University of Southern California Dissertations and Theses

Enhancing thermal comfort: data-driven approach to control air temperature based on facial skin temperature

(USC Thesis Other)

Enhancing thermal comfort: data-driven approach to control air temperature based on facial skin temperature

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content 1

Enhancing Thermal Comfort: Data-Driven Approach to Control Air
Temperature Based on Facial Skin Temperature

By

Mengqi Jia

Presented to the
FACULTY OF THE
SCHOOL OF ARCHITECTURE
UNIVERSITY OF SOUTHERN CALIFORNIA
In partial fulfillment of the
Requirements of degree
MASTER OF BUILDING SCIENCE

May 2019

2

ACKNOWLEDGEMENTS

This research acknowledges National Science Foundation (NSF) for funding the project, entitled “Human-Building
Integration: Bio-Sensing Adaptive Environmental Controls” under grant number #1707068. This research has been
conducted with the support of the University of Southern California (USC) School of Architecture by providing the
environmental chamber as the place for experiments. And the Burohappold also supported the experiment with their
office.
I would like to express my sincere gratitude to my committee chair, Professor Joon-Ho Choi. He organized the
experiment and provided great idea and suggestion for the thesis and directed the research in the right direction. I
would also like to express my sincere appreciation to my committee members, Professor Karen Kensek, and Yolanda
Gill. They spent a lot of time for revising the paper and add the comments. I really appreciate their valuable advice on
improving the project and my writing skills. I would like to express my thanks to my lovely MBS family. Without
your support, I would not complete this thesis project while enjoying such wonderful and memorable time.

3

COMMITTEE MEMBERS

Joon-Ho Choi, Ph.D. LEED AP BD+C
Assistant Professor
USC School of Architecture
joonhoch@usc.edu

Karen M. Kensek, LEED AP BD+C
Professor of the Practice of Architecture
USC School of Architecture
kensek@usc.edu

Yolanda Gil
Research Professor
USC School of Engineering
gil@isi.edu

4

ABSTRACT
Heating, ventilation, and air-conditioning (HVAC) systems play an essential role in supporting building functions.
However, they consume an enormous amount of energy in buildings. Energy consumption and thermal comfort are
two significant factors in the design process of HVAC systems. Although reducing energy consumption is
important, one must balance that with thermal comfort. Accurate HVAC systems control can provide a better
thermal environment and save energy. The limitations of the current thermal models promote the study of thermal
comfort predictions.
A tool was created with the help of an electric engineering student to achieve automatically control the HVAC system
based on an occupant’s facial skin temperature to provide a better thermal environment. The program lasted for three
years. In the first stage, research assistants build up foundation of the environment chamber and data collection systems.
After the fundamental research of facial skin temperature and thermal comfort, this stage aimed to use machine
learning method to achieve the automatic HVAC system control based on facial skin temperature.
Human subject experiments were conducted in both an environment chamber and a real office building. Different
parameters about the occupants were collected during the experiments by sensors, such as the facial skin
temperature, heart rate, and electrode activity. The parameters of the environment were also collected, including air
temperature, relative humidity, and carbon dioxide level. Two air conditioners and four heaters controlled the
experimental conditions; the temperature change ranged from 20 ℃ to 30 ℃. There were 20 subjects in the first-
round experiments and 6 subjects in each validation experiment turn. For the first-round experiments, the
experiment time was 90 minutes, and every 15 minutes, the subject was given the thermal perception survey. For the
validation experiments in the real office, the time was 135 minutes, and the survey interval was 5 minutes.
Two different machine learning algorithms were used. The gradient boosting was used to select the input features
for the control models. The back propagation neural network was used to make the predictions for the thermal
perceptions and control the HVAC systems.
For the data analysis of the first-round experiments, the 4-cross-validation analysis reported that the average
prediction accuracy for thermal comfort is 95.6% and 95.2% for thermal sensation in the gradient boosting
algorithm. Moreover, in the validation experiments, the data analysis proved that the model prediction performance
was about 80.4%. The system control validation experiments showed the average time of neutral condition was
79.7%, and it performed better for male subjects than female subjects. A prediction accuracy above 80% is an
acceptable number for the validation experiments. The results of the analysis showed that the potential of the
accurate predictions of the thermal condition is reliable.

KEYWORDS: thermal condition, occupants’ feedback, machine learning algorithm, automatic control
HYPOTHESIS
It is possible to improve the indoor thermal environment through automatic HVAC system control based on the
occupant’s facial skin temperature.
RESEARCH OBJECTIVES
1. Develop an automatic system by utilizing physiological signals to help control indoor air temperature.
2. Determine the relationship between air temperature and the thermal comfort of occupants.
3. Determine the relationship between facial skin temperature and thermal comfort.
5

Contents
ACKNOWLEDGEMENTS .......................................................................................................... 2
COMMITTEE MEMBERS ......................................................................................................... 3
ABSTRACT ............................................................................................................................. 4
1. INTROUDUCTION ............................................................................................................... 8
1.1 Problems in the Current Thermal Comfort Evaluation Method ............................................ 8
1.2 The Relationship between Facial Skin Temperature and Thermal Comfort ......................... 10
1.3 Introduction of HVAC Systems Control .............................................................................. 11
1.4 Applications of Machine Learning Method in the Building Domain .................................... 12
1.5 Summary ......................................................................................................................... 13
2. LITERATURE REVIEW .................................................................................................... 15
2.1 The Development of Thermal Comfort Evaluation ............................................................. 15
2.2 Important Factors of Thermal Environment ...................................................................... 16
2.3 The Study of the Relationship between Facial Skin Temperature and Thermal Comfort ..... 17
2.4 Automatic Control Techniques in HVAC Systems ............................................................... 18
2.5 Applications of Machine Learning Methods in Building Domain ........................................ 18
2.6 Review of the Human Building Integration Lab in Building Science .................................... 19
2.7 Summary ......................................................................................................................... 19
3. METHODOLOGY ........................................................................................................... 21
3.1 Environment Chamber ..................................................................................................... 21
3.1.1 Previous Design Introduction ............................................................................................... 22
3.1.2 HVAC Systems ....................................................................................................................... 22
3.1.3 Data Collection Sensors ........................................................................................................ 22
3.1.4 Data Acquisition System....................................................................................................... 24
3.1.5 Control System Device.......................................................................................................... 26
3.2 Thermal Perception Survey Design ................................................................................... 26
3.2.1 Personal Information Question ............................................................................................ 26
3.2.2 Thermal Preference Question .............................................................................................. 27
3.3 Design of the Experiment Process ..................................................................................... 28
3.3.1 Training Data Collection Experiment in the Environment Chamber .................................. 28
3.3.2 Model Validation Experiment in a Real Office .................................................................... 30
3.3.3 Automatic HVAC System Control Validation Experiment in the Experiment Chamber ..... 31
6

3.4 Data Analysis ................................................................................................................... 31
3.4.1 Data Analysis Software ........................................................................................................ 31
3.4.2 Data Preprocessing ............................................................................................................... 34
3.4.3 Machine Learning Model Training ....................................................................................... 35
3.4.4 Machine Learning Model Validation .................................................................................... 35
3.5 Summary ......................................................................................................................... 35
4. TRAINING DATA COLLECTION EXPERIMENT DATA AND RESULTS ................................... 36
4.1 Moving Average Time Window Selection .......................................................................... 36
4.2 Group Data Analysis ......................................................................................................... 37
4.2.1 Thermal Perception Data Analysis ....................................................................................... 37
4.2.2 General Machine Learning Model Comparison ................................................................... 41
4.2.3 HRV Data Analysis ................................................................................................................ 44
4.3 Individual Data Analysis ................................................................................................... 45
4.3.1 Facial Skin Temperature and Air Temperature Analysis ..................................................... 45
4.3.2 Electrode Activity (EDA) Data Analysis ................................................................................ 51
4.4 Input Features Selection and Importance Comparison ...................................................... 55
4.4.1 Input Features Comparison for the Thermal Comfort Prediction ....................................... 55
4.4.2 Input Features Comparison for the Thermal Sensation Prediction .................................... 57
4.4.3 Input Features Performance Testing ................................................................................... 58
4.5 Machine Learning Algorithm Comparison ......................................................................... 60
4.5.1 Machine Learning Algorithm Comparison for Thermal Comfort Prediction ...................... 60
4.5.2 Individual Machine Learning Model Training ...................................................................... 62
4.6 Summary ......................................................................................................................... 63
5. VALIDATION EXPERIMENTS DATA AND RESULTS ........................................................... 64
5.1 Model Validation Experiments in the Real Office .............................................................. 64
5.2 System Validation Experiments in the Environment Chamber ........................................... 79
5.3 Summary ......................................................................................................................... 80
6. Conclusion ................................................................................................................... 81
6.1 Conclusion of the Human Subject Experiments ................................................................. 81
6.2 Limitations....................................................................................................................... 82
6.3 Future Work .................................................................................................................... 82
6.4 Conclusion ....................................................................................................................... 83
7

REFERENCE .......................................................................................................................... 84
APPENDIX A: CODE OF MACHINE LEARNING ALGORITHM .................................................... 89

8

1. INTROUDUCTION
Buildings play an essential role in human life. Most people spent more than 80% of their daily time in the indoor
environment (Casillas & Cordon, 2003). Indoor environment quality (IEQ) is significant for the occupants. The IEQ
is determined by different factors inside buildings, such as air quality, lighting, thermal conditions, and acoustics.
Thermal comfort is a significant factor in IEQ. Heating, cooling, and air-conditioning (HVAC) systems are essential
to provide the thermal environment inside buildings.
Energy usage of HVAC systems accounts for more than half of the total energy usage in the building domain
(Wickramaratne, 2015). An effective method in HVAC system control is necessary to reduce energy consumption
and at the same time to provide a better thermal environment. When thermal comfort becomes an important factor
for adjusting HVAC systems, predicting and controlling thermal comfort also becomes important.
The conventional thermal condition evaluation method is the predicted mean vote (PMV)–predicted percentage of
dissatisfied (PPD) model. Individual features are not considered. This is the main deficiency of the PMV–PPD
model. The limitations promote the development of thermal preferences predictions based on individual features
such as gender and even skin temperature.
Another consideration is the HVAC systems control. The primary function of HVAC systems is to provide a
suitable indoor thermal condition. The problem of the HVAC system control is offering a better thermal
environment while reducing energy usage. The conventional HVAC system temperature setpoint is based on the
building code and standards. The setpoint of the HVAC system is the operating temperature of the HVAC systems.
However, the studies show weak correlations between the standards and the occupant’s reported temperature
(Barlow & Fiala, 2007).
Based on the previous discussion, it is necessary to find a more effective way of measuring the thermal comfort of
the occupants and controlling HVAC systems. Machine learning has the vast potential to provide accurate control
for HVAC systems.
This chapter documents the introduction of problems in the current thermal evaluation method, the relationship
between facial skin temperature and thermal sensation, the HVAC system control strategy, and the applications of
machine learning methods in building domain. All technical terms are introduced at the beginning of the section.
1.1 Problems in the Current Thermal Comfort Evaluation Method
Thermal comfort is the satisfaction level of the thermal environment in the occupant’s subjective evaluation. It can
be influenced by air temperature, air speed, relative humidity, and other personal factors of the occupants
(ASHRAE, 2010). Thermal comfort is an obvious factor that influences the IEQ. It influences the occupants’ health
and productivity. Some studies reported the robust correlation between thermal environment and the productivity of
occupants. The occupants have higher productivity in the satisfied thermal environment (Tarantini, Pernigotto, &
Gasparella, 2017).
There are several thermal comfort evaluation models, some developed for the indoor environment. The conventional
model of thermal comfort prediction is the PMV–PPD method. The PMV–PPD method is a mathematical model
driven from the experiment to define the thermal comfort level. The calculation theory is based on heat balance
equations (de Dear et al., 2013). The equation of the PMV–PPD model includes air temperature, mean radiant
temperature, relative humidity, air speed, metabolic rate, and clothing insulation.
The PMV model calculates the thermal comfort level of occupants in a specific environment combination (air speed,
air temperature, relative humidity, and other factors are included in the PMV equation). In the initial experiments for
developing the PMV model, the survey has a seven-point scale for occupants to report their thermal comfort level.
The range is from -3 (cold) to 3 (hot).
9

The result of the PMV model is a number between -3 and 3. The ideal outcome is 0, which indicates the occupant in
the neutral thermal condition.
The PPD model was developed based on the PMV model. The PPD model is an equation to predict the percentage
of a large group of people dissatisfied with a specific indoor environment. The PMV index reveals the mean
response of a large number of people for thermal preference. The PPD index indicates the number of people who
have a dissatisfied thermal condition. The higher PPD corresponds with the number far from 0 in the PMV index
(Fig. 1.1) (Pourshaghaghy & Omidvari, 2012).

Figure 1. 1 The relation between PMV index and PPD index (Pourshaghaghy & Omidvari, 2012)
The PMV–PPD tool developed by UC Berkeley illustrates the main concepts of the PMV–PPD method. The left
part includes the parameters that have influence on thermal preferences. Operative temperature, air speed, humidity,
metabolic rate, and clothing level can be changed into real conditions. The right part shows the results of
calculations. The pink square indicates the comfort zone in setting-up situations (Fig. 1.2).

10

Figure 1. 2 Thermal comfort tool based on the PMV–PPD Model
With further studies on thermal comfort, researchers have investigated the limitation of the PMV–PPD model. On
one hand, the indoor environment is not a steady situation in most buildings because of the outside weather
conditions and inside equipment adjusting to it. The PMV–PPD model regards it as a constant situation, and this
causes inaccuracy in predictions. On the other hand, the human body is a complex machine; different people have
different metabolic rate even if they are in the same activity and are of the same gender and age. The PMV–PPD
model uses the average sensation level of people and cannot make accurate predictions during the personal thermal
comfort experiments (Barlow, 2002).
With the increasing requirement for the accurate thermal comfort model, the dynamic thermal condition converts to
the new research topic in building science (Tanabe, Kobayashi, Nakano, Ozeki, & Konishi, 2002). Some studies
have suggested an individual control model for thermal states. The tendency changes from stable thermal comfort
study to dynamic thermal condition, which emphasizes the function of occupants and the dynamic environment
change in the building. The study reported high potential for energy usage reduction for HVAC systems based on
real-time occupant measurements (Goyal, Barooah, & Middelkoop, 2015). The thermal condition is not always
optimal because the settings of the HVAC systems include limited participants of occupants’ need. An interactive
system for controlling the HVAC systems based on the feedback of the occupants can save 20% more energy
compared with the standard driven HVAC system control method. Moreover, accounting the feedback from
occupants can improve their satisfaction (Murakami, Terano, Mizutani, Harada, & Kuno, 2007).
1.2 The Relationship between Facial Skin Temperature and Thermal Comfort
Thermoregulation plays an essential role in the thermal preferences of the subject. Human thermoregulation is the
process for humans to maintain a stable core internal temperature. The normal range for human internal temperature
is 98° F (37° C) to 100° F (37.8° C) (Zhang, Huizenga, Arens, & Yu, 2001). The mechanism of the body helps people
keep this temperature. If it is higher or lower than the baseline temperature, the human body will adjust it through
vasodilatation, vasoconstriction, sweating, or other actions. Human skin temperature plays an important role in the
thermoregulation process because the body has heat transfer through the skin to the environment.
11

The temperature in the face varies from area to area. The tissues of the face have low metabolism rate, and the heat
transfer from the skin mainly depends on the blood flow rate. Higher temperature is reported in the forehead and
lower in the cheek. The possible reason for this is that the metabolism of the brain keeps the forehead at a higher
temperature, and the fat of cheek keeps the heat convection with the environment (Ariyaratnam & Rood, 1990).
Some studies showed that the local temperature of humans correlates with their thermal conditions (Fiala, Lomas, &
Stohrer, 2001). A three-part series studies reported that the domestic thermal comfort and sensation correlates with
the whole-body thermal conditions in both a uniform and nonuniform environment (Zhang, Huizenga, Arenas, &
Wang, 2004). Besides, some studies indicated the potential of using body skin temperature to reflect the thermal
condition of the occupants (Choi & Yeom, 2017b).
Different body segments have different performance in thermal condition predictions (Choi & Yeom, 2017b).
Clothing condition has an influence on the local thermal sensation of the subjects (Dai, Zhang, Arens, & Lian,
2017). Consider the convenience and the prediction’s accuracy; the human face was widely selected as the
monitored part in recent studies (Ghahramani, Castro, Becerik-Gerber, & Yu, 2016a). According to the data analysis
using the Random Forest classification model, the prediction accuracy based on facial skin temperature is nearly
80% (Li, Menassa, & Kamat, 2018).
The main heat transfer between the human body and the environment is by adjusting blood speed and status. The
human facial skin has a high index of blood vessels and is uncovered with clothing (Ghahramani et al., 2016a). It is
a suitable testing point for predicting the thermal comfort of people. The study reported that using facial skin
temperature can be predicted for uncomfortable cold and uncomfortable warm conditions with 90% accuracy
(Ghahramani et al., 2016a).
1.3 Introduction of HVAC Systems Control
The mechanical system of heating, cooling, and ventilation is used to provide a satisfied indoor environment in
buildings (Homod, Mohamed Sahari, Almurib, & Nagi, 2012). Heating systems are usually used to increase the
environment temperature, while cooling systems decrease the environment temperature. On the other hand, the
ventilation systems are used to provide fresh air from the outside, which can ensure the indoor air quality.
An HVAC system equipment requires a control system to set up the operating time and temperature. The purpose of
the HVAC system control is to control the air temperature, relative humidity, ventilation, and pressure. The
thermostat is used to control the HVAC system to provide a stable indoor environment.
The design and control of HVAC systems is a significant part of building design. The operation energy usage of
HVAC systems accounts for a significant amount of the total energy consumption in commercial and residential
buildings (Martin, Federspiel, & Auslander, 2002). The fact is that overcooling in summer and overheating in winter
is a normal phenomenon in the United States. This phenomenon indicates that HVAC systems waste a lot of energy
with inefficient operating. The American Society of Heating, Refrigerating and Air-Conditioning Engineers
(ASHRAE) standard uses the PMV–PPD model to conform to the thermal comfort of the occupants (Kosonen &
Tan, 2004). However, it has limitations in predicting thermal comfort (Razmara, Maasoumy, Shahbakhti, &
Robinett, 2015). The improvement of HVAC systems control is a meaningful step in reducing the energy
consumption of the building domain. An automatic HVAC system control is a potential strategy to reduce energy
usage significantly. The development of computer science makes it possible to control HVAC systems in an
intelligent way. Some studies have reported that the machine learning method can help enhance the thermal
satisfaction of the occupants (Dai et al., 2017).
The manual method for HVAC systems control is to use the thermostat. The thermostat is part of the HVAC system,
which allows occupants to set up the temperature for the system (Wang, Zhang, Arens, & Huizenga, 2007). The
method is still widely used in commercial buildings and residential buildings. However, it has drawbacks, and it
needs to be set when the thermal comfort changes. So every time occupants want to change the indoor environment,
they need to adjust the temperature (Razmara et al., 2015).
12

