internet-of-things.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Battery-less detection and recording of tamper activity along with wireless interrogation

PDF

Piezoelectric MEMS for acoustic sensing, contactless RF switching, and resonant mass sensing

PDF

MEMS piezoelectric resonant microphone arrays and their applications

PDF

Wineglass mode resonators, their applications and study of their quality factor

PDF

Piezoelectric ultrasonic and acoustic microelectromechanical systems (MEMS) for biomedical, manipulation, and actuation applications

PDF

Audio and ultrasound MEMS built on PZT substrate for liquid and airborne applications

PDF

Integrated wireless piezoelectric ultrasonic transducer system for biomedical applications

PDF

Additive manufacturing of piezoelectric and composite for biomedical application

PDF

Experimental and analytical studies of infrastructure systems condition assessment using different sensing modality

PDF

Highly integrated 2D ultrasonic arrays and electronics for modular large apertures

PDF

Microfluidic cell sorting with a high frequency ultrasound beam

PDF

Nano-fabricated devices in electrochemistry and cancer therapy

PDF

Design and use of a biomimetic tactile microvibration sensor with human-like sensitivity and its application in texture discrimination using Bayesian exploration

PDF

Resonant light-matter interactions in nanophotonic structures: for manipulating optical forces and thermal emission

PDF

Learning and control for wireless networks via graph signal processing

PDF

Slurry based stereolithography: a solid freeform fabrication method of ceramics and composites

PDF

Memristor for parallel and analog data processing in the era of big data

PDF

Low-dimensional asymmetric crystals: fundamental properties and applications

Asset Metadata

Creator Shkel, Anton Andreevich (author)

Core Title Zero-power sensing and processing with piezoelectric resonators

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Electrical Engineering

Publication Date 10/09/2018

Defense Date 07/02/2018

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag acoustic signature,Acoustics,energy harvesting,film bulk acoustic resonator,health monitoring,health sensing,lung auscultation,MEMS,Microphones,OAI-PMH Harvest,piezoelectric,resonators,sensors,signal classification

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Kim, Eun Sok (committee chair), Narayanan, Shrikanth (committee member), Shung, Kirk (committee member)

Creator Email anton.shkel@gmail.com,shkel@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-87668

Unique identifier UC11675341

Identifier etd-ShkelAnton-6804.pdf (filename),usctheses-c89-87668 (legacy record id)

Legacy Identifier etd-ShkelAnton-6804.pdf

Dmrecord 87668

Document Type Dissertation

Format application/pdf (imt)

Rights Shkel, Anton Andreevich

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