See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/239399796

The Use of Kalman Filter and Neural Network Methodologies in Gas Turbine
Performance Diagnostics: A Comparative Study

Article in Journal of Engineering for Gas Turbines and Power · October 2003
DOI: 10.1115/1.1419016

CITATIONS READS

197 1,250

4 authors, including:

Ranjan Ganguli
Indian Institute of Science
345 PUBLICATIONS 6,651 CITATIONS

SEE PROFILE

All content following this page was uploaded by Ranjan Ganguli on 16 May 2014.

The user has requested enhancement of the downloaded file.

 The Use of Kalman Filter
and Neural Network
A. J. Volponi
H. DePold
Methodologies in Gas Turbine
R. Ganguli Performance Diagnostics:
Pratt & Whitney,
400 Main Street,
A Comparative Study
East Hartford, CT 06108
The goal of gas turbine performance diagnositcs is to accurately detect, isolate, and
assess the changes in engine module performance, engine system malfunctions and in-
C. Daguang strumentation problems from knowledge of measured parameters taken along the engine’s
Beijing University of Aeronautics gas path. The method has been applied to a wide variety of commercial and military
and Astronautics, engines in the three decades since its inception as a diagnostic tool and has enjoyed a
Beijing, China reasonable degree of success. During that time many methodologies and implementations
of the basic concept have been investigated ranging from the statistically based methods
to those employing elements from the field of artificial intelligence. The two most publi-
cized methods involve the use of either Kalman filters or artificial neural networks (ANN)
as the primary vehicle for the fault isolation process. The present paper makes a com-
parison of these two techniques. 关DOI: 10.1115/1.1419016兴

Introduction
The goal of gas turbine performance diagnostics is to accurately
detect, isolate, and assess the changes in engine module perfor-
mance, engine system malfunctions, and instrumentation prob-
lems from knowledge of measured parameters taken along the
engine’s gas path. Discernable shifts in engine speeds, tempera-
tures, pressures, fuel flow, etc., provide the requisite information
for determining the underlying shift in engine operation from a
presumed nominal state. Historically, this type of analysis was
performed through the use of a Kalman filter or one of its deriva-
tives to simultaneously estimate a plurality of engine faults. In the
past decade, artificial neural networks 共ANN兲 have been em-
ployed as a pattern recognition device to accomplish the same
task. Both methods have enjoyed a reasonable success.
The purpose of this paper is to outline the two methodologies,
discuss their relative merits and weaknesses, and provide a direct
comparison of the two techniques via a controlled computer simu-
lation study. In the sequel, we will provide a brief general descrip-
tion of both the Kalman filter and ANN as applied to the diagnos-
tic problem. For the purpose of conducting a comparison, we will
limit the framework of the diagnostic system to the problem of
isolating a single fault to the component level. The single faults
under consideration will be comprised of engine module, engine
will briefly describe this procedure, however, for a more detailed
discussion we refer the reader to the following sources: 关4 –13兴.
The general approach taken for engine fault diagnostics typi-
cally involves the use of a linearized model approximation evalu-
ated at a selected engine operating point. This provides a matrix
relationship between changes in engine component performance
共independent parameters兲 and the attendant changes in typically
measured engine parameters such as spool speeds, internal tem-
peratures and pressures, fuel flow, etc. 共dependent parameters兲.
This relationship may be succinctly represented as
z⫽Hx⫹ ␪ (1)
where z is a vector of measured parameter deltas, x is a vector of
fault deltas, H is a matrix of fault influence coefficients, and ␪ is
a random vector representing the uncertainties inherent in the
measurement process. In addition to the precision of the indi-
vidual sensors, it has been customary to address the potential for
sensor bias and drift. Consequently, the fault vector given in the
model above is often configured to contain components directly
related to sensor error in addition to engine fault deltas.
The fault vector x given in the model can be thought of as the
concatenation of an engine fault vector (x e ) and a sensor 共error兲
fault vector x S , i.e., x⫽ 关 x e ]x s 兴 T where

冋 册 冋 册
system, and instrumentation faults.
⌬ ␩ FAN ⌬N1 err
Kalman Filter Approach ⌬⌫ FAN ⌬ P3 err
Kalman filter methods were introduced as a fault isolation and x e ⫽ ⌬ ␩ CH and x s ⫽ ⌬W f err . (2)
assessment technique for relative engine performance diagnostics ] ]
in the late 1970s and early 1980s 共关1–3兴兲. The success enjoyed in ⌬A 5 ⌬T49err
these early programs promoted the use of these techniques in sub-
sequent years to become the central methodology utilized in many We may rewrite Eq. 共1兲 as

冋册
current engine performance analysis programs. In the sequel we
xe
Contributed by the International Gas Turbine Institute 共IGTI兲 of THE AMERICAN z⫽H e x e ⫹H s x s ⫹ ␪ ⫽ 关 H e ] H s 兴 ¯ ⫹ ␪ ⫽Hx⫹ ␪ . (3)
SOCIETY OF MECHANICAL ENGINEERS for publication in the ASME JOURNAL OF xs
ENGINEERING FOR GAS TURBINES AND POWER. Paper presented at the Interna-
tional Gas Turbine and Aeroengine Congress and Exhibition, Munich, Germany, May
8 –11, 2000; Paper 2000-GT-0547. Manuscript received by IGTI November 1999;
The matrix H has been partitioned into two parts; a matrix of
final revision received by ASME Headquarters February 2000. Associate Editor: D. engine fault influence coefficients (H e ) and a matrix of sensor
Wisler. fault influence coefficients (H S ). The generation of these matrices

