VOL. 3, NO.

1 HVAC&R RESEARCH JANUARY 1997

19
A Statistical, Rule-Based Fault Detection
and Diagnostic Method for Vapor
Compression Air Conditioners
Todd M. Rossi, Ph.D. James E. Braun, Ph.D., P.E.
Member ASHRAE Member ASHRAE
This paper presents a method for automated detection and diagnosis of faults in vapor compres-
sion air conditioners that only requires temperature measurements, and one humidity measure-
ment. The differences between measured thermodynamic states and predicted states obtained
from models for normal performance (residuals) are used as performance indices for both fault
detection and diagnosis. For fault detection, statistical properties of the residuals for current and
normal operation are used to classify the current operation as faulty or normal. A diagnosis is
performed by comparing the directional change of each residual with a generic set of rules unique
to each fault. This diagnostic technique does not require equipment-specific learning, is capable
of detecting about a 5% loss of refrigerant, and can distinguish between refrigerant leaks, con-
denser fouling, evaporator fouling, liquid line restrictions, and compressor valve leakage.
INTRODUCTION
Automated fault detection and diagnostics (FDD) for HVAC systems has the potential to
reduce energy and maintenance costs and improve comfort and reliability. Sensors are chosen
and located strategically within the system so that their outputs are sensitive to the faults for
which the FDD system is designed. The appropriate number and type of sensors results from a
tradeoff between initial cost and performance of the FDD system. A microprocessor is used to
process the measurements in order to provide the fault detection and diagnostic capability, and
also may be used to provide communications with a central service facility.
FDD Background
As described by Isermann (1984) and depicted in Figure 1, a FDD system may include the
following steps: fault detection, fault diagnosis, and fault evaluation. Fault detection indicates a
deviation of performance from expectation, diagnosis determines the cause of the fault, and
evaluation assesses whether the impact is severe enough to justify service. In each of these steps,
it is necessary to define criteria or thresholds for establishing appropriate outputs. The outputs
would be fault or no fault for fault detection, the type of fault for diagnosis, and repair or dont
repair for the fault evaluation step.
Fault detection is accomplished by comparing performance determined from measurements
with some expectation of performance. If the deviation exceeds a threshold, then a fault would
be indicated. As depicted in Figure 2, this process may be described in two steps: preprocessing
and classification. The preprocessor takes measurements from sensors and manipulates them to
generate features for classification. Classifiers then operate on the features to determine whether
the system contains a fault.
Todd M. Rossi is with Field Diagnostic Services, Ivyland, Pennsylvania. James E. Braun is an Associate Professor in the School of
Mechanical Engineering. Purdue University, West Lafayette, Indiana.
Table of Contents
20 HVAC&R RESEARCH
Simple transformations, characteristic quantities, and models are three types of preprocessors.
Simple transformations include the identity transformation (i.e., no preprocessing) and trend
generation (i.e., time derivatives). Characteristic quantities are features that are computed
directly from measurements and that are indicative of component performance. Examples
include overall system efficiencies and heat exchanger effectiveness. Model-based preproces-
sors utilize mathematical models of the monitored system to generate features. Model parame-
ters could be learned from measurements when the system is operating normally, or determined
using physical models. The features used by the classifier from model-based preprocessors
could be the differences between measured and modeled performance (i.e., residuals) or physi-
cal parameters of the model.
In a broad sense, the classifier is an expert system. The knowledge necessary to make a fault
decision can be stored in a number of forms, including: a set of production rules (i.e. IF, THEN,
ELSE rules), a fault tree, and conditional probabilities for statistical pattern recognition classifi-
ers. Typically, it is necessary to assign the thresholds for deviations between current and normal
performance that constitute faults. In selecting thresholds, there is a tradeoff between detection
sensitivities and false alarm rates. Tighter thresholds result in greater sensitivities (detection of
smaller faults), but will lead to more false alarms (an indication of a fault that doesnt exist).
Thresholds are often determined based upon heuristics, although better performance (lower ratio
of false alarms to correct diagnoses) is achieved when statistical thresholds are employed.
In general, preprocessing simplifies the classification and improves overall performance of
the FDD system. In the absence of any preprocessing (identity transformation), the FDD system
is a classic expert system. All fault detection is then based upon rules that act directly on the
measurements. Consider a vapor compression system that uses condenser head pressure to
Figure 1. Supervision of HVAC&R Equipment
Figure 2. Sequential Steps in Fault Detection and Diagnosis
VOLUME 3, NUMBER 1, JANUARY 1997 21
detect faults. Without preprocessing, a simple check for excessive head pressure could be: If the
head pressure is greater than 425 psig (2.93 MPa), then a fault exists. Since the head pressure
varies under normal operation with the ambient temperature, the fault detection threshold must
be greater than the highest head pressure associated with normal operation. A more complex
expert system might contain a set of rules with different head pressure limits for different ambi-
ent temperatures.
Alternatively, a model-based preprocessor could model the relationship between head pres-
sure and ambient temperature under normal operation. Then, a fault would be identified if the
deviations between measured and modeled head pressures exceeded a specified threshold. The
FDD system with the model-based preprocessor would be significantly more sensitive to abnor-
mal behavior than the single rule system and easier to implement than the expert system with
many rules. The thresholds for allowable deviations could be established by evaluating the sta-
tistical properties of the measurements, and how well the model for normal operation fits the
measurements.
The structure of Figure 2 can also be used to describe fault diagnosis. Measurements are pro-
cessed in order to simplify the classification required to identify the particular component at
fault. The overall classification problem is different for fault diagnosis than fault detection in
that the decision is not binary (i.e., fault/no fault): the classifier must choose the specific fault
from a list of possibilities. However, the diagnostics problem can be reduced to a series of fault
detection problems through fault isolation.
With fault isolation, fault detection methods are applied to individual components for which
diagnoses are desired. For instance, condenser fouling in an air conditioner could be detected by
estimating the heat exchanger effectiveness from measurements on the condenser. The fault is
diagnosed as soon as it is detected and no additional classification is necessary. The disadvan-
tage of fault isolation is the large number of measurements required. The diagnosis of heat
exchanger fouling would require measurements of all states entering and leaving the heat
exchanger.
Another diagnostic approach involves comparing physical parameters determined from mea-
surements with values representative of normal operation. For instance, heat exchanger conduc-
tance could be estimated from entering and leaving conditions and used to diagnose fouling.
Here again, fault detection and diagnosis are combined and no separate diagnostic classification
is necessary.
A more common diagnostic approach that requires fewer measurements involves the use of
fault models. For each type of fault to be diagnosed, a fault model predicts the outputs associ-
ated with the occurrence of that fault for a current set of inputs. The fault is diagnosed through
the use of a classifier that attempts to find the fault model with the best representation for the
current behavior. The advantage of fault modeling for diagnosis is that fewer measurements are
required. However, it is necessary to have fault models for each fault and combinations of faults
to be diagnosed.
Fault evaluation follows fault detection and diagnosis and requires an evaluation of the
impact of a fault on system performance. Without this step, it must be obvious that the benefit of
servicing the fault justifies its expense. This is the case for many hard failures, such as broken
fan belts or seized compressors. However, fault evaluation is necessary for many performance
degradations, such as heat exchanger fouling, where the fault could be detected and diagnosed
well before the need for service. For HVAC applications, Rossi and Braun (1996) defined four