Thermostat is a component to detect and control the temperature of any equipment by heating or cooling. It is
widely used in HVAC systems to set up the temperature. There are four categories of the thermostat in HVAC
systems: non-programmable digital thermostats, programmable digital thermostats, programmable mechanical
thermostats, and data logging thermostats. The non-programmable digital thermostats use thermistor to detect the
temperature and are easy to add in the HVAC system. The programmable digital thermostats have multiple functions
depending on the complexity of the programming. The programmable mechanical thermostat has mercury inside the
equipment, and it is seldom used in recent years. The data logging thermostats can store previous data and are used
in complex systems.
The advanced programmable digital thermostat is occupancy-response adaptive thermostat. It can change the
setpoint of the HVAC systems based on the different occupancy situation. The limitation is that it can only detect if
there is any occupancy. But it cannot detect the occupants’ requirements directly. For example, when occupants are
sleeping, they may need higher temperature. The smart thermostat cannot detect the occupants’ activities directly.
The development of computer science makes the automatic control of HVAC systems possible. The automatic
system control is a control method based on sensor data collection. The control system can get feedback from the
sensor data and adjust the operating of systems (Lin, Federspiel, & Auslander, 2002). The consideration of building
energy management systems has the potential to save 20% on the energy consumption of buildings (Homod et al.,
2012). A study found that use of the fuzzy logic controller of HVAC systems results to good energy performance
(Casillas & Cordon, 2003).
1.4 Applications of Machine Learning Method in the Building Domain
Machine learning is one of the aspects of artificial intelligence (AI) in the computer science domain. AI describes
the function of the algorithm that can manage the data and achieve specific tasks. Machine learning, natural
language processing, expert systems, vision, speech, and robotics are different aspects of AI. The machine learning
method uses the algorithms to do the tasks without step-by-step instructions. Machine learning algorithms build the
mathematical model based on the existing data, which is called training data. The new data in the same format as the
training data can be identified with the machine learning model. The significant step of machine learning is to decide
the input features and output features. The steps for building up the machine learning model is training the model
and testing the performance. One dataset is separated into two groups: the training dataset and the testing dataset.
The training dataset is used to calculate the algorithm, and the testing dataset is used to test the performance of the
algorithm. For example, the machine learning algorithm is training to do image recognition. The task is recognizing
the picture with dogs. There are 1000 pictures. It divided into training dataset and testing dataset. The machine
learning algorithm will learn the pattern of dog from the training dataset. The testing dataset won’t put into the
algorithm at the algorithm training stage. After the machine learning algorithm trained by the training dataset. The
testing dataset will put into the algorithm to test if the algorithm can find the pictures with dog in the training
dataset.
There are two categories of the machine learning methods: supervised learning and unsupervised learning.
Supervised learning needs to identify both the input features and output features. Unsupervised learning only needs
the input features. The algorithm itself decides the output features. The classification task is a supervised learning
approach. It is proving the feature to the right classes. The popular classification algorithms are decision tree,
random forest, logistic regression, gradient boosting, and artificial neural network (ANN) (Alpaydin, 2014).
In the building domain, machine learning methods are sometimes used. Machine learning is used to predict the
thermal comfort of the occupants and the energy consumption of the system. The ANN is used in advanced building
system control. One of its uses is improving the performance of building systems, for example, the operation of
HVAC systems. Fuzzy logic is used to combine the complaint level of the occupants and the PMV factors
(environmental and personal comfort) to achieve the setpoint control task (Fig. 1.3)(Martin et al., 2002).
13

Figure 1. 3 Supervisory controller block diagram (Martin et al., 2002)
The ANN algorithm was adopted in control air temperature, relative humidity, and PMV of the residential building.
The logical had two steps decided to turn on/off the heating or cooling. The ANN algorithm was used to predict the
temperature difference and the humidity differences to control the system. The modeling test illustrated the
acceptable prediction accuracy and thermal environment control (Fig.1.4) (Moon, Jung, Lee, & Choi, 2015).

Figure 1. 4 Algorithm of air temperature control with ANN
1.5 Summary
Thermal comfort plays a significant role in the indoor building environment. Evaluating the thermal comfort of the
occupants is necessary. The widely used thermal comfort prediction method is the PMV–PPD model. However, one
limitation of the PMV–PPD model is its lack of specific individual data. Another critical factor of thermal comfort is
the HVAC system. An intelligent HVAC system control is important for providing a better thermal environment and
saving energy.
14

One can develop a tool to enhance the thermal comfort in office buildings by automatic control of the HVAC
systems according to the facial skin temperature of the occupants. First, one must investigate the relationship
between facial skin temperature and thermal comfort. The explanation of the correlation can provide a better
application of facial skin temperature in thermal condition research. Moreover, exploring the influence of air
temperature on thermal comfort is necessary. Air temperature is easy to monitor and apply. The most important
objective is to determine the suitable machine learning model for both prediction process of thermal comfort and
thermal sensation and controlling process of HVAC systems.

15

2. LITERATURE REVIEW
This chapter introduces the development of thermal comfort evaluation, the important factors of thermal
environment, the study of the relationship between facial skin temperature and thermal comfort, the HVAC system
control technologies, and the application of machine learning method in building domain.
2.1 The Development of Thermal Comfort Evaluation
Thermal comfort is an essential factor for the occupants of a building. It has an influence on both human health and
the productivity of the occupants. The improving thermal comfort requirement promotes the development of thermal
comfort evaluation. The development of technology makes it possible to do complicated measurements and data
analysis of thermal comfort.
The PMV–PPD model is widely used in thermal condition prediction. However, many studies have investigated the
limitations of environment-based thermal prediction methods: The calculated PMV range is different from the
feedback of the occupants. All calculated PMV values are located in the neutral range and close to zero, which
means ideal thermal condition, but the survey data shows that some values are outside the neutral range and far from
zero (Fig. 2.1) (de Dear et al., 2013).

Figure 2. 1 Comparison of the thermal sensation vote and PMV values (de Dear et al., 2013)
The PMV–PPD model also disregards the differences in weather conditions and occupants. It assumes all occupants
have the same thermal preferences. This assumption limited its performance. The adaptive thermal comfort has three
aspects: behavioral, physiological, and psychological (de Dear, Brager, & Cooper, 1997). There are two ways of
thinking about adaptive comfort. The first is that occupants can adapt to the environment. Another meaning is that
the indoor comfort temperature correlates with the outdoor environment, which indicates a different outdoor
environment will influence the indoor thermal environment (Ferrari & Zanotto, 2012).
16

The study of adaptive thermal comfort approach reported the linkage between the PMV–PPD model and the
adaptive thermal comfort model. The adaptive thermal comfort included the outdoor environment change. It was
approved that the adaptive thermal comfort approach had better performance in evaluating the thermal comfort of
the occupants in the office building (Linden, Loomans, & Hensen, 2008).
The PMV–PPD model is mainly based on environmental factors. However, occupants can be important parameters
of thermal comfort predictions. One tendency is considering the function of occupants in the thermal comfort
evaluation. For thermal comfort evaluation, the method changed from the physically based model to the adaptive
comfort model, with more attention concentrated on the occupants (de Dear et al., 2013).
Several studies had discussed occupant-based thermal prediction methods. One illustrated the thermal comfort
prediction method based on the feedback of occupants. The survey data supported the potential of using the personal
comfort model in office buildings. The Internet of Things (IoT) and machine learning methods were used in the
personal comfort model. The data was collected from occupants’ daily feedback of the thermal environment. The
result showed smaller root mean square error (RMSE) in the personal model compared with the PMV model A
smaller RMSE indicates better results. It means the prediction of personal model are more similar with the
occupants’ feedback (Kim, Schiavon, & Brager, 2018).
In another study, occupants-based thermal sensation and thermal comfort models were developed according to the
human subjects testing data in a nonuniform environment—the thermal predictions involved with the local thermal
condition and the overall body condition. During the experiments, different local parts were heated or cooled
separately while the other body part (e.g. head, face, neck, back and many others) was explored in a stable thermal
environment. The result reported the potential to use the local body part to make the prediction of overall thermal
status (Zhang et al., 2004).
The research method for thermal comfort is also changing. The previous methods were experiments including
climate chambers experiments and field experiments in real buildings. However, technology promotes research
method improvements. The comfort model simulation is becoming popular in the academic field. It uses a software
to do the simulation of the thermal environment. The comfort simulation for the office occupants becomes a more
popular research area in the recent years (Li et al., 2018). The computational fluid dynamics (CFD) model was used
to analyze the thermal environment of the building, for example, the air temperature distribution and air flow inside
the building. The analysis of these parameters helped the thermal comfort analysis (Tap et al., 2011).
2.2 Important Factors of Thermal Environment
Six features are necessary to be considered when defining the thermal comfort of a building: air temperature, air
speed, relative humidity, mean radiant temperature, metabolic rate, and clothing insulation (Földváry Ličina et al.,
2018).
Air temperature indicates the dry-bulb temperature of the air. Dry-bulb temperature doesn’t change with the
humidity conditions. For the indoor air temperature, it means the average air temperature in a specific area in the
building (Zhong, 2017). Air temperature is one of the crucial features of thermal comfort. Occupant dissatisfaction
with air temperature is the main reason for an unsatisfactory indoor environment in office buildings (ASHRAE,
2010). Air temperature is the main index for HVAC systems control. The proper air temperature setting is an
efficient way to reduce the energy consumption of HVAC systems. The recommended air temperature range is
different based on the function of the room. The widely acceptable air temperature range for the neutral condition is
from 23℃ to 26℃ (Hoyt, Lee, Zhang, Arens, & Webster, 2009).
Air speed is the rate of the movement of the air at one point in a room (ASHRAE, 2010). For the thermal comfort
condition, air speed is another important parameter. It is also considered as one of the control features for HVAC
systems. Air speed that is too high will lead to an uncomfortable experience in the indoor environment, while air
speed that is too low cannot supply enough air volume and can cause deficient air quality. The limitation of air speed
for ASHRAE standard is less than 0.2m/s (ASHRAE, 2010).
17

Relative humidity reflects the amount of water vapor in the air. It is the percentage between actual vapor density and
the saturation vapor density at a specific temperature. The relative humidity has effects on the thermal comfort of
occupants. At the same temperature, the higher relative humidity will increase the heat loss of the human body. But
extremely low relative humidity will also cause discomfort of desiccation for people. The ASHRAE has a
recommended range of relative humidity between 60% and 70%. Furthermore, relative humidity influences the
occupant’s health and is associated with infections and allergies (Arundel, Sterling, Biggin, & Sterling, 1986).
Mean radiant temperature is the average radiant temperature of the surrounding surface. The human body has heat
transfer with surroundings through radiation.
Metabolic rate is a human factor that influences thermal comfort. Because it represents the energy generated by the
human body, it is associated with the gender, age, and activities of the person. Metabolic rate is a dynamic
parameter, and it is difficult to get a specific number when measuring its influence on thermal comfort. Metabolic
rate has a range that depends on people’s activity level. There is heat transfer between the human body and the
indoor environment (Huizenga, Hui, Duan, & Arens, 1999).
For the part of the body covered by clothing, the transfer between the human body and the indoor environment will
be influenced by the clothing (Kelly, 2014). The clothing insulation indicates the thermal influence of clothing.
Adjusting the clothing insulation is an effective way for people to adjust their thermal comfort (Newsham, 1997).
Hight clothing insulation will help people stay in warm condition. Otherwise, clothing insulation leads people to
lose more heat. The clothing insulation rate varies with different clothing situations. The clothing insulation rate for
normal office wearing is 1.1 (Zhong, 2017). The unit of measure is the clot.
2.3 The Study of the Relationship between Facial Skin Temperature and Thermal Comfort
Skin is an important organ for thermoregulation. There are many studies that have investigated the local body
thermal conditions and their influences on the overall thermal sensation and thermal comfort. The correlation
between body skin temperature and thermal preference supported the thermal environment studies for a new way
(Berkeley, 2004). It indicated that the skin temperature has the potential to evaluate the thermal comfort of the
occupants.
Many studies supported that thermal comfort is correlated with skin temperature. Moreover, the unstable indoor
environment influences the body skin temperatures in different ways for different body parts. In a warmer condition,
the head has a warmer thermal sensation than the other body parts, and the head is insensitive for the cold situation
(Ã, Zhang, & Huizenga, 2006). Local thermal sensation was found as a function of mean skin temperature, and it
had influences on the overall thermal sensation of the occupants. Computer-simulated experiments show that skin
and head core temperature impact regulatory responses (Fiala et al., 2001).
Experiments in the environment chamber have validated the potential of using body skin temperature as a prediction
for thermal comfort and thermal sensation. The data analysis reported that the combination of forehead, neck, wrist
back, belly, and chest has the highest prediction accuracy of up to 93% (Fig. 2.2) (Choi & Yeom, 2017a).

Figure 2. 2 Cross-validation results (Choi & Yeom, 2017a)
18

A multi-node thermoregulation model was used for thermal comfort predictions (Tanabe et al. 2002). There were 16
body segments based on the 65 multi-node thermoregulation model. The 65 multi-mode is a way to divide the whole
body into different parts based on the experiments and calculations. The evaluation of thermal comfort based on
analysis with the 65 multi-node model considered the solar radiation to improve the prediction accuracy. The model
can be used in complex thermal conditions like airplanes and cars (Tanabe et al., 2002).
Moreover, the human face has more blood vessels and has stronger heat transfer with ambient temperature
(Ghahramani et al., 2016a). An infrared camera was used during the human subject experiments, and the data
analysis provided reliable evidence for the correlation between facial skin temperature and thermal preferences. The
main attribution was a novel framework for using the thermal camera for data collection (Li et al., 2018). The
temperature of local facial parts (cheekbone, ear, the front face, and nose) was measured. The facial skin
temperature indicated the differences between genders, and females have lower facial skin temperature in both cold
and hot conditions (Ghahramani, Castro, Becerik-Gerber, & Yu, 2016b).
2.4 Automatic Control Techniques in HVAC Systems
Building energy management systems (BEMS) can play an essential role in energy, saving up to 20% of the building
energy consumption (Argiriou, Bellas-Velidis, Kummert, & André , 2004). Using an automatic control system for
the passive system can result in great energy saving (Casillas & Cordon, 2003). The development of technology has
driven the new direction for the HVAC systems design. One thermal comfort–based control system was validated by
energy simulation and case study. The results indicated that the thermal comfort–based control system has better
performance in nonuniform conditions (Freire, Oliveira, & Mendes, 2006).
Some studies of HVAC system control aim to add the feedback of occupants to the control system. The supervisory
control system provided the potential to include the thermal complaints of the occupants to the control system in the
HVAC control strategy. The digital control HVAC systems provided the training data of the learning algorithm
(Martin et al., 2002). The data comparison between two occupant-based HVAC control systems and baseline control
system reported that the occupant-based system had the less energy consumption, and it indicated the potential of
using simple feedback from the occupants to reduce energy usage (Goyal et al., 2015).
Artificial intelligence has been used in the HVAC system operation in some commercial buildings. It helped to
reduce the operating energy consumption of the building. The energy consumption prediction was reliable when the
training data was coming from the previous month. However, if the training data and testing data had a long time
interval, the prediction accuracy was decreased (Kreider, Xing An Wang, Anderson, & Dow, 1992).
Model predictive control technology was adopted in the experiments, which indicated that the model predictive
control technology saved up to 36% energy compared with turning on and off the controller. The model predictive
control calculated the model based on the operating data of the HVAC system, and it controlled the cooler or heater
based on the prediction results of the thermal level (Razmara et al., 2015).
2.5 Applications of Machine Learning Methods in Building Domain
Artificial intelligence (AI) has been supplied in building systems control since the 1990s. The development of
computer science optimized the energy performance of HVAC systems (Ló pez, Sá nchez, Doctor, Hagras, &
Callaghan, 2004). In the previous study, the machine learning algorithm was mostly used to predict building load
and energy consumption. The standard method was using the digital data from the HVAC systems to train the
algorithm and test the model performance in the real operating process (Argiriou et al., 2004). The new direction is
using the machine learning algorithm to make thermal comfort predictions. The previous studies worked on
predicting building load and energy consumption. The prediction of building load is using the previous operating
data of HVAC systems to make prediction for the future building load.
The artificial neural network (ANN) algorithm is one of the robust machine learning models. It can perform real-
time data processing, fast response, and adaptive model instantly. For the load prediction application, the input
19

features included the critical factors that influenced the heating/cooling load of the building, and the output feature
was the heating/cooling load. To get an accurate prediction, a large amount of training data was needed. The
prediction accuracy was more than 90% (Fig. 2.3) (Kalogirou, n.d.).