Journal of Engineering for Gas Turbines and Power OCTOBER 2003, Vol. 125 Õ 917
Copyright © 2003 by ASME
 Table 1 Sample root cause influences

T49C2 WF N2C2 N1C2 P25Q2 T25C2 T3C2 P3Q2

°C percent percent percent percent °C °C percent
FAN 3.86 0.70 0.30 ⫺0.68 ⫺2.00 ⫺1.95 ⫺1.58 ⫺0.03
LPC ⫺4.54 ⫺0.66 ⫺0.29 ⫺0.14 1.18 0.11 ⫺2.62 0.01
HPC ⫺6.80 ⫺0.80 0.06 ⫺0.05 ⫺0.83 ⫺0.71 ⫺3.66 0.17
HPT ⫺10.88 ⫺1.29 0.57 ⫺0.08 ⫺1.29 ⫺1.14 4.03 1.26
LPT ⫺1.19 0.96 ⫺0.63 0.98 3.40 3.45 ⫺1.42 0.11
2.5 BLD ⫺3.07 ⫺0.49 ⫺0.16 0.00 1.04 0.85 ⫺0.86 0.00
FP14 1.22 0.21 0.07 ⫺0.24 ⫺0.67 ⫺0.73 0.15 ⫺0.01
FP8 ⫺0.61 ⫺1.39 ⫺0.17 ⫺0.64 ⫺1.06 ⫺1.31 ⫺2.62 ⫺1.09
2.9 BLD ⫺4.22 ⫺1.06 ⫺0.29 ⫺0.06 0.68 0.63 0.60 0.02
TCC 17.75 2.10 ⫺0.90 0.12 2.14 1.86 ⫺4.38 ⫺1.11
HPCSVM ⫺0.95 ⫺0.11 0.39 0.00 ⫺0.08 ⫺0.09 0.34 0.09
P49 Error ⫺0.33 ⫺1.70 ⫺0.25 ⫺0.46 ⫺0.55 0.63 ⫺0.21 ⫺1.22

and their interpretation have been discussed in great detail in
关4,14兴 and elsewhere in the literature and will not be pursued in
this paper.
The model, as configured above, has often been used for the
purpose of tracking slowly occurring changes in engine perfor-
mance from revenue flight data, through the use of a Kalman
filter-based methodology. An estimate for these performance
shifts, x̂, would be given by
Single Fault Isolation. The Kalman filter can be configured
to emulate a single fault isolator (SFI). This is a snapshot type of
analysis in the sense that it operates on a set of measurement
deltas without any a priori information or pre-history. The object
of the analysis is to flag a root cause on the basis of a single
measurement delta set. Root causes are single fault occurrences
and are pre-defined for the system. They consist of coupled faults
within the major modules of the engine, certain system faults such

冋册 冋 册 冋 册冉 冋 册冊
as handling and ECS bleed leaks and failures, variable stator vane
x̂ e x̄ e De x̄ e malfunctions, TCC malfunctions as well as certain instrumenta-
¯ ⫽x̂⫽x̄⫹D 共 z⫺Hx̄ 兲 ⫽ ¯ ⫹ ¯ z⫺ 关 H e ]H s 兴 ¯ tion faults. These faults are assumed to occur in isolation, i.e.,
x̂ s x̄ s Ds x̄ s there will be one and only one root cause occurrence at any given
time. The purpose of the single fault isolator is to identify the
correct root cause once a trend shift is detected.
x̂⫽prediction⫹GAIN共residual兲⫽prediction⫹correction (4)
Root Causes. Root Causes can be thought of as state vari-
where x̄ represents an a priori estimate of the engine/sensor fault ables, x 1 ,x 2 , . . . ,x n . They are represented within this system as
deltas and D is the 共Kalman兲 gain matrix referred to as the diag- vectors of measurement deltas, z * 1 ,z *
2 , . . . ,z n* , which are calcu-
nostic matrix. The diagnostic matrix is computed as a function of lated by applying influence coefficients. As an illustration, we can
several quantities; the engine/sensor influence coefficients H, the consider the following 12 root causes. These root causes may exist
measurement covariance matrix R, and a positive semidefinite
at varying levels. The state representation of the root cause will be
weighting matrix P0 . The diagnostic matrix is computed as
in the form of a one percent cause to be consistent with other