criteria for evaluating the need for service: comfort, safety, environmental, and economic. In
general, service should be performed whenever: (1) comfort cannot be maintained, (2) equip-
ment or personal safety is compromised, (3) environmental damage is occurring (e.g., refriger-
ant leakage), or (4) reduced energy costs justify the service expense.
22 HVAC&R RESEARCH
Applications of FDD to Vapor Compression Equipment
There is a large body of literature on fault detection and diagnostic techniques for applications
in critical processes. As the cost of hardware (e.g., sensors, micro-processors) has gone down,
interest in developing FDD systems for HVAC&R applications has increased. Most of the liter-
ature for HVAC&R applications has focused on hard failures for large central chilled water
distribution systems and air handling systems. The literature for fault detection and diagnosis for
vapor compression equipment is relatively sparse but includes contributions by McKellar
(1987), Stallard (1989), Yoshimua and Noboru (1989), Kumamaru et al. (1991), Wagner and
Shoureshi (1992), Hiroshi et al. (1992), Grimmelius et al. (1995), and Rossi and Braun (1996).
McKellar (1987) identified many of the common faults for home refrigerators and investigated
the effects of several faults on thermodynamic measurements within the vapor compression cycle.
The faults that he considered were compressor valve leakage, heat exchanger fan failures, frost on
the evaporator, partially blocked capillary tube, and refrigerant charging failures. McKellar found
that each of these faults had unique effects on three measures: suction pressure (or temperature),
discharge pressure (or temperature), and discharge-to-suction pressure ratio and concluded that
these measures were sufficient for developing a FDD system. He did not develop a general
approach to characterizing expectations for these measurements (i.e., a model-based preproces-
sor) nor did he discuss thresholds for fault detection and diagnostic classifiers.
Based upon the work of McKellar, Stallard (1989) developed an expert system for automated
FDD applied to refrigerators. Condensing temperature, evaporating temperature, condenser inlet
temperature, and the ratio of discharge-to-suction pressure were used directly as classification
features (i.e., no model-based preprocessor). Feature limit checking, the simplest of rule-based
classifiers, was used for both detection and diagnostic classification. Fault diagnoses were per-
formed by evaluating the direction in which classification features changed from expected values
and matching these changes to expected directional changes associated with each fault (when they
exceeded the fixed threshold). Different rules were used for each of three discrete ranges of ambi-
ent temperature.
Wagner and Shoureshi (1992) used a different approach to perform FDD for refrigerator faults.
Dynamic, nonlinear state estimation techniques were used to generate residuals between current
and expected states. Compressor shell temperature, condensing temperature, and compressor
power were measured system responses, and ambient temperature was a measured model input.
Experiments were used to develop dynamic fault models for each of the faults considered. On-line
measurements were statistically compared with normal and fault models in order to perform diag-
nostic classification.
Kumamaru et al. (1991) used characteristic curves to obtain quantitative expectations for heat
pump performance as a function of cooling water temperature and loading. Diagnostics were per-
formed using residuals as input features. The method did not utilize statistically based thresholds
and did not detect performance degradations.
Yoshimua and Noboru (1989) used a combination of seven temperature and pressure mea-
surements to perform FDD for packaged air conditioners. Their method used rules with fixed
thresholds to perform detection and diagnosis. They did not use any preprocessing or statistical
rule evaluation.
Hiroshi et al (1992) developed a refrigerant leak detection method for automotive air condi-
tioners that used a measurement of the liquid to gas flow ratio in the liquid line. The method did
not utilize any model-based preprocessing to account for the effects of ambient temperature and
load conditions on expectations for this measurement and used fixed thresholds. As a result, the
method could only detect refrigerant loss with a sensitivity of about 50% of full charge.
VOLUME 3, NUMBER 1, JANUARY 1997 23
Grimmelius et al. (1995) used differences between measurements and outputs of steady-state
models for expected behavior as input features for detection and diagnoses of chiller faults. The
method used approximately 20 measurements, including temperatures, pressures, power con-
sumption, and compressor oil level. Diagnoses were performed using a pattern recognition tech-
nique applied to the current residuals and a matrix of expected residual changes associated with
each possible fault. The fault matrix was determined using experiments on a chilled water sys-
tem to which faults had been introduced.
Rossi and Braun (1996) addressed the issue of fault evaluation for air conditioning equip-
ment. They developed a simplified method for estimating the optimal service times that mini-
mizes combined energy and service costs for cleaning condensers and evaporators in air
conditioners. They also compared the costs associated with optimal maintenance scheduling
with those associated with regular maintenance and with a procedure where service was per-
formed based only on the comfort and safety criteria (constrained service). Savings of between 5
and 15% of operating costs were possible through optimal maintenance scheduling.
Although each of the previous studies provided contributions, all of the methods have limita-
tions. In some cases, expensive measurements are required (e.g., mass flow, power consump-
tion, pressure), while in others, fault detection and diagnostic sensitivity could be significantly
improved with the use of model-based preprocessing and statistically-based thresholds. The use
of system specific fault models could require extensive experimentation for each possible fault,
particularly when performance degradation faults are considered. None of the previous studies
included results for sensitivities of the FDD methods in detecting and diagnosing faults.
Scope of This Study
This paper describes the development and evaluation of a new method for detecting and diag-
nosing faults in air conditioning equipment that only requires temperature and humidity mea-
surements. The diagnostic approach is based on generic rules and does not require equipment
specific experimentation. Thresholds for both fault detection and diagnosis are based upon sta-
tistical analysis of on-line measurements. Five distinct faults were considered: (1) refrigerant
leakage; (2) liquid line restriction; (3) leaky compressor valves; (4) fouled condenser coil; and
(5) dirty evaporator filter. Only the fault detection and diagnostic steps were considered. Further
work would be necessary to establish methods for fault evaluation. One primary goal was to
identify the sensitivity of the algorithm in detecting and diagnosing each of these faults.
Refrigerant loss detection could be an immediate application for the proposed FDD method,
since fault evaluation is not necessary for this fault. A refrigerant leak should be repaired as soon
as it is detected and diagnosed. As a step towards implementation, the impact of the number of
sensors on the sensitivity for detecting and diagnosing refrigerant leakage was also studied.
The performance of FDD methods for this study was evaluated using a combination of simu-
lations and laboratory experiments. Diagnostic rules were developed through simulation and
checked within the laboratory. The sensitivities of the FDD method for detecting and diagnosing
each of the five faults were estimated through simulation. Both simulations and experiments
were used to investigate the impact of the number of sensors on FDD performance for detecting
and diagnosing refrigerant leaks.
EVALUATION TOOLS
For the simulations, a vapor compression system model developed and validated by Rossi
(1995) was used. It is a modular, steady-state model that solves mass, energy, and momentum bal-
ances for any set of entering air conditions. A steady-state model is appropriate because the fluid
flow and heat transfer dynamics are generally much faster than the dynamics of the load and ambi-
ent conditions. The model allows the introduction of all of the faults considered in this study.
24 HVAC&R RESEARCH
A fully-instrumented, three ton rooftop air conditioner was used for testing FDD algorithms.
The system has fixed-speed condenser and evaporation fans, a fixed orifice expansion device,
and a single-stage, on/off controlled reciprocating compressor. Figure 3 illustrates the manner
with which faults were simulated in the test unit: (1) condenser fouling: paper was placed on the
air-side of the coils, (2) evaporator filter fouling: paper was placed on the air-side filter, (3)
leaky compressor valves: this effect was represented using a hot gas bypass line with a manual
valve, (4) liquid line restriction: a manual valve was located in the liquid line before the expan-
