Measurement Selections for

Multicomponent Gas Path

Diagnostics Using Analytical
Approach and Measurement
Mohd Shahrizal Jasmani Subset Concept

Downloaded from http://asmedigitalcollection.asme.org/gasturbinespower/article-pdf/133/11/111701/5685391/111701_1.pdf by Universiti Teknologi Petronas user on 05 July 2024

Maintenance Engineering Department,
PETRONAS Carigali Sdn. Bhd.,
24300 Kerteh, Terengganu, Malaysia Gas path analysis (GPA) is a powerful tool to predict gas turbine degradations based on
measurement parameters of gas turbine engines. Accordingly, prudent measurement se-
Yi-Guang Li lections are crucial to ensure accurate GPA predictions. This paper is intended to inves-
Department of Power and Propulsion, tigate the influence of measurement parameter selection toward the effectiveness of GPA
Cranfield University, algorithm. An analytical methodology for measurement selection, combined with mea-
Cranfield, Bedfordshire, MK43 0AL UK surement subset concept, is developed to properly select measurements for multiple com-
ponent fault diagnosis. The effectiveness of GPA using the measurement sets selected with
Zaharudin Ariffin the introduced measurement selection method are then compared with the results of using
Group Technology Solutions, standard measurements installed on existing gas turbine engines. A case study applying
PETRONAS, the new measurement selection method to GPA diagnostic analysis is demonstrated on a
50090 Kuala Lumpur, Malaysia three-shaft aeroderivative industrial gas turbine model based on similar unit installed
onboard an offshore platform operated by PETRONAS. The engine is modeled and simu-
lated using PYTHIA, a gas turbine performance and diagnostics analysis tool developed by
Cranfield University. To validate the findings, nonlinear GPA prediction errors are evalu-
ated in various cases of single and multicomponents faults. As a result, the selected
measurements have successfully produced much superior diagnostics accuracies in the
fault cases when compared with the standard measurements. These findings proved that
proper measurement selection for better GPA diagnostic analysis can be achieved by
using the proposed analytical methods. Several engine sensor enhancements are also
discussed to accommodate the unique sensor requirements for health diagnostics using
GPA. 关DOI: 10.1115/1.4002348兴

1 Introduction measurement parameters, as these parameters provide crucial in-

formation for accurate diagnostic results. Traditionally, a produc-
Components of a gas turbine engine may degrade over time
tion gas turbine engine is equipped with limited sensors to provide
during its operation. Hence, gas turbine health monitoring is es-
measurements for the purpose of engine control and protection as
sential for reliable, safe, and cost effective operation of gas tur- discussed by Mathioudakis et al. 关5兴. However, effective diagnos-
bine engines and may be achieved by analyzing the deviations in tic system may require different quantity and quality of measure-
gas path measurements, such as pressures and temperatures with ments, which must be carefully selected to ensure reliable diag-
the use of appropriate diagnostic algorithm. Different gas turbine nostic results. Hence, the main goal of GPA measurement
diagnostic methods have been developed over the years to accu- selection process is to determine the engine measurement sets that
rately quantify the degradations. One of the earliest methods is by produce high accuracy in predictions of gas turbine gas path com-
using gas path analysis 共GPA兲, which was introduced by Urban ponent degradation.
关1兴 in 1969. GPA is able to predict gas turbine health parameters The number of possible gas path measurements on a gas turbine
via linear performance models and the method has been evolved may be quite extensive. Rudimentarily, the desired combinations
into a powerful tool for gas turbine health diagnostics as described can be determined using “brute-force” selection approach de-
by Urban 关2兴 and Volponi 关3兴. scribed by Mushini and Simon 关6兴. In this method, all possible
The prediction accuracy of gas path diagnostics had also been combinations are sought and evaluated using “n-choose-k” com-
improved over the years with the advent of better computational binatorial approach, i.e., to choose k number of possible combi-
facilities. This provides opportunities for more complex and pre- nations from n number of total measurements available:
cise diagnostic algorithms such as genetic algorithm 共GA兲, artifi-
cial neural network 共ANN兲, and fuzzy logic techniques as thor-
oughly reviewed by Li 关4兴. The implementation of predictive and
prognostic maintenance philosophies in the industries has also
冉冊
n
k
=
n!
k!共n − k兲!
共1兲