冋册 冋 册
influences.
De P e0 H Te 共 H e P e0 H Te ⫹R 兲 ⫺1 1. FAN coupled FAN 共⫺1 percent ␩, ⫺1.25 percent FC兲
D⫽ ¯ ⫽ P 0 H 共 H P 0 H ⫹R 兲 ⫽
T T ⫺1 ¯¯ 2. LPC coupled LPC 共⫺1 percent ␩, ⫺1.10 percent FC兲
3. HPC coupled HPC 共⫺1 percent ␩, ⫺0.80 percent FC兲
Ds P s0 H sT 共 H s P s0 H sT ⫹R 兲 ⫺1 4. HPT coupled HPT 共⫺1 percent ␩, ⫹0.75 percent FP4兲
(5) 5. LPT coupled LPT 共⫺1 percent ␩, ⫹1.65 percent FP45兲
6. 2.5 BLD stability bleed leak 共one percent兲
7. 2.9 BLD start bleed 共one percent兲
where P (0e )andP 0(s )
are weighting submatrices for the engine and 8. FP14 fan discharge area 共one percent兲
sensor fault estimation, respectively. An in-depth report on the 9. FP8 core discharge area 共one percent兲
generation of the P 0 and R matrices can be found in references 10. TCC turbine case cooling 共on兲
11. HPCSVM HPC stator vane misrigging
关4,5兴. 12. P49 Error P49 indication problem 共two percent兲
The use of the predictor/corrector methods like the Kalman An example of the influences for these root causes is depicted in
filter to estimate sensor error has made possible a more reliable
Table 1.
and consistent gas turbine module performance analysis. The pro-
cedure outlined above can be applied in a snapshot analysis or as The actual root cause may appear as some multiple of the in-
a continuing recursive analysis as new engine data are made avail- fluences represented in Table 1. For instance, a 2.5 bleed root
able over time. In either scenario, the simultaneous determination cause may manifest itself as a stuck open bleed 共⬇15 percent 2.5
of both engine faults and measurement errors by this methodology bleed fault兲 or a partial bleed leak, say two percent 2.5 bleed fault.
has been successfully applied to a large number of commercial These two faults will be treated as the same by the Kalman esti-
and military families of engines with varying instrumentation mator. The difference between the two faults is one of magnitude.
suites for two decades. The magnitude is estimated by the Kalman filter. Its ability to
In the sequel, we shall alter our emphasis to consider the prob- estimate correctly will depend on the signal to noise ratio for the
lem of fault isolation, given the premise that a fault event has been given fault. In this particular case, a two percent 2.5 bleed is
detected. The problem of detection becomes one of recognizing a
sometimes confused with an LPC fault or a 2.9 Bleed fault. On the
step or rate change in a gas path parameter or a collection of
parameters. The problems associated with fault detection and the other hand, a stuck 2.5 Bleed 共14.77 percent兲 has a significantly
mechanisms which can be applied to accomplish this task have higher S/N ratio and is estimated correctly. Another factor which
been reported by 关15兴. The types of faults that will be considered impacts accuracy of the estimator is the number of measured pa-
in this discussion include engine performance faults, engine sys- rameters. We will consider systems that have between four and
tem faults, and instrumentation faults. eight measurements in flight.

918 Õ Vol. 125, OCTOBER 2003 Transactions of the ASME

 Given the definitions above for root cause influences 共one per- Table 2 Single fault problem set
cent兲, a typical set of possible single faults may be constructed.
Fault percent
For purposes of this study, we will consider the following single
fault definitions, depicted in Table 2. FAN ⫺2
If we represent these 11 faults as, x 1 ,x 2 , . . . ,x 11 , and the ma- LPC ⫺2
HPC ⫺2
trix of influences by H * ,(9⫻8) 1 matrix depicted above then a set HPT ⫺2
of expected measurement delta vectors, z * 1 ,z 2* , . . . ,z *
11 , would LPT ⫺2
be calculated by 2.5 BLD 共low and high兲 2 and 14.77
2.9 BLD 共low and high兲 6.74 and 15.45
z i* ⫽ 共 H * 兲 T x i for i⫽1,2, . . . ,11. (6) HPCSVM ⫺6
P49 Error 2
Each z i* would be an 8⫻1 vector of measurement deltas.
The general form of the discrete Kalman filter estimator at time
k⫹1 is as follows:
State extrapolation: As mentioned above, the normalized measurement error will
take into consideration the measurement nonrepeatabilities of the
x̂ 共 k⫹1 兩 k 兲 ⫽⌽ 共 k⫹1 兲 x̂ k . system. It is assumed that these are known a priori or computed
Covariance extrapolation: from data during initialization of the diagnostic system. However
the values are obtained, they are assumed to be known and are
P 共 k⫹1 兩 k 兲 ⫽⌽ 共 k⫹1 兲 P k ⌽ T 共 k⫹1 兲 ⫹Q k⫹1 . passed to the Kalman filter in the form of a positive definite ma-
trix R. The diagonal elements of this matrix represent the vari-
Kalman gain:
ances of the measurement deltas 共corrected quantities兲. Thus, if
D k⫹1 ⫽ P 共 k⫹1 兩 k 兲 H k⫹1
T
共 H k⫹1 P 共 k⫹1 兩 k 兲 H k⫹1
T
⫹R k⫹1 兲 ⫺1 . we represent the measurement delta-delta vector 共between time k
and k⫹1, assuming that a trend shift has been detected during this
State update: time period兲 by Z⫽ 关 z 1 ,z 2 , . . . ,z 8 兴 , then the diag(R)
x̂ k⫹1 ⫽x̂ 共 k⫹1 兩 k 兲 ⫹D k⫹1 关 z k⫹1 ⫺H k⫹1 x̂ 共 k⫹1 兩 k 兲兴 ⫽关␴21,␴22, . . . , ␴ 28 兴 , represents the individual variances. For defi-
niteness, assume we have calculated the nth single fault estimate.
Covariance update: The associated normalized measurement error norm, e n , is calcu-

冋 册
lated as follows:x̂ n ⫽SFI estimate for the nth root cause
P k⫹1 ⫽ 关 I⫺D k⫹1 H k⫹1 兴 P 共 k⫹1 兩 k 兲 .