sion device, and (5) refrigerant leakage: refrigerant charge removal was controlled by a valve
located on the high pressure side of the unit and monitored with a scale.
All refrigerant cycle temperature measurements were made using K-type thermocouples
mounted to the exterior surfaces of refrigerant piping and insulated. On the air side, tempera-
tures of entering and leaving air were measured using platinum resistance temperature devices
(RTDs) for both the evaporator and condenser streams. The thermocouples and RTDs were cali-
brated by immersion in a mixed ice bath. The measured error for each sensor was used for an
offset correction and to estimate the accuracy of the measurement. The error of all the thermo-
couple measurements was less than 0.4 K, while the RTDs were within 0.3 K. The relative
humidity of air entering the evaporator was determined from measurements of dew point using a
chilled mirror dewpoint hydrometer. This device is accurate to within 0.5 K and results in an
error in relative humidity that is less than 0.05 for the range of conditions considered in this
study.
FDD TECHNIQUE
Figure 4 shows a block diagram of the fault detection and diagnostic system in relation to
input and output measurements from a vapor compression cycle. For fixed-speed fans, an air-
to-air vapor compression cycle is driven only by the air inlet states into the evaporator and con-
denser (U). These states are characterized by the ambient temperature (T
amb
) and the return air
temperature (T
ra
) and relative humidity (
ra
) of air stream entering the evaporator. In steady-
state, every internal state of a normally operating cycle depends solely on the driving conditions.
In this investigation, the performance of the cycle was characterized using a vector of tempera-
ture measurements (Y) only. At most, seven temperatures were considered: (1) evaporating tem-
perature (T
evap
), (2) suction line superheat (T
sh
), (3) condensing temperature (T
cond
), (4) liquid
Figure 3. Rooftop Air Conditioner with Simulated Faults
VOLUME 3, NUMBER 1, JANUARY 1997 25
line subcooling (T
sc
), (5) hot gas line or compressor outlet temperature (T
hg
), (6) air temperature
rise across condenser (T
ca
), and (7) air temperature drop across evaporator (T
ea
).
The preprocessor portion of the FDD system contains two major components: a steady-state
model and the preprocessor portion of a steady-state detector. The measured inputs are used by
the model to predict each measured output for the vapor compression cycle under normal oper-
ation (i.e., no faults). The difference between measured and model predictions of the operating
states (residuals) are used by the fault detection and diagnostic classifiers for decision making.
A steady-state model is appropriate for the faults in this study because the fault development
rate is generally slower than the equilibration time of the plant and it is easier to implement
than dynamic modeling. Dynamic signatures would be useful for abrupt failures such as broken
fan belts.
A steady-state detector is necessary to determine when the models predictions are valid (e.g.,
predictions are not valid during startup and shutdown transients). The steady-state detectors
preprocessor evaluates the variation in the output measurements for use by a classifier.
The classifier consists of fault detection, diagnostic, and steady-state classifiers. The fault
detection classifier operates on the residuals and provides a binary output indicating when the
current operating state deviates from expectation with greater than a prescribed statistical confi-
dence. The diagnostic classifier provides the most likely cause of the fault after it is detected.
The steady-state detection classifier provides a binary output that is an input to a switch (SW)
that controls the output of the FDD system. The FDD system will only indicate a fault and pro-
vide a diagnosis when the system is in steady state.
This study has focused on the development and evaluation of the fault detection and diagnos-
tic classifiers only. The steady-state model was a lookup table that produced perfect predictions
Figure 4. Fault Detection and Diagnostic System
26 HVAC&R RESEARCH
of plant output measurements when given the correct measured inputs. Errors in measurements
were propagated through the model and used to assess the sensitivity of the FDD system in
detecting faults. In practice, imperfect modeling would decrease the sensitivity of the FDD sys-
tem. Further work would be necessary to develop appropriate on-line models and assess their
impact on overall FDD performance.
All simulations and experiments were performed at steady-state conditions in this study.
Steady-state detection could be implemented using techniques developed by Glass et al. (1994).
These methods are based on estimating the sample standard deviation about the mean of output
measurements over a moving exponentially weighted window. Alternatively, the time rate of
change in temperature measurements during a moving window could be used in steady-state
detection.
Fault Detection Classifier
The features used for classification are the residuals between current state variable measure-
ments and outputs from the models. The classifier identifies a fault when the current measure-
ments are statistically different than the expected values. Figure 5 illustrates how this classifier
works for a one-dimensional example. In this case, the classifier feature is the residual of the
suction line superheat. The probabilities of obtaining a specific residual are shown for both nor-
mal and faulty operation. Under normal operation, there is a distribution of residuals that results
from measurement noise and modeling errors. In the absence of modeling errors and with ran-
dom noise, the distribution for normal operation would have zero mean. The introduction of a
fault changes both the mean and standard deviation of the residuals.
If both normal and faulty distributions were known a priori, then this would be a classical
classification problem. In this case, the current measurement residual would be compared to a
threshold value that depended upon the normal and faulty distributions. The minimum error in
classifying either normal behavior as faulty or fault behavior as normal would occur for a deci-
Figure 5. One Dimensional Example of Detection Classifier
VOLUME 3, NUMBER 1, JANUARY 1997 27
sion threshold located at the intersection of the two distributions. The probability of making an
erroneous classification using this threshold, termed the classification error, would then be the
overlap of the two probability distributions, shown as the shaded area in Figure 5. When using
this threshold, the classification error is the sum of the probability that a normal classification
would be faulty and the probability that a faulty classification would be normal. It could be
important to consider each type of classification error separately if the consequence of one is
more serious than the other.
This problem differs from a classic classification problem in that the faulty distribution is not
known a priori. As a fault progressively becomes worse, the current distribution of measurement
residuals moves further away from the distribution for normal operation. A fault should be indi-
cated when the classification error (shaded area) is small enough for the false alarm rate to be
acceptable.
An optimal linear classifier (Fukunaga 1990) was used for estimating the classification error
associated with current and normal distributions of residuals. This is one of several possible sta-
tistical classifier designs (Fukunaga, 1990) and assumes that measurement noise is caused by
independent random processes, and has a Gaussian distribution. The residuals for both normal
and current operation are fitted to a Gaussian model that is completely described in terms of a
mean vector and covariance matrix. The assumption of a Gaussian distribution significantly
reduces the memory and computational requirements for fault detection. A Bayes classifier for a
Gaussian distribution is in general quadratic (Fukunaga, 1990, p.125) and results in the follow-
ing inequality for evaluating whether a fault should be indicated.
(1)
where Y is a vector of current residuals, M
N
and
N
are the mean vector and covariance matrix
that describe the distribution of residuals in the absence of any faults (i.e., normal operation),
and M
C
and
C
are the mean vector and covariance matrix that describe the current distribution
of residuals determined using recent measurements. For a perfect model, the mean of the residu-
als would be zero. However, an actual model would likely give biased estimates under certain
operating conditions.
A fault would be indicated whenever the left-hand side of Equation (1) was greater than zero.
This formulation minimizes the probability of making a wrong decision, assuming that there is
no built-in bias that favors a faulty or normal output. The quadratic classifier is reduced to a
linear classifier by assuming that the separation between the distributions for current and normal
operation is dominated by mean vector differences as opposed to covariance matrix differences.
If it were assumed that