Figure 2. 3 Results of the ANN model prediction (Kalogirou, n.d.)
Compared with the conventional load calculation, the advantages are that the ANN model is faster and has more
accurate results (Kalogirou, n.d.). Cooling energy consumption was predicted by the ANN-based prediction model
in different setpoint temperature. The prediction results were compared with energy simulation data, and the root-
mean-square deviation was in the acceptable range. It indicated that the ANN-based model can generate useful data
in cooling energy consumption (Moon et al., 2015). The ANN-based model was applied in the residential building’s
thermal control. Four combination conditions were compared: temperature and humidity control without ANN,
temperature and humidity control with ANN, PMV control without ANN, and PMV control with ANN. Based on
the comparison, the ANN model generally increases the prediction accuracy for temperature and PMV, and the
ANN based-model reduces the period of over- or under-satisfied thermal conditions (Moon, 2012).
2.6 Review of the Human Building Integration Lab in Building Science
The Human Building Integration Lab in Building Science program in School of Architecture at University of
Southern California. The director is Joon-Ho Choi. The main research direction is synthesizing the study of
advanced building technologies and promoting the indoor environment quality. The recent research included thermal
comfort, visual comfort, and office environment study. The previous study included using data-driven approach to
evaluate the thermal sensation of the occupants, the fundamental research of facial skin temperature and thermal
comfort, and to investigate the factors influenced the occupants in the building.
In this stage, there were three research topic, thermal comfort, visual comfort and office environment. The thermal
comfort study focused on using skin temperature to make predictions for the occupant’s thermal comfort. The visual
comfort study focused on using pupil size to make predictions for the occupant’s visual comfort. The office
environment study tried to find the relationship between different IEQ factors.
2.7 Summary
Thermal comfort is an important factor of the indoor environment. The previous widely used prediction model of
thermal comfort is the PMV–PPD model. However, the development of the thermal comfort evaluation promotes the
research direction from environmental parameters to human parameters. Many studies have proved that the feedback
of the occupants has a difference with the outcome of PMV–PPD model. Moreover, recent studies have reported that
thermal comfort correlates with body skin temperature. HVAC systems have the potential in reducing the energy
consumption of buildings, particularly the operating control of the HVAC systems. There are some methods adopted
20

in the automatic HVAC system control process. The intelligent control method has decreased the energy usage of
the HVAC systems. The machine learning method is widely used in the industry and the academic field in building
domain. Its primary use is estimating the energy performance of the building and predicting the thermal comfort
conditions of the occupants.
It is possible to use the facial skin temperature to predict the thermal comfort of the occupants. A machine learning
method can learn the data pattern and give predictions based on facial skin temperature.

21

3. METHODOLOGY
This chapter documents the underlying methodology; setting up the environment chamber, creating thermal
perception surveys, and designing the experiment’s process are the three tasks that were done before the human
subject experiments. The data analysis step was completed after finishing the tests (Fig. 3.1). Setting up the
environment chamber includes the introduction of the previous experiment chamber, HVAC systems, data collection
sensors, data acquisition systems, and the automatic control device of HVAC systems. There are two parts in the
thermal perception survey to collect the personal information and thermal preference of subjects. The process of
training data collection experiments, model validation experiments, and system control validation experiments are
designed in the preparation stage. The data analysis step consists of data process platform introduction, data
preprocessing, machine learning model training, and machine learning model validation.

Figure 3. 1 Tasks before human subject experiment
3.1 Environment Chamber
The environment chamber is located in the basement of Watt Hall at the University of Southern California. The
length is 6 m, the width is 3 m, and the total area is 18 m
2
(Fig. 3.2). There are four main systems in the chamber:
HVAC systems, data collection systems, data acquisition systems, and the automatic control device. The walls are
covered with a thermal insulation board to avoid heat transfer with the outside environment and provide better
thermal environment control, and there are no windows in the experiment chamber.
22

Figure 3. 2 The layout of the environment chamber
3.1.1 Previous Design Introduction
The experiment chamber was chosen by the previous Master of Building Science students. There were three things
set up by the former students: the thermal insulation board, the HVAC systems, and the data collection sensors. The
thermal insulation board was kept in the next-stage experiments. However, the HVAC systems and the data
collection sensors were updated in the experiments.
3.1.2 HVAC Systems
The environment chamber has seasonal HVAC systems to provide a stable thermal environment: the supply air
diffuser located at the center of the ceiling and the return air outlet located at the corner of the room. To have
accurate control of the thermal environment in the chamber, the diffusers of seasonal HVAC systems were blocked.
Additional heating and cooling systems were essential parts to provide the specific conditions required for the
experiment. The air temperature was controlled in the range of 20℃ to 30℃, and the air velocity was less than 0.2
m/s. Overall, relative humidity was controlled at 35% (± 5%), while CO2 density was maintained between 700 ppm
and 900 ppm.
The additional heating and cooling systems consisted of two air conditioners and two heaters. The capacity of each
air conditioner is 1400 BTU, and the capacity of each heater is 1200 Watts. The air conditioners were set up at the
corner of the chamber to supply constant cool air. The insulated ducts were used to eliminate the exhaust air of the
air conditioners. The heaters were set up at another corner to heat the indoor air. One humidifier was used during the
experiments for the relative humidity control.
3.1.3 Data Collection Sensors
The data collection sensor system consisted of the relative humidity sensor, carbon dioxide gas sensor, air
temperature sensor, surface temperature sensor, and heart rate sensor. The environment data collection included air
temperature at different heights (0.1 m, 0.6 m, 1.1 m, and 1.6 m), radian mean temperature, carbon dioxide level,
and relative humidity. All the environmental sensors were installed in a sensor holder (Fig. 3.3). Subjects in the
experiments wore the skin temperature sensor and heart rate sensor.
23

Figure 3. 3 Sensor Holder
The accuracy of the sensors is necessary for the experiments. The specifications of the sensor were listed for
identification. The specifications included the information about the manufacturer, function, testing range,
resolution, and accuracy (Tab. 3.1). During the experiments, the subjects were required to wear the sensors to make
the record.
Table 3.1 Specifications of the sensor (Vernier products, 2019)
Device Manufacturer
/
Model
Function Specifications
Relative
Humidity
Sensor
Vernier/RH-
BTA
Relative Humidity
Measurement
Range: 0% to 95% Resolution: 0.04% RH
Accuracy:± 3.5%
Carbon
Dioxide
Gas Sensor
Vernier/CO2-
BTA
CO2 Density
Measurement
Range:0 to 10,000 ppm Resolution: 3ppm Accuracy:
± 100 ppm
Temperature
Sensor
Vernier/TMP-
BTA
Air Temperature
Measurement
Range: –40 to 135° C Resolution: 0.03° C (0 to 40° C)
Accuracy: ± 0.2° C
Data
Acquisition
Board
Vernier Data Acquisition
from
Environmental
Sensors
Maximum Sampling Rate: 48,000 samples per second
24

Surface
Temperature
Sensor
Vernier/STS-
BTA
Facial Skin
Temperature
Measurement
Range: –25 to 125° C Resolution:0.03° C (0 to 40° C)
Accuracy: ± 0.2° C at 0° C
Heart Rate
Sensor
Pola/T31 Occupants’ Heart
Rate and
Variability
Measurement
Heart rate sensor material Polyurethane
Electrode
Activity
Interface
Device
Biopac/MP160 Physiological
Signal
Recording and
Analysis
AcqKnowledge software
Thermal
Camera
FLIR/E5 Electrode Activity
Sensor
Resolution 80*60 Thermal Sensitivity <0.15℃ Field to
View 45° *34°
Skin temperature is an important parameter measured in the experiments. The infrared camera and surface
temperature sensor were chosen for facial skin temperature collection. The infrared camera is more convenient for
users because no touching is needed. However, the experiments need accurate skin temperature to support the
prediction. Surface temperature sensor provides more accurate data and a more specific measurement point.
3.1.4 Data Acquisition System
The data acquisition system consisted of sensors, a data acquisition box, and a computer with software for
programming. The programmable software used in the experiment was LabVIEW. The data acquisition hardware
was the data acquisition box. There were four data acquisition boxes in the data acquisition system. Each data
acquisition box had four channels that devices can plug into. The function of the channel was to connect the
hardware with software. The data acquisition box transferred the voltage signals from the sensors into parameters
needed in the experiments. Data acquisition box number 1 collected the data from air temperature data at four
heights. Data acquisition box number 2 collected the data from carbon dioxide density, relative humidity, and mean
radian temperature sensors. Data acquisition box number 3 and number 4 received the data from surface temperature
sensors and heart rate sensor. The software used in the data acquisition was LabVIEW. The interface included three
parts: heart rate graph, human skin temperature, and environment parameters (Fig. 3.4).
25

Figure 3. 4 The interface of the data acquisition file
Another data acquisition system was developed for automatic HVAC system control validation experiments. Only
one data acquisition box was used in the validation experiments. It included three parameters collection: cheek left
(on face), cheek right (on face), and air temperature. The interface shows the input data collected by sensors and the
output data by the machine learning algorithm. The other part is for manual data testing (Fig. XX).

Figure 3. 5 Interface of the validation experiment file
26

3.1.5 Control System Device
The function of the control device is controlling the ACs and heaters based on the output signal of the prediction
results. The control device consists of one data acquisition box and two power relays. The data acquisition box
transfers the output results from the machine learning model into the voltage signal. The voltage signal can trigger
the equipment. The power relay hardware is connected to a data acquisition box and the outlets.
3.2 Thermal Perception Survey Design
A survey was provided to the subjects during the experiments. The thermal perception survey was divided into two
parts: personal information questions and thermal preference questions.
3.2.1 Personal Information Question
The personal information questions were designed for collecting personal information about the subjects. The
personal information has the potential for use in the group data analysis. For example, gender, age, body mass index
(BMI), and ethnicity can be categories in the data analysis process. The other questions were designed for collecting
the subject’s daily behaviors. The sleeping conditions, physical conditions, emotion conditions, and outside
environmental conditions can also influence the thermal perception of the subject. This data was collected to support
data analysis. The personal information part was provided by the subject at the beginning of the experiment (Fig.
3.6).
27

Figure 3. 6 Part 1 of the thermal perception survey
3.2.2 Thermal Preference Question
The second part was the thermal preference survey, which included questions about thermal comfort, thermal
sensation, additional cooling/heating preference, and stress level (Fig. 3.7). The first three questions were designed
to get integrated feedback about thermal preference from the subject. The last question was designed to get feedback
about the stress level. This part was the main feedback for the experiments; it was repeated during the experiments.
28

Figure 3. 7 Experiment survey (part 2)
3.3 Design of the Experiment Process
There were three different experiments: training data collection experiments, model validation experiments, and
system control validation experiments. All the subjects in the experiments were volunteers from the University of
Southern California (USC). A consent form was required to provide to each subject before the experiment.
Volunteers were given $10 Amazon gift cards per hour. The subjects were required to wear long pants and long
shirts to keep the same clothing rate at 1.1. The setting of the clothing condition also helped to simulate a normal
wearing in the office situation because the automatic system was aimed at an office application.
Moreover, the subjects were required to sit and do basic computer-based work, which was also simulated for office
conditions. The detailed experiments process was designed differently based on the distinct purpose of the
experiments. There were three rounds of experiments in the whole process.
3.3.1 Training Data Collection Experiment in the Environment Chamber
29

The training data collection experiment was the first step. A large amount of data should train the prediction and
control model before application. The first-round experiment was collecting training data.
The training data collection experiment was done in the environmental chamber. There were 20 subjects who
participated in the experiments. During the experiment, two people were staying in the environment chamber, one
tested subject and one experiment assistant. There were five points on the face measured in the experiments: right
forehead, left forehead, right cheek, left cheek, and chin (Fig 3.8).

Figure 3. 8 Measure spot on face
Each experiment lasted for 90 minutes, 15 minutes for subject adaption and 75 minutes for the data measurement.
The first 15 minutes were adaptation time for the subjects. During the adaptation time, the thermal condition of the
chamber was stable. The air temperature was 24℃, which was reported as the neutral air temperature for the
occupants (Choi & Yeom, 2017b). The adaptation was ensuring the status of the subjects were the same before the
experiments because the outdoor thermal environment and activities that the subjects were doing before the
experiment have an influence on the thermal perception of the subjects. During the adaptation period, the subjects
were required to fill in the first part of the survey and wear the devices.
The first thermal comfort survey was given at the end of the adaptation time. For the whole experiment series, the
thermal comfort survey was required to be done every 15 minutes because the temperature settings changed every
15 minutes. The time interval of the temperature change was 15 minutes because the temperature change in the
environment chamber needed about 2 minutes and the remaining time was to allow the subjects to adapt to the new
thermal condition. The temperature changing step was between 2℃ and 4℃. A huge temperature changing step
could cause discomfort to the subjects. The temperature change range was from 20℃ to 30℃, which was the normal
range of indoor temperature. During the experiment, the temperature changing was irregular, which indicated that
the temperature did not change from low to high or high to low because the previous study already discussed the
thermal experiments in the regular temperature change settings (Zhong, 2017). The regular temperature change
indicated an increase or decrease with certain steps. Temperature changing was irregular. The irregular temperature
changing indicated that the temperature can be changed in any order. However, the step was between 2℃ and 4℃.
The larger temperature change step would cause a dramatic environment change, and the smaller step was difficult
to be recognized by the subject. The air temperature change needed 4–5 minutes to reach the new temperature
settings. The blue section indicates the stable temperature settings during the experiments. The temperature changes
were set manually by the experiment assistant (Fig. 3.9).
30

Figure 3. 9 Main process of the experiment
3.3.2 Model Validation Experiment in a Real Office
The purpose of the model validation experiment was to test the performance of the individual machine learning
model. According to the data analysis of the training data collection experiments, 13 subjects had all levels of the
thermal sensation feedback (cold, hot, neutral) during the experiments. The other 7 subjects only had two feedback
of the thermal sensation question (cold and hot, cold and neutral, hot and neutral). The machine learning algorithm
required all the situations to appear in the training dataset. Only the 13 subjects had the individual prediction model.
Six of the 13 subjects participated in the model validation experiments (3 males and 3 females).
The model validation experiment was an important step to validate the performance of the prediction models in a
real office space, the BuroHappold Engineering office located in Los Angeles, which is a climate zone 8 as defined
by the Department of Energy. The office is on the top floor of the building. One of the typical workstations in the
open office area was used for the validation test (Fig. 3.10).

Figure 3. 10 Workstation at BuroHappold
The workstation chosen as the experiment spot is in the west side of the office. The seasonal HVAC systems
maintain the indoor temperature between 22℃ and 24℃. The subjects were required to wear long pants and short
shirts to keep the same clothing insulation. The consent form needed to be given to the subjects before the
experiments. Left cheek, right cheek, and the air temperature were measured during the experiments. The
experiment process lasted for 135 minutes. The first 15 minutes were adaptation time for the subject. The first part
of the survey which collected the personal information was provided during the adaptation time, and the second part
31

of the survey which collected the thermal preference was delivered every 5 minutes. The survey time interval of the
training data collection experiment was 15 minutes. The shorter time interval allowed to get more feedback during
the experiments. The surveys were more frequent in validation experiments because more survey points were
needed for validation.
3.3.3 Automatic HVAC System Control Validation Experiment in the Experiment Chamber
The third stage was the system control validation experiment. The system control validation experiments were
conducted in the same experiment chamber with the data collection experiment. The experiment validated the
performance of the automatic HVAC systems based on the machine learning predictions. The same subjects from
the model validation experiment participated in the automated HVAC system control validation experiments. The
subjects were required to wear long pants and short shirts to keep the same clothing insulation rate.
The consent form was required to be provided before the experiments. The automatic HVAC system control
validation experiment lasted for 60 minutes. The first 15 minutes were adaptation time. The personal information
part of the survey was provided at the adaptation time, and the thermal preference question was repeated every 5
minutes. Left cheek, right cheek, and air temperature were measured during the experiments.
The individual model (the machine learning model) of the subjects had been trained before. Surface temperature
sensors detected the facial skin temperature of the subject, and the signal was transferred into LabVIEW. The
machine learning algorithm made predictions of the subject’s thermal comfort and thermal sensation. The output
signal for the model controlled the HVAC devices.
Data cleaning was the first step in the data preprocessing. The raw data was included in the abnormal data and
missing data. The low-quality data influenced the performance of the machine learning model. So the data cleaning
was important. The first step was defined the abnormal data (Table. XX).
For the missing data, use the average number to fill in the blank.
3.4 Data Analysis
Data analysis is a significant step to get a conclusion and explain the experiments’ findings. This section documents
the software used in the data analysis process, the data preprocessing steps, the stages of machine learning model
training, and the method of machine learning model validation.
3.4.1 Data Analysis Software
There were four software programs used in the data analysis process. Excel was adapted to organize the raw data, do
the statistical data analysis (sum, difference, average, standard deviation), and illustrate the table, line chart, and bar
chart. The detailed statistical data analysis was done in the Minitab, such as one-way ANOVA, correlation, and
interval plot. The majority machine learning algorithm was coded in Python (decision tree, logistic regression,
random forest, and gradient boosting.). LabVIEW was used in the automatic HVAC system control process. The
detailed data analysis steps in this software were discussed in chapters 4 and 5.
Excel is a spreadsheet tool with many functions and formulas. It turns the data into an organized format and
generated charts. In the data analysis process, Excel was used to do the simple statistic calculations, for example,
finding the maximum number and calculating the sum, average, and mean absolute errors. One data analysis step in
Excel is displayed in Fig. XX. The standard deviation of facial skin temperature and air temperature was calculated
using the formula in Excel.
32

Figure 3. 11 Part of the table in data analysis
Minitab is a powerful statistical data analysis tool. It is widely used in proceeding the statistical data analysis. One-
way ANOVA was done in the Minitab of the data analysis. One-way ANOVA is used to identify the difference
between two factors. It helps to find the statistical significance of two variables. One example in the data analysis
step was identifying the differences in skin temperature between genders (Fig. 3.12). For example, if the p-value of
the one-way ANOVA is 0.000 (small p-value < 0.05 represents strong evidence against the null hypothesis, and P >
0.05 represents weak evidence against the null hypothesis) for the skin temperature difference between males and
females, it indicates the statistical significance of the assumption.
33

Figure 3. 12 One-way ANOVA test of genders (p-value = 0.000)
Python is a coding platform. There were four machine learning algorithms used in the data analysis steps coded in
Python (decision tree, logistic regression, random forest, and gradient boosting). The gradient boosting algorithm
was used to calculate the data collected from the subjects (Fig. 3.13).
Male Female
34.45
34.40
34.35
34.30
34.25
34.20
Gender
Facial Skin Temperature
Interval Plot of Facial Skin Temperature vs Gender
95% CI for the Mean
The pooled standard deviation is used to calculate the intervals.
34

