Fault Detection Methods For Vapor-Compression Air Conditioners Using Electrical Measurements

Fault detection methods for vapor-compression air conditioners
using electrical measurements

by
Christopher Reed Laughman
S.B., Massachusetts Institute of Technology (1999)
M.Eng., Massachusetts Institute of Technology (2001)
Submitted to the Department of Architecture
in partial fulfillment of the requirements for the degree of
Doctor of Philosophy in Architecture: Building Technology

at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
September 2008
c Massachusetts Institute of Technology 2008. All rights reserved.

Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Department of Architecture
8 August 2008
Certified by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Leslie K. Norford
Professor of Building Technology
Thesis Supervisor
Certified by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Steven B. Leeb
Professor of Electrical and Mechanical Engineering
Thesis Supervisor
Accepted by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Julian Beinart
Professor of Architecture
Chair of the Department Committee on Graduate Students
Thesis Committee
Thesis Supervisor: Leslie K. Norford

Title: Professor of Building Technology
University: Massachusetts Institute of Technology
Thesis Supervisor: Steven B. Leeb

Title: Professor of Electrical and Mechanical Engineering
Thesis Reader: Leon R. Glicksman

Thesis Reader: James L. Kirtley

Title: Professor of Electrical Engineering
Thesis Reader: Steven R. Shaw

Title: Associate Professor of Electrical Engineering
University: Montana State University
Thesis Reader: Peter R. Armstrong

Title: Professor of Mechanical Engineering (?)
University: Masdar Institute of Science and Technology
Fault detection methods for vapor-compression air conditioners using electrical
measurements
by
Christopher Reed Laughman
Submitted to the Department of Architecture

on 8 August 2008, in partial fulfillment of the
requirements for the degree of
Doctor of Philosophy in Architecture: Building Technology
Abstract
Abstract goes here. Yadda yadda yadda.
Thesis Supervisor: Leslie K. Norford

Thesis Supervisor: Steven B. Leeb

Title: Professor of Electrical and Mechanical Engineering
for Katie
Acknowledgements
Acknowledgments
”You need the willingness to fail all the time.
You have to generate many ideas and then you have to work very hard only to
discover that they don’t work. And you keep doing that over and over until
you find one that does work.”
- John W. Backus
thanks to steve leeb, les norford. also thanks to prof. k and leon.
thanks to peter armstrong, rob, and steve shaw.
thanks to jim, warit, roderick lafoy, ryan bavetta.
thanks to lau fan, dudes at abco, dudes at uri, dudes at purdue
thanks to friends, glass lab.
thanks to family
thanks to katie.
main.tex 9 Rev: 6701, July 16, 2008

Table of Contents
Contents
1 Introduction 17
1.1 Fault Detection and Diagnostic Systems . . . . . . . . . . . . . . . . . . . . . 20
1.1.1 Fault Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.1.2 Fault Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.1.3 Fault Evaluation and Decision . . . . . . . . . . . . . . . . . . . . . . 29
1.2 Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.2.1 Air Conditioning System Overview . . . . . . . . . . . . . . . . . . . 31
1.2.2 Common Faults in Air-Conditioners . . . . . . . . . . . . . . . . . . . 33
1.2.3 FDD Methods for Air-Conditioners . . . . . . . . . . . . . . . . . . . 38
1.3 Electrically-based FDD techniques . . . . . . . . . . . . . . . . . . . . . . . . 42
1.4 Research Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
main.tex 11 Rev: 6701, July 16, 2008

List of Figures
List of Figures
1-1 The basic structure of an FDD system (adapted from [29]). . . . . . . . . . . 20

1-2 Fault detection using a signal model-based approach (adapted from [29]). . 23
1-3 Fault detection using a system model-based approach (adapted from [29]). . 23
1-4 caption. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1-5 caption. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1-6 Schematic diagram of the air-conditioner. . . . . . . . . . . . . . . . . . . . . 32
main.tex 13 Rev: 6701, July 16, 2008

List of Tables
List of Tables
main.tex 15 Rev: 6701, July 16, 2008

Chapter 1
Introduction
Whether the motivation is environmentalism, money, or just an interest in thermody-

namics, the topic of energy conservation is presently receiving a great deal of attention.
Such considerations can arise in a wide variety of different contexts: the amount of en-
ergy needed to manufacture a product, the amount of energy needed to operate a piece of
equipment, and the amount of energy needed for transportation are just a few of the the
different types of concerns which can arise. Moreover, one common concern at a time dur-
ing which the price of oil exceeds $140 a barrel is to account for and minimize the amount
of energy consumed by a given task, allowing energy to be better allocated or saved.
Buildings represent one particularly important and common consumer of energy. Even
neglecting the myriad uses of energy which take place inside of buildings, the energy re-
quired to heat, cool and provide light in buildings amounts to billions of kilowatt-hours
every year. Especially apropos to considerations of energy conservation is the fact that
the installation of air conditioning in buildings is becoming increasingly commonplace.
Over half of the industrial and commercial buildings currently in existence are air condi-
tioned, while approximately 90% of newly constructed homes have central air installed.
Similar statistics are cited in [63], in which 14% of the U.S. primary energy consumption
is dedicated to HVAC systems for commmercial and industrial buildings, and 32% of the
electricity generated is consumed to heat, cool, ventilate, and light commercial buildings.
This profusion of air conditioners uses a great deal of energy; by one estimate, 76 billion
kWh per year are consumed by packaged air conditioning equipment [51]. Of the types
of air conditioners sold, packaged rooftop units (RTUs) form the largest segment of the
main.tex 17 Rev: 6701, July 16, 2008

Chapter 1 : Introduction
market, cooling 45% of the 58.7 billion square feet of commercial space in the U.S.
While air-conditioning has long been popularly viewed as a luxury [16, 45], it serves
many essential functions in buildings today. Perhaps the application which comes to mind
most readily is that of increasing the level of thermal comfort in buildings via a reduction in
air temperature and/or humidity, resulting in benefits both to the health and the produc-
tivity of the building occupants. Commercial and industrial applications also abound, as
air-conditioning is often essential to process control in chemical or manufacturing plants,
and it is also needed for many businesses, such as those which deal with the preparation,
storage, and sale of food products. Medical applications of air-conditioning are also very
common, as hospitals require it for the comfort and well-being of their patients.
This dependence of so many applications on air-conditioning brings a corresponding
need for the air-conditioners to operate within specified bounds on its performance. The
many potential consequences of faulty operation depend on the application and the sever-
ity of the fault; for faults which result in a loss of air-conditioning, the effects can range
from discomfort for home or business occupants, to ruining foodstuffs or temperature-
dependent products, to directly affecting the health of hospital patients [10]. Less dramatic
faults might not have such extreme short-term consequences, but they can still cause size-
able reductions in equipment efficiency and corresponding increases in energy consump-
tion, or the shortening of equipment life, resulting in increases in service costs. Factors
which can cause or contribute to such poor behavior are legion, including neglect and
improper installation and maintenance, as well as unforeseen electrical and mechanical
failures. Field studies suggest that such faults are widespread, as a study of 4,168 air-
conditioners found that 72 percent of the air-conditioners had the incorrect refrigerant
charge, and 44 percent had improper airflow [?].
Methods for monitoring and evaluating the operation of air-conditioning systems are
gaining in importance as air-conditioning becomes more common. Systems which are
designed to implement these methods, often referred to as condition monitoring or fault
detection and diagnostic (FDD) systems, have a variety of potential uses, including mon-
itoring the performance of the equipment over time, alerting users of the presence of a
fault, improving the quality of service by identifying the cause of faulty operation, or even
changing the operation of the equipment to reduce the effect of the fault. The functional-
Rev: 6701, July 16, 2008 18 main.tex

ity of these FDD systems can thereby greatly improve the performance of the underlying
air-conditioning system and assist the service technician in expeditiously fixing and main-
taining equipment.
While the users of air-conditioning systems which incorporate FDD methods will see a
direct benefit due to the reduced downtime and improved performance of the air-conditioner,
these FDD systems also benefit a number of other parties. For example, an effective FDD
method can help a service technician to quickly identify malfunctioning components and
reduce the time required to troubleshoot ambiguous problems. This will assist both users
and service technicians in ensuring that the system was installed correctly, and provid-
ing prognostic capabilities in identifying and monitoring faults which can be addressed
during scheduled service calls, rather than in emergency situations. Equipment manufac-
turers also benefit from the installation of such an apparatus, in that it could enable them
to better understand how and when faults occur.
Interest in the theoretical and experimental development of FDD methods in general

is growing, due to the dependence of society on systems of increasing complexity. These
systems can be found nearly everywhere in the early 21st century, from buildings to cars
to airplanes to consumer electronics to traffic control systems. The traditional attitude of
the designer, that the system will function reliably for a long period of time if the design
is “good enough”, is slowly giving into the realization that many systems will be operated
in conditions unforeseen by the designer, and that the cost of faults in such systems can
sometimes be greater than the system itself. The implementation of FDD and condition
monitoring systems is vital in such circumstances, to ensure that the system will continue
to function well in the face of considerable operative uncertainty.
The research presented here investigates and develops FDD methods for residential
air-conditioners using electrical measurements. Such an approach inherently has a variety
of both benefits and challenges, which will be discussed in this introductory chapter. In
order to provide context for this development of the FDD method, however, many of the
factors and possible approaches which must be considered when designing FDD systems
will be reviewed in the following section.
main.tex 19 Rev: 6701, July 16, 2008