=
N
=
C
, then Equation (1) would reduce to
(2)
In practice, an average covariance matrix is determined as the weighted average of
N
and
C
according to
(3)
where s is determined by minimizing the classification error (i.e., probability of making an erro-
neous decision). The classification error is determined by integrating the overlapping areas asso-
ciated with the multi-dimensional normal and fault distributions using (Fukunaga 1990, p. 137)
Y M
N
( )
T
Y M
N
( )
N
1

Y M
C
( )
T
Y M
C
( )
C
1

ln

N

C
--------- + 0 >
M
C
M
N
( )
T

1
Y
1
2
--- M
N
T

1
M
N
M
C
T

1
M
C
( ) 0 >
s
N
1 s ( )
C
+ =
28 HVAC&R RESEARCH
(4)
where
In this study, a fault was indicated whenever the classification error () was below a threshold
equal to 0.001. This threshold for classification error gives a small false alarm rate and was
found to provide acceptable fault detection sensitivities. The value of s that minimizes the classi-
fication error as determined with Equation (4) was found using a golden-section search optimi-
zation method (Rao, 1984). The mean vector and covariance matrix that describe the
distributions for normal and current operation were determined, as outlined in the section on
testing and evaluation.
Diagnostic Classifier
Fault diagnosis is performed using the residual features as inputs to a rule-based classifier.
The set of rules relates each fault to the direction that each measurement changes when the fault
occurs. Table 1 gives the diagnostic rules for the five faults and seven measurements considered
in this study. These rules were developed and tested through simulation over a range of operat-
ing conditions and tested using experiments at a single operating point. Each of the faults results
in a different combination of increasing or decreasing measurements. The rules are effectively
fault models. However, they are generic for all similar types of air conditioners and do not
require on-line learning.
Figure 6 illustrates the fault diagnostic classification method for two possible faults (refriger-
ant leakage and liquid line restriction) with two input features (superheat and subcooling resid-
uals). The progression of changes in the contours of two-dimensional probability distributions
are shown as the two different faults are slowly introduced. Normal operation is shown as the
distribution centered at the zero point. As a fault develops, the contour moves along a curve.
When the overlap between the normal performance distribution and the current distribution (as
indicated by the classification error ) is small enough for the false alarm rate to be acceptable
(e.g., < 0.001), then a fault is signaled by the fault detector. The different diagnostic classes
Table 1. Rules for Diagnostic Classifier
Fault T
evap
T
sh
T
cond
T
sc
T
hg
T
ca
T
ea
Refrigerant leak
Compressor valve leakage
Liquid-Line restriction
Condenser fouling
Evaporator fouling
erfc
V
T
M
N
v
o