冉 冊
8 1/2
z k ⫺z k* 2
For the single fault isolator we make the following adjustments:
兺 ␴k
⌽ k⫹1 ⫽I e n⫽
k⫽1
(7)
8
Q k⫹1 ⫽0 兺
k⫽1
z 2k
H k⫹1 ⫽M c 共 H * 兲 T
k ⫽kth element of the vector M c (H * ) x̂ n .
T
z*
M c ⫽measurement configuration vector⫽diag共 m 1 ,m 2 , . . . ,m 8 兲 It is possible for the estimated values for some single faults to
be opposite in polarity than what would be reasonably expected.2
共diagonal matrix assuming eight potential measurements兲 For example, an HPT SF of magnitude ⫺2 percent might yield an

where m j ⫽ 再 1
0
if jth measurement is available
otherwise.
The effect of the measurement configuration matrix is to zero
the rows of the root cause influence matrix corresponding to the
measurements that are NOT available. The resulting matrix of
influences is 8⫻9 for this example.
The single fault isolation is obtained by processing the general
Kalman equations iteratively to provide a snapshot analysis for
each of the root causes under consideration 共11 in this example兲.
Each ‘‘call’’ to the Kalman filter will be made with a different P 0
matrix chosen to accentuate the kth root cause. Since these are
snapshot analyses, the covariance update calculation is not re-
quired. The a priori state estimate is also assumed to be zero. An
SFI analysis is performed typically after a trend shift has been
detected in the measurement deltas at some discrete time, say,
between time k and k⫹1. The ‘‘delta-delta’’ (Z k⫹1 ⫺Z k ) consti-
tutes the input measurement delta, Z, to the SFI.
The SFI is evaluated iteratively for each single fault 共11 in the
above example兲. This process will yield estimates for each single
fault under consideration. It is necessary to rank each of these
estimates and determine the top two or three single faults. The
measure used to compare estimates for this purpose of ranking is
a normalized measurement error norm which we will describe
below. The single fault admitting the minimum error is deemed
the most likely, the second smallest error, the next likely, and so
forth.
1
The row dimension of H * is 9 共and not 11兲 since the BLD25 and BLD29 root
causes are used to model the two different magnitude faults being considered.
estimated value of ⫹0.5 percent for an LPC SFI. Given that a
sudden shift in observed gas path parameters has taken place dur-
ing engine operation, it is not likely that the condition of any
given module has improved and thus a positive shift in perfor-
mance would not be given serious consideration. Thus, some pre-
processing is mandated before the error ranking is performed.
This takes the form of perusing each of the SFI estimates and
considering only those which admit a reasonable polarity. Order-
ing the errors from min to max we obtain
e i 1 ⭐e i 2 ⭐¯⭐e i n (8)
corresponding to the single fault estimates
x̂ i 1 ⭐x̂ i 2 ⭐¯⭐x̂ i n (9)
in order of likelihood. In most cases, the first of these ranked
single faults is deemed to be the underlying fault and is reported
to the user. In some instances, however, it may be prudent to
report the first and second SFs, since there exists the possibility
for an erroneous fault identification, especially if the associated
error norms are ‘‘close.’’
Depending on the circumstances, the SFI may experience a
confusion between two single faults. The aliasing of SFs depends
on the single faults themselves, their relative magnitude with re-
spect to the nonrepeatability of the measurements and the number
of measurements available for the analysis. All of these factors
can contribute to a confounding of the underlying SF with another
2
Although this is not a common occurrence, it is possible for an SF estimate
indicating an improvement in performance to admit the minimum measurement error
norm and hence win the isolation selection.