1.1 Fault Detection and Diagnostic Systems
In developing an FDD method, it is useful to identify the constituent tasks which are per-
formed. A useful and intuitive general framework for FDD methods is outlined by Iser-
mann in [28, 29] in which the FDD process is broken down into four steps: fault detection,
fault diagnosis, fault evaluation, and decision, as seen in Figure 1-1.
inputs system sensors outputs
fault
detection
fault
diagnosis
fault evaluation
& decision
Figure 1-1: The basic structure of an FDD system (adapted from [29]).
In the fault detection block, measurements of the system are made and then processed
in order to determine if a fault has occurred. Since the reported faults could potentially
be caused by a variety of circumstances, the fault diagnosis block processes the informa-
tion from the fault detector and determines the cause of the fault. Once this has been
determined, the fault evaluation and decision block assesses the significance of the fault
determines the proper course of action, be it to continue operating in a faulty state, shut
down, change operation so that the effect of the fault is mitigated or eliminated, or any
other option.
While this is a useful general framework, the construction of the particular blocks will
vary widely in any specific implementation of an FDD method. For example, many such
methods only implement the fault detection block in hardware and software, and rely
Rev: 6701, July 16, 2008 20 main.tex

Section 1.1 : Fault Detection and Diagnostic Systems
upon the expert knowledge of a system user to perform the fault diagnosis and evaluation
steps. Due to such concerns, the performance of FDD system is often dependent on both
the performance of the individual blocks as well as the interactions between the blocks.
Each of the constituent components of the overall FDD system must therefore be designed
while considering the operation of the other components.
A variety of issues which can affect the performance of the FDD method must also be
considered during the design stage. For example, the reliability of the FDD system itself
is critical to its performance, as situations in which the FDD system erroneously reports
a fault can often be as problematic as those in which faults are reported correctly. The
FDD method must correspondingly be designed and specified so that its reliability and
performance in the field environment can be assessed and understood.
The cost of developing and implementing an FDD method is also an important con-
sideration. These costs may be justified by examining the savings collected by keeping
the system running efficiently, servicing faults before they increase in magnitude, or by
the convenience of scheduling service visits on a non-emergency basis. They might also
be justified because of the consequences of not detecting the faults; even if the faults are
inexpensive and easy to repair, situations in which system failure could have significant
ramifications for the health of individuals or a business can be a powerful incentive for in-
stalling an FDD system. As many of these issues shaped the design and implementation of
the FDD methods discussed in this research, each block of the general FDD method will be
discussed briefly in the following sections in order to better provide context for the design
decisions which were made.
1.1.1 Fault Detection
The structure of a fault detector can generally be broken down into three different pieces.
The first of these takes a set of input measurements of the system which are obtained from
a sensor or set of sensors and uses these measurements to characterize a model or gen-
erate an expected response of the system. The output of this model is then processed to
generate a set of fault signatures, which make it easier to observe the effect of the fault
and which are known to be directly related to the fault or faults of interest. These fault
signatures effectively project the observations onto a space which enhances the detectabil-
main.tex 21 Rev: 6701, July 16, 2008

ity of the fault. Once these fault signatures have been obtained, they are processed by a
change detector which compares the observed fault signatures to the expected signatures
that are generated by normal behavior, and generates an output which indicates the exis-
tence and/or size of the mismatch, also called a symptom. These symptoms are output to
the fault diagnosis block, which assesses the presence or the severity of the fault.
While there are a number of different types of models which may be used in the fault
detector, they can generally be grouped into one of two different types: signal-based mod-
els, and system-based models. Signal model-based fault detection techniques do not in-
corporate any knowledge about the monitored system itself; instead, the sensor output is
characterized solely by a model of the signal, and the presence of a fault is identified by
looking at attributes of the signal model. This approach is illustrated in Figure 1-2. In
comparison, system model-based fault detection techniques incorporate models of the un-
derlying system that generate the observed sensor data, and use this data to characterize
the proposed system model and to examine model characteristics which are indicative of
a fault. This alternative strategy for fault detection is illustrated in Figure 1-3. As each
of these approaches has relative strengths and weaknesses, so both will be used in this
research depending on the needs of the particular application.
Signal model-based fault detection methods are often the most straightforward fault
detection methods to implement, since little additional information about the current state
of the system is needed for the operation of the method. Two common approaches look
strictly at the time-series data obtained from the sensors; the first of these analyzes the ob-
servations as they are obtained and uses a simple thresholding or limit-checking algorithm
for fault detection. Observed changes in the signal which crosses an established limit are
deemed to be indicative of a fault. Such simplicity in a fault detector is attractive because
it minimizes the fault detector’s complexity, which has benefits for both the FDD system
reliability and cost. An alternative approach is to analyze trends in the time-series history
of the observed data, rather than looking solely at the instantaneous values of the data.
Such an approach is useful when the observations are expected to change over time in a
well-understood manner.
While the simplicity of this type of fault detector is attractive, it also imposes limitations
on the performance of the detection method. These fault detectors are typically sensitive
Rev: 6701, July 16, 2008 22 main.tex

signal
model
fault
detector
fault
signature
normal change
behavior detector
symptoms
Figure 1-2: Fault detection using a signal model-based approach (adapted from [29]).
system
model
fault
detector
fault
signature
normal change
behavior detector
symptoms
Figure 1-3: Fault detection using a system model-based approach (adapted from [29]).
main.tex 23 Rev: 6701, July 16, 2008

to noise and bias, as there is no apparatus for differentiating between a spurious event
which has no appreciable effect on the system and a fault condition. These detectors are
also highly dependent on the equipment installer and FDD method designer to ensure that
the sensor accurately characterizes the system performance over its lifetime, and that the
threshold is set appropriately for the particular instance of the system, with its attendant
variations in performance.
An additional limitation of such threshold-checking fault detectors is that faults may
not appear in an abrupt manner, so that changes in the system as it “wears in” over time
must be distinguished from slowly appearing faults. In drawing a distinction between
faults which appear over a long period of time and faults which occur abruptly, it is help-
ful to borrow terminology from [11] in which the slowly developing faults will be referred
to as “soft” faults, while the abrupt faults will be labeled “hard” faults. Using this termi-
nology, hard faults tend to be conducive to detection with a threshold detector because
they represent an abrupt change in the variable of interest, while soft faults tend to be
much more difficult to identify with a threshold detector because they are difficult to dif-
ferentiate from characteristics of the system, which will gradually change with continuing
use.
Alternative methods for processing the sensor data can also increase the amount of
fault information obtained from the sensor array. While many extant fault detection sys-
tems rely solely on data obtained when the system is operating in quasi-steady-state, data
acquired during transient changes in the state of the system can provide a great deal of
information about its behavior. Analyzing the sensor data via other signal representa-
tions can also help to identify changes in the signal which would be difficult to observe
in the time-series data. Common techniques include analyzing the signal with the Fourier
transform or a multi-resolution basis such as wavelets, or fitting the data with radial basis
functions.
In contrast to the signal modeling approach, the system modeling approach to fault de-
tection involves constructing a mathematical or computational model of the system being
monitored. Models are constructed with a given structure and a set of parameters, which
are either set when the model has been chosen, or are initialized with a set of character-
ization data which is known to accurately represent the behavior of the working system.
Rev: 6701, July 16, 2008 24 main.tex

After the model has been initialized, faults can be identified by driving the model with
inputs that are related to the inputs of the real system and comparing the output of the
model of the simulated system with the output of the real system. One of the principal ad-
vantages of this approach is that it makes it possible to draw inferences about the system
which are not directly visible, but are instead tied to the the relationship between a number
of different changes which are related by an unobservable variable or fault condition.
Naturally, these models can take a wide variety of different forms. One common class
of models is referred to as black-box models, which require no information about the sys-
tem being modeled. Perhaps the simplest type of black-box model is a simple polynomial,
using sensor observations as the independent variable and the coefficients of the sensor
observation terms being the model parameters. Other common types of black-box models
are based upon machine learning techniques, such as neural nets and fuzzy logic-based
models. The parameters for these models are generally obtained by a learning process that
collects training data which is known to characterize the normal operating behavior; this
step either occurs at the time that the model is constructed with a verified set of data, or
after installing the system in the field, when the system is assumed to be fault-free.
One of the shortcomings of these types of models is that they do not take advantage
of any of the physical understanding of the system which is based upon fundamental
constitutive laws. There are many benefits to using these physically-motivated, or grey-
box, models, such as the ability to incorporate intuitive knowledge about the behavior of
the system into the mathematical dynamics of the model, as well as set the parameters of
the model ahead of time, rather than relying solely upon the data. Linear and nonlinear
Kalman filtering [56, ch. 4] are perhaps the best known application of this approach to fault
identification. These type of models can also be applied to processes, rather than physical
systems, allowing fault detection methods to be implemented on systems which are driven
by random inputs [17].
The system-modelling approach is also useful when used to track system parameters as
the system performance changes over time. Rather than driving the model with the same
input as the physical system and then analyzing the differences between the model output
and the system output, this alternative approach drives the model with the same input, but
instead adjusts the parameters of the model to successfully reproduce the system output,
main.tex 25 Rev: 6701, July 16, 2008