2
N
2
--------------------------------




erfc
V
T
M
C
v
o

2
C
2
--------------------------------




+ =
V s
N
1 s ( )
C
+ ( )
1
M
C
M
N
( ) =
v
o

s
N
2
V
T
M
C
1 s ( )
C
2
V
T
M
N
+
s
N
2
1 s ( )
C
2
+
------------------------------------------------------------------------- =

N
2
V
T

N
V =

C
2
V
T

C
V =
VOLUME 3, NUMBER 1, JANUARY 1997 29
are separated by the axis. The overlap of the current distribution with each of the modeled
classes is calculated and represents the probability that the fault class is the correct diagnosis. A
diagnosis is indicated when the probability (overlap) of the most likely class is larger than the
second most likely class by a specified threshold. As the fault becomes more severe, confidence
in the fault detection and diagnosis increases as the current distribution moves further from the
normal distribution, and from the axis separating the classes. Again, the choice of a diagnositic
threshold results from a tradeoff between diagnostic sensitivity and the rate of false diagnoses.
In order to perform the classification for diagnostics, the probability that each rule applies to
the current operation is evaluated. The probability of each hypothesis is determined by the
degree to which the distribution characterizing the current residuals overlaps each class. The
overlap is evaluated by integrating the area under the m-dimensional Gaussian probability distri-
bution that falls within each class region of the domain as given by:
(5)
where d
jk
is the integration limit associated with the domain of fault j in dimension k (), Y
k
is
the k
th
residual, and P(M
c
M
N
,
c
) is the m-dimensional probability distribution function for
the current residuals. For diagnoses, the current distribution has been shifted in order to give
zero mean for normal operation. A nonzero residual mean could occur for normal operation with
an imperfect model.
The calculation of the overlap within each class is simplified by assuming that each dimen-
sion is independent. In this case, the probabilities in each dimension can be ANDed together
such that:
(6)
Figure 6. Fault Diagnostic Classifier (Two-Dimensional Example)
w
j

P M
c
M
N

c
, ( )dY
1
dY
2

dY
m
0
d
j1

0
d
j2

0
d
jm

=
w
j
P
0
d
jk

M
c
k ( ) M
N
k ( )
c
k k , ( ) , [ ]dY
k
k=1
m

=
30 HVAC&R RESEARCH
where P[M
c
(k) M
N
(k),
c
(k,k)] is the probability distribution function for the current residuals
in the k
th
dimension. For Gaussian distributions, the integration reduces to:
(7)
where C
jk
= +1 if [M
c
(k) M
N
(k)] falls within the domain for the j
th
fault (i.e., [M
c
(k) M
N
(k)]
has the same sign as defined in Table 1 for the appropriate fault) and C
jk
= 1, otherwise.
In this study, the probability associated with each diagnosis was estimated using Equation (7).
A diagnosis was considered valid when the ratio of the probability of the most likely class to the
second most likely class was greater than 2. This diagnostic threshold was found to provide
good sensitivity and no misdiagnoses.
TESTING AND EVALUATION
In the simulation studies, outputs of the detailed physical model of an air conditioner were
used to represent the plant. Sensors were modeled by adding zero mean, independent, and iden-
tically distributed Gaussian noise to known values of the plant inputs and modeled outputs. The
standard deviations of the temperature and humidity measurements (
T
and

) were inputs to
the sensor model. For most of the results presented in this study,
T
= 0.5 K and

= 0.05.
In the experimental investigations, the driving conditions were fixed and the sensor values
were recorded for no fault, and then for progressively larger degrees of fault introduction. The
no fault measurement was used as the reference for generating the residuals for each fault level.
Since the driving conditions were fixed and the system was allowed sufficient time to achieve
steady state, there was no need for a steady-state detector.
The input and output measurements were also characterized using Gaussian distributions with
both on-line measurements and manufacturers specifications. The full extent of a sensors vari-
ation is often not captured by sampling a sensor for several minutes to an hour. Examples of
issues affecting sensor performance that may not be noticed within an hour include: ambient
temperature variations, power cycling, long-term electrical noise and drift, and calibration
errors. Therefore, measured variances are often much smaller than the specified accuracies of
the sensors. To give more realistic measurement error estimates, the variances were modified by
combining the estimated variance and the specified accuracy as if they were independent, nor-
mally distributed noise sources, as follows:
(8)
where
2
is the overall measurement variance,
meas
is the standard deviation of measurements
estimated from the data, and
spec
is the specified accuracy of each sensor. In most cases,
spec
dominates
meas
.
The preprocessors steady-state model incorporated a lookup table that produced perfect plant
predictions when given inputs with no measurement errors. With Gaussian distributed measure-
ment noise, the distribution for the i
th
temperature residual (difference between measured and
predicted outputs) is defined by
(9)
w
j
1
2
--- 1 C
jk
erf
M
c
k ( ) M
N
k ( )
2
C
k k , ( )
--------------------------------------