Journal of Engineering for Gas Turbines and Power OCTOBER 2003, Vol. 125 Õ 919
 Table 3 SFI accuracy: Kalman filter, 8 measurements entries indicate the percentage of time the implanted fault is cor-
rectly annunciated as first, second, and third choice.
8 Measurement Set
The results demonstrate a 96.9 percent and 91.1 percent accu-
FAULT 1st Top 2 Top 3 racy for the first choice fault isolation for eight and four measure-
ment systems, respectively. If we weaken the accuracy criteria to
FAN 100% 100% 100%
LPC 90% 100% 100% that of being correct within the top two choices, the relative pre-
HPC 100% 100% 100% cision of the isolation increases to 100 percent and 96.1 percent.
HPT 100% 100% 100% In either case, the isolation accuracy is quite good.
LPT 100% 100% 100% The fact that we can obtain a 91 percent hit rate with only four
2.5 BLD 85% 100% 100%
2.9 BLD 96.7% 100% 100% measurements raises the question as to why one would ever con-
HPCSVM 100% 100% 100% sider adding more instrumentation. There are several answers to
P49err 100% 100% 100% this question, one being that a multiple fault performance assess-
Total 96.9% 100% 100% ment would not perform as well with fewer measurements. But
even within the constraints of this study 共i.e., single fault annun-
ciation兲, the ‘‘more the better’’ argument still holds if we consider
Table 4 SFI accuracy: Kalman filter, 4 measurements the robustness of the system. One measure of robustness would be
to vary the module couplings 共flow capacity versus efficiency,
4 Measurement Set hard coupled relationship兲 in the implanted faults beyond that
FAULT 1st Top 2 Top 3 which is assumed in the numerical model 共i.e., the influences co-
efficients H * ). Figures 1 and 2 depict the impact on first choice
FAN 100% 100% 100% accuracy for eight and four measurement systems, respectively,
LPC 90% 90% 100%
HPC 100% 100% 100% when the coupling factor for the compression modules 共FAN,
HPT 100% 100% 100% LPC, HPC兲 are randomly varied.4
LPT 100% 100% 100% The LPC and HPC clearly exhibit greater robustness to model-
2.5 BLD 50% 75% 95% ing assumptions in the 8 measurement system.
2.9 BLD 80.0% 100.0% 100%
HPCSVM 100% 100% 100% In the sequel, we will briefly discuss artificial neural networks
P49err 100% 100% 100% 共ANNs兲, and their usage in engine performance diagnostics. The
Total 91.1% 96.1% 99.4% discussion will be intentionally brief since there already exists
ample documentation of this methodology in the literature, and
the reader is directed to the references for additional detail. For
the purpose of making a direct comparison with the Kalman
SF. To mitigate the possibility of a false identification, some post- filter methodology, we will confine much of the discussion to the
processing of the SFI results may be required. This will improve single fault isolation problem introduced in the prelude. The iden-
the overall accuracy of the system. The rules to be applied would tical computer simulation test will be used as the vehicle for the
be empirically motivated and would be suggested by computer evaluation.
simulation test cases. For the problem set considered in this study
共Table 2兲, it was not necessary to undergo this type of post- Artificial Neural Network Approach
processing.
The fault isolation problem can be considered to be a pattern
Computer Simulation Results. Using Equation 共1兲, we can classification problem. N-dimensional vectors in an
generate a set of hypothetical ‘‘noisy’’ measurements for each of N-dimensional space represent the system response. The system
the 11 single faults under consideration (x i ,i⫽1,2, . . . 11). These response for different faults tends to be partitioned into different
noisy measurement vectors (z k ) are then passed through the SFI regions of this space and can be regarded as patterns. Pattern
process, the results ranked and tabulated for first, second, and recognition involves learning these partitions, from simulated or
third ranked faults. These results appear in Table 3 and 4 for real data, so that a given system response can be classified as a
eightand four measurement set3 systems, respectively, where the particular fault. Neural networks represent a powerful pattern rec-
four bleed leak faults have been combined into two faults. The ognition technique, and have been applied for fault detection of

3
4
The turbine modules’ 共HPT and LPT兲 isolation accuracy was essentially unaf-
The four measurements set used in this study consisted of the flight required fected and hence were not plotted.
parameters (T49, Wf, N1 and N2).

Fig. 1 Coupling factor impact on 8 measurement system

920 Õ Vol. 125, OCTOBER 2003 Transactions of the ASME

 Fig. 2 Coupling factor impact on 4 measurement system

complex systems such as aerospace vehicles 共关16兴兲, nuclear power
plants 共关17兴兲, and chemical process facilities 共关18兴兲 among others.
A key advantage of neural networks over other methods is their
ability to recognize relationships between patterns despite the
presence of noise contamination and/or partial information 共关19兴兲.
Most applications of ANN to fault diagnostics follow a com-
mon process. The ANN is trained off-line on fault signatures re-
lating the changes in system measurements from a ‘‘good’’ base-
line to system faults. Typically, faults are embedded into the
computer simulations, or real fault data is used, or a combination
of both. In case simulated data is used, noise must be added to
make the simulations realistic. Such a training process where the
ANN is presented input and output data by the system designer is
known as supervised learning. Once the ANN has been properly
trained using this process of supervised learning, it can analyze
data that are different from those it was originally exposed to
during the training sessions. When the trained ANN is placed
on-line, it recognizes a similar response from the actual system.
The discussion below involves the use of two types of neural
networks for engine fault diagnostics, a feed forward ANN trained
using back-propagation 共BP兲 algorithm and a hybrid neural
network.
Back-Propagation „BP… Algorithm. While there are several
types of neural networks, multilayer feedforward networks trained
using the backpropagation algorithm have emerged as the most
widely used. Figure 3 illustrates the schematic of a feedforward
neural network, which consists of an input layer, an output layer,
and one or more hidden layers. The number of neurons in the
input and output layers is determined by the number of input
measurements and output parameters. The number of hidden layer
nodes is selected based on convergence criterion and the charac-
teristic of the input-output mapping relationship.
A three-layer feedforward network is used for the present work,
as shown in Fig. 1. The feedforward network is trained using
supervised learning, which involves presenting input-output pairs
to the ANN and then using the BP algorithm to learn the relation-
ships between the inputs and outputs by minimizing the following
error measure:
N

E⫽ 兺E
k⫽1
k (10)

in which E k represents the root mean square error associated with

the kth training sample and where N represents the number of
samples that are used for training the network. The BP algorithm
uses a gradient search to perform the nonlinear optimization
needed to minimize the error. Further details about the BP algo-
rithm are outlined in standard texts on neural networks 共关20兴兲.
To improve the ANN’s ability to deal with data scatter, the input
data were normalized using the following formula:

Table 5 SFI accuracy: ANN, 8 measurements

8 Measurement Set
FAULT 1st Top 2 Top 3
FAN 100% 100% 100%
LPC 60% 80% 90%
HPC 100% 100% 100%
HPT 100% 100% 100%
LPT 100% 100% 100%
2.5 BLD 80% 85% 85%
2.9 BLD 86.7% 86.7% 86.7%
HPCSVM 90% 100% 100%
P49err 100% 100% 100%
Total 90.7% 94.6% 95.7%
Fig. 3 Architecture of three layer feedforward ANN

Journal of Engineering for Gas Turbines and Power OCTOBER 2003, Vol. 125 Õ 921
 Table 6 SFI accuracy: ANN, 4 measurements Table 8 SFI accuracy: hybrid NN, 4 measurements

4 Measurement Set 4 Measurement Set

FAULT 1st Top 2 Top 3 FAULT 1st Top 2 Top 3
FAN 100% 100% 100% FAN 90% 100% 100%
LPC 90% 90% 90% LPC 90% 90% 100%
HPC 90% 100% 100% HPC 100% 100% 100%
HPT 100% 100% 100% HPT 100% 100% 100%
LPT 100% 100% 100% LPT 100% 100% 100%
2.5 BLD 75% 75% 75% 2.5 BLD 70% 90% 100%
2.9 BLD 86.7% 86.7% 86.7% 2.9 BLD 86.7% 86.7% 86.7%
HPCSVM 100% 100% 100% HPCSVM 70% 100% 100%
P49err 100% 100% 100% P49err 100% 100% 100%
Total 93.5% 94.6% 94.6% Total 91.1% 97.8% 100%

Y in ⫽Y im / 共 Y i max• ␴ i 兲 training data based on influence coefficients, the Hybrid ANN

uses the influence coefficients as part of the network model.
where Y i is the ith monitoring parameter; n, m, and max are the A Gaussian nearest neighbor function was substituted for the
normalized, measured, and maximum possible value, respectively; ANN feature identification function and the fault mapping func-
␴ i is the standard deviation of the ith monitoring parameter. tion was ignored for this Hybrid NN algorithm. The output of the
The standard deviations and the influence coefficients used for network was the root-sum-of-the-squares number of standard de-
the ANN testing are the same as those used for the Kalman filter viations from a perfect match of the fault pattern. The output of
described previously. For comparison of fault isolation results the network was used directly to rank the faults.
with four inputs and eight inputs, 20 training cases and 50 testing A test was made to determine if the neural network mapping to
cases were generated. The training cases were used for the BP the faults could also be optimized. Up to 72 weightings were
algorithm to train the neural network. Once the neural network available between the eight measurement features and the nine
was trained, the test cases were used to evaluate the performance faults. Thirty-six weightings were available with four measure-
of the neural network. Data used for training was not used for ments. While it was demonstrated that the weightings could be
testing the neural network. The diagnostic results were considered optimized, the tests were run with all the weightings set to unity.
for the three highest outputs, which represent the three most likely The results from the Hybrid neural net are shown in Tables 7
faults. and 8, for eight and four measurements, respectively. The accu-
The test results for eight and four measurements using the BP racy of the Hybrid neural network compares favorably with the
ANN are shown in Table 5 and Table 6, respectively. For both the Kalman SFI. For the eight measurements case, both the Kalman
eight and four measurement cases, the Kalman SFI results shown SFI and the Hybrid ANN show a fault isolation accuracy among
in Tables 5 and 6 are better than the BP ANN results. For the eight the top three choices of 100 percent. For the four measurement
measurements case, the Kalman SFI has a 100 percent accuracy in case, the Kalman SFI shows a fault isolation accuracy among the
fault isolation among the top three choices, compared to 95.7 top three choices of 99.4 percent compared to 100 percent for the
percent for the BP ANN. For the four measurements case, the Hybrid ANN.
Kalman SFI has an accuracy of 99.4 percent in fault isolation In general, the Hybrid ANN gives better results than the BP
among the top three choices, compared to 94.6 percent for the ANN. This may be because the hybrid uses the influence coeffi-
BP ANN. cients for each fault within the network, whereas the BP ANN has
Hybrid Neural Network Algorithm. The BP ANN is handi- to learn the influence coefficients from the training data.
capped relative to the Kalman SFI in some ways. For example,
while the Kalman SFI uses influence coefficients in the form of
the H matrix to define the model, the BP ANN uses data generated Kalman Filter and Neural Network Methods
from influence coefficients to ‘‘learn’’ the model. Feedforward neural networks are typically made of intercon-
A Hybrid NN is a network architecture where one or more nected nonlinear neurons and are particularly useful where the
ANN functions are replaced by an algorithm that includes domain input-output relationship is nonlinear. The Kalman filter and the
knowledge 共关21兴兲. The objective is to substitute features in the Hybrid neural network assume a linearized model of the system.
neural network architecture that are already analytically under- In the present application, the linear model takes the form of
stood. This avoids the need for training the ANN to learn infor- influence coefficients. For a feedforward ANN, BP learning for
mation which is already known. For example, instead of using linear problems in engine fault diagnostics will generally result in
poorer performance compared to a Kalman filter or hybrid ANN
approach.
Table 7 SFI accuracy: hybrid NN, 8 measurements Learning in neural networks involves mapping an input to an
output. The input and outputs can be generated from models, from
8 Measurement Set real data, or a combination of both. ANNs can therefore also be
model-free estimators, a quality that is very useful if modeling
FAULT 1st Top 2 Top 3
information such as influence coefficients are not available. Kal-
FAN 100% 100% 100% man filters, on the other hand, are model-based estimators, and are
LPC 90% 100% 100% suitable for problems in engine performance diagnostics where
HPC 100% 100% 100%
HPT 100% 100% 100% influence coefficients are available as the model.
LPT 100% 100% 100% Neural networks are not limited to multilayer neurons with BP
2.5 BLD 90% 100% 100% training. Self-organizing maps based on competitive learning
2.9 BLD 100% 100% 100% 共关22兴兲, simulated annealing based on statistical thermodynamics
HPCSVM 100% 100% 100%
P49err 100% 100% 100% 共关23兴兲, Boltzmann learning 共关24兴兲 and radial basis functions 共关25兴兲
Total 97.8% 100% 100% are some of the other developments in ANN which may be appli-
cable to engine diagnostics. Combining the Kalman filter ap-