Figure 3. 13 Part of the coding in gradient boosting
LabVIEW is a system engineering tool. It can connect the hardware and software. It was used to develop the ANN
model and control the HVAC systems in the data analysis and experiment.
3.4.2 Data Preprocessing
There were three steps for data preprocessing: cleaning the raw data, choosing the suitable data size, and deciding
the appropriate time window for moving average calculation.
Sensors collected all the data, but there were some missing, and abnormal data should be cleaned with a suitable
data cleaning method. The air temperature range in the experiment was 20℃ to 30℃, and the skin temperature
range of the normal person is 28℃ to 36℃ in the environment temperature range from 16℃ to 32℃ (Academy,
Academy, & States, 2016). The abnormal data should be replaced with normal data, and the missing data should
also be added with standard data. The data collected in the experiments was time series data. Linear interpolation
was used to add or correct the abnormal data for the time series data (Raudys & Pabarskaite, 2018).
Another critical step in data preprocessing is moving average calculation. The simple moving average calculation
was used in the data analysis. It is the unweighted mean of the previous n data (Raudys & Pabarskaite, 2018). One
example is the moving average of temperature. The formula is
TMa =
𝑇𝑀 + 𝑇𝑀 − 1 + ⋯ + 𝑇𝑀 − ( 𝑛 − 1 )
𝑛
35

where TMa is the moving average of temperature, TM is the last temperature of the calculation, TM-1 is the
temperature before TM, and n is the total number of the temperature.
3.4.3 Machine Learning Model Training
Decision tree, logistic regression, random forest, gradient boosting, and ANN algorithm were compared in the data
analysis. The most important step in using a machine learning algorithm was deciding the input and output features.
Facial skin temperature and air temperature were the input features of the machine learning model, and thermal
comfort and thermal sensation were the output features of the machine learning model. The data collected in the
training data collection was used for the machine learning model training. The method used to test the performance
of the machine learning model was 4-cross-validation. The cross-validation divided the dataset into a training group
and a testing group. The training group data was used to build up the algorithm, and the testing group was used to
test the performance of the model. The 4-cross-validation divided the dataset into 75% training data and 25% testing
data.
3.4.4 Machine Learning Model Validation
The machine learning model has the potential to be overfitting. Overfitting in the computer science domain indicates
that the machine learning model works well with the training dataset but cannot perform well with the new dataset.
To better test the performance of the machine learning model, two validation experiments were conducted. Data
analysis of the model validation experiment used prediction accuracy. Prediction accuracy calculated the data of
predictions from the machine learning model and the feedbacks from the subjects. The formula is
A =
𝑁𝑟
𝑁𝑡

where A is the accuracy of the machine learning model, Nr is the number of right predictions (right prediction
indicates that the prediction results of the machine learning model and the feedback of the subject are the same), and
Nt is the total number of predictions.
3.5 Summary
Chapter 3 introduced the four main steps, and the four steps were classified into three groups: experiments
preparation (set up the environment chamber and design the subject survey), three round of experiments, and data
analysis. There were four systems in the environmental chamber. HVAC systems provided the thermal environment
of the experiments, and data collection sensor systems collected the human parameters (facial skin temperature and
heart rate) and environmental parameters (air temperature, relative humidity, and CO 2 level). In the data analysis
stage, three steps were done: data preprocessing, machine learning model training, and machine learning validation.
The detailed discussion of data analysis is documented in chapters 4 and 5.

36

4. TRAINING DATA COLLECTION EXPERIMENT DATA AND RESULTS
As discussed in the methodology, the purpose of the first-stage experiments was to collect the data for the machine
learning model. The initial data analysis steps were divided into five parts: moving average time window selection,
general data analysis, individual data analysis, input features selection, and learning algorithm selection (Fig. 4.1).

Figure 4. 1 Training data analysis workflow
4.1 Moving Average Time Window Selection
The sensor data was collected every 5 seconds. Each experiment lasted for 105 minutes. Except for the first 15
minutes, there were more than 1,000 data points for each subject. The accuracy of skin temperature sensor and air
temperature sensor was 0.2℃. To reduce measuring errors and noise data, a suitable moving average data was
necessary. The data format for 5 measure points in the face and air temperature was same. The left cheek data was
taken as the example to discuss the moving average time window selection steps. The first 10 minutes of data was
taken from one subject’s data set. Four different lines represent the raw data and the different moving average data
(Fig. 4.2). The 10-second moving average line has limited difference with the raw data; this means not much data
noise is eliminated. The 60-second moving average data makes the line too smooth. It indicates the dataset lost data
sensitivity and detailed information. The 30 seconds moving average data has the best performance for eliminating
data noise and keeping the data sensitive at the same time in these groups.
37

Figure 4. 2 Skin temperature moving average comparison
4.2 Group Data Analysis
Twenty subjects participated in the training data collection experiments. The group data analysis has several aspects,
including gender, age, and BMI. However, the variety of subjects was limited in the experiments (Tab. 4.1). There
were 19 subjects in the junior group and 1 subject in the middle-age group. Five subjects were reported underweight,
and 15 were in the normal weight group. So only gender can be the category for the general data analysis because of
the similar number of subjects in both male and female groups. Moreover, gender was proved as an important factor
in thermal perception evaluation (Choi & Yeom, 2017b) . The analysis included three parts: thermal perception
comparison, machine learning model discussion, and HRV data analysis.
Table 4.1 Demographic Information about Subjects
Range

Junior
21-29
Mid-
age
30-39
Sub
Total
Under
Weight
(<18.5 ）
Normal
Weight
(<24.9 ）
Over
Weight
(>30 ）
Sub
Total
Number Male 9 1 10 - 9 1 10

Female 10 - 10 5 5 0 10

Total 19 1 20 5 14 1 20
4.2.1 Thermal Perception Data Analysis
There are two main factors in the thermal perception data analysis: facial skin temperature and environment
temperature. The thermal perception also includes two different evaluation aspects: thermal comfort and thermal
sensation. In the experiments, there were 5 points measured on the face. The average facial skin temperature was
used for the group data analysis to simplify the data analysis steps. In the experiments, the feedback of thermal
comfort and thermal sensation concentrated on the range from -1 to 1; only 4 feedback from the subjects reported -2
or 2. To make the analysis simple and prediction accurate, the thermal perception scale was changed from 5-point
scale to 3-point scale in the data analysis process. In the thermal comfort evaluation, -1 represented uncomfortable, 0
33.6000
33.7000
33.8000
33.9000
34.0000
34.1000
34.2000
7:33:27
7:33:52
7:34:17
7:34:42
7:35:07
7:35:32
7:35:57
7:36:22
7:36:47
7:37:12
7:37:37
7:38:02
7:38:27
7:38:52
7:39:17
7:39:42
7:40:07
7:40:32
7:40:57
7:41:22
Skin Temperature
Time
Skin Temperature Moving Average Comparison
Cheek left
10S
30S
60S
38

represented neutral, and 1 represented comfortable. In the thermal sensation evaluation, -1 served cool, 0 represented
neutral, and 1 represented hot.
In the thermal comfort comparison, males and females show different facial skin temperature in the same thermal
comfort conditions (Fig. 4.3). The p-value of the thermal comfort and genders is an extremely small number
rounded to 0.000; this value indicates that the statistic is significant. In the neutral state, females have a higher skin
temperature than males. In the other two conditions, females have lower skin temperature than males, especially in
an uncomfortable situation, where the skin temperature of females have 1.5℃ difference. For neutral and
comfortable conditions, males show almost the same skin temperature. For the uncomfortable state, females have a
much lower facial skin temperature than the other two conditions.

Figure 4. 3 Facial skin temperature versus thermal comfort by genders
(Legend: -1 uncomfortable, 0 neutral, 1 comfortable)
In the thermal sensation comparison, males and females also show different facial skin temperature in the same
thermal sensation level (Fig. 4.4). The p-value of the thermal sensation and genders is also an extremely small
number rounded to 0.000. Males have higher facial skin temperature in all thermal sensation conditions. However,
the temperature difference is small, which is 0.25℃ in hot and neutral conditions and 0.5℃ in hot conditions. The
facial skin temperature increases the same way for both males and females. The tendency is that the facial skin
temperature increases as the thermal sensation changes from cold to hot.
Thermal Comfort
Gender
1 0 -1
Male Female Male Female Male Female
34.7
34.6
34.5
34.4
34.3
34.2
34.1
34.0
33.9
33.8
Facial Skin Temperature
Interval Plot of Facial Skin Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
39

Figure 4. 4 Facial skin temperature versus thermal sensation by genders
(Legend: -1 cold, 0 neutral, 1 hot)
For the thermal perception analysis, another factor is air temperature. In the thermal comfort comparison, males and
females report different thermal comfort level at the same temperature (Fig. 4.5). The p-value of the thermal comfort
and genders is also 0.000. It indicates that in the thermal comfort evaluation, different genders have different air
temperatures in the same thermal comfort level. At 25.5℃, females felt uncomfortable, and males indicated
comfortable. For the neutral condition, females reported a temperature of 27℃, and males reported 25℃. For all
thermal comfort conditions, females preferred a higher temperature than males. Females reported the neutral
sensation at a temperature higher than 27℃, while the hot sensation was recorded at 26℃. It is common sense that
people tend to feel warmer in a higher temperature thermal environment. The possible explanation is the order of
temperature settings. Subjects had different thermal comfort and thermal sensation feedback, even if they were in
the same temperature sometime. For example, the environment temperature is 26℃. If the previous temperature
setting is 28℃, the subject may feel cold at 26℃ because the previous temperature setting is much higher. And if
the previous temperature setting is 20℃, the subject may feel hot at 26℃. The thermal perception is not always
objective. The sequence of the temperature settings may influence the result. Another possible reason is the
individual differences between females, who have more influence in the data analysis. Thermal comfort and thermal
sensation vary between different people. The data analysis was based on 19 subjects. Each subject influenced the
result. Maybe several females reported the unusual data, and the result was influenced by these data.
Thermal Sensation
Gender
1 0 -1
Male Female Male Female Male Female
35.25
35.00
34.75
34.50
34.25
34.00
33.75
33.50
Facial Skin Temperature
Interval Plot of Facial Skin Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
40

Figure 4. 5 Environment temperature versus thermal comfort by genders
(Legend: -1 uncomfortable, 0 neutral, 1 comfortable)
In the thermal sensation comparison, females show better endurance for higher temperature, and males show better
persistence for lower temperature (Fig. 4.6). The p-value of the thermal sensation and genders is also small number
rounded to 0.000 in comparing the environment temperature. Females reported cold at 24℃, and males reported
cold at 23℃. When the temperature is higher than 26℃, females indicated neutral. When the temperature is lower
than 26℃, males reported neutral. When the temperature is higher than 28℃, both males and females felt hot. The
data shows a suitable tendency for thermal sensation feedback; the thermal sensation changes from cold to hot when
the temperature is increases.
Thermal Comfort
Gender
1 0 -1
Male Female Male Female Male Female
27.5
27.0
26.5
26.0
25.5
25.0
Environment Temperature
Interval Plot of Environment Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
41

Figure 4. 6 Environment temperature versus thermal sensation by genders
(Legend: -1 cold, 0 neutral, 1 hot)
According to the data analysis of thermal perception, there are distinct differences between males and females in
both facial skin temperature and environment temperature. Females prefer higher temperature and have lower skin
temperature than males.
4.2.2 General Machine Learning Model Comparison
According to previous data analysis, males and females show differences in both thermal comfort and thermal
sensation evaluations. Gender is a meaningful category for thermal perception predictions. As discussed in the
methodology, there are two steps in achieving automatic HVAC systems control. A good performance for both
thermal comfort and thermal sensation is necessary. For the initial group prediction model (thermal comfort
prediction and thermal sensation prediction), the input features were facial skin temperature (right forehead, left
forehead, right cheek, left cheek, and chin) and air temperature, and the output features were thermal sensation and
thermal comfort. There were four classification machine learning algorithms used in the initial data analysis. The 4-
cross-validation was used in the data analysis steps. It divided the dataset into 2 groups: 75% of the data was used
for training the algorithm, and 25% of the data was used for testing the performance of the trained algorithm. For the
group data analysis, the machine learning model represents a low prediction accuracy in both thermal comfort
predictions and thermal sensation predictions (Fig. 4.7 and Fig 4.8). The average prediction accuracy for the model
of the male is 59%. The highest accuracy is 67.2% in the decision tree model, and the lowest accuracy is 46.7% in
the logistic regression. The prediction accuracy for the female is even worse than the male’s results. The average
accuracy is 43.2%, the highest is 46%, and the lowest is 41.7%. The general model combined all the data. But the
prediction accuracy does not improve. The prediction accuracy of the thermal comfort is lower than the thermal
sensation in males and the general group. All machine learning models have a prediction accuracy more moderate
than 50% in the thermal comfort prediction.

Thermal Sensation
Gender
1 0 -1
Male Female Male Female Male Female
29
28
27
26
25
24
23
Environment Temperature
Interval Plot of Environment Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
42

Figure 4. 7 Thermal comfort prediction comparison

Figure 4. 8 Thermal sensation prediction comparison
In the data mining domain, a good classification accuracy depends on different situations because different
classification tasks have different number of classes and different difficult levels (Kononenko & Kukar, n.d.). For
example, in the face recognition task, the main factor that will influence the result is the picture. It is objective. The
prediction accuracy can be more than 90%. Another example is the prediction of the customer behaviors. The
customer behaviors depend on many factors, not only on the data fact itself. The prediction accuracy of 50% or 60%
is also regarded as a good prediction performance. The thermal comfort and thermal sensation classification task is
also decided by the class number. There were 3 classes in both thermal comfort (uncomfortable, neutral,
41.4%
22.0%
50.0%
44.7%
39.5%
41.5%
35.5%
48.8%
45.8%
42.9%
35.2%
30.1%
35.8%
31.3%
33.1%
0.0% 10.0% 20.0% 30.0% 40.0% 50.0% 60.0%
Decisicon
Tree
Logistic
Regression
Random
Forest
Gradient
Boosting
Average
Thermal Comfort Prediction Comparison
General Female Male
67.2%
46.7%
63.3%
58.7%
59.0%
42.8%
42.5%
46.0%
41.7%
43.2%
49.5%
52.5%
51.8%
51.6%
51.3%
0.0% 10.0% 20.0% 30.0% 40.0% 50.0% 60.0% 70.0% 80.0%
Decisicon
Tree
Logistic
Regression
Random
Forest
Gradient
Boosting
Average
Thermal Sensation Prediction Comparison
General Female Male
43

comfortable) and thermal sensation (cold, neutral, hot). It is also influenced by the subject status. The randomly
chosen accuracy is 33%. Higher than 80% is an acceptable prediction accuracy, and above 90% is good accuracy.
Lower than 80% is not a meaningful prediction accuracy.
The prediction accuracy of males, females, and all gender groups is less than 80%. A prediction accuracy of less
than 80% is not acceptable in the thermal comfort and thermal sensation predictions. The results indicate that it is
difficult to build the group machine learning model for thermal perception predictions. The possible reason is that
the input features (facial skin temperature and air temperature) and the output features (thermal comfort and thermal
sensation) have dramatic differences among subjects. For example, some subjects feel cold at 26℃, and some
subjects feel neutral at 26℃. They also have different facial skin temperature at the air temperature, which makes
the classification task difficult. The simple logic of the machine learning algorithm to do the classification task for
thermal comfort is matching specific combinations of facial skin temperature and air temperature with the thermal
comfort level.
The results report that the prediction accuracy of thermal sensation differs between genders. Males have more
accurate prediction results in the four algorithms than males (Fig. 4.9). However, in the thermal comfort predictions,
there are no obvious differences between genders, except in the logistic regression algorithm. The logistic regression
algorithm has better predicting performance in the female group than in the male group (Fig. 4.10).