so that both the model input and output match the system input and output. The resulting
changes in the model parameters will reflect the underlying changes in the system, and
analysis of these changes can be used to determine the presence of a fault. The appeal of
such an approach is that fault signatures can be generated for unmeasured parameters of
the model which are of interest, rather than the measured variables which, by themselves,
may not provide enough information to identify a fault. Physical models are particularly
useful in this context, as it allows the designer to detect faults based upon his intuition
regarding the changes in the physical model of the system.
In designing the fault detector, due consideration must also be given to the possibility
that aspects of the FDD method will not function as expected, either by not reporting faults
which are present, or by reporting faults which are not present. Both of these scenarios are
problematic because the service requirements of the system will not be addressed correctly,
and because they will cause the user to lose confidence in the accuracy of the FDD method.
It is therefore imperative that the reliability of the FDD method be well understood.
Failure of or faults in the FDD method can arise for a few different reasons. Malfunc-
tioning sensors clearly have significant ramifications for the FDD method, both in the case
that the sensors stop providing data to the fault detector altogether, or in the case that
the output of the sensors acquires noise or bias which prevents the sensor output from
accurately representing the true behavior of the system. Model inaccuracies also repre-
sent a significant challenge. For example, if the behavior of the fault-free system cannot
be described by the model due to unplanned changes in the system, discrepancies be-
tween the expected and observed data could be reported as a fault. Moreover, many fault
detection methods rely upon baseline data collected after the system is installed and its
functionality is tested and verified (also referred to as commissioning) by a technician. If
the technician does not install or commission the system accurately, the baseline data for
the FDD system will represent faulty, rather than fault-free, behavior of the system. The
possibility of emergent behavior which arises in complex systems must also be considered,
as unforeseen circumstances might interact with highly complex FDD systems to produce
unexpected results. These phenomena can only be mitigated or avoided through extensive
testing.
Rev: 6701, July 16, 2008 26 main.tex

fault 1 symptom 1 fault 1 symptom 1
fault 2 symptom 2
fault 3 symptom 3
Figure 1-4: caption.
fault 1? symptom 1
fault 2? symptom 2
fault 3? symptom 3
Figure 1-5: caption.
1.1.2 Fault Diagnosis
Historically speaking, most FDD systems have been designed to monitor a small number
of variables and alert the system user to the presence of a fault when a fault symptom
is generated by the fault detection module. This can be an effective tool in some types
of systems, where the user is interested only in knowing the presence of a fault, rather
than knowing what the actual cause of the fault is. For many types of complex systems
in use today, however, such systems are insufficient. Fault diagnosis modules can be used
to determine the cause of faults in such systems by analyzing the observed set of fault
symptoms.
The central challenge which must be addressed by any fault diagnosis module is that
faults are not related to fault symptoms in a one-to-one relationship. This fact is illustrated
in Figures 1-4. As can be seen on the right hand side of this figure, a given fault may
be manifested via a number of different symptoms. While some of these symptoms may
be easier to distinguish from normal operating behavior than the others, each fault could
main.tex 27 Rev: 6701, July 16, 2008

manifest itself in multiple ways. The diagram on the left hand side of this figure illustrates
the other complicating factor for fault diagnosis, which is the fact that multiple different
faults can manifest themselves with the same set of fault symptoms. This fact can make it
difficult to guarantee the occurrence of a particular fault if only one set of fault symptoms
is monitored.
The extent of the potential difficulties facing the fault diagnosis module can be seen
in Figure 1-5, in which multiple faults could be occurring simultaneously. Often referred
to as the multiple-simultaneous fault problem, the FDD method must ideally be able to
analyze a given set of fault symptoms and identify which faults, of all possible options, are
affecting the system at any given point in time. This is particularly challenging because the
fault symptoms are not necessarily binary quantities, so that the diagnostic module must
infer the relative contributions of different faults to the observed magnitude of a given
symptom. A pseudo-mathematical analogy to this situation suggested by [40, 41] takes the
following form:
    
a a a fault 1 symptom 1
 11 12 13     
a21 a22 a23  fault 2 = symptom 2 . (1.1)
    
    
a31 a32 a33 fault 3 symptom 3
In general, the technique used by FDD methods solve this problem is called fault iso-
lation. In the context of (1.1), this can be understood as diagonalizing the fault diagnostic
matrix containing the terms aii , though it is important to note that this diagonalization is
not carried out computationally, but rather is only performed analogically. This step can be
greatly assisted by properly designing the sensor network for the FDD system; by carefully
selecting the type and location of these sensors, it is often possible to capture fault signa-
tures which are strongly correlated to the fault of interest and only weakly correlated with
other faults. Similarly, the proper sequencing of the fault diagnoses can also be helpful by
using the process of elimination to reduce the number of conflicting fault symptoms.
A number of different approaches can be taken to implementing this fault isolation

and diagnostic strategy. These approaches generally fall under the umbrella of classifica-
tion problems, since the problem is essentially one of classifying a set of observed fault
Rev: 6701, July 16, 2008 28 main.tex

symptoms as corresponding to one or multiple faults. There is a rich body of literature

on the different types of classification techniques, including Bayes classification, geomet-
ric classifiers, decision trees, neural networks, and clustering algorithms. Each of these
methods has its own strengths and weaknesses, and the “right” technique for a given FDD
method can only be chosen by evaluating them in the context of the system of interest.
Two other factors must also be considered when designing an FDD method. One sit-
uation which can arise occurs when the faulty behavior of one system component causes
faults in other system components, so that the faults cascade. An FDD method must there-
fore be able to diagnose both the proximate faults as well as the source fault; if only the
proximate faults are diagnosed and repaired, the system will continue to operate in a sub-
optimal manner because undiagnosed fault will cause the repaired components to fail as
well. The design of the FDD system should also account for the fact that the complete set
of all possible system faults may not be known. In this case, the attribution of a fault symp-
tom to a particular cause will not be accurate, because of model inaccuracies. While the
only information which can be given to the system user in such a circumstance is that there
is a fault, but that it is not one of the faults implemented in the method, this information
can still be of assistance to a service technician.
1.1.3 Fault Evaluation and Decision
Once the existence of a fault has been detected and the cause of the fault has been iden-
tified, the proper course of action must be determined. In many situations, this is chosen
by the system user or service technician. As an assessment of the severity of the fault is
essential to this decision process, the principal function of the fault evaluation system is
making this assessment, which is commonly called the “fault level”. The type of fault level
can vary, as some faults are binary in nature, while other vary in a continuous manner. The
fault evaluation module must therefore analyze the information provided by the fault de-
tection module, which tells it that there is a fault, as well as the information provided by
the fault diagnostic module, which tells it what the fault is caused by, in order to estimate
how bad the fault is.
Once the fault level has been evaluated, the FDD method must select of the proper
course of action. In the described FDD framework, this step is designated as the fault
main.tex 29 Rev: 6701, July 16, 2008

decision module. The following options provide a sense of the possible choices which
could be made by this module, depending on the exigencies of the particular situation.
1. Notify the user of the existence of a faulty condition, as well as the particular fault
which has been diagnosed. Information about the fault detection signature and/or
other potential faults may also be communicated to the user to properly convey in-
formation about the relative uncertainty of the diagnostic, in order to properly char-
acterize the chance that the fault detection or the fault diagnostic modules did not
correctly identify the fault. This fault evaluation procedure might be useful if the
equipment owner or user has direct control over the system, and is able or desires to
make his own evaluation of the status of the equipment as well as how to proceed
with repairing it.
2. Record the existence of the fault, as well its diagnosis, in a logfile. This logfile might
simply reside in the onboard computational framework of the FDD method, or it
could be located in a remote location. Information about the confidence of the fault
diagnosis could also be recorded. This procedure might be useful if the fault level is
not high, or if it is either unnecessary or not practical to inform the user of the fault;
when the service technician or the manufacturer examines the output of the logfile,
they will be able to use this information most effectively.
3. Contact a service company directly, via a remote connection. This approach could
dovetail with the previous approach, in that the service company could receive the
information otherwise logged in the datafile regarding the nature and severity of the
identified fault. Under such a framework, an FDD method which contacted a main-
tenance company directly for service could ensure that the system being monitored
would be operating in a faulty state for a minimal amount of time. Obviously, such
an approach would require a high degree of confidence in the FDD method’s ability
to correctly detect and diagnose faults, so as to prevent unneeded service calls.
4. Directly change the operation of the system in response to the presence of the diag-
nosed fault. Such a course of action might only be necessary or justifiable under the
most extreme of fault conditions, e.g. if the system will undergo irreparable damage
Rev: 6701, July 16, 2008 30 main.tex

Section 1.2 : Previous Work
if allowed to continue operating in its present state. The FDD method could also
modify the system performance if it was integrated with the control system, so that
the system performance is modified to compensate for the effects of the fault.
Clearly, a variety of other courses of action exist, depending on the system being mon-
itored, the particular application of the system, and the structure and implementation of
the FDD method. Nevertheless, consideration of the particular approach taken to evaluate
the effects of the fault and provide the appropriate response is an important element in the
job facing the designer of the FDD method.
1.2 Previous Work
While the previously discussed framework is useful in developing a general understand-

ing of FDD methods, it will be primarily applied to FDD methods for air-conditioning
systems in this research. As a wide variety of FDD methods have been implemented for
air-conditioning and refrigeration systems, it is helpful to survey this body of work to gain
a sense of the advantages and disadvantages of these approaches. In order to provide
context for these studies, the general operation of air-conditioning systems will first be re-
viewed, and will be followed by a summary of a number of surveys which examined the
frequency and types of faults which occur in field-installed air-conditioning equipment.
1.2.1 Air Conditioning System Overview
A schematic diagram of a vapor-compression air-conditioner is illustrated in Figure 1-6.