+
k=1
m

meas
2

spec
2
+ =
Y
i
Y
i
T
amb
T
ra

ra
, , ( ) f
i
T
amb
w
T
T
ra
, w
T

ra
, w

+ + + ( ) w
T
+ =
VOLUME 3, NUMBER 1, JANUARY 1997 31
where Y
i

is the i
th
output of the plant, f
i
is the preprocessor models prediction of the plant out-
put for normal operation, w
T
is zero mean noise added to the temperature measurements, and w

is zero mean noise added to the relative humidity measurement. Sensor noise was propagated
through the steady-state preprocessor using a first-order Taylor series about the known operat-
ing point.
(10)
The mean vector and covariance matrix were estimated in order to characterize a normal distri-
bution of the residuals. The i
th
entry in the mean vector is
(11)
where E is the expected value operator. The covariance matrix can be determined by using a
Taylor series approximation (Gelb 1989). Diagonal entries in the i
th
row of the covariance
matrix are
(12)
whereas off-diagonal entries are
(13)
and where
The Taylor series approximation was reasonable since only small measurement noise was
propagated through the model. The partial derivatives in Equations (10), (12), and (13) were
evaluated numerically using model predictions from the vapor compression system model
described by Rossi (1995).
FDD SENSITIVITIES
In this section, simulated results are presented for FDD performance using all of the seven
output temperature measurements depicted in Figure 4. Before a fault is detectable, the classifi-
cation error between the current observation and the estimate provided by the model must be
less than 0.001. Table 2 shows the effect of simulated charge leakage on the output of the fault
detection and diagnostic system. When 2% of the charge is removed, the classification error is
below 0.001 and a fault is indicated. The class probability indicating a refrigerant leak is already
nearly a factor of 10 greater than the next most likely explanation. As the fault becomes worse,
the certainty in the diagnosis improves dramatically.
Table 3 shows the effect of simulated compressor suction valve leakage on the output of the
fault detection and diagnostic system. Valve leakage was modeled as a decrease in the compres-
sors volumetric efficiency. A fault is indicated (classification error is below 0.001) when the
efficiency is reduced by 5%. At this point, the class probability indicating valve leakage is
Y
i
Y
i
T
amb
T
ra

ra
, , ( ) f
i
T
amb
T
ra

ra
, , ( )
f
i
T
amb
---------------w
T
f
i
T
ra
----------- w
T
f
i

ra
----------- w

w
T
+
E Y
i
[ ] Y
i
T
amb
T
ra

ra
, , ( ) f
i
T
amb
T
ra

ra
, , ( )
E Y
i
E Y
i
( ) ( )
2
[ ]
T
2
f
i
T
amb
---------------


2

T
2
f
i
T
ra
-----------


2

T
2
f
i

ra
-----------


2

2
+ + +
E Y
i
E Y
i
( ) ( ) Y
j
E Y
j
( ) ( ) [ ]
f
i
T
amb
---------------
f
j
T
amb
---------------
T
2
f
i
T
ra
-----------
f
j
T
ra
-----------
T
2
f
i

ra
-----------
f
j

ra
-----------

2
+ +

T
2
E w
T
2
[ ]

2
E w

2
[ ] = =
32 HVAC&R RESEARCH
already a factor of 7 greater than the next most likely explanation. Once again, the evidence
becomes stronger as the fault level increases.
The effect of a simulated liquid line restriction on the output of the fault detection and diag-
nostic system is shown in Table 4. The restriction was modeled by inserting a valve in the liquid
line and decreasing its cross-sectional area. In this case, a fault is not detected until the diameter
of the valve opening is reduced by about 80%. At this point, the class probability indicating liq-
uid line restriction is overwhelming. The FDD system appears to be insensitive to liquid line
restrictions because the restriction is modeled as a decrease in the diameter of the pipe just
before the expansion device and the restriction is not noticeable when it is much larger than the
expansion device opening. For the test unit, the diameter of the pipe is about 5 times larger than
the expansion device diameter. In practice, a liquid line restriction would occur at the site of the
expansion device or within the filter-dryer and would be detected with a much greater sensitivity
than demonstrated in Table 4.
Tables 5 and 6 show the effects of simulated condenser and evaporator fouling on the output
of the fault detection and diagnostic system, respectively. Condenser and evaporator fouling
Table 2. FDD Performance for Refrigerant Leakage
Fault Size, %
Classification
Error () Diagnosis
Class Probabilities
[leak comp rest cond evap]
0.1 3.71 E01 Normal
0.5 1.21 E01 Normal
1.0 1.72 E02 Normal
2.0 1.59 E05 Refrigerant leak [0.26 0.00 0.03 0.00 0.00]
5.0 3.76 E15 Refrigerant leak [0.62 0.00 0.00 0.00 0.00]
10.0 0 Refrigerant leak [0.91 0.00 0.00 0.00 0.00]
Table 3. FDD Performance for Compressor Valve Leakage
Fault Size, %
Classification
Error () Diagnosis
Class Probabilities
[leak comp rest cond evap]
0.1 4.20 E01 Normal
0.5 3.71 E01 Normal
1.0 2.43 E01 Normal
2.0 7.62 E02 Normal
5.0 1.47 E04 Compressor valve [0.00 0.15 0.00 0.00 0.02]
10.0 8.13 E10 Compressor valve [0.00 0.48 0.00 0.00 0.01]
15.0 1.09 E15 Compressor valve [0.00 0.65 0.00 0.00 0.00]
20.0 3.19 E29 Compressor valve [0.00 0.77 0.00 0.00 0.00]
30.0 0 Compressor valve [0.00 0.93 0.00 0.00 0.00]
Table 4. FDD Performance for Liquid Line Restriction
Fault Size, %
Classification
Error () Diagnosis
Class Probabilities
[leak comp rest cond evap]
5.0 4.44 E01 Normal
10.0 4.47 E01 Normal
15.0 4.42 E01 Normal
20.0 4.40 E01 Normal
30.0 3.85 E01 Normal
40.0 2.87 E01 Normal
60.0 1.96 E03 Normal
80.0 0 Restriction [0.00 0.00 0.98 0.00 0.00]
VOLUME 3, NUMBER 1, JANUARY 1997 33
were modeled as decreases in the air flow rate across the coils. For condenser fouling, a fault is
indicated when the flow rate is reduced by 20%, whereas the evaporator flow rate is reduced by
40% before a fault is detected. However, the evidence for evaporator fouling is much stronger at
the point when a fault is detected.
The sensitivities of the FDD system to the five faults would probably be adequate for this appli-
cation. Of the five faults considered, only refrigerant leakage would require repair as soon as it
was detected (environmental criteria). In this case, the FDD system detected changes of less than
2% in refrigerant charge. The other four faults would be serviced only if they affected comfort,
economics, or safety. Since the cooling capacity and power consumption of an air conditioner are
strongly coupled to the thermodynamic states used in the FDD method, the comfort and economic
criteria are not likely to be violated if a fault is not detectable. Furthermore, unsafe operation
would normally be associated with large deviations from normal operating states.
SENSOR REQUIREMENTS FOR REFRIGERANT LEAKAGE
It is not necessary to use all of the seven output temperature measurements depicted in Figure
4 if the goal is to distinguish refrigerant leakage from the other possible faults. Figure 7 shows
the simulated sensitivities for detecting refrigerant leaks as a function of the number of sensors
for a measurement noise of
T
= 0.5 K and