Figure 4. 9 Interval plot of gender (thermal comfort)
Category
Average Gradient Boosting Random Forest Logistic Regression Decisicon Tree
Male Female Male Female Male Female Male Female Male Female
50.0%
45.0%
40.0%
35.0%
30.0%
25.0%
20.0%
Data
Interval Plot of Decisicon T, Logistic Reg, Random Fores, ...
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
44

Figure 4. 10 Interval plot of gender (thermal sensation)
4.2.3 HRV Data Analysis
The heart rate was measured in the experiments. As discussed in the methodology, the HR data was transformed into
HRV by Kubios. There were two stress-level-related HRV indexes in the Kubios HRV report: PNS index and SNS
index. Parasympathetic nervous system (PNS) index and sympathetic nervous system (SNS) index are factors
explained in the HRV data. They are indexes without units. The PNS and SNS have a complex linear and nonlinear
relationship. The PNS index reports the rest situation of the subject, and the SNS index indicates the nervous status.
The normal range of the PNS index and SNS index is between -3 and 3. The number less than -3 is low and more
than 3 is high (Shaffer & Ginsberg, 2017). The analysis method is one-way ANOVA. The one-way ANOVA
calculations have p-value larger than 0.05 (p-value < 0.05 indicates statistic significant) (Tab. 4.2). The results
reveal that there is no significant relationship between HRV and stress level and thermal comfort and thermal
sensation. There are two possible reasons for that. First, the stress level was reported by the subjects. Their meaning
of stress level may be different with the physiological definition by Kubios (Niskanen & Ranta-aho, 2017).
Moreover, during the whole experiment, the only environmental change was air temperature, and the air temperature
was changed slowly with narrow range. There are no dramatic environment changes to lead the heart rate change
and stress level change. The HRV is not a suitable input feature for the thermal perception predictions.
Table 4.2. Interval plot of HRV index
PNS Index SNS index
Category
Average Gradient Boosting Random Forest Logistic Regression Decisicon Tree
Male Female Male Female Male Female Male Female Male Female
70.0%
65.0%
60.0%
55.0%
50.0%
45.0%
40.0%
Data
Interval Plot of Decisicon Tr, Logistic Reg, Random Fores, ...
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
45

Stress Level

p-value: 0.713

p-value: 0.626
Thermal Comfort

p-value: 0.451

p-value: 0.502
Thermal Sensation

p-value: 0.353

p-value: 0.411
4.3 Individual Data Analysis
There were 20 subjects in the data collection experiments. The data of five subjects’ data were taken as an example
for individual data analysis. Different measure points have different facial skin temperature even in the same thermal
comfort or thermal sensation level.
4.3.1 Facial Skin Temperature and Air Temperature Analysis
The data of subject A shows that the forehead has a higher temperature than the cheek and chin in all situations (Fig.
4.11 and Fig. 4.12). The left cheek and right cheek have a slight temperature difference of 0.5℃. The data of subject
B shows that in the uncomfortable and neutral thermal conditions, the cheek and chin have almost the same
temperature (Fig. 13 and Fig. 4.14). However, in a comfortable situation, the temperature difference becomes larger.
The data of subject C shows the continuous temperature increase from cheek left to forehead right in all thermal
conditions (Fig. 4.15 and Fig. 4.16). It indicates that for subject C, the forehead always has the highest temperature,
and the cheek still has the lowest temperature. A narrow facial skin temperature range of subject D was revealed by
the data analysis (Fig. 4.17 and Fig 4.18). Subject D only gave the hot and cold feedback in the thermal sensation
evaluation. Moreover, chin has the lowest temperature in all conditions. The data of subject E suggests that the left
1 0 -1
3.0
2.5
2.0
1.5
1.0
0.5
0.0
-0.5
Stress Level
PNS index
Interval Plot of PNS index vs Stress Level
95% CI for the Mean
The pooled standard deviation is used to calculate the intervals.
1 0 -1
0.5
0.4
0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
-0.4
Stress Level
SNS index
Interval Plot of SNS index vs Stress Level
95% CI for the Mean
The pooled standard deviation is used to calculate the intervals.
1 0 -1
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
Thermal Comfort
PNS index
Interval Plot of PNS index vs Thermal Comfort
95% CI for the Mean
The pooled standard deviation is used to calculate the intervals.
1 0 -1
0.6
0.5
0.4
0.3
0.2
0.1
0.0
-0.1
-0.2
Thermal Comfort
SNS index
Interval Plot of SNS index vs Thermal Comfort
95% CI for the Mean
The pooled standard deviation is used to calculate the intervals.
1 0 -1
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
Thermal Sensation
PNS index
Interval Plot of PNS index vs Thermal Sensation
95% CI for the Mean
The pooled standard deviation is used to calculate the intervals.
1 0 -1
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
-0.1
-0.2
Thermal Sensation
SNS index
Interval Plot of SNS index vs Thermal Sensation
95% CI for the Mean
The pooled standard deviation is used to calculate the intervals.
46

cheek and right cheek have almost the same temperature in all states (Fig. 4.19 and Fig 4.20). The left forehead and
right forehead have practically the same temperature in all situations. Chin has the lowest temperature in cold
thermal condition and the highest temperature in hot thermal state. It indicates that chin is more sensitive with the
environment change. Because when the air temperature changes, the skin temperature of chin also changes. The skin
temperature of the chin is much higher in the high environment temperature and much lower in the low environment
temperature. Other facial parts also have temperature change when air temperature changes, but not dramatic.

Figure 4. 11 Thermal comfort versus facial skin temperature for subject A
(Legend: -1 uncomfortable, 0 neutral, 1 comfortable)
Thermal Comfort
Measure Points
1
0
-1
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHea d left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
35.5
35.0
34.5
34.0
33.5
33.0
32.5
32.0
Facial Skin Temperature
Interval Plot of Facial Skin Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
47

Figure 4. 12 Thermal sensation versus facial skin temperature for subject A
(Legend: -1 cold, 0 neutral, 1 hot)

Figure 4. 13 Thermal comfort versus facial skin temperature for subject B
(Legend: -1 uncomfortable, 0 neutral, 1 comfortable)
Thermal Sensation
Measure Points
1
0
-1
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
35
34
33
32
31
Facial Skin Temperature
Interval Plot of Facial Skin Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
Thermal Comfort
Measure Point
1
0
-1
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
35.5
35.0
34.5
34.0
33.5
Facial Skin Temperature
Interval Plot of Facial Skin Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
48

Figure 4. 14 Thermal sensation versus facial skin temperature for subject B
(Legend: -1 cold, 0 neutral, 1 hot)

Figure 4. 15 Thermal comfort versus facial skin temperature for subject C
(Legend: -1 uncomfortable, 0 neutral, 1 comfortable)

Thermal Sensation
Measure Point
1
0
-1
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
36.0
35.5
35.0
34.5
34.0
33.5
33.0
Facial Skin Temperature
Interval Plot of Facial Skin Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
Thermal Comfort
Measure Point
1
0
-1
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
35.5
35.0
34.5
34.0
33.5
33.0
32.5
32.0
Facial Skin Temperature
Interval Plot of Facial Skin Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
49

Figure 4. 16 Thermal sensation versus facial skin temperature for subject C
(Legend: -1 cold, 0 neutral, 1 hot)

Figure 4. 17 Thermal comfort versus facial skin temperature for subject D
(Legend: -1 uncomfortable, 0 neutral, 1 comfortable)
Thermal Sensation
Measure Point
1
0
-1
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
36
35
34
33
32
31
Facial Skin Temperature
Interval Plot of Facial Skin Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
Thermal Comfort
Measure Point
1
0
-1
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
36.5
36.0
35.5
35.0
34.5
34.0
Facial Skin Temperature
Interval Plot of Facial Skin Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
50

Figure 4. 18 Thermal sensation versus facial skin temperature for subject D
(Legend: -1 cold, 0 neutral, 1 hot)

Figure 4. 19 Thermal comfort versus facial skin temperature for subject E
(Legend: -1 uncomfortable, 0 neutral, 1 comfortable)
Thermal Sensation
Measure Point
1
-1
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
ForeHead right
ForeHead left
Chin
Cheek right
Cheek left
36.0
35.5
35.0
34.5 Facial Skin Temperature
Interval Plot of Facial Skin Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
Thermal Comfort
Measure Point
1
0
-1
Forehead right
ForeHead left
Chin
Cheek right
Cheek left
Forehead right
ForeHead left
Chin
Cheek right
Cheek left
Forehead right
ForeHead left
Chin
Cheek right
Cheek left
36
35
34
33
32
31
30
29
28
Facial Skin Temperature
Interval Plot of Facial Skin Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
51

Figure 4. 20 Thermal sensation versus facial skin temperature for subject E
(Legend: -1 cold, 0 neutral, 1 hot)
There are three main conclusions for the individual data analysis:
(1) Different facial parts can have different skin temperatures in the same thermal comfort/thermal sensation
conditions for the same subject.
(2) Different people have distinct facial skin temperature change tendency in different thermal situations. When the
thermal sensation changes from cold to hot, the facial skin temperature tends to increase but not in the linear way.
Some subjects have the highest skin temperature in the forehead in cold situations, but in hot situations, the highest
skin temperature changes to the cheek. Some subjects have the same highest skin temperature in the forehead in both
cold and hot situations.
(3) The average temperature difference between different facial parts is around 3℃ in the same thermal conditions
for individuals. The environment temperature change range is 10℃, but the facial skin temperature change for
individuals is 3℃ because the human body has a thermoregulation mechanism to maintain the skin temperature.
4.3.2 Electrode Activity (EDA) Data Analysis
EDA is the electrical characteristics of the skin. It indicates the psychological change of humans. The EDA data was
collected for 20 subjects. The subjects were required to do simple computer works to avoid the extra influence,
except thermal environment change. The peaks in the line indicate the psychological change of the subjects.
At the beginning and end of the experiments, there were more peaks in the EDA lines. They indicate more
psychological changes of the subjects. In the middle of the experiments, the figure shows fewer peaks (Fig. 4.21).
The EDA data indicates more psychological changes at the beginning and end of the experiments, and the
psychological states in the middle of the experiments are stable. The air temperature was decreased with a flat way
Thermal Sensation
Measure Point
1
0
-1
Forehead right
ForeHead left
Chin
Cheek right
Cheek left
Forehead right
ForeHead left
Chin
Cheek right
Cheek left
Forehead right
ForeHead left
Chin
Cheek right
Cheek left
36
35
34
33
32
31
30
29
Facial Skin Temperature
Interval Plot of Facial Skin Temperature
95% CI for the Mean
Individual standard deviations are used to calculate the intervals.
52

at the beginning of the experiment, and it was fluctuated after the first 30 minutes. The stress level reported by the
subject was “relaxed” during the entire experiment (Fig. 4.22).
The data reported that the air temperature change was not a signal for the psychological change of the subjects.
Because the air temperature change happened in the middle of the experiments, the psychological change mainly
happened at the beginning and end of the experiments. If the psychological change happens after the air temperature
changes or at the same time, it may indicate relationship between these two factors.

Figure 4. 21 EDA data tendency (subject 10)

Figure 4. 22 Stress level versus air temperature (subject 10)
The EDA data of subject number 13 shows fluctuations at the first 20 minutes and from 30 to 40 minutes. At the end
of the experiment, it is stable (Fig. 4.23). The air temperature was decreased from 28℃ to 22℃. The thermal
sensation and comfort level changed with the air temperature change (Fig. 4.24). The stress level was reported as
“nervous” for the entire experiment.
-2
-1
0
1
2
15
17
19
21
23
25
27
29
31
1
11
21
31
41
51
61
71
81
91
101
111
121
131
141
151
161
171
181
191
201
211
221
231
Stress Level
Air Temperature
Axis Title
Stress Level VS Air Temperature
Air1.1 Stress Level
53

Figure 4. 23 EDA data tendency (subject 13)

Figure 4. 24 Stress level versus air temperature (subject 13)

The EDA data of subject number 7 was stable. The peaks show more at 20 minutes and at the end of the
experiments. Other times, the line is smooth, with seldom peaks of wave (Fig. 4.25). It indicates that most of the
time, the psychological status of subject 7 is stable. The air temperature decreased at the beginning and increased the
rest of the time. The stress level reported by the subject was reported as “nervous” most of the time (Fig. 4.26).

Figure 4. 25 EDA data tendency (subject 7)
-2
-1
0
1
2
20
22
24
26
28
30
32
1
16
31
46
61
76
91
106
121
136
151
166
181
196
211
226
241
256
271
286
301
316
331
Stress Level
Air Temperature
Axis Title
Stress Level Versus Air Temperature
Air1.1 Stress Level
54

Figure 4. 26 Stress level versus air temperature (subject 7)
The EDA data of subject number 14 is fluctuating, especially near the beginning of the experiment, at 25 minutes,
45 minutes, and 70 minutes (Fig. 4.27). However, the stress level of the subject was reported as “nervous” during
the entire experiment (Fig. 4.28).

Figure 4. 27 EDA data tendency (subject 14)
-2
-1
0
1
2
15
17
19
21
23
25
27
29
31
1
15
29
43
57
71
85
99
113
127
141
155
169
183
197
211
225
239
253
267
281
295
309
323
Stress Level
Air Temperature
Axis Title
Stress Level Versus Air Temperature
Air1.1 Stress Level
55

Figure 4. 28 Stress level versus air temperature (subject 14)
The hypothesis is that the air temperature change can be an external factor that causes the psychological change of
the subjects. However, the discussion above indicated that the psychological status revealed by the EDA data is
more sensitive. It means the EDA changed a lot. Even the subjects stay in a stable environment. If they think some
sad or scary thing or talking, the EDA will change. For subject number 10, 13, and 14, the reported stress level by
the subject survey are stable during the entire experiment. It means that the subject thought they were in the same
stress status during the whole experiment. The thermal environment change didn’t change their stress status.
However, the EDA data changed frequently during the experiments. It indicated that the psychology of the subjects
changed during the experiments. But the psychological change was not correlated with the air temperature change.
They were changing in a different way. It changed in a different way with the air temperature change and the stress
level reported by the subjects in the survey.
4.4 Input Features Selection and Importance Comparison
As discussed in the methodology, the suitable input features are the foundation for the accurate predictions of the
thermal perception predictions. Meanwhile, to minimize the measurement points is also important. Several tested
points influence the thermal perceptions of the subjects, and more measured points will have a higher cost. The
purpose of the input features selection is finding the most critical two input features in the face. The integrated input
features selection steps included a comparison of all the different combinations for all measured parameters.
However, there were five measured points in the face, and the test time for one subject will be 20, and the total test
time will be 400 for all subjects. The first step used a gradient boosting model to generate the importance of each
feature. Choosing the first two essential features and testing the prediction accuracy of the chosen features are the
data analysis steps.
4.4.1 Input Features Comparison for the Thermal Comfort Prediction
In the thermal comfort prediction, the tested input features were forehead right, forehead left, chin, cheek left, and
cheek right, and the output feature was thermal comfort. There were 20 subjects in the experiments. The data of
subject 11 had too many missing data points, so it was removed from the data set. The data of 19 subjects were
tested in the gradient boosting algorithm, and the importance of the features was ranked with 1-5 by the algorithm
calculations. 1 indicates the most important, and 5 means the least concern (Tab. 4.3). The prediction accuracy of a
thermal sensation is different for different subjects. The average accuracy is 93.5%. Although, different subjects
have a different preference for the most critical point. The method was to count the number of voting to 1 and 2.
-2
-1
0
1
2
20
22
24
26
28
30
32
34
1
17
33
49
65
81
97
113
129
145
161
177
193
209
225
241
257
273
289
305
321
337
353
369
Stress Level
Air Temperature
Axis Title
Stress Level Versus Air Temperature
Air1.1 Stress Level
56

Based on the reasonable prediction accuracy and calculations, the most important two points in thermal comfort
prediction are cheek left and cheek right.
Table 4.3. Individual Thermal Comfort Prediction Comparison
Subject No.
Forehead
Right
Forehead
Left
Chin Cheek Left Cheek Right Accuracy
1 1 2 5 4 3 83.5%
2 5 3 4 2 1 97.3%
3 4 5 1 3 2 87.0%
4 5 4 1 3 2 100.0%
5 3 4 2 5 1 99.3%
6 3 2 5 4 1 98.7%
7 5 2 4 1 3 90.6%
8 5 1 4 2 3 80.2%
9 1 5 2 4 3 100.0%
10 1 5 4 3 2 96.6%
11 2 5 3 4 1 88.5%
12 3 5 4 2 1 100.0%
13 2 4 5 3 1 95.3%
14 3 4 1 2 5 99.0%
15 2 1 5 4 3 90.0%
16 4 5 3 1 2 92.9%
17 4 3 5 2 1 96.5%
18 3 1 5 2 4 98.4%
19 2 5 3 1 4 81.8%
Average

93.5%
Count for
No.1 3 3 3 3 7

57

Count for
No.2 4 3 2 6 4

sum 7 6 5 9 11

4.4.2 Input Features Comparison for the Thermal Sensation Prediction
The input features selection for the thermal sensation prediction has the same process with the thermal comfort
predictions. The only difference is the output feature. It is replaced with thermal sensation level (cold, neutral, hot).
The average prediction accuracy is 93.5%. The most important two parts are cheek right and cheek left (Tab. 4.4).
Table 4.4 Individual Thermal Comfort Prediction Comparison
Subject No.
Forehead
Right
Forehead
Left
Chin
Cheek
Left
Cheek
Right
Accuracy
1 2 4 5 3 1 86.9%
2 5 3 2 4 1 93.6%
3 5 1 4 2 3 100.0%
4 5 3 4 2 1 99.3%
5 5 3 2 1 4 96.2%
6 3 4 5 1 2 98.2%
7 3 5 4 1 2 100.0%
8 5 4 3 2 1 93.8%
9 2 5 1 3 4 100.0%
10 1 4 2 5 3 96.2%
11 1 3 4 5 2 88.0%
12 3 1 5 2 4 100.0%
13 4 1 5 2 3 92.6%
14 3 4 5 1 2 100.0%
15 2 5 4 3 1 90.0%
16 5 4 2 3 1 97.0%
17 4 1 5 3 2 98.5%
18 4 1 3 5 2 86.2%
58

19 3 5 2 1 4 59.8%
Average

93.5%
Count for
No.1 2 5 1 5 6

Count for
No.2 3 0 5 5 6

sum 5 5 6 10 12

4.4.3 Input Features Performance Testing
It is necessary to test the performance of the input features chosen by gradient boosting. For the testing model, the
input features were cheek left and cheek right; the output features were thermal comfort/thermal sensation. The data
of 19 subjects were tested in thermal comfort predictions and thermal sensation predictions separately. The
algorithm was gradient boosting. The results indicate a lower prediction accuracy compared with five input features
(Tab. 4.5). The average prediction accuracy for thermal comfort is 92.9%. It is 0.6% lower than the prediction of 5
input features. However, the average prediction accuracy for thermal sensation is 88.14%, which is 5.36% lower
than the projection of 5 input features, and there are only 4 subjects’ data that have a higher prediction accuracy than
95%. The result reports that reducing the input features decreases the prediction accuracy in thermal perception
prediction accuracy, especially in thermal sensation predictions. The prediction accuracy is higher when the input
features are 5 compared with the 2 input features. Because the input features of the machine learning algorithm are
the classification evidence, more suitable input features can help the algorithm classify the features.
Table 4.5 Prediction Results with Two Input Features (left cheek, right cheek)
Subject No. Thermal Comfort Thermal Sensation
1 86.10% 66.2%
2 99.80% 95.5%
3 100.00% 88.2%
4 100.00% 94.3%
5 92.60% 91.3%
6 93.50% 94.1%
7 100.00% 80.8%
8 86.84% 79.5%
9 99.60% 90.7%
10 89.05% 83.9%
11 83.40% 93.5%
59