The overall objective of the ideal vapor compression refrigeration cycle is the transfer of
thermal energy from the air passing through the evaporator to the air passing through the
condenser via the refrigerant. To see how this occurs, consider the refrigerant leaving the
compressor, travelling toward the condenser in the direction of the boldface arrow. Having
just been compressed, this vapor is at both a high temperature and a high pressure. As the
vapor travels through the condenser and is cooled by the air blowing over the coils, it
gradually gives up enough of its thermal energy to cause it to change from 100% vapor at
the top of the condenser to 100% liquid by the bottom of the condenser.
main.tex 31 Rev: 6701, July 16, 2008

compressor refrigerant
flow direction
evaporator
expansion valve
condenser
Figure 1-6: Schematic diagram of the air-conditioner.
Once the refrigerant has fully condensed, it has the capacity to absorb more thermal
energy via conversion back to a gas. This process is accomplished in two stages; first, the
the refrigerant passes through an expansion valve into the evaporator coil. The expansion
valve (typically a thermostatic expansion valve, or TXV) serves the dual purposes of reg-
ulating the flow and reducing the pressure of the refrigerant. This drop in pressure causes
a corresponding reduction in the saturation temperature of the refrigerant, or the temper-
ature at which the liquid refrigerant evaporates. As this low pressure refrigerant flows
through the evaporator coil, the process of its evaporation causes it to absorb thermal en-
ergy from the air traveling past the evaporator coil. This process causes the air flowing
over the coil to become cooler. After leaving the evaporator as 100% low pressure vapor,
the refrigerant is compressed and returned into the condenser, completing the refrigeration
cycle [6].
Due to the wide variety of requirements and applications which exist, many different
types of air-conditioning systems have been constructed which implement this basic re-
frigeration cycle or variations thereof. Window air-conditioners satisfy the requirements
for small room-size residential applications. Air-conditioning systems which serve the en-
tire house are often constructed with either centralized air-conditioning systems, which
place an evaporator at a central location and distribute the cool air through a ventilation
system, or split air-conditioning systems, which distribute refrigerant lines throughout the
house and cool the air locally. A variety of types of compressors are used in these units,
including reciprocating compressors, scroll compressors, and rolling piston compressors;
Rev: 6701, July 16, 2008 32 main.tex

the particular compressor used also depends on the size of the cooling load and the appli-
cation. Still other types of mechanical cooling systems exist for larger buildings, including
systems which incorporate single and multi-stage chillers and cooling towers [5].
While the function of real refrigeration equipment is theoretically equivalent to that
of the ideal refrigeration cycle described above, the behavior of the physical realization
of such a system differs from that of the idealized system in a number of aspects which
can have a significant effect upon the unit’s performance. Some of these differences are
expected, and their effects can be minimized or otherwise compensated for in the design of
the unit. Other differences may be unforeseen by the design engineer, and are the result of
aberrant behavior due to atypical environmental conditions or malfunctioning equipment.
These undesirable behaviors are precisely what FDD methods seek to characterize and
eliminate.
1.2.2 Common Faults in Air-Conditioners
Faults typically develop in air-conditioners via three different mechanisms: improper com-
missioning, premature component failure, and wear due to normal usage. Each of these
fault mechanisms results in a degradation of the air-conditioner’s performance, either be-
cause the air-conditioner ceases to function (hard faults), because of a reduction in the ef-
ficiency of the air-conditioner, or because the increased rate of wear for the air-conditioner
components (soft faults). These faults can be manifested in the electrical, mechanical, or
thermal interactions of the air-conditioner.
In order to avoid the consequences of malfunctioning and broken air conditioning
equipment, studies have been performed to enumerate the different types of faults to
which packaged air-conditioners are susceptible. These studies classify the faults accord-
ing to a range of different criteria, such as cost to repair, importance to the overall function
of the unit, and the tendency of a given fault to accelerate the onset of other faults. While
their methods of comparison might differ, all of these studies are useful when attempting
to understand the scope of faults which occur in package air conditioning units. Given
the near ubiquity of RTUs, it may be surprising to find that only three major studies have
been published in enumerating these faults. One possible reason for this fact is that failure
information is often quite valuable, and this proprietary information is therefore not made
main.tex 33 Rev: 6701, July 16, 2008

public by equipment manufacturers. These studies were authored by Cunniffe, James, and
Dunn in 1986 [18], Stouppe and Lau in 1989 [54], and Breuker and Braun in 1998 [11]. An
additional study which examined both heating and air-conditioning equipment was pub-
lished by Hasan in 1974 [25]; as this study focuses on larger types of equipment, such as
cooling towers and chillers, it will not be considered here.
In [18], Cunniffe, James and Dunn conducted a survey of vapor compression refrig-
eration plants in a variety of different applications, such as low temperature refrigeration
applications, air conditioning systems, and vehicle refrigeration systems, in order to quan-
tify and enumerate the common faults which occur. Rather than quantify the statistics of
particular faults, the ultimate causes of the faults were classified; of the 851 systems which
were studied, 32% of the mechanical faults and 29% of the electrical faults were caused
by unforeseen operating conditions. In comparison, 38% of the mechanical faults and 44%
of the electrical faults were caused by faulty materials, manufacture, design, installation,
service, commissioning, or maintenance. The remainder of the faults were either due to
normal deterioration or remained unclassified.
As these statistics were collected, the authors observed that a number of operational
faults overloaded the compressor, which consequently failed and was diagnosed as the
faulty component. This type of failure is problematic because the compressor is often re-
paired without fixing the underlying fault. In order to address this concern, the authors
identify a number of fault classes which cause overloading. High discharge temperatures
represent one class of faults which account for 9-15% of the failures in reciprocating com-
pressors, which are in turn caused by faults such as high intermittent plant operation,
fouled condenser surfaces, non-condensible gases, or other unforeseen operating condi-
tions. Liquid return to the compressor housing and refrigerant migration, in which liquid
refrigerant accumulates in the compressor shell and then boils up into the compressor
cylinders, can also cause a host of problems, ranging from increased discharge temper-
atures, to major mechanical damage, to the degradation of winding insulation. Poorly
controlled expansion valve or evaporator dynamics, excessive equipment cycling, and in-
adequate system cleanliness also caused compressors to fail prematurely.
Another set of researchers [54] who systematically investigated the faults which are re-
sponsible for malfunctioning air conditioning equipment were Stouppe and Lau [54]. They
Rev: 6701, July 16, 2008 34 main.tex

examined 15,760 failures in units with a cooling capacity up to 50 tons (600,000 BTU/hr)
over the course of eight years (1980-1987) by scrutinizing insurance claims which were
filed for the repair of air conditioning equipment. Many different types of refrigeration sys-
tems were examined, incorporating both reciprocating and centrifugal compressors. While
some compressor technology has changed in the intervening fifteen years, their basic find-
ings are still quite useful in considering the range of failures which occur in refrigeration
systems.
Of the 15,760 failures examined, 12,518 had their specific cause of failure denoted on
the claim. The authors found that these units generally failed when they were approxi-
mately 10 years old, and that such units had no major inspections or overhauls over that
entire period of time. Furthermore, 11,349 of the failures were electrical (comprising mo-
tors, controls, and electrical apparatus), while 4,411 were mechanical (compressor bodies,
system piping, or vessels). It is important to reëmphasize the fact that compressors were
separated into their two different functional halves, being the mechanical section of the
compressor, including the pistons, valves, and associated hardware, and the compressor
motor, which drives the mechanical apparatus that performs the compression.
The authors specifically examined the causes of failure for hermetically sealed air con-
ditioning and refrigeration units: 76.6% of the failures were electrical (motor windings,
control equipment, and other associated issues), 18.9% were mechanical (compressor valves,
springs, bearings, connecting rods, pistons, crankshafts, lubrication), and 4.5% of the fail-
ures were caused by malfunctions in the refrigerant circuit. A few particular notes were
also made trends which were observed in the causes of compressor motor failure. Insula-
tion deterioration on the motor windings due to age and/or service, as well as unbalanced
voltage and single phase operation were some of the main causes of the failure; degra-
dation in motor stator windings was by far the most prevalent cause of motor failure,
accounting for 84% of failures in hermetic motors and 74% in the non-hermetic motors.
The remaining causes of the failures were generally broken rotor bars, bearings, and mo-
tor control equipment. Of the motor control failures, 168 failures were directly attributed
to short cycling, or the repeated starting and stopping of the motor many times in rapid
succession.
A more detailed study of mechanical compressor failures for reciprocating freon-based
main.tex 35 Rev: 6701, July 16, 2008

units showed that the bulk of the failures could be attributed to two main causes. The
first of these causes is that of metal fatigue in the internal suction and discharge valves
and springs, due to the hundreds of millions of operating cycles that these parts see over
the average lifetime. Liquid slugging, or the unintentional injection of liquid refrigerant
into the cylinder, was also the direct cause of 20% of the mechanical failures, due to the
hydraulic forces from the attempt to compress the incompressible liquid refrigerant which
acts on the valves, valve plates, pistons, and connecting rods. Under normal operating
conditions, gaseous refrigerant is pulled into the cylinder during the intake cycle, but when
liquid refrigerant enters, these hydraulic forces can cause broken valves and valve plates,
as well as cascading damage to the remaining mechanical components of the compressor.
While the information collected in [54] is valuable, it does not quantify the incidence
of faults which would allow the RTU to continue operating in a less efficient manner, nor
does it attempt to classify failures on the basis of their repair cost. This is a significant
shortcoming in their report, as it only identifies the components which ultimately failed,
not the proximate cause of those failures. Though many different faults can occur in an
RTU, a significant portion of them can cause the compressor to fail; the identification of
the compressor as the component which failed most often does not help to determine the
type or frequency of problems in the RTU system which caused that failure. The cost
of repair is also an important figure-of-merit in examining failures, as failures which are
expensive to repair should have priority when choosing which faults to include in an FDD
system.
Braun and Breuker [11] attempted to compensate for these shortcomings by examining
two different types of faults in package rooftop air conditioners: those which caused the
air conditioner to run inefficiently, but did not cause the equipment to cease operation, and
those in which the RTU could no longer operate mechanically or electrically. By analyzing
a statistically representative subset of a database containing 6000 separate faults observed
from 1989 to 1995, the authors found that, of the fault classifications which resulted in
“no air conditioning,” or faults in which the unit was running inefficiently but was still
electrically and mechanically operational, 60% were caused by electrical problems (such
as motors and control problems), while the remaining 40% were caused by mechanical
faults. On the basis of cost, however, compressors represented the most significant portion
Rev: 6701, July 16, 2008 36 main.tex