= 0.05. For a given number of sensors, the combi-

nation providing the best sensitivity was selected. The numbers above each bar indicate which
sensors were selected (see Figure 4 for a key). The results show that at least two measurements,
T
sh
and T
sc
, are required to distinguish refrigerant leaks from the other four faults. However, add-
ing a measurement of T
hg
significantly improved sensitivity, while additional sensors did not
provide much better performance. With three sensors (T
sh
, T
sc
, and T
hg
), less than 2% reduction
Table 5. FDD Performance for Condenser Fouling
Fault Size, %
Classification
Error () Diagnosis
Class Probabilities
[leak comp rest cond evap]
1.0 4.22 E01 Normal
2.0 3.62 E01 Normal
5.0 2.08 E01 Normal
10.0 3.31 E02 Normal
15.0 2.64 E03 Normal
20.0 4.96 E05 Condenser fouling [0.00 0.00 0.00 0.14 0.00]
30.0 3.68 E10 Condenser fouling [0.00 0.00 0.00 0.26 0.00]
40.0 2.84 E19 Condenser fouling [0.00 0.00 0.00 0.32 0.00]
60.0 0 Condenser fouling [0.00 0.00 0.00 0.35 0.00]
Table 6. FDD Performance for Liquid Line Restriction
Fault Size, %
Classification
Error () Diagnosis
Class Probabilities
[leak comp rest cond evap]
5.0 3.50 E01 Normal
10.0 1.73 E01 Normal
15.0 9.53 E02 Normal
20.0 3.80 E02 Normal
30.0 2.25 E03 Normal
40.0 4.20 E06 Evaporator fouling [0.00 0.00 0.00 0.00 0.74]
60.0 5.62 E19 Evaporator fouling [0.00 0.00 0.00 0.00 0.99]
80.0 0 Evaporator fouling [0.00 0.00 0.00 0.00 0.98]
90.0 0 Evaporator fouling [0.00 0.00 0.00 0.00 0.98]
34 HVAC&R RESEARCH
in charge was detected. When the measurement noise was reduced to
T
= 0.2 K and

= 0.02,
then refrigerant loss of less than 0.7% of full charge was detectable.
Figure 8 shows the classification error and fault probability ratio versus percent charge reduc-
tion for three sensors. The fault probability ratio is the ratio of the most probable fault to the sec-
ond most probable fault. The horizontal line in the center of the plot is the threshold for
classification error (left axis) and fault probability ratio (right axis). A fault is indicated when the
classification error is below the threshold and a valid diagnosis is indicated when the fault prob-
ability ratio is above the threshold. For this study, the minimum detectable fault was associated
with both values exceeding their respective thresholds. This point occurs at the intersection of
the vertical line separating the fault and normal regions and the abscissa in Figure 8, and is equal
to the value of the three sensor bar in Figure 8.
Figure 9 provides the minimum detectable charge reduction versus number of sensors for the
experimental investigation. The results confirm that the simulation tool correctly predicted the
relative importance of the output measurements. Three sensors were sufficient for detecting
refrigerant loss in addition to the three input measurements required for the preprocessor model.
However, the simulated performance was more sensitive to refrigerant charge than actually
occurs. The experiment indicates that a 5.0% reduction in charge can be detected.
CONCLUSIONS
A rule-based, statistical fault detection and diagnostic system was developed and evaluated
for vapor compression equipment. The method only requires temperature measurements and one
humidity measurement to distinguish between the following five faults: (1) refrigerant leakage,
(2) liquid-line restriction, (3) compressor valve leakage, (4) condenser fouling, (5) evaporator
fouling. The diagnostic approach is based on generic rules and does not require equipment spe-
cific experimentation. Thresholds for both fault detection and diagnosis are both based upon sta-
tistical analysis of on-line measurements.
Figure 7. Refrigerant Leak Detection Sensor SensitivitySimulation Results
(Numbers indicate sensors selected from the list in Figure 4.)
VOLUME 3, NUMBER 1, JANUARY 1997 35
The sensitivities of the FDD method for detecting and diagnosing each of the five faults were
estimated through simulation in the presence of typical measurement errors. This method is
most sensitive for detecting refrigerant loss. Simulated results showed that a 2% loss in refriger-
ant could be detected using five temperature measurements (superheat, subcooling, hot gas line,
Figure 8. Refrigerant Leak Detection Classification Error and
Diagnostic Fault Probability Ratio (3 Sensors)
Figure 9. Refrigerant Leak Detection Sensor SensitivityExperimental Results
(Numbers indicate sensors selected from the list in Figure 4.)
36 HVAC&R RESEARCH
condenser air inlet, and evaporator air inlet) and one humidity measurement (evaporator air
inlet) Experimental results confirmed the use of these measurements, but only a 5% reduction in
refrigerant was detectable.
Refrigerant loss detection could be an immediate application for the proposed FDD method,
since fault evaluation is not necessary for this fault. A refrigerant leak should be repaired as soon
as it is detected and diagnosed. Since the other four faults should only be serviced if they affect
comfort, economics, or safety, less detection sensitivity is required for them.
In this study, the thresholds for fault detection (classification error) and diagnoses (fault prob-
ability ratio) were chosen heuristically to give a low false alarm rate. The sensitivity of the FDD
method in detecting and diagnosing each fault was then evaluated for the fixed thresholds. Ide-
ally, the selection of thresholds should consider the tradeoff between the sensitivity of the
method and the false alarm rate. One would like to choose thresholds that allow the method to
detect small faults (high sensitivity), but that rarely lead to an indication of a fault that doesnt
exist (low false alarm rate). In theory, the principle of least risk could be used to help decide
what are appropriate sensitivities and false alarm rates. It would be possible to evaluate the sen-
sitivity and false alarm rate of the FDD method as a function of threshold values. This informa-
tion could be used in conjunction with costs associated with misdiagnoses (choosing faulty
behavior thats normal or normal behavior thats faulty) to determine appropriate thresholds.
However, these costs would be very difficult to obtain for this particular application.
This study also did not consider the effect of modeling errors on the sensitivity of the FDD
method. Modeling errors do reduce sensitivity and should be considered in future studies.
ACKNOWLEDGEMENTS
The research described in this paper was supported by grants from the National Institute of
Standards and Technology, Johnson Controls, Inc., and Purdue University. The research bene-
fited from many interactions with an International Energy Agency (IEA) working group on fault
detection and diagnostics (Annex 25).
NOMENCLATURE
E[
.
] expected value operator
f
i
plant model for i
th
output
M
C
mean vector that describes the current distri-
bution of residuals determined using recent
measurements
M
N
mean vector that describes the distribution of
residuals determined using measurements for
normal operation
P(
.
) probability density function
s parameter that minimizes classification error
for optimal linear classifier
T
amb
ambient temperature (inlet to condenser)
T
cond
condensing temperature
T
evap
evaporating temperature
T
hg
hot gas line temperature (compressor outlet)
T
ra
return air temperature (inlet to evaporator)
T
sc
liquid line subcooling
T
sh
suction line superheat
U vector of inputs that affect plant performance
Y vector of measured plant outputs
T
ca
air temperature rise across condenser
T
ea
air temperature drop across evaporator
w
j
estimate of probability that jth fault represents
the correct diagnosis
Y vector of residuals between measured and
modeled plant outputs
classification error (probability of making an
erroneous classification for normal or faulty
behavior)