922 Õ Vol. 125, OCTOBER 2003 Transactions of the ASME

proach with some of the ANN methods may yield superior results P3 ⫽ HPC exit pressure
for the engine diagnostic problem, than those obtainable within P49 ⫽ exhaust pressure
either methodology acting alone. HPCSVM ⫽ variable stator vane
SFI ⫽ single fault isolator
Concluding Remarks H ⫽ influence coefficient matrix
Test results have suggested that the back-propagation neural D ⫽ Kalman gain matrix
network, the hybrid neural network, and the Kalman filter method R ⫽ measurement covariance matrix
are highly accurate for isolating single gas turbine fault symp- P ⫽ state covariance matrix
toms. Furthermore, the results also indicate that these methodolo- ⌽ ⫽ state transition matrix
gies compare favorably in terms of accuracy, with a very slight x ⫽ state ⌬ vector
advantage going to the Kalman filter approach. z ⫽ measurement ⌬ vector
Each method has its own advantages and disadvantages. ␩ ⫽ efficiency ⌬
The ANN is inherently nonlinear and can be used in applica-
⌫ ⫽ flow capacity ⌬
tions where model information is scarce or lacking altogether. The
ANNs are, however, data driven and therefore must be trained. FC ⫽ flow capacity
The training is typically performed offline in a supervised fashion, A ⫽ turbine nozzle area ⌬
meaning that the input-output relationship is known. In the present Subscripts
application this means that the underlying fault共s兲 in the training
data are already known. This could be a drawback if real engine e ⫽ relating to engine parameters
data was used for training, since the precise disposition of the s ⫽ relating to sensor parameters
fault may or may not be known. If the engine configuration and/or 0 ⫽ initial condition
instrumentation noise levels change, the ANN approach would k ⫽ discrete time k
require that a re-training be performed. Once trained, the ANN Superscripts
architecture provides a numerically simple, and hence fast diag-
nostic operator suitable for real-time application. T ⫽ matrix transpose
The Kalman filter is a linear model-based estimator and is suit- ⫺1 ⫽ matrix inverse
able in those instances where a linear model is available and is
known to be a reasonably accurate representation of the input-
output relationship. In the engine performance diagnostics appli-
cation, influence coefficients have, historically, been fairly accu-
rate and robust linear models. The Kalman filter approach utilizes
References
all model information available 共a priori estimate information, 关1兴 Urban, L. A., 1974, ‘‘Parameter Selection for Multiple Fault Diagnostics of
measurement noise information, etc.兲 and can be easily configured Gas Turbine Engines,’’ AGARD Conference Proceedings, No. 165, NATO,
Neuilly-sur-Seine, France.
to operate with different measurement suites, and fault configura-
关2兴 Urban, L. A., 1972, ‘‘Gas Path Analysis Applied to Turbine Engine Condition-
tions: single fault 共as in this present study兲 or multiple fault iso- ing Monitoring,’’ AIAA/SAE Paper 72-1082.
lation systems. Adaptive measures are also available to allow real- 关3兴 Volponi, A., 1983, ‘‘Gas Path Analysis: An Approach to Engine Diagnostics,’’
time reconfiguration of the Kalman filter to changing Time-Dependent Failure Mechanisms and Assessment Methodologies, Cam-
measurement noise levels. bridge University Press, Cambridge, UK.
The hybrid neural network, like the Kalman filter, is a model- 关4兴 Volponi, A. J., and Urban, L. A., 1992, ‘‘Mathematical Methods of Relative
based method utilizing influence coefficients as the primary linear Engine Performance Diagnostics,’’ SAE Trans., 101; Journal of Aerospace,
Technical Paper 922048.
model in a neural network architecture framework. For the single-
关5兴 Doel, D. L., 1992, ‘‘TEMPER—A Gas Path Analysis Tool for Commercial Jet
fault isolation problem referenced in this study, the hybrid NN Engines,’’ ASME Paper 92-GT-315.
closely resembles a weighted least squares solution. This explains 关6兴 Doel, D. L., 1993, ‘‘An Assessment of Weighted-Least-Squares Based Gas
the close agreement between the hybrid NN and Kalman filter in Path Analysis,’’ ASME Paper 93-GT-119.
the single-fault isolation computer simulation study. 关7兴 Stamatis, A. et al., 1991, ‘‘Jet Engine Fault Detection With Discrete Operating
Points Gas Path Analysis,’’ J. Propul. Power, 7, No. 6.
关8兴 Merrington, G. L., 1993, ‘‘Fault Diagnosis in Gas Turbines Using a Model
Acknowledgment Based Technique,’’ ASME Paper 93-GT-13.
The authors wish to thank Zhang Jin, Zhang Mingchuan, and 关9兴 Glenny, D. E., 1988, ‘‘Gas Path Analysis and Engine Performance Monitoring
In A Chinook Helicopter,’’ AGARD: Engine Condition Monitoring, NATO,
Zhu Zhili of the Department of Jet Propulsion and Power at the Neuilly-sur-Seine, France.
Beijing University of Aeronautics and Astronautics as well as Lu 关10兴 Winston, H., et al., 1991, ‘‘Integrating Numeric and Symbolic Processing For
Pong-jeu and Hsu Tzu-cheng of the Institute of Aeronautics and Gas Path Maintenance,’’ AIAA Paper 91-0501.
Astronautics, National Cheng Kung University in Taiwan for their 关11兴 Luppold, R. H., et al., 1989, ‘‘Estimating In-Flight Engine performance Varia-
work in developing the ANN Fault Analysis Application under tions Using Kalman Filter Concepts,’’ AIAA Paper 89-2584.
contract with Pratt & Whitney. 关12兴 Gallops, G. W., et al., 1992, ‘‘In-Flight Performance Diagnostic Capability Of
An Adaptive Engine Model,’’ AIAA Paper 92-3746.
关13兴 Kerr, L. J., et al., 1991, ‘‘Real-Time Estimation of Gas Turbine Engine Dam-
Nomenclature age Using a Control Based Kalman Filter Algorithm,’’ ASME Paper 91-GT-
ANN ⫽ artificial neural network 216.
关14兴 ‘‘Sensor Error Compensation in Engine Performance Diagnostics,’’ 1994,
FAN ⫽ fan module ASME Paper No. 94-GT-058.
LPC ⫽ low pressure compressor 关15兴 DePold, H., and Gass, F. D., 1998, ‘‘The Application of Expert Systems and
HPC ⫽ high pressure compressor Neural Networks to Gas Turbine Prognostics and Diagnostics,’’ ASME Paper
HPT ⫽ high pressure turbine 98-GT-101.
LPT ⫽ low pressure turbine 关16兴 McDuff, R. J., and Simpson, P. K., 1990, ‘‘An Investigation of Neural Network
2.5 BLD ⫽ stability bleed for F-16 Fault Diagnosis,’’ Proceedings of the SPIE Technical Symposium on
Intelligent Information Systems, The International Society for Optical Engi-
2.9 BLD ⫽ start bleed
neering, Bellingham, WA.
TCC ⫽ turbine case cooling 关17兴 Guo, Z., and Uhhrig, R. E., 1992, ‘‘Using Modular Neural Networks to Moni-
N1 ⫽ low spool speed tor Accident Conditions in Nuclear Power Plants,’’ Proceedings of the SPIE
N2 ⫽ high spool speed Technical Symposium on Intelligent Information Systems, The International
T49 ⫽ exhaust gas temperature Society for Optical Engineering, Bellingham, WA.