12 99.41% 99.4%
13 91.34% 88.8%
14 98.37% 87.0%
15 99.72% 87.6%
16 84.43% 96.3%
17 97.44% 78.5%
18 69.16% 96.7%
19 94.34% 82.3%
Average 92.90% 88.1%
More Than 90% 13/19 9/19
STDEV. 0.080 0.079
The air temperature was another parameter measured during the experiments. Compared with facial skin
temperature, the air temperature is more accessible to measure. The next test added air temperature as one of the
input features; the output features were kept the same (thermal comfort and thermal sensation). The average
prediction accuracy in thermal comfort is 95.57%, which is even higher than the result of 5 input features, and the
average prediction accuracy in thermal sensation is also higher than the 5 input features. Moreover, there are 68%
prediction results that are higher than 90% (Tab. 4.6).
The prediction performance for 3 input features is best compared with other combinations. The prediction accuracy
of 3 input features in thermal comfort and thermal sensation predictions is higher than 2 input features and 5 input
features. The 2 input features and 5 input features only have facial skin temperature. The 3 input features have both
facial skin temperature and air temperature. Air temperature is an important factor that has influences on thermal
comfort and thermal sensation. It is also easier to collect compared with skin temperature. Additional air
temperature can increase the prediction accuracy.
Table. 4.6 Prediction Results with 3 Input Features (Cheek left, cheek right, and air temperature)
Subject No. Thermal Comfort Thermal Sensation
1 92.00% 88.8%
2 100.00% 99.8%
3 100.00% 97.1%
4 100.00% 100.0%
5 93.10% 94.0%
6 100.00% 100.0%
60

7 100.00% 82.0%
8 83.55% 99.0%
9 100.00% 100.0%
10 100.00% 84.7%
11 95.40% 96.3%
12 100.00% 100.0%
13 92.89% 92.5%
14 100.00% 97.7%
15 99.72% 95.3%
16 89.19% 96.3%
17 98.46% 93.7%
18 71.49% 98.4%
19 100.00% 93.2%
Average 95.57% 95.2%
More than 90% 16/19 16/19
STDEV. 0.073 0.051
4.5 Machine Learning Algorithm Comparison
The prediction of thermal perception is a classification task. Classification task means when the machine learning
algorithm have the input features, it can recognize the class that the input features belong to. For example, the new
combination of facial skin temperature and air temperature is put into the algorithm, it will calculate and identify this
data point belongs to “hot”, “cold”, or “neutral” category. I It can recognize the class that the input features belong
to. In the thermal perception prediction, it means when the specific facial skin temperature and air temperature input
into the algorithm, it can define the thermal perception based on the input features. Many machine learning
algorithms are used for the classification task. Five common machine learning algorithms were chosen in the
machine learning algorithm comparison. As discussed in the last section, the input features were cheek left, cheek
right, and air temperature. The output features were thermal comfort and thermal sensation.
4.5.1 Machine Learning Algorithm Comparison for Thermal Comfort Prediction
For the individual thermal comfort predictions, the same algorithm has different prediction performance for different
subjects. There are three indexes compared with different algorithms (Tab.4.7). The ANN algorithm has the best
performance in the average prediction accuracy, and the prediction results of all subjects are more than 95%.
Moreover, the standard deviation is also the lowest in the five algorithms. The logistic regression does not perform
well in this classification task because the average prediction accuracy is 73.1%, which is not acceptable in the
61

thermal perception predictions. Only 3 subjects have a prediction accuracy more than 90%, which indicates that
logistic regression does not perform well for most subjects. The other three algorithms have almost the same
prediction performance on average. In the ideal condition, each subject should use the best algorithm for their data.
However, the system used for testing different subjects. It is significant to choose a better algorithm according to all
subjects’ data. Synthesizing the data analysis result, the ANN model is suitable for thermal comfort prediction. All
the prediction results of the ANN model are higher than 90%, which indicates good prediction accuracy, and it has
the highest average prediction accuracy compared with the other four algorithms. Meanwhile, the standard deviation
is 0.01. Small standard deviation reports that the ANN model performs well in most of the subjects.
Table. 4.7 The Results of Thermal Comfort Prediction
Subject No.
Decision Tree Logistic Regression Random Forest Gradient Boosting ANN
1 88.4% 61.5% 90.3% 84.7% 97.3%
2 100.0% 60.6% 100.0% 100.0% 100.0%
3 100.0% 52.5% 100.0% 100.0% 100.0%
4 99.6% 77.6% 100.0% 99.7% 100.0%
5 92.7% 64.1% 92.7% 94.3% 98.9%
6 98.5% 56.1% 100.0% 100.0% 99.0%
7 100.0% 85.4% 100.0% 100.0% 100.0%
8 98.0% 77.6% 87.2% 83.6% 100.0%
9 100.0% 92.7% 100.0% 100.0% 100.0%
10 93.6% 76.5% 91.2% 100.0% 100.0%
11 95.4% 84.9% 96.4% 95.4% 100.0%
12 89.5% 71.8% 100.0% 100.0% 97.1%
13 97.1% 67.1% 94.5% 92.9% 100.0%
14 98.7% 75.3% 100.0% 100.0% 100.0%
15 99.7% 61.4% 99.7% 99.7% 100.0%
16 86.9% 84.1% 85.1% 89.2% 100.0%
17 98.5% 100.0% 98.5% 98.5% 100.0%
18 69.8% 39.3% 72.6% 71.5% 94.4%
19 100.0% 100.0% 98.9% 100.0% 100.0%
Average 95.1% 73.1% 95.1% 95.2% 99.3%
More Than
90% 74.0% 11.0% 68.0% 74.0% 100.0%
STDEV. 7.3% 15.7% 7.1% 7.6% 1.4%
In the thermal sensation prediction, the results are almost the same with the thermal comfort prediction (Tab. 4.8).
The ANN has the highest average prediction accuracy and the lowest standard deviation. Only one subject’s
prediction accuracy is less than 95%. The logistic regression has the lowest prediction accuracy. The best machine
learning algorithm for thermal sensation prediction is also ANN. All the prediction results of the ANN model are
higher than 90%, which indicates good prediction accuracy, and it has the highest average prediction accuracy
compared with the other four algorithms. Meanwhile, the standard deviation is 0.01. Small standard deviation
reports that the ANN model performs well in most of the subjects.
Table. 4.8 The Results of Thermal Sensation Prediction
Subject No.
Decision Tree Logistic Regression Random Forest Gradient Boosting ANN
1 86.2% 64.5% 88.6% 89.4% 94.7%
62

2 99.8% 1.0% 99.8% 100.0% 100.0%
3 86.2% 99.1% 86.2% 86.2% 100.0%
4 100.0% 96.6% 99.7% 100.0% 100.0%
5 94.3% 73.9% 91.7% 91.3% 100.0%
6 97.0% 72.2% 100.0% 100.0% 100.0%
7 80.8% 79.6% 81.7% 82.0% 100.0%
8 97.7% 78.8% 94.4% 99.0% 100.0%
9 100.0% 94.4% 99.7% 100.0% 100.0%
10 84.7% 99.2% 84.7% 84.7% 100.0%
11 97.4% 96.4% 100.0% 96.4% 100.0%
12 99.4% 86.8% 100.0% 100.0% 100.0%
13 82.2% 66.9% 90.1% 92.5% 100.0%
14 87.0% 99.7% 100.0% 97.7% 100.0%
15 95.3% 93.1% 89.0% 95.3% 100.0%
16 90.4% 81.1% 100.0% 96.3% 100.0%
17 80.5% 80.1% 91.6% 93.7% 100.0%
18 89.1% 98.9% 98.3% 98.4% 100.0%
19 92.8% 58.5% 91.3% 93.2% 100.0%
Average 91.6% 80.0% 94.0% 94.5% 99.7%
More Than 95% 11/19 9/19 14/19 15/19 19/19
STDEV. 0.07 0.23 0.06 0.05 0.01
4.5.2 Individual Machine Learning Model Training
The ANN algorithm has the best performance in both thermal comfort and thermal sensation predictions. For the
machine learning application, the model was trained before the validation experiments. There were two concerns for
the training dataset. First, for the thermal comfort and thermal sensation feedback, it was necessary to have a
response in all levels (cold, hot, and neutral for thermal sensation; comfortable, uncomfortable, and neutral for
thermal comfort). For the thermal sensation feedback, there were 13 subjects that have all three responses. For the
thermal comfort feedback, there were 15 subjects have all three responses. The number of both is 13. Second, the
control should be accurate and fast. The separate prediction of thermal comfort and thermal sensation needs two
loops for controlling the HVAC systems. The first step is thermal comfort prediction when the prediction result is
uncomfortable. The thermal sensation loop is triggered to control the HVAC systems based on the thermal sensation
prediction results. Considering these two aspects, the training model combined the feedback of thermal comfort and
thermal sensation. The combined thermal perception was the output of the machine learning model. The input
features were cheek left, cheek right, and air temperature. The rule for combining the thermal comfort and thermal
sensation was according to the survey feedback. The neutral and comfortable feedback was regarded as 0,
uncomfortable and hot regarded as 1, and uncomfortable and cold regarded as -1. The thermal perception used a
three-point scale, with -1 representing extra heating needed, 1 representing additional cooling required, and 0
serving as no change for the thermal settings.
Only 13 subjects have an integrated response in the experiments, and the data of these subjects are used for the
machine learning model training. The individual model performance in 4-cross-validation is good (Tab.4.9). The
prediction accuracy for all the subjects is 100%. The result of the ANN model predictions is 100% for each subject.
The results are meaningful. Because the ANN can do multiple calculations, it has multiple layers to fit the training
data.

63

Table 4.9. Prediction Accuracy of Individual Thermal Perception
Subject No. Prediction Accuracy
1 100.0%
2 100.0%
3 100.0%
4 100.0%
5 100.0%
6 100.0%
8 100.0%
12 100.0%
13 100.0%
14 100.0%
15 100.0%
17 100.0%
19 100.0%
4.6 Summary
The data analysis for the training data collection experiment was significant. Data was collected through 20 subjects.
Since the data of one subject was missing, the total number of subjects analyzed was 19. Moving average was the
first step in data analysis. With the comparison of the different time window, the 30s was chosen for the moving
average calculations. According to the comparison of 19 subjects’ thermal perception predictions, left cheek and
right cheek were the most suitable facial parts for the machine learning model.
Moreover, the air temperature was also a reasonable input feature for the machine learning model. The tested
method is 4-cross-validation for the machine learning algorithms. The tested average prediction accuracy with three
input features (left cheek, right cheek, and air temperature) is 95.57%, and the standard deviation is 0.073 for
thermal comfort. For the thermal sensation prediction, the average accuracy is 95.20%, and the standard deviation is
0.051. The comparison of different algorithms supported that the ANN was the most reliable algorithm for thermal
perception predictions. Combining thermal comfort and thermal sensation as the output feature improved the
prediction accuracy to 100% for all subjects. There were 13 machine learning models trained for 13 participated
subjects because of the limitations of the thermal feedback. The disadvantage of the machine learning model is
overfitting. The overfitting means the algorithm can do complex calculations based on the training data, and the
results will fit the training data perfectly, which may cause poor performance of the testing data. The validation
experiments are necessary for the prediction models.

64

5. VALIDATION EXPERIMENTS DATA AND RESULTS
Two experiments were conducted to validating the machine learning model performance. Six subjects were chosen
from the 13 who have the individual machine learning model (3 males and 3 females). One validation experiment
was tested at the BuroHappold Los Angeles office. Another validation experiment was conducted in the experiment
chamber at USC. The real office validation experiments aimed to check the performance of the machine learning
model, and the validation experiments in the experiment chamber tested the performance of the automatic control
HVAC systems. The discussion is divided into two parts: the results’ discussion of the real office experiments, and
the results’ discussion of the chamber experiments (Fig. 5.1).

Figure 5. 1 Workflow of the validation experiment
5.1 Model Validation Experiments in the Real Office
The model validation experiments were done at the BuroHappold Los Angeles office. As mentioned in the
methodology, each experiment lasted for 135 minutes. There were 24 feedbacks for each subject. The data analysis
compared the prediction results with feedback from the subjects. There were 3 males and 3 females. The data
analysis results show that the prediction accuracy is different for different models (Tab. 5.1). The model of subjects
C and D has better prediction performance than the other four. However, subject F has the lowest prediction
accuracy. Mean absolute error (MAE) is the measurement of the difference between two variables.
Table 5.1 The Results of Validation Experiments
Subject No. Prediction Accuracy MAE
A(M1 ）
70.8% 0.291
B(F1) 79.2% 0.208
C(F2) 100.0% 0.000
65

D(M2) 100.0% 0.000
E(M3) 78.3% 0.217
F(F3) 54.2% 0.458
Average 80.4%

STDEV. 0.161

Subject A is a male, and the model prediction accuracy is 70.8%. The air temperature change during the experiment
was less than 1℃ (Tab.5.2). The real office building had good control of the environment, and the skin temperature
of subject A was also minimum. The difference is 0.4℃, and the standard deviation is almost 0.1. The data indicates
that the status of the subject and the environment are both stable. However, the subjects gave neutral for most of the
time and cold for 2 survey points. The incorrect prediction points are survey 6 and 7. The machine learning
predicted the continues results with the previous, and the subject changed the feedback from neutral to cold. The air
temperature only had 0.1℃ change, and for the survey points 12, 13,14,17, and18, the subject gave continuous
feedback in neutral, but the machine learning algorithm gave predictions in the “cold” category.
Table 5.2. Prediction and Feedback of subject A
Survey
Order
Cheek
Left
Cheek
Right
Air
Temperature Prediction Feedback
Absolute
prediction-
Feedback
1 33.8 33.9 23.4 0 0 0
2 33.8 33.7 23.5 0 0 0
3 33.8 33.8 23.6 0 0 0
4 33.8 33.7 23.6 0 0 0
5 33.7 33.6 23.3 0 0 0
6 33.6 34.0 23.6 0 -1 1
7 33.5 33.9 23.7 0 -1 1
8 33.4 33.7 23.5 0 0 0
9 33.4 33.7 23.8 0 0 0
10 33.5 33.9 23.9 0 0 0
11 33.5 33.7 23.9 0 0 0
12 33.4 33.7 23.9 -1 0 1
13 33.4 33.7 24.2 -1 0 1
66

14 33.4 33.8 23.7 -1 0 1
15 33.4 33.9 23.8 0 0 0
16 33.5 34.0 24.0 0 0 0
17 33.4 33.7 23.8 -1 0 1
18 33.4 33.7 23.6 -1 0 1
19 33.5 33.7 23.6 0 0 0
20 33.6 33.9 23.4 0 0 0
21 33.6 33.8 23.4 0 0 0
22 33.6 33.7 23.5 0 0 0
23 33.7 33.7 23.8 0 0 0
Min 33.4 33.6 23.3 - - -
Max 33.8 34.0 24.2 - - -
Average 33.6 33.8 23.7 - - -
Difference 0.4 0.4 0.8 - - -
STDVE 0.14 0.10 0.21 0.41 0.28 0.46
The skin temperature change line is flat, which reveals the skin temperature is stable during the experiments (Fig.
5.2). As discussed above, the air temperature change range is less than 1℃ (Fig. 5.3). However, it changed
frequently within the 1℃ range. It reports that the subject was in a stable thermal environment.
67

Figure 5. 2 Skin temperature change (subject A)

Figure 5. 3 Air temperature change (subject A)
Subject B is a male, and the model prediction accuracy is 79.16%. The air temperature changing range is 1.5℃, and
the skin temperature change range is 2℃ (Tab. 5.3). The subject B gave continuous feedback during the entire
experiment. However, the ANN model changed the predictions in surveys 17–21. Because in this period, the right
cheek skin temperature increased to more than 33℃, and the prediction results changed from cold to neutral.
Table 5.3. Prediction and Feedback of subject B
Survey Order Cheek Left
Cheek
Right
Air
Temperature Prediction Feedback
Absolute
prediction-
Feedback
1 31.0 31.3 22.4 -1 -1 0
2 30.9 31.8 22.7 -1 -1 0
30.00
31.00
32.00
33.00
34.00
35.00
36.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Skin Temperature Change
Cheek Left Cheek Right
22.50
23.00
23.50
24.00
24.50
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Air Temperature
Survey Order
Air Temperature Change
Air Temperature
68

3 30.7 32.2 22.7 -1 -1 0
4 31.5 32.6 22.9 -1 -1 0
5 31.4 32.4 22.8 -1 -1 0
6 31.4 32.6 23.0 -1 -1 0
7 31.4 32.6 23.0 -1 -1 0
8 31.4 32.7 23.1 -1 -1 0
9 31.4 32.6 23.1 -1 -1 0
10 31.5 32.8 23.2 -1 -1 0
11 31.5 32.6 23.4 -1 -1 0
12 31.5 32.6 23.3 -1 -1 0
13 31.5 32.8 23.3 -1 -1 0
14 31.5 32.4 23.7 -1 -1 0
15 31.7 32.8 23.4 -1 -1 0
16 31.7 32.8 23.3 -1 -1 0
17 31.9 33.0 23.6 0 -1 1
18 31.8 33.0 23.5 0 -1 1
19 31.9 33.0 23.6 0 -1 1
20 32.0 33.2 23.7 0 -1 1
21 31.9 33.0 23.9 0 -1 1
22 31.9 32.4 23.7 -1 -1 0
23 31.9 32.2 23.7 -1 -1 0
24 31.9 32.1 23.7 -1 -1 0
Min 30.7 31.3 22.4 - - -
Max 32.0 33.2 23.9 - - -
Difference 1.3 1.9 1.5 - - -
Average 31.5 32.6 23.3

69

STDEV. 0.3 0.4 0.4 0.4 0.0 0.4

The skin temperature increased slowly during the experiments. The right cheek always had a higher temperature
than the left cheek (Fig. 5.4). The air temperature increased slowly from 22.5℃ to 24℃ (Fig.5.5).