ra
return air relative humidity (inlet to evaporator)
standard deviation of measurements
weighted average of
N
and
C
:

C
covariance matrix that describes the current
distribution of residuals determined using
recent measurements

N
covariance matrix that describes the distribu-
tion of residuals in absence of any faults (i.e.,
normal operation)
Subscripts
exp model prediction (expected performance pre-
dictions)
meas measured
spec manufacturers specification
T temperature
relative humidity
VOLUME 3, NUMBER 1, JANUARY 1997 37
REFERENCES
Fukunaga, K. 1990. Introduction to Statistical Pattern Recognition. W. Lafayette, IN: Academic Press.
Gelb, A. 1989. Applied Optimal Estimation, MIT, Cambridge, MA: M.I.T. Press.
Glass, A.S., P. Gruber, M. Roos, and J. Todtli. 1994. Preliminary Evaluation of a Qualitative Model-Based
Fault Detector for a Central Air Handing Unit. Proceedings of the Third IEEE Conference on Control
Applications pp. 1873-1882: IEEE Catalogue No. 94CH3420-7.
Grimmelius, H.T., J. Klein Woud, and G. Been. 1995. On-Line Failure Diagnosis for Compression Refrig-
eration Plants. International Journal of Refrigeration 18(1): 31-41.
Hiroshi, I.H., K. Matsuo, Fujiwara, K. Yamada, and K. Nishizawa. 1992. Development of Refrigerant
Monitoring Systems for Automotive Air-Conditioning Systems. Society of Automotive Engineers
Paper No. 920212.
Isermann, R. 1984. Process Fault Detection Based on Modeling and EstimationA Survey. Automatica
20(4): 387-404.
Kumamaru, T., T. Utsunomiya, Y. Yamada, Y. Iwasaki, I. Shoda, and M. Obayashi. 1991. A Fault Diagno-
sis System for District Heating and Cooling Facilities, Proceedings of the International Conference on
Industrial Electronics, Control, and Instrumentation, Kobe, Japan (IECON 91), pp. 131-136.
McKellar, M.G. 1987. Failure Diagnosis for a Household Refrigerator. M.S. Thesis, West Lafayette, IN:
Purdue University.
Rao, S. 1984. Optimization, 2nd ed. New Delhi, India: Wiley Eastern Limited.
Rossi, T.M. 1995. Detection, Diagnosis, and Evaluation of Faults in Vapor Compression Cycle Equipment.
Ph.D. Thesis, West Lafayette, IN: Purdue University.
Rossi, T.M., and J.E. Braun. 1996. Minimizing Operating Costs of Vapor Compression Equipment with
Optimal Service Scheduling, International Journal of Heating, Ventilating, Air-Conditioning and
Refrigerating Research 2(1): 23-47.
Stallard, L.A. 1989. Model Based Expert System for Failure Detection and Identification of Household
Refrigerators. M.S. Thesis. West Lafayette, IN: Purdue University.
Wagner, J., and R. Shoureshi. 1992. Failure Detection Diagnostics for Thermofluid Systems, Journal of
Dynamic Systems, Measurement, and Control 114(4): 699-706.
Yoshimura, M., and I. Noboru. 1989. Effective Diagnosis Methods for Air-Conditioning Equipment in
Telecommunications Buildings. In Proceedings of INTELEC 89: The Eleventh International Telecom-
munications Energy Conference, 21(1): 1-7.