of the total unit repair cost (24%), even though they break much less frequently.
In examining the compressor failures, the authors found that approximately 70% of the
compressor faults were due to internal motor problems, such as shorted windings, open
windings, or locked rotors. One interesting fact which they point out is that, although
the immediate cause of compressor failure was usually attributed to motor problems, the
proximate cause was often a mechanical fault which overloaded the motor. Furthermore,
they reiterate one of the conclusions from [54], that the prevalent cause of this mechanical
overloading condition was again liquid refrigerant in the compressor cylinder. This speaks
directly to the difficulty of the fault diagnostic problem, as the diagnosis of the faults in the
compressor motor is difficult to directly ascribe to the presence of liquid slugging.
The authors also specify the main contributors to the costs for faults in the condenser,
evaporator, and general air handling systems. In analyzing repairs to the condenser unit,
defective fan motors accounted for 50% of the cost, while fouled coils represented 30% of
the cost. In comparison, fouling represented 61% of the cost of repairing an evaporator
unit, and coil damage represents 25% of the cost. One interesting sidenote is that only
18% of the evaporator fouling and 14% of the condenser fouling faults resulted in a loss
of comfort. Other general costs required to fix miscellaneous electrical components were
dominated by contactor failure, which represented 40% of the cost, followed by 27% for
general damaged components, and 13% for wiring errors and short circuits.
Though much of the stated motivation in developing FDD methods for air-conditioning
systems focuses on the effect on system owner and user, equipment manufacturers are also
affected by the incidence of faults in equipment. As documented in [13, 34], the difficulty
of properly identifying and diagnosing faults in field-installed equipment results in air-
conditioning technicians generating a large number of misdiagnoses, either because the
technician does not know that a given fault can be fixed in the field or because the failure
is attributed to an incorrect cause. According to Copeland Corp., half of the compressors
which are returned to the factory under warranty are victim of this tendency to misdiag-
nose; when these diagnoses are checked by factory technicians, no defect is found in nearly
half of them. This is often referred to as the NDF (no defect found) problem, and is another
strong motivation for the development of accurate and cost-effective FDD methods for air-
conditioning systems.
main.tex 37 Rev: 6701, July 16, 2008

1.2.3 FDD Methods for Air-Conditioners
While it is clear from §1.2.2 that the incidence of faults in HVAC equipment is sufficiently
high to warrant the development and implementation of FDD systems, there are a num-
ber of other factors which can have important consequences for the design of such systems.
First among these is the fact that many buildings are not owner-occupied. This fact makes
the market for air-conditioning system very first-cost sensitive, as the owner has little in-
centive to provide more efficient and expensive equipment when less expensive equip-
ment will suffice [33]. The air-conditioning industry has responded to this fact, producing
little equipment which exceeds the federal standard; as of 2000, less then 10% of the pack-
aged air-conditioners with a cooling capacity between 5.5 and 11.25 tons1 exceeded the
federal standard for energy consumption by 20%, 21% of the packaged air-conditioners
with a cooling capacity between 11.25 and 20 tons exceeded the federal standard by the
same amount [51].
When buildings are owner-occupied, higher-efficiency air-conditioners are more fre-
quently used, but they are still adopted more slowly than might be expected (only 6.7%
of residential air-conditioners exceed federal standards by 40%). A variety of reasons con-
tribute to this phenomenon; for example, the air-conditioning system is often one of the last
systems to be installed in a house, and cost overruns make less expensive air-conditioning
units attractive, even despite knowledge of higher operating costs [61]. In addition, as the
cost of energy per month is much lower than most other commercial expenses (such as
salaries or manufacturing costs), little attention is often paid to the cost of operating an
inefficient air-conditioning system unless energy prices are extremely high [33].
Motivated both by the need reliability in air-conditioning equipment and the potential
for increased interest in high-efficiency units, there has been a surge over the last 10 years
in the investigation and development of FDD methods for air-conditioning units. Much
of this work has focused on using temperature and pressure measurements in order to
evaluate the relative health of the unit as well as forecasting any incipient problems which
have yet to degrade the performance seriously, as can be seen by surveying the published
literature on such systems, some of which is treated in detail in [15].
1
1 refrigeration ton = 12,000 BTU/hr = 3.52 kW
Rev: 6701, July 16, 2008 38 main.tex

A number of papers have been published that address the problem of casting FDD
methods into the general framework for FDD systems discussed in 1.1. Haves, et al [26]
presents a particularly thorough perpective on some of the challenges inherent in con-
structing these methods. The authors of [21, 44] also discuss the variety of different ap-
proaches for implementing FDD methods on air-conditioners, and also address concerns
and potential strategies for achieving higher market penetration with such systems. Stylianou [55]
also discusses ways that these FDD methods might be integrated into building energy
management systems. Thybo, et al [57] also provide some interesting perspectives on im-
plementing FDD methods in air-conditioning units after their research in developing FDD
methods for refrigerated display cases for supermarkets. Some important considerations
identified in designing an FDD method include the method’s scalability for different types
of equipment, reducing the number of sensors in the system due to their cost and added
complexity, and minimizing the procedures required for a technician to commission the
FDD system. The authors also emphasize the fact that FDD methods are generally more
effective when incorporated into the system during the design stage. A set of important
faults is identified via a survey in [64], which determined that many of these faults occur
because of improper design and commissioning. The authors of this paper also surmise
that fault detection in air-conditioners is increasingly difficult because the high complexity
of the systems can obscure the effects of the faults and the large number of components
can make it difficult locate the source of a fault.
Many signal-based FDD methods have been developed for air-conditioning systems.
One popular approach involves identifying a set of relationships between measurements
of steady-state quantities (typically temperatures and pressures) which are indicative of
particular faults. Such methods are typically referred to as rule-based methods, as the
relationships between variables are expressed in the form of rules derived from physical
knowledge of the system. notes on the training of the system.
how many measurements do they need?
Jim Braun and his graduate students at Purdue University have been using and refining
this rule-based approach for FDD methods since 1995, and have experimentally demon-
strated a system which can diagnose six common faults: loss of volumetric efficiency of
the compressor, fouled evaporator coils and filters, fouled condenser coils and filters, liq-
main.tex 39 Rev: 6701, July 16, 2008

uid line restriction, incorrect refrigerant charge, and the presence of noncondensible ma-
terials in the refrigerant loop. This system was originally developed for air-conditioners
with fixed orifice expansion (FOX) devices [12, 43], and was also investigated in systems
with TXVs [14]. Improvements to this FDD method resulted in the ability to successfully
detect and diagnose multiple simultaneous faults [37, 38, 40, 41], and this method was ex-
perimentally demonstrated to identify faults and increase the operating efficiency of air-
conditioners at a number of locations in California [36,39]. The detector used to ensure that
the air-conditioner is operating in steady-state was subsequently studied in [30] where it
was proposed that the two standard measurements which are used to determine steady-
state (the evaporator superheat Tsh and condenser subcooling Tsc ) were not sufficient to
determine the steady-state in all tests, and that all of the variables should be analyzed
before steady-state operation can be declared.
Other research also incorporates similar rule-based approaches. An object-oriented
FDD method is developed in [24], which uses rule-based methods to evaluate the per-
formance of the air-conditioning unit, and also evaluate the performance of the sensors
by implementing redundancy checks on the sensors. Rule-based FDD methods are also
used in [31] to identify faults in a variable speed vapor compression system, and a similar
system is also evaluated in [22], albeit with a larger number of measurements.
A variety of black-box system-modeling FDD methods have also been tested. These
methods do not require any physical knowledge of the air-conditioning system, but in-
stead use training data to help the model characterize the non-faulty behavior of the sys-
tem. The authors of [?] develop a method which combines polynomial regression and
locally-weighted regression techniques with statistical machine learning techniques to iden-
tify faults in the air-conditioning system, and test this method on detecting incorrect refrig-
erant charge in an experimental unit. Genetic algorithms are used in [1] to identify faults
in boilers, while the performance of rule-based methods is experimentally compared with
the performance of a method which uses fuzzy models and classifiers to detect faults in a
three zone HVAC system in [9]. A similar fuzzy-modeling approach is also taken in [65],
along with the use of vibration analysis to identify faults in pumps and motors. Rule-
based knowledge is also used with artificial neural networks in [23] to identify fouling in
an air-handling unit.
Rev: 6701, July 16, 2008 40 main.tex