Figure 5. 4 Skin temperature change (subject B)

Figure 5. 5 Air temperature change (subject B)
Subject C is a female, and the model prediction accuracy is 100%. The environment temperature changed in the
range of 0.3℃ during the entire experiments (Tab. 5.4). The left cheek had a large temperature change range, which
was 3.2℃. The right cheek only had a temperature change in 0.8℃. Only the survey point 7 had the temperature in
30.0
31.0
32.0
33.0
34.0
35.0
36.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Skin Temperature
Survey Order
Skin Temperature Change
Cheek Left Cheek Right
22.0
22.5
23.0
23.5
24.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Air Temperature
Survey Order
Air Temperature Change
Air Temperature
70

32.3℃. The other data was within a 1℃ difference. The prediction accuracy is 100%. The feedback of the subject
was constant in neutral, and the prediction results were also constant in neutral.
Table 5.4 Prediction and Feedback of subject C
Survey Order Cheek Left Cheek Right Air Temperature Prediction Feedback
Absolute
prediction-
Feedback
1 34.7 33.8 23.9 0 0 0
2 34.6 33.9 24.1 0 0 0
3 34.6 34.2 24.1 0 0 0
4 34.8 34.2 24.0 0 0 0
5 34.5 34.2 23.9 0 0 0
6 33.8 34.2 23.9 0 0 0
7 32.3 34.2 24.0 0 0 0
10 35.1 34.2 24.1 0 0 0
11 35.4 34.0 24.0 0 0 0
12 35.4 34.1 24.0 0 0 0
13 35.4 34.4 23.9 0 0 0
14 35.3 34.4 23.9 0 0 0
15 35.3 34.2 24.0 0 0 0
16 35.0 34.6 24.0 0 0 0
17 35.0 34.5 23.9 0 0 0
18 34.9 34.5 24.0 0 0 0
19 35.1 34.4 24.1 0 0 0
20 34.9 34.6 24.1 0 0 0
21 35.0 34.6 23.9 0 0 0
22 34.9 34.3 23.9 0 0 0
23 34.8 34.2 24.0 0 0 0
24 34.9 34.2 24.2 0 0 0
71

Min 32.3 33.8 23.9 - - -
Max 35.4 34.6 24.2 - - -
Difference 3.2 0.8 0.3 - - -
STDEV. 0.7 0.2 0.1 0.0 0.0 0.0
The skin temperature change of the subject change was fluctuant, especially at the beginning and end of the
experiments (Fig. 5.6). The air temperature was stable at the beginning of the experiment, with a little fluctuation at
the end (Fig. 5.7).

Figure 5. 6 Skin temperature change (subject C)

Figure 5. 7 Air temperature change (subject C)
30.0
31.0
32.0
33.0
34.0
35.0
36.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Skin Temperature
Survey Order
Skin Temperature Change
Cheek Left Cheek Right
22.0
22.5
23.0
23.5
24.0
24.5
25.0
25.5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Air Temperature
Survey Order
Air Temperature Change
Air Temperature
72

Subject D is a male, and the prediction accuracy is 100%. The skin temperature change range of the left cheek was
1.2℃, and the right cheek was 1.6 ℃ (Tab. 5.5). The air temperature change range was 0.5℃. The feedback was
constant in cold, and the prediction was also constant in cold.
Table 5.5. Prediction and Feedback of subject D
Survey Order
Cheek
Left
Cheek
Right Air Temperature Prediction Feedback
Absolute
prediction-
Feedback
1 34.0 34.3 24.0 -1 -1 0
2 34.4 34.5 23.9 -1 -1 0
3 34.4 34.4 23.8 -1 -1 0
4 34.8 34.5 23.7 -1 -1 0
5 34.9 34.9 23.8 -1 -1 0
6 35.0 35.3 23.8 -1 -1 0
7 35.0 35.0 23.6 -1 -1 0
8 35.2 35.2 23.7 -1 -1 0
9 35.0 35.4 23.7 -1 -1 0
10 35.0 35.4 23.8 -1 -1 0
11 35.1 35.1 23.6 -1 -1 0
12 35.1 34.7 23.6 -1 -1 0
13 34.9 34.5 23.7 -1 -1 0
14 34.8 34.6 23.5 -1 -1 0
15 34.8 34.7 23.5 -1 -1 0
16 34.8 34.4 23.6 -1 -1 0
17 34.7 34.5 23.7 -1 -1 0
18 34.8 34.9 24.0 -1 -1 0
19 34.8 34.6 24.0 -1 -1 0
20 34.8 34.5 23.8 -1 -1 0
21 34.7 34.4 23.7 -1 -1 0
73

22 34.5 34.2 23.6 -1 -1 0
23 34.6 33.8 23.8 -1 -1 0
24 34.6 34.0 23.8 -1 -1 0
Min 34.0 33.8 23.5 - - -
Max 35.2 35.4 24.0 - - -
Difference 1.2 1.6 0.5 - - -
STDEV. 0.3 0.4 0.1 0.0 0.0 0.0
The skin temperature increased slowly at the beginning of the experiment and decreased at the end of the experiment
(Fig. 5.8). The air temperature was stable during the experiment (Fig. 5.9).

Figure 5. 8 kin temperature change (subject D)
30.0
31.0
32.0
33.0
34.0
35.0
36.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Skin Temperature
Survey Order
Skin Temperature Change
Cheek Left Cheek Right
74

Figure 5. 9 Air temperature change (subject D)
Subject E is a female, and the prediction accuracy is 100%. The skin temperature change range of the left cheek was
1.2℃, and the right cheek was 1.6℃ (Tab. 5.6). The air temperature change range was 0.5℃. The feedback was
constant in cold, and the prediction was also constant in cold.
Table 5.6. Prediction and Feedback of subject E
Survey
No. Cheek Left
Cheek
Right
Air
Temperature Prediction Feedback
Absolute
prediction-
Feedback
1 34.6 32.9 24.0 -1 0 1
2 34.4 32.5 23.9 -1 0 1
3 34.4 34.0 23.9 0 0 0
4 34.3 33.0 24.0 -1 0 1
5 34.8 34.6 23.9 0 0 0
6 35.0 34.6 23.9 0 0 0
7 34.9 34.7 24.1 0 0 0
8 34.8 34.5 24.1 0 0 0
9 34.5 34.1 24.0 0 0 0
10 34.5 34.2 24.0 0 0 0
11 34.2 34.2 24.0 0 0 0
22.0
22.5
23.0
23.5
24.0
24.5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Air Temperature
Survey Order
Air Temperature Change
Air Temperature
75

12 34.0 34.0 24.1 0 0 0
13 33.3 33.8 24.0 0 0 0
14 33.2 33.5 24.0 0 0 0
15 32.6 33.5 24.0 -1 0 1
17 34.3 33.9 24.9 0 0 0
18 34.1 33.9 24.7 0 0 0
19 34.0 34.0 24.9 0 0 0
20 33.8 32.8 24.8 -1 0 1
21 33.4 33.8 25.0 0 0 0
22 34.1 33.6 25.0 0 0 0
23 34.5 33.7 24.2 0 0 0
24 33.8 33.8 23.5 -1 -1 0
Max 35.0 34.7 25.0

Min 32.6 32.5 23.5

Difference 2.4 2.1 1.4

Average 34.2 33.8 24.2

STDEV. 0.583 0.569 0.413 0.439 0.204 0.412

76

Figure 5. 10 Skin temperature change (subject E)

The model for subject F has the lowest prediction accuracy of only 54.2%.
The air temperature change range was 0.4℃, which means the environment temperature is stable (Tab. 5.6). The
skin temperature change range was 2.4℃. During the experiment, the feedback of subject F was not constant .
However, the prediction of the thermal perception was constant in cold. Compared with survey points 8 and 9, the
left cheek had a 0.1℃ difference, and the right cheek and the air temperature were the same. The prediction results
of these two points were cold (-1). However, the feedback of the subject F in survey 8 was “cold” and “neutral” in
survey 9. Because this data indicated that why machine learning algorithm didn ’t give the right prediction, the air
temperature in these two points were same, and the facial skin temperature only had 0.1 ℃ difference, so the
algorithm give the same prediction of these two points. but the subject gave different feedback.
Table 5.7. Prediction and Feedback of subject F
30.0
31.0
32.0
33.0
34.0
35.0
36.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Skin Temperature
Survey Order
Skin Temperature Change
Cheek Left Cheek Right
22.0
22.5
23.0
23.5
24.0
24.5
25.0
25.5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Air Temperature
Survey Order
Air Temperature Change
Air Temperature
Figure 5. 11 Air temperature change (subject E)
77

Survey Order
Cheek
Left
Cheek
Right Air Temperature Prediction Feedback
Absolute
prediction-
Feedback
1 32.7 33.7 23.9 0 -1 1
2 32.8 34.2 23.8 -1 -1 0
3 33.2 34.0 23.9 -1 -1 0
4 34.5 34.5 23.9 -1 -1 0
5 33.9 34.5 23.9 -1 -1 0
6 33.6 34.1 23.8 -1 -1 0
7 33.7 34.1 23.8 -1 -1 0
8 33.9 34.4 23.7 -1 -1 0
9 34.0 34.6 23.7 -1 0 1
10 33.9 34.4 23.8 -1 -1 0
11 34.3 34.3 23.8 -1 -1 0
12 34.6 34.3 23.8 -1 -1 0
13 34.2 34.3 23.7 -1 -1 0
14 34.4 34.2 23.5 -1 0 1
15 34.6 33.5 23.6 -1 0 1
16 34.7 33.3 23.6 -1 0 1
17 35.0 32.2 23.6 -1 0 1
18 34.9 34.2 23.6 -1 0 1
19 35.0 34.4 23.7 -1 0 1
20 34.8 34.4 23.7 -1 0 1
21 34.9 34.5 23.7 1 0 1
22 34.2 34.2 23.6 -1 -1 0
23 34.0 33.9 23.7 -1 0 1
24 34.5 33.7 23.7 -1 -1 0
78

Min 32.7 32.2 23.5 - - -
Max 35.0 34.6 23.9 - - -
Difference 2.4 2.4 0.4 - - -
Average 34.2 34.1 23.7

STDEV. 0.630 0.518 0.104 0.439 0.493 0.498

The skin temperature of subject F was changed in a narrow range less than 1℃, and the air temperature was also
stable with 0.5℃ temperature change. This indicates that the subject stayed in the stable thermal environment and
the physiology change of the skin temperature is not dramatic (Fig. XX and Fig. XX).

Figure 5. 12 Skin temperature change (subject F)
30.0
31.0
32.0
33.0
34.0
35.0
36.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Skin Temperature
Survey Order
Skin Temperature Change
Cheek Left Cheek Right
79

Figure 5. 13 Air temperature change (subject F)
According to the discussion above, there were several conclusions of the validation experiments in the real office.
The air temperature change range was narrow during all the experiments. The skin temperature change range had a
variety of different subjects. When the feedback of the subjects is constant, the prediction accuracy is high. When
the feedback of the subjects changes with the minimum air temperature and skin temperature change, the prediction
accuracy is decreased. There are only three classes of the classification model. It is difficult for the machine learning
model to identify the tiny temperature change.
5.2 System Validation Experiments in the Environment Chamber
The validation experiments for the automatic control of the HVAC systems were done at the experiment chamber in
USC. The subjects were the same as in the model validation experiments. The system validation experiments were
used to evaluate the performance of the model and control systems. During the tests, the input features were real-
time data of cheek left, cheek right, and air temperature. The real-time data was preprocessed in real time during the
experiments. The evaluation method was calculating the satisfied time during the whole test. The average satisfied
time is 79.7%. Half of the subjects gave a neutral response at 88.9% of the entire experiment period (Tab.5.8 and
5.9). In the system validation experiments, the control performance of males is better compared with the
performance of females. The average satisfied time of males is 85.2% and females 74.1%.
Table. 5.8 Results of Experiments
Subject No. Satisfied Time
A(M1 ）
88.9%
B(F1) 55.6%
C(F2) 77.8%
D(M2) 77.8%
E(M3) 88.9%
22.0
22.5
23.0
23.5
24.0
24.5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Air Temperature
Survey Order
Air Temperature Change
Air Temperature
80

F(F3) 88.9%
Average 79.7%
STDEV. 0.118
Table 5.9. Survey Results of Experiments
Survey No. A B C D E F
1 0 -1 -1 0 0 0
2 0 0 -1 0 0 -1
3 0 -1 -1 0 0 0
4 0 0 -1 0 0 0
5 0 -1 0 0 1 0
6 -1 0 0 0 0 0
7 0 -1 0 -1 0 0
8 0 0 0 -1 0 0
9 0 0 0 0 0 0
Neutral 8/9 5/9 5/9 7/9 8/9 8/9

5.3 Summary
The validation experiments evaluated the performance of the individual thermal perception prediction models and
the control of the HVAC systems. The results reported that different models had different performance in the
validation experiments. For the model validation experiments, the average prediction accuracy was 80.4%. The
prediction accuracy for subjects C and D was 100%. The thermal environment in the office was pretty stable, and
the air temperature difference was less than 1℃. The individual model had accurate overall thermal predictions. It
indicated the overall thermal perception for the occupants. However, when the air temperature change and facial
skin temperature change were less than 2℃, the prediction model cannot identify it. In the validation experiments of
the automatic control of the HVAC system, the HVAC systems were controlled by the output signal of the ANN
model. The automatic control step was achieved successfully because of the prediction accuracy and different body
status of the subjects. The average satisfied time of 6 subjects was 79.7%.
81

6. CONCLUSION
It could be possible to use facial skin temperature to predict the thermal comfort and thermal sensation level of the
occupants. The human subject experiments were conducted to this possibility (Fig. XX).

Figure 6. 1 Workflow of Human Subject Experiment
There were 3 rounds of experiments: training data collection experiments, machine learning model validation
experiments, and HVAC system control validation experiments. Twenty subjects participated in the training data
collection experiments. The experiment time for each subject was 105 minutes, and the thermal perception survey
was given every 15 minutes. For the validation experiments, 6 subjects chosen from the first-round experiments re-
participated in both validation experiments. The model validation experiments lasted for 135 minutes for each
subject in the real office. The HVAC system control validation experiments lasted for 60 minutes. In both validation
experiments, the survey time interval was 5 minutes. Facial skin temperature and air temperature were collected as
input features for the machine learning models. Heart rate was collected as reference human parameters. Carbon
dioxide index and relative humidity were monitored during the experiments to eliminate the influence of other
thermal environment factors, but these were not used.
The data processing steps included manual data preprocessing and machine learning model training. The machine
learning model training was based on suitable input features selection, machine learning algorithm selection, and
machine learning model performance testing. The input features for thermal perception prediction were cheek left,
cheek right, and air temperature. The output features were thermal perception and thermal comfort. The ANN
algorithm had the best performance among the 5 different classification machine learning algorithms.
6.1 Conclusion of the Human Subject Experiments
The data analysis of the training data collection experiments is divided into two parts: general data analysis and
individual data analysis. The general data analysis shows that males and females have different facial skin
temperature in the same thermal conditions, and the air temperature of the same thermal condition is also different
between genders. Moreover, the machine learning model of genders and the group have the low prediction accuracy.
The machine learning prediction model cannot be used for group thermal perception predictions. The one-way
ANOVA analysis of HRV index and thermal perception has a large p-value, which means there is no statistical
significance between HRV and thermal perception. HRV is not a suitable parameter for thermal perception
prediction.
The 30s was the suitable time window for the moving average calculation. It means before the whole calculation, the
average calculation was based on 30 seconds data The moving average was important in data processing steps. For
82

the individual data analysis, the input feature importance comparison indicated that the cheek left and cheek right
were the most important two parts in the face for both thermal comfort and thermal sensation predictions. The
average prediction accuracy for 5 input features was 93.5% for both thermal comfort and thermal sensation
predictions. The 2 input features model had lower thermal perception predictions compared with the 5 input
features. The 3 input features had the best prediction accuracy in both thermal comfort and thermal sensation
conditions. The average prediction accuracy for thermal comfort is 95.57%, and for thermal sensation, it is 95.2%.
The ANN algorithm had the best performance compared with decision tree, logistic regression, random forest, and
gradient boosting.
The validation experiments show that machine learning models have different prediction performance for different
subjects. However, the average prediction accuracy for individual models is acceptable, which is 80.4%. For the
system control validation, the average satisfied time period is 79.7%.
6.2 Limitations
There are four main limitations in the research:
The number of survey points. In the first-round experiments, there were 6 survey points for each subject. In the
validation experiments, there were 24 survey points and 9 survey points. The machine learning model required a
large amount of data. The limited survey points reduced the prediction accuracy and the reliability of the prediction
results. There is no strict rule to calculate the survey points. But more than 50 survey points will be good for the data
analysis (Choi & Yeom, 2017b).
The number of subjects. Though the machine learning model is designed for individuals, a large number of subjects
is necessary. There were only 20 subjects in the first-round experiments and 6 subjects in the validation experiments.
Different individual models have different performance in thermal perception predictions, especially in the
validation experiments. It is necessary to increase the size to get robust statistical analysis results. The suitable
number for the first round is 10 to 20 (Choi & Yeom, 2017b). However, some studies have less than 10 subjects. In
the thermal study domain, there is no strict rule for the number of subjects.
The accuracy of the facial skin temperature sensor. The range of facial skin temperature change is 3℃ in the
experiments. The narrow temperature changing required the accuracy of the skin temperature sensor. The typical
skin temperature range for the subject in normal thermal environment is 32℃ to 34℃. The normal temperature
change of the subject in limited air temperature is less than 1℃. The accuracy of the skin temperature sensor is
0.2℃. This indicates a large measurement error of up to 20% of 1℃ temperature change. A skin temperature sensor
with higher accuracy reduces the measurement errors and improves the thermal perception prediction accuracy.
The limitation of the subject survey. In the subject survey no clear definitions of stress level mentioned. The missing
definitions of stress level may cause inaccurate response of the stress level feedback.
6.3 Future Work
Seven things can be done in the future:
More subjects in the validation experiments. There should be 10 subjects for each validation experiments. Only 6
subjects of the validation experiments finished because of the limitation of time and volunteers. 4 more subjects will
add for the validation experiments in April 2019.
Have a larger sample size that can be considered in future experiments. There were limited subject variety in the
experiments. The future work should expand the subject varieties, for example, people in different age groups, BMI
conditions, and ethnicities.
83

Make the integrated automatic control system more sophisticated. The automatic control system is simply to turn
on/off the device. More complicated operations should be considered in the future work, for example, to set up the
temperature points on the ACs or heaters.
Other thermal factors could be tested. The experiments only tested the air temperature. Air speed, radiant
temperature, and relative humidity could be used. The relationship between thermal comfort/thermal sensation and
other factors should be illustrated.
Multiple places could be tested. The experiments only tested in the environment chamber and office building. The
classroom, gymnasium, and even outdoor environment can be tested to monitor the performance of the machine
learning algorithm predictions.
Multiple activities’ status could be tested. The experiments only tested the office working situation. The changed
status can be one of the future tasks. The thermal comfort and thermal sensation feedback are different in different
activities. For example, when eating or exercising, people may like lower temperature than usual because of the
different metabolism.
The functional infrared camera could be used in future experiments. If the funding is enough, the functional infare
camera will be a good choice. There are some infare camera can detect the detail skin temperature of the face
without touching the subject.
6.4 Conclusion
Facial skin temperature has the potential to make predictions for thermal comfort and thermal sensation. Compared
with the chin and forehead, the cheek has the big potential to support the thermal comfort and thermal sensation
predictions. The ANN algorithm has the reasonable prediction accuracy for thermal perceptions in the real office
conditions and experiment chamber for individuals. The group thermal perception prediction based on facial skin
temperature is difficult to achieve because the prediction accuracy is less than 80%. It is not meaningful to do the
group machine learning model.