Journal of Engineering for Gas Turbines and Power OCTOBER 2003, Vol. 125 Õ 923
 关18兴 Wantanabe, K. et al., 1989, ‘‘Incipient Fault Diagnosis of Chemical Processes
via Artificial Neural Network,’’ AIChE J., 35, No. 11.
关19兴 Holmstrom, L., and Koistinen, P., 1992, ‘‘Using Additive Noise in Back Propa-
gation Training,’’ IEEE Trans. Neural Netw., 3, No. 1.
关20兴 Haykin, S., 1994, Neural Networks—A Comprehensive Foundation, Macmil-
lian, New York.
关21兴 Dasarathy, B. V., 1991, Nearest Neighbor (NN) Norms: NN Pattern Classifi-
cation Techniques, IEEE Computer Society Press, New York.
关22兴 Kohonen, T., 1990, ‘‘The Self-Organizing Map,’’ Proc. IEEE, 78, IEEE, New
York.
924 Õ Vol. 125, OCTOBER 2003
View publication stats
关23兴 Kirkpatrick, S., 1984, ‘‘Optimization by Simulated Annealing: Quantitative
Studies,’’ J. Stat. Phys., 34.
关24兴 Ackley, D. H., Hinton, G. E., and Sejnowski, T. J., 1985, ‘‘A Learning Algo-
rithm for Boltzmann Machines,’’ Cognitive Science, Vol. 9, Cognitive Science
Society, Cincinnati, OH.
关25兴 Broomhead, D. S., and Lowe, D., 1988, ‘‘Multivariate Function Interpolation
and Adaptive Networks,’’ Complex Systems, Vol. 2, Complex Systems Publi-
cations, Champaign, IL.
关26兴 Cybenko, G., 1989, ‘‘Approximations of Superposition of a Sigmoidal Func-
tion,’’ Mathematics of Control, Signals and Systems, Vol. 2.
Transactions of the ASME