Grey-box system-modeling techniques, in which information about the physical pro-

cesses governing the system are incorporated into the model, have also been extensively
used. While these methods typically also require some level of initial training to dis-
tinguish faulty behavior from nonfaulty behavior, much of this training only requires
the specification of thresholds, and does not require the full characterization of the field-
installed equipment. Physical models of the system are formulated in [63], which uses re-
cursive least squares to identify the model parameters from training data, while physical
models are also used in tandem with nonlinear observers and threshold-checking methods
to detect condenser fan failure and capillary tube blockage in [60]. State-space models of
an air-handling unit are integrated into a Kalman filter to identify and update the model
parameters in [59], and airflow and sensor faults are experimentally identified based upon
features in the residual waveforms generated by the Kalman filter. A state machine-like
approach is also used in [?] to identify when valves and actuators get into faulty states, as
when they are stuck open.
Hybrid grey-box approaches, in which machine learning techniques are used to iden-
tify parameters of physical model, have also been used extensively. Neural network and
bond-graph techniques are used to identify parameters of a physical model of a refrigera-
tion system in [57, 58], allowing faults to be diagnosed via fixed and trending thresholds.
This method also integrates the fault detection system with a controller for the equipment,
to mitigate the effect of the faults. The challenging problem of distinguishing the slowly
changing characteristics of the air-conditioning system from soft faults is discussed in [27],
in which radial basis functions and direct-search algorithms are used to identify the phys-
ical parameters of a system and identify the presence of fouling and valve leakage faults
in an air-handling unit.
The identification of sensor faults, as distinct from system faults, is addressed in [62]
by analyzing discrepancies between sensor readings which are related through a model.
The model outputs are compared with the sensor outputs, and genetic algorithms are used
to determine the existence of sensor faults in the air-conditioning unit.
Research into non-temperature sensor based FDD methods is another ongoing area of
research. A preliminary system is developed in [20] which detected changes in the system
due to a variety of faults from measurements of the motor current, but this system was
main.tex 41 Rev: 6701, July 16, 2008

not developed into afull FDD system which could diagnose separate faults. FDD meth-
ods based solely upon electrical measurements are also tested in [2, 3] for the purposes
of identifying a variety of types of faults, such as incorrect refrigerant charge and fan im-
balance. Copeland Corporation has also developed a FDD system for compressors based
upon measurements of electrical power, named the ComfortAlert system [13]. This sys-
tem is designed to identify basic compressor faults, such as long run time, short cycling,
system pressure trip, locked rotor, open circuit, or low control voltage. This system has
also been demonstrated to be effective and has been marketed and distributed since 2002.
Other relevant literature regarding the detection of motor faults includes work published
in [7, 8, 46, 48], among numerous others.
1.3 Electrically-based FDD techniques
Previous work in developing FDD systems sets the bar high for making new contribu-
tions to FDD systems in packaged and unitary air-conditioning units. Existing systems
have detection thresholds which are sufficiently low that they can detect faults before a 5%
drop in either the capacity or coefficient of performance (COP). As these systems have also
been developed with an eye toward commercial implementation, FDD systems like that
developed at Purdue University are projected to cost as little as $250-300/unit [39].
One notable characteristic of many of the FDD methods discussed in §1.2.3 is that they
rely upon a relatively extensive network of sensors. Many of the methods have five or
more temperature sensors, as well as pressure sensors, mass-flow sensors, and relative
humidity sensors. While these mechanical sensors can provide valuable information about
the state of an air-conditioning system, their use in a FDD method must be considered in
the context of their cost and rate of failure. For example, mass flow sensors are typically
very expensive, with prices exceeding a few hundred dollars per unit. Other sensors, such
as temperature sensors, are prone to bias if they are not properly mounted to the measured
surface, or if a radiation shield is not installed properly. Much like air conditioners, it
is clear that sensors are vulnerable not only to hard faults, but also to soft faults, such
as sensor bias due to environmental or unforeseen manufacturing defects. Though the
ability of the FDD system to successfully identify faults certainly may outweigh the sensor
Rev: 6701, July 16, 2008 42 main.tex

Section 1.3 : Electrically-based FDD techniques
reliability concerns, the effects of both cost and reliability must be considered in the design
of FDD methods for air-conditioners.
Another interesting observation which might be made of most of the extant FDD meth-
ods is that many of the common faults identified in the surveys of §1.2.2 are not identified
by the FDD methods. In the case of many of the electrical faults, a variety of methods have
already been developed by the community of engineers who study motor faults which
could be easily adapted to the case of identifying faults in the compressor or fan motors.
Other faults simply are simply not conducive to the types of measurements made by the
majority of the FDD systems. For example, even though liquid slugging is identified by
all three fault surveys as a primary cause of damage in compressors, no published FDD
method includes a means for identifying this fault. This is probably due to the fact that a
substantial amount of sensitive mechanical instrumentation must be installed in order to
detect this phenomenon [42, 52, 53].
Mechanical diagnostic techniques based upon the output of electrical sensors represent
an interesting and potentially viable alternative fault detection approach for air-conditioning
systems which addresses both of the above concerns. In effect, these methods measure the
currents and the voltages at the terminals of the electromechanical device, and then iden-
tify faulty mechanical behavior on the basis of these observations. These techniques are
fundamentally rooted in the first law of thermodynamics: in many situations, the electri-
cal power flowing into an electromechanical device is directly to the mechanical power
flowing out of it, so that changes in the mechanical load of the motor which are related to
fault behavior can be observed in corresponding changes in the electrical power flowing
into the motor. FDD methods using this approach either use a signal-based approach, in
which a priori knowledge is used to relate mechanical faults to particular features observed
in the sensor output, or a model-based approach, in which the observed inputs and out-
puts of the system are used to characterize the behavior of the system and determine the
existence of any faults.
This electrically-based approach to fault detection, suggested by [2, 3] has a variety of
benefits, not the least of which is the fact that electrical diagnostics, such as broken ro-
tor bars and shorted windings, can be performed without using any additional sensors
(unlike the FDD methods which rely upon mechanical measurements.) The electrical in-
main.tex 43 Rev: 6701, July 16, 2008

formation can also be used to measure the energy consumption of the air-conditioning
system for end-use load information purposes. Perhaps the greatest benefit, however, is
that electrical sensors are generally easier to install, less expensive, and more reliable than
the equivalent mechanical sensors. This is due in part to the fact that electrical sensors do
not need to be in close physical proximity to the unit to function well, allowing them to be
placed indoors or in a similarly protected location, while at least part of the physical air-
conditioning apparatus (the condensing unit) usually resides outside. These features have
not gone unnoticed by industry, as the ComfortAlert product manufactured by Copeland
Corp relies upon electrical measurements to perform its diagnostic procedures.
Another appealing characteristic of electrically-based FDD methods is that diagnostics
can conceivably be performed on an aggregated electrical signal which contains electrical
power information for multiple loads. Generally referred to as non-intrusive load moni-
toring (NILM) [19,32,35,47,49,50], this technique is useful because it can further reduce the
number of sensors needed to identify the operation of loads and perform control and di-
agnostic operations for those loads, since multiple loads can be monitored simultaneously.
There are limits to the number of loads which can be effectively monitored with a NILM, as
well as the types of diagnostics which can be successfully implemented, due to bandwidth
and noise limitations; nevertheless, the synergy between electrically-based FDD methods
and potential NILM applications makes the development of such FDD methods that much
more attractive.
Naturally, this alternative approach to fault detection and diagnostics also has its lim-
itations. One such limitation is apparent in the use of electrical measurements to detect
changes in a motor’s mechanical load; the motor is being used for the dual purposes of
both a transducer and an energy conversion device. While motors are typically designed
to be good energy conversion systems, they are not usually intended to be used as trans-
ducers, so that variations in the motor’s operating characteristics, while not problematic
for their role as energy conversion systems, make it much harder to develop reliable fault
detection methods. In some cases, there are mechanical faults which have no observable
effect on the electrical variables because of effects like noise or scaling problems, such as
arise for loads coupled through gearboxes. Electrically-based methods can also require
substantial computational support, although the decreasing cost of computation and the
Rev: 6701, July 16, 2008 44 main.tex

Section 1.4 : Research Overview
prevalence of similar requirements among other FDD methods for air-conditioners suggest
that this is less of a barrier than might otherwise be perceived. The final disadvantage of
this approach is caused by the fact that the wide variety of mechanical changes, which can
be measured via temperatures, pressures, forces, and so forth, are being mapped into the
relatively small number of electrical variables, i.e. currents and voltages. This mapping
can make fault isolation more difficult, since different types of mechanical faults some-
times have identical effects on the electrical variables.
The observation that different mechanical faults may not be individually distinguish-
able because they have identical effects on the electrical variables has a direct impact on
the design of the FDD method. It is therefore incumbent upon the FDD method designer
to carefully model the physical behavior underlying the different mechanical faults, so that
distinguishing characteristics of each fault might be identified and utilized for fault isola-
tion. The high bandwidth of electrical sensors is of considerable assistance in this process,
as it makes it possible to identify fault signatures which contain relatively high-frequency
behavior. Fault signatures which have timescales ranging from fractions of a second to
many hours can therefore be used in these electrically-based FDD techniques; faults can
thus be isolated not only by looking for different types of behavior on the same interval of
time in the signal or model, but also by analyzing the input data on a number of very dif-
ferent timescales. Such an approach differs markedly from that used by many other FDD
methods, which rely upon steady-state features of the measured variables partly because
of the low bandwidth of many mechanical sensors, e.g. thermocouples.
1.4 Research Overview
The main objective of this research is the development of electrically-based FDD meth-
ods for a set of common faults in air-conditioners. Though most extant FDD methods for
air-conditioners rely upon mechanical sensors, the opportunity to identify a range of me-
chanical faults with a small number of reliable eletrical sensors is quite compelling, as is
the fact that such a FDD system would be able to identify both electrical and mechanical
faults with a single sensor array. Moreover, the relatively low cost for the sensors and pro-
cessing system makes such an approach attractive in the cost-sensitive air-conditioning
main.tex 45 Rev: 6701, July 16, 2008