84

REFERENCE
Ã, E. A., Zhang, H., & Huizenga, C. (2006). Partial- and whole-body thermal sensation and
comfort — Part I : Uniform environmental conditions, 31, 53–59.
https://doi.org/10.1016/j.jtherbio.2005.11.028
Academy, N., Academy, N., & States, U. (2016). Identification of Insulin in Rat Brain Author
( s ): J . Havrankova , D . Schmechel , J . Roth and M . Brownstein Source : Proceedings of
the National Academy of Sciences of the United States of America , Published by : National
Academy of Sciences Stable, 75(11), 5737–5741.
Argiriou, A. A., Bellas-Velidis, I., Kummert, M., & André , P. (2004). A neural network
controller for hydronic heating systems of solar buildings. Neural Networks, 17(3), 427–
440. https://doi.org/10.1016/j.neunet.2003.07.001
Ariyaratnam, S., & Rood, J. p. (1990). Measurement of facial skin temperature. Journal of
Dentistry, 18(5), 250–253. https://doi.org/10.1016/0300-5712(90)90022-7
Arundel, A. V., Sterling, E. M., Biggin, J. H., & Sterling, T. D. (1986). Indirect health effects of
relative humidity in indoor environments. Environ Health Perspect, 65(3), 351–361.
https://doi.org/10.1289/ehp.8665351
ASHRAE. (2010). Thermal environmental conditions for human occupancy. ASHRAE Inc.,
2010, 42. https://doi.org/ISSN 1041-2336
Barlow, S. (2002). How adaptive comfort theories might influence future low energy office
refurbishment strategies Existing building stock, 7730, 1–12.
Barlow, S., & Fiala, D. (2007). Occupant comfort in UK offices — How adaptive comfort
theories might influence future low energy office refurbishment strategies, 39, 837–846.
https://doi.org/10.1016/j.enbuild.2007.02.002
Berkeley, U. C. (2004). Predict Comfort. ASHRAE Journal, (August). Retrieved from
http://repositories.cdlib.org/cedr/cbe/ieq/OlesenBrager2004_comfort/
Casillas, J., & Cordon, O. (2003). Fuzzy Control of HVAC Systems Optimized by Genetic
Algorithms Fuzzy Control of HVAC Systems Optimized by Genetic Algorithms, (June
2014). https://doi.org/10.1023/A
Choi, J. H., & Yeom, D. (2017a). Investigation of the relationships between thermal sensations
of local body areas and the whole body in an indoor built environment. Energy and
Buildings, 149, 204–215. https://doi.org/10.1016/j.enbuild.2017.05.062
Choi, J. H., & Yeom, D. (2017b). Study of data-driven thermal sensation prediction model as a
function of local body skin temperatures in a built environment. Building and Environment,
121, 130–147. https://doi.org/10.1016/j.buildenv.2017.05.004
Dai, C., Zhang, H., Arens, E., & Lian, Z. (2017). Machine learning approaches to predict thermal
demands using skin temperatures: Steady-state conditions. Building and Environment, 114,
1–10. https://doi.org/10.1016/j.buildenv.2016.12.005
Darshana Wickramaratne. (2015). Electronic Theses and Dissertations UC Merced. Ph.D Thesis:
Electronic, Vibrational and Thermoelectric Properties of Two-Dimensional Materials.
85

De Dear, R. J., Akimoto, T., Arens, E. A., Brager, G., Candido, C., Cheong, K. W. D., … Zhu,
Y. (2013). Progress in thermal comfort research over the last twenty years. Indoor Air,
23(6), 442–461. https://doi.org/10.1111/ina.12046
Dear, R. De, Brager, G., & Cooper, D. (1997). Developing an Adaptive Model of Thermal
Comfort and Preference, (March).
Ferrari, S., & Zanotto, V. (2012). Adaptive comfort : Analysis and application of the main
indices. Building and Environment, 49, 25–32.
https://doi.org/10.1016/j.buildenv.2011.08.022
Fiala, D., Lomas, K. J., & Stohrer, M. (2001). Computer prediction of human thermoregulatory
and temperature responses to a wide range of environmental conditions, 143–144.
Földváry Ličina, V., Cheung, T., Zhang, H., de Dear, R., Parkinson, T., Arens, E., … Zhou, X.
(2018). Development of the ASHRAE Global Thermal Comfort Database II. Building and
Environment, 142, 502–512. https://doi.org/10.1016/j.buildenv.2018.06.022
Freire, R. Z., Oliveira, G. H. C., & Mendes, N. (2006). Non-Linear Predictive Controllers for
Thermal Comfort Optimization and Energy Saving. IFAC Proceedings Volumes, 39(19),
87–92. https://doi.org/https://doi.org/10.3182/20061002-4-BG-4905.00015
Ghahramani, A., Castro, G., Becerik-Gerber, B., & Yu, X. (2016a). Infrared thermography of
human face for monitoring thermoregulation performance and estimating personal thermal
comfort. Building and Environment, 109, 1–11.
https://doi.org/10.1016/j.buildenv.2016.09.005
Ghahramani, A., Castro, G., Becerik-Gerber, B., & Yu, X. (2016b). Infrared thermography of
human face for monitoring thermoregulation performance and estimating personal thermal
comfort. Building and Environment, 109, 1–11.
https://doi.org/10.1016/j.buildenv.2016.09.005
Goyal, S., Barooah, P., & Middelkoop, T. (2015). Experimental study of occupancy-based
control of HVAC zones. Applied Energy, 140, 75–84.
https://doi.org/10.1016/j.apenergy.2014.11.064
Homod, R. Z., Mohamed Sahari, K. S., Almurib, H. A. F., & Nagi, F. H. (2012). RLF and TS
fuzzy model identification of indoor thermal comfort based on PMV/PPD. Building and
Environment, 49(1), 141–153. https://doi.org/10.1016/j.buildenv.2011.09.012
Hoyt, T., Lee, K. H., Zhang, H., Arens, E., & Webster, T. (2009). Indoor Environmental Quality
(IEQ) Title Energy savings from extended air temperature setpoints and reductions in room
air mixing Publication Date ENERGY SAVINGS FROM EXTENDED AIR
TEMPERATURE SETPOINTS AND REDUCTIONS IN ROOM AIR MIXING.
International Conference on Environmental Ergonomics. Retrieved from
https://escholarship.org/uc/item/28x9d7xj
Huizenga, C., Hui, Z., Duan, T., & Arens, E. (1999). Indoor Environmental Quality (IEQ) Title
An improved multinode model of human physiology and thermal comfort AN IMPROVED
MULTINODE MODEL OF HUMAN PHYSIOLOGY AND THERMAL COMFORT.
Retrieved from
86

https://cloudfront.escholarship.org/dist/prd/content/qt1ms313wz/qt1ms313wz.pdf?t=p3hm8
2
Kelly, M. (2014). UC Berkeley, D. https://doi.org/10.1111/ina.12046
Kim, J., Schiavon, S., & Brager, G. (2018). Personal comfort models – A new paradigm in
thermal comfort for occupant-centric environmental control. Building and Environment,
132, 114–124. https://doi.org/10.1016/j.buildenv.2018.01.023
Kononenko, I., & Kukar, M. (n.d.). Machine learning and data mining : introduction to
principles and algorithms.
Kosonen, R., & Tan, F. (2004). Assessment of productivity loss in air-conditioned buildings
using PMV index. Energy and Buildings, 36(10 SPEC. ISS.), 987–993.
https://doi.org/10.1016/j.enbuild.2004.06.021
Kreider, J. F., Xing An Wang, Anderson, D., & Dow, J. (1992). Expert systems, neural networks
and artificial intelligence applications in commercial building HVAC operations.
Automation in Construction, 1(3), 225–238. https://doi.org/10.1016/0926-5805(92)90015-C
Li, D., Menassa, C. C., & Kamat, V. R. (2018). Non-intrusive interpretation of human thermal
comfort through analysis of facial infrared thermography. Energy and Buildings, 176, 246–
261. https://doi.org/10.1016/j.enbuild.2018.07.025
Lin, C., Federspiel, C. C., & Auslander, D. M. (2002). Multi-sensor single-actuator control of
HVAC systems. International Conference for Enhanced Building Operations.
Linden, W. Van Der, Loomans, M., & Hensen, J. (2008). Adaptive thermal comfort explained by
PMV. Indoor Air, (August), 8.
Ló pez, A., Sá nchez, L., Doctor, F., Hagras, H., & Callaghan, V. (2004). An evolutionary
algorithm for the off-line data driven generation of fuzzy controllers for intelligent
buildings. Conference Proceedings - IEEE International Conference on Systems, Man and
Cybernetics, 1, 42–47. https://doi.org/10.1109/ICSMC.2004.1398270
Martin, R. A., Federspiel, C. C., & Auslander, D. M. (2002). Supervisory Control for Energy
Savings and Thermal Comfort in Commercial Building {HVAC} Systems. Proceedings of
the AAAS 2002 Spring Symposium, 67–74.
Moon, J. W. (2012). Performance of ANN-based predictive and adaptive thermal-control
methods for disturbances in and around residential buildings. Building and Environment,
48(1), 15–26. https://doi.org/10.1016/j.buildenv.2011.06.005
Moon, J. W., Jung, S. K., Lee, Y. O., & Choi, S. (2015). Prediction performance of an artificial
neural network model for the amount of cooling energy consumption in hotel rooms.
Energies, 8(8), 8226–8243. https://doi.org/10.3390/en8088226
Murakami, Y., Terano, M., Mizutani, K., Harada, M., & Kuno, S. (2007). Field experiments on
energy consumption and thermal comfort in the office environment controlled by
occupants’ requirements from PC terminal. Building and Environment, 42(12), 4022–4027.
https://doi.org/10.1016/j.buildenv.2006.05.012
87

Newsham, G. R. (1997). Clothing as a thermal comfort moderator and the effect on energy
consumption. Energy and Buildings, 26(3), 283–291. https://doi.org/10.1016/S0378-
7788(97)00009-1
Niskanen, J., & Ranta-aho, P. O. (2017). USER ’ S GUIDE HRV Standard.
https://doi.org/10.1016/j.enconman.2005.10.016
Pourshaghaghy, A., & Omidvari, M. (2012). Examination of thermal comfort in a hospital using
PMV-PPD model. Applied Ergonomics, 43(6), 1089–1095.
https://doi.org/10.1016/j.apergo.2012.03.010
RAUDYS, A., & PABARŠKAITĖ, Ž. (2018). Optimising the Smoothness and Accuracy of
Moving Average for Stock Price Data. Technological and Economic Development of
Economy, 24(3), 984–1003. https://doi.org/10.3846/20294913.2016.1216906
Razmara, M., Maasoumy, M., Shahbakhti, M., & Robinett, R. D. (2015). Optimal exergy control
of building HVAC system. Applied Energy, 156, 555–565.
https://doi.org/10.1016/j.apenergy.2015.07.051
Shaffer, F., & Ginsberg, J. P. (2017). An Overview of Heart Rate Variability Metrics and Norms.
Frontiers in Public Health, 5(September), 1–17. https://doi.org/10.3389/fpubh.2017.00258
Soteris A. Kalogirou, C. C. N. and C. N. S. 1. (n.d.). BUILDING HEATING LOAD
ESTIMATION USING ARTIFICIAL NEURAL NETWORKS Soteris A. Kalogirou 1 ,
Constantinos C. Neocleous 1 and Christos N. Schizas 2, 1–8.
Tanabe, S. I., Kobayashi, K., Nakano, J., Ozeki, Y., & Konishi, M. (2002). Evaluation of thermal
comfort using combined multi-node thermoregulation (65MN) and radiation models and
computational fluid dynamics (CFD). Energy and Buildings, 34(6), 637–646.
https://doi.org/10.1016/S0378-7788(02)00014-2
Tap, M., Kamar, H. M., Kadir, A., Nazri, M., Amry, K., & Salimin, M. (2011). Simulation of
Thermal Comfort of a Residential House. International Journal of Computer Science
Issues, 8(5), 200–208.
Tarantini, M., Pernigotto, G., & Gasparella, A. (2017). A Co-Citation Analysis on Thermal
Comfort and Productivity Aspects in Production and Office Buildings. Buildings, 7(2), 36.
https://doi.org/10.3390/buildings7020036
Wang, D., Zhang, H., Arens, E., & Huizenga, C. (2007). Observations of upper-extremity skin
temperature and corresponding overall-body thermal sensations and comfort. Building and
Environment, 42(12), 3933–3943. https://doi.org/10.1016/j.buildenv.2006.06.035
Zhang, H., Huizenga, C., Arenas, E., & Wang, D. (2004). Thermal sensation and comfort in
transient non-uniform thermal environments. European Journal of Applied Physiology,
92(6), 728–733. https://doi.org/10.1007/s00421-004-1137-y
Zhang, H., Huizenga, C., Arens, E. A., & Yu, T. (2001). Considering individual physiological
differences in a human thermal model. Journal of Thermal Biology, 26(4–5), 401–408.
Retrieved from http://www.sciencedirect.com/science/article/B6T94-43PBM1P-
W/2/59b2b8ce9e60fe3a3439b7f1c2f9ead1
88

Zhong, C. (2017). Data-Driven Approach for User-Centered Environmental Control by, (May).

89

APPENDIX A: CODE OF MACHINE LEARNING ALGORITHM

Figure. 1 Decision Tree
90

Figure. 2 Gradient Boosting
91

Figure. 3 Logistic Regression
92

Figure. 4 Random Forest

Abstract (if available)

Abstract Heating, ventilation, and air-conditioning (HVAC) systems play an essential role in supporting building functions. However, they consume an enormous amount of energy in buildings. Energy consumption and thermal comfort are two significant factors in the design process of HVAC systems. Although reducing energy consumption is important, one must balance that with thermal comfort. Accurate HVAC systems control can provide a better thermal environment and save energy. The limitations of the current thermal models promote the study of thermal comfort predictions. ❧ A tool was created with the help of an electric engineering student to achieve automatically control the HVAC system based on an occupant’s facial skin temperature to provide a better thermal environment. The program lasted for three years. In the first stage, research assistants build up foundation of the environment chamber and data collection systems. After the fundamental research of facial skin temperature and thermal comfort, this stage aimed to use machine learning method to achieve the automatic HVAC system control based on facial skin temperature. ❧ Human subject experiments were conducted in both an environment chamber and a real office building. Different parameters about the occupants were collected during the experiments by sensors, such as the facial skin temperature, heart rate, and electrode activity. The parameters of the environment were also collected, including air temperature, relative humidity, and carbon dioxide level. Two air conditioners and four heaters controlled the experimental conditions

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Enhancing thermal comfort: air temperature control based on human facial skin temperature

PDF

Exploration for the prediction of thermal comfort & sensation with application of building HVAC automation

PDF

Developing environmental controls using a data-driven approach for enhancing environmental comfort and energy performance

PDF

Human–building integration: machine learning–based and occupant eye pupil size–driven lighting control as an applicable visual comfort tool in the office environment

PDF

Multi-occupancy environmental control for smart connected communities

PDF

Economizer performance and verification: effect of human behavior on economizer efficacy and thermal comfort in southern California

PDF

Developing a data-driven model of overall thermal sensation based on the use of human physiological information in a built environment

PDF

Exploring participatory sensing and the Internet of things to evaluate temperature setpoint policy and potential of overheating/overcooling of spaces on the USC campus

PDF

Learning personal thermal comfort and integrating personal comfort requirements into HVAC system control loop

PDF

Visualizing thermal data in a building information model

PDF

Energy savings by using dynamic environmental controls in the cavity of double skin facades

PDF

Mitigating thermal bridging in ventilated rainscreen envelope construction: Methods to reduce thermal transfer in net-zero envelope optimization

PDF

Real-time simulation-based feedback on carbon impacts for user-engaged temperature management

PDF

Adaptive façade controls: A methodology based on occupant visual comfort preferences and cluster analysis

PDF

Human-environmental interaction: potential use of pupil size for office lighting controls

PDF

Nursing comfort study: are hospitals performing for nurses? IEQ analysis of working conditions inside hospitals

PDF

Human-building integration based on biometric signal analysis: investigation of the relationships between human comfort and IEQ in a multi-occupancy condition

PDF

Night flushing and thermal mass: maximizing natural ventilation for energy conservation through architectural features

PDF

Bridging performance gaps by occupancy and weather data-driven energy prediction modeling using neural networks

PDF

Development of data-driven user-centered building façade design guideline models: machine learning-based approaches to predict user preferences

Asset Metadata

Creator Jia, Mengqi (author)

Core Title Enhancing thermal comfort: data-driven approach to control air temperature based on facial skin temperature

School School of Architecture

Degree Master of Building Science

Degree Program Building Science

Publication Date 04/24/2021

Defense Date 03/20/2019

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

Tag automatic control,machine learning algorithm,OAI-PMH Harvest,occupants’ feedback,thermal condition

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Choi, Joon-ho (committee chair), Gil, Yolanda (committee member), Kensek, Karen (committee member)

Creator Email Mandy05271@hotmail.com,MENGQIJ@USC.EDU

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-149832

Unique identifier UC11660625

Identifier etd-JiaMengqi-7278.pdf (filename),usctheses-c89-149832 (legacy record id)

Legacy Identifier etd-JiaMengqi-7278.pdf

Dmrecord 149832

Document Type Thesis

Format application/pdf (imt)

Rights Jia, Mengqi

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