market. This research represents a continuation and dramatic expansion of the early ef-
forts initiated by Peter Armstrong and published in [2–4], which suggested the potential
inherent in these approaches. Three particularly common and problematic classes of faults
will be studied in this research: airflow faults, liquid slugging, and refrigerant leaks. Each
of these faults is investigated theoretically and experimentally, and this document will
address both the fundamental characteristics of these faults as well as the many experi-
mental considerations which arose in inducing and detecting these faults in commercially-
available equipment.
With regards to the structure of the remainder of this document, the airflow diagnos-
tic methods will be discussed first in Chapter ?? due to the fact that many of the issues
which surround the simulation and parameter estimation of induction machine models
are treated extensively in this chapter. The liquid slugging FDD methods, which also rely
upon the induction machine model of Chapter ?? will then be addressed in Chapter ??.
Chapter ?? will address the diagnostic techniques used to identify the presence of refriger-
ant leaks in an air-conditioning system with electrical measurements and two temperature
measurements, and then the range of FDD methods developed over the course of this re-
search will be reviewed and placed in the larger context of potential air-conditioning FDD
systems in Chapter ??. Each of these chapters will begin by motivating the development
of FDD methods for the particular fault of interest and reviewing the previous work done
in identifying these faults. The theoretical background used to develop the electrically-
based approach to identifying these faults will then be discussed; this will be followed by
a survey of the experimental apparatus used to simulate and develop the particular diag-
nostic techniques. Finally, the experimental results for the class of faults studied will be
presented and discussed at the end of each chapter.
Rev: 6701, July 16, 2008 46 main.tex

Bibliography
Bibliography
[1] P.P. Angelov, R.A. Buswell, V.I. Hanby, and J.A. Wright. A methodology for modeling
HVAC components using evolving fuzzy rules. In 26th Annual Conference of the IEEE
Industrial Electronics Society, volume 1, pages 247–252, Nagoya, 28 October 2000.
[2] P. R. Armstrong. Model Identification with Application to Building Control and Fault De-
tection. PhD thesis, Massachusetts Institute of Technology, September 2004.
[3] P.R. Armstrong, C.R. Laughman, S.B. Leeb, and L.K. Norford. Fault detection based
on motor start transients and shaft harmonics measured at the RTU electrical service.
In International Refrigerigation and Air Conditioning Conference at Purdue, number R137,
pages 1–10, 12 July 2004.
[4] P.R. Armstrong, C.R. Laughman, S.B. Leeb, and L.K. Norford. Detection of rooftop
cooling unit faults based on electrical measurements. HVAC+R Research Journal,
12(1):151–175, January 2006.
[5] ASHRAE. ASHRAE Handbook: HVAC Systems and Equipment. ASHRAE, Atlanta, GA,
2004.
[6] A. Bejan. Advanced Engineering Thermodynamics. Wiley-Interscience, second edition,

1997.
[7] M.E.H. Benbouzid and G.B. Kliman. What stator current processing-based technique
to use for induction motor rotor faults diagnosis? IEEE Transactions on Energy Conver-
sion, 18(2):238–244, June 2003.
main.tex 49 Rev: 6701, July 16, 2008

Bibliography
[8] M.E.H. Benbouzid, M. Veiera, and C. Theys. Induction motors’ faults detection and
localization using stator current advanced signal processing techniques. IEEE Trans-
actions on Power Electronics, 14(1):14–22, January 1999.
[9] M. Benouarets, A.L. Dexter, R.S. Fargus, P. Haves, T.I. Salsbury, and J.A. Wright.
Model-based approaches to fault detection and diagnosis in air-conditioning systems.
In Proceedings of the Fourth International Conference on System Simulation in Buildings,
pages 529–547, Liege, Belgium, 1995.
[10] J.E. Braun. Automated fault detection and diagnostics for vapor compression cooling
equipment. ASME Journal of Solar Energy Engineering, 125:266–274, August 2003.
[11] M.S. Breuker and J.E. Braun. Common faults and their impacts for rooftop air condi-
tioners. International Journal of HVAC+R Research, 4(3):303–318, June 1998.
[12] M.S. Breuker and J.E. Braun. Evaluating the performance of a fault detection and
diagnostic system for vapor compression equipment. International Journal of HVAC+R
Research, 4(4):401–425, October 1998.
[13] B. Checket-Hanks. Compressor problem: ’No Defect Found’. Air Conditioning, Heat-
ing, and Refrigeration News, pages 10,12, 19 May 2003.
[14] B. Chen and J.E. Braun. Simple rule-based methods for fault detection and diagnostics
applied to packaged air conditioners. ASHRAE Transactions, 107(1):847–857, 2001.
[15] M.C. Comstock and J.E. Braun. Literature Review for Application of Fault Detec-
tion and Diagnostic Methods to Vapor Compression Cooling Equipment. Technical
Report HL 99-19, Report #4036-2, Purdue University, December 1999. Sponsored by
ASHRAE Deliverable for Research Project 1043-RP.
[16] G. Cooper. Air-Conditioning America: Engineers and the Controlled Environment, 1900-
1960. Johns Hopkins University Press, 1998.
[17] R.W. Cox. Minimally intrusive strategies for fault detection and energy monitoring. PhD
thesis, Massachusetts Institute of Technology, September 2006.
Rev: 6701, July 16, 2008 50 main.tex

Bibliography
[18] M. Cunniffe, R.W. James, and A. Dunn. An analysis of fault occurrence in refrigeration
plant and the effect on current practice. Australian Refrigeration, Air Conditioning, and
Heating, 40(7):36–43, July 1986.
[19] T. DeNucci, R.W. Cox, S.B. Leeb, J. Paris, T.J. McCoy, C.R. Laughman, and W.C.
Greene. Diagnostic indicators for shipboard systems using non-intrusive load moni-
toring. In IEEE Electric Ship Technologies Symposium, Philadelphia, PA, July 2005.
[20] D. Dragomir-Daescu, A.A. Al-khalidy, M. Osama, and G.B. Kliman. Damage detec-
tion in refrigerator compressors using vibration and current signatures. In SDEMPED
2003 – Symposium on Diagnostics for Electric Machines, Power Electronics, and Drives,
pages 355–360, Atlanta, GA, 24 August 2003.
[21] Michael Y. Feng, Kurt W. Roth, Detlef Westphalen, and James Brodrick. Packaged
rooftop units: automated fault detection and diagnostics. ASHRAE Journal, 47(4):68–
72, April 2005.
[22] H.T. Grimmelius, J.K. Woud, and G. Been. On-line failure diagnosis for compression
refrigeration plants. International Journal of Refrigeration, 18(1):31–41, 1995.
[23] A.K. Halm-Owoo and K.O. Suen. Applications of fault detection and diagnostic tech-
niques for refrigeration and air conditioning: a review of basic principles. Proceedings
of the Institution of Mechanical Engineers Part E - Journal of Process Mechanical Engineer-
ing, 216(3):121–132, August 2003.
[24] C.Y. Han, Y. Xiao, and C.J. Ruther. Fault detection and diagnosis of HVAC systems.
ASHRAE Transactions, 105(1):568–578, 1999.
[25] A. Hasan. The occurrence of faults in heating and air conditioning equipment. The
Heating & Ventilating Engineering and Journal of Air Conditioning, 47(561):447–455, April
1974.
[26] P. Haves. Fault modelling in component-based HVAC simulation. In Building Simu-

lation ’97; Fifth International IBPSA Conference, volume 1, pages 119–127, 8 September
1997.
main.tex 51 Rev: 6701, July 16, 2008

Bibliography
[27] P. Haves, T.I. Salsbury, and J.A. Wright. Condition monitoring in HVAC subsystems
using first principles models. ASHRAE Transactions, 102(1):519–527, 1996.
[28] R. Isermann. Process fault detection based on modeling and estimation methods – a
survey. Automatica, 20(4):387–404, July 1984.
[29] R. Isermann. Fault-Diagnosis Systems: An Introduction from Fault Detection to Fault Tol-
erance. Springer, 2006.
[30] M. Kim, S.H. Yoon, P.A. Domanski, and W.V. Payne. Design of a steady-state detector
for fault detection and diagnosis of a residential air conditioner. International Journal
of Refrigeration, 31:790–799, 2008.
[31] Minsung Kim and Min Soo Kim. Performance investigation of a variable speed vapor
compression system for fault detection and diagnosis. International Journal of Refriger-
ation, 28:481–488, 2005.
[32] C.R. Laughman, K.D. Lee, R.W. Cox, S.R. Shaw, S.B. Leeb, L.K. Norford, and P.R.
Armstrong. Power signature analysis. IEEE Power and Energy Magazine, 1(2):56–63,
March 2003.
[33] T.M. Lawrence, J.D. Mullen, D.S. Noonan, and J. Enck. Overcoming barriers to effi-
ciency. ASHRAE Journal, 47:S40–S47, September 2005.
[34] S. Lee. Don’t bury the compressor before it’s dead. Air Conditioning, Heating, and
Refrigeration News, pages 01,10, 4 April 2005.
[35] S.B. Leeb, S.R. Shaw, and J.L. Kirtley. Transient event detection in spectral enve-
lope estimates for nonintrusive load monitoring. IEEE Transactions in Power Delivery,
10(3):1200–1210, July 1995.
[36] H. Li and J.E. Braun. On-line models for use in automated fault detection and diagno-
sis for HVAC&R equipment. In Proceedings of the 2002 ACEEE Conference, 7.147-7.158,
Monterey, CA, 2002.
[37] H. Li and J.E. Braun. An improved method for fault detection and diagnosis applied
to air conditioners. ASHRAE Transactions, 109(2):683–692, 2003.
Rev: 6701, July 16, 2008 52 main.tex

Bibliography
[38] H. Li and J.E. Braun. Application of automated fault detection and diagnostics for
rooftop air conditioners in california. In Proceedings of the 2004 ACEEE Conference,
pages 3.190–3.201, Monterey, CA, 2004.
[39] H. Li and J.E. Braun. An economic evaluation of automated fault detection and di-
agnosis for rooftop air conditioners. In Proceedings of the International Refrigeration and
Air Conditioning Conference at Purdue, R145, pages 1–10, 12 July 2004.
[40] H. Li and J.E. Braun. A methodology for diagnosing multiple-simultaneous faults in

rooftop air conditioners. In Proceedings of the International Refrigeration and Air Condi-
tioning Conference at Purdue, R131, pages 1–10, 12 July 2004.
[41] Haorong Li. A Decoupling-Based Unified Fault Detection and Diagnosis Approach for Pack-
aged Air Conditioners. PhD thesis, Purdue University, August 2004.
[42] Z. Liu and W. Soedel. An investigation of compressor slugging problems. In 1994

International Compressor Engineering Conference at Purdue, volume 2, pages 433–440,
July 1994.
[43] T. M. Rossi and J.E. Braun. A statistical, rule-based fault detection and diagnostic
method for vapor compression air conditioners. International Journal of HVAC+R Re-
search, 3(1):19–37, January 1997.
[44] K. Roth, D. Westphalen, and J. Brodrick. Residential central ac fault detection and
diagnostics. ASHRAE Journal, 48(5):96–97, May 2006.
[45] A. Salkin. Shivering for luxury. The New York Times, 26 June 2005.
[46] R.R. Schoen and T.G. Habetler. Effects of time-varying loads on rotor fault detection
in induction machines. IEEE Transactions on Industry Applications, 31(4):900–906, July
1995.
[47] S.R. Shaw, C.B. Abler, R.F. Lepard, D. Luo, S.B. Leeb, and L.K. Norford. Instrumenta-
tion for high performance nonintrusive electrical load monitoring. Transactions of the
ASME Journal of Solar Energy Engineering, 120(3):224–229, August 1998.
main.tex 53 Rev: 6701, July 16, 2008

Bibliography
[48] S.R. Shaw and S.B. Leeb. Identification of induction motor parameters from transient
stator current measurements. IEEE Transactions on Industrial Electronics, 46(1):139–149,
February 1999.
[49] S.R. Shaw, S.B. Leeb, L.K. Norford, and R.W. Cox. Nonintrusive load monitoring and
diagnostics in power systems. IEEE Transactions on Instrumentation and Measurement,
57(7):1445–1454, July 2008.
[50] S.R. Shaw, L.K. Norford, D. Luo, and S.B. Leeb. Detection and diagnosis of HVAC
faults via electrical load monitoring. International Journal of HVAC+R Research, 8(1):13–
40, January 2002.
[51] J. Shugars, P. Coleman, C. Payne, and L.V.W. McGrory. Bridging the efficiency gap:
commercial packaged rooftop air conditioners. In 2000 ACEEE Summer Study, 2000.
[52] R. Singh, J.J. Nieter, and G. Prater Jr. An investigation of the compressor slugging
phenomenon. ASHRAE Transactions, 92(4):250–258, 1986.
[53] R. Singh, G. Prater Jr, and J.J. Nieter. Prediction of slugging-induced cylinder over-
pressure. In 1986 International Compressor Engineering Conference at Purdue, volume 2,
pages 444–459, August 1986.
[54] D.E. Stouppe and T.Y.S. Lau. Air conditioning and refrigeration equipment failures.
National Engineer, 93:14–17, 1989.
[55] M. Stylianou. A response for performance degradation in rooftop units. Technical Re-
port TR 1999-19, CEDRL - Natural Resources Canada, 1615 Lionel-Boulet, Varennes,
Quebec, Canada, J3X 1S6, 1999.
[56] TASC Technical Staff. Applied Optimal Estimation. MIT Press, 1974.
[57] C. Thybo and R. Izadi-Zamanabadi. Development of fault detection and diagnosis

schemes for industrial refrigeration systems - lessons learned. In Proceedings of the
2004 IEEE International Conference on Control Applications, pages 1248–1253, 2 Septem-
ber 2004.
Rev: 6701, July 16, 2008 54 main.tex

Bibliography
[58] C. Thybo, R. Izadi-Zamanabadi, and H. Niemann. Toward high performance in in-

dustrial refrigeration systems. In Proceedings of the 2002 IEEE International Conference
on Control Applications, pages 915–920, Glasgow, Scotland, U.K., 18 September 2002.
[59] P.B. Usoro, I.C. Schick, and S. Negahdaripour. An innovation-based methodology for
HVAC system fault detection. Journal of Dynamic Systems, Measurement, and Control,
107(4):284–289, December 1985.
[60] J. Wagner and R. Shoureshi. Failure detection diagnostics for thermofluid systems.
Journal of Dynamic Systems, Measurement, and Control, 114(4):699–706, December 1992.
[61] M.L. Wald. New math for summer: the cost of chill. The New York Times, 23 June 2005.
[62] Shengwei Wang and Jin-Bo Wang. Robust sensor fault diagnosis and validation in
HVAC systems. Transactions of the Institute of Measurement and Control, 24(3):231–262,
2002.
[63] J. Wen and T.F. Smith. Development and validation of online parameter estimation
for HVAC systems. Journal of Solar Energy Engineering, 125:324–330, August 2003.
[64] H. Yoshida, H. Yuzawa, T. Iwami, and M. Suzuki. Typical faults of air-conditioning

systems and fault detection by ARX model and extended Kalman filter. ASHRAE
Transactions, 102(1):557–564, 1996.
[65] M. Yoshimura and N. Ito. Effective diagnosis methods for air-conditioning equipment
in telecommunications buildings. In 11th Annual Telecommunications Energy Confer-
ence, pages 21.1/1–21.1/7, Florence, 15 October 1989.
main.tex 55 Rev: 6701, July 16, 2008

Fault Detection Methods For Vapor-Compression Air Conditioners Using Electrical Measurements

Uploaded by

Copyright:

Available Formats

Fault Detection Methods For Vapor-Compression Air Conditioners Using Electrical Measurements

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fault Detection Methods For Vapor-Compression Air Conditioners Using Electrical Measurements

Uploaded by

Copyright:

Available Formats

Fault detection methods for vapor-compression air conditioners

using electrical measurements

Doctor of Philosophy in Architecture: Building Technology

Thesis Supervisor: Leslie K. Norford

Thesis Supervisor: Steven B. Leeb

Thesis Reader: Leon R. Glicksman

Thesis Reader: James L. Kirtley

Thesis Reader: Steven R. Shaw

Thesis Reader: Peter R. Armstrong

Submitted to the Department of Architecture

Thesis Supervisor: Leslie K. Norford

Thesis Supervisor: Steven B. Leeb

”You need the willingness to fail all the time.

main.tex 9 Rev: 6701, July 16, 2008

main.tex 11 Rev: 6701, July 16, 2008

1-1 The basic structure of an FDD system (adapted from [29]). . . . . . . . . . . 20

main.tex 13 Rev: 6701, July 16, 2008

main.tex 15 Rev: 6701, July 16, 2008

Whether the motivation is environmentalism, money, or just an interest in thermody-

main.tex 17 Rev: 6701, July 16, 2008

Rev: 6701, July 16, 2008 18 main.tex

Interest in the theoretical and experimental development of FDD methods in general

main.tex 19 Rev: 6701, July 16, 2008

1.1 Fault Detection and Diagnostic Systems

inputs system sensors outputs

Rev: 6701, July 16, 2008 20 main.tex

1.1.1 Fault Detection

main.tex 21 Rev: 6701, July 16, 2008

Rev: 6701, July 16, 2008 22 main.tex

inputs system sensors outputs

inputs system sensors outputs

main.tex 23 Rev: 6701, July 16, 2008

Rev: 6701, July 16, 2008 24 main.tex

main.tex 25 Rev: 6701, July 16, 2008

Rev: 6701, July 16, 2008 26 main.tex

fault 1 symptom 1 fault 1 symptom 1

Figure 1-4: caption.

Figure 1-5: caption.

1.1.2 Fault Diagnosis

main.tex 27 Rev: 6701, July 16, 2008

A number of different approaches can be taken to implementing this fault isolation

Rev: 6701, July 16, 2008 28 main.tex

symptoms as corresponding to one or multiple faults. There is a rich body of literature

1.1.3 Fault Evaluation and Decision

main.tex 29 Rev: 6701, July 16, 2008

Rev: 6701, July 16, 2008 30 main.tex

1.2 Previous Work

While the previously discussed framework is useful in developing a general understand-

1.2.1 Air Conditioning System Overview

A schematic diagram of a vapor-compression air-conditioner is illustrated in Figure 1-6.

main.tex 31 Rev: 6701, July 16, 2008

Figure 1-6: Schematic diagram of the air-conditioner.

Rev: 6701, July 16, 2008 32 main.tex

1.2.2 Common Faults in Air-Conditioners

main.tex 33 Rev: 6701, July 16, 2008

Rev: 6701, July 16, 2008 34 main.tex

main.tex 35 Rev: 6701, July 16, 2008

Rev: 6701, July 16, 2008 36 main.tex

main.tex 37 Rev: 6701, July 16, 2008