emphasized the importance of gas path diagnostics toward equip- Based on Eq. 共1兲, the number of candidates for feasible sensor
ment availability and reliability. placement increased exponentially with higher number of diag-
One of the essential aspects in GPA is proper selections of nosable components, especially in gas turbines with multishaft
configurations. Therefore, the number of possible measurement
sets is very large and would not be feasible for individual assess-
Contributed by International Gas Turbine Institute 共IGTI兲 of ASME for publica-
tion in the JOURNAL OF ENGINEERING FOR GAS TURBINES AND POWER. Manuscript re-
ment. To highlight the importance of the selection, Ogaji et al. 关7兴
ceived April 28, 2010; final manuscript received August 5, 2010; published online demonstrated that different selection of measurements would
May 13, 2011. Editor: Dilip R. Ballal. make big differences to the diagnostic results.

Journal of Engineering for Gas Turbines and Power NOVEMBER 2011, Vol. 133 / 111701-1
Copyright © 2011 by ASME
 To avoid the strenuous combinatorial selection process, optimal
selection techniques are employed by using heuristic methods
such as GA 关8兴. Sowers et al. 关9兴 used optimal selection to opti-
mize the figure of merit 共FOM兲 as the objective function in their
selection of measurement sets for turbofan engine diagnostics.
However, these optimal selection techniques require complex it-
erative scheme and still computationally demanding to seek for
the optimum solution.
One of the alternatives is to qualitatively tackle the problem
using analytical approach. The optimal measurement set can ac-
tually be selected by observing the behaviors and influences of the
measurements and speculating their effects toward diagnostic re-
sults. Analytical studies were conducted using sensitivity analysis

Downloaded from http://asmedigitalcollection.asme.org/gasturbinespower/article-pdf/133/11/111701/5685391/111701_1.pdf by Universiti Teknologi Petronas user on 05 July 2024

by Stamatis et al. 关10兴 who also demonstrated the use of singular
value decomposition 共SVD兲 to analyze the influence matrices.
Provost 关11兴 used correlations and observability analysis to select
appropriate measurement set for turbofan engine diagnostics. The
effects of interferences caused by the measurements toward GPA
matrix are also being evaluated using matrix condition number, as Fig. 1 Measurement selection methodology
demonstrated by Cotae et al. 关12兴, Borguet and Leonard 关13兴, and
Mathioudakis and Kamboukos 关14兴 in their measurement selec-
tion processes.
In this study, a new analytical measurement selection method 3 Measurement Selection
based on sensitivity and correlation analysis for GPA is introduced To effectively select the gas path measurements for diagnostic
with the introduction of measurement subset concept for multi- analysis, several measurement selection criteria are hereby intro-
component fault diagnosis. The method is tested in a case study duced and described in detail in this paper. The general overview
and the advantages and disadvantages of the method are dis- of the measurement selection process is shown in Fig. 1.
cussed. Prior to measurement selection process, an accurate model of
gas turbines have to be available and able to replicate the design
2 Gas Path Analysis and off-design performance of actual engines. All possible gas
Prior to elaboration on the measurement selection methodology, path measurement parameters at inlet and outlet of each compo-
the basic principles of GPA based on Li and Singh 关15兴, are con- nent can be included for consideration. However, as the number of
veniently presented in this paper to supplement further discus- possible engine measurements may be large, they need to be
sions. In GPA, it is assumed that the linear relationship between clearly classified in accordance to sensor types, locations and
gas turbine health parameters and gas path measurements for a gas characteristics. This will effectively prescreen the initial measure-
turbine engine is expressed as ments and reduce the number of potential candidates for further
analysis. The classification can also help to differentiate between
⌬z៝ = H · ⌬x៝ 共2兲 the existing and potential measurements as well as identifying
where ⌬z៝ 苸 R is the deviation of M number of gas path mea-
M advanced sensor technologies, which may or may not be consid-
surement parameters, ⌬x៝ 苸 RN is the deviation of N number of gas ered in gas path diagnostics. The measurement parameters may be
turbine component health parameters, and H 苸 R M⫻N is the influ- classified into the following categories:
ence coefficient matrix 共ICM兲 that represents the linear functional i. ambient and operating conditions
relationship between ⌬z៝ and ⌬x៝ . Component health parameters ii. derived
are the indicators or indices that can describe the conditions of a iii. advanced
gas turbine component, for example, efficiency index and flow iv. standard
capacity index that describe the shift of component characteristic v. provisional
map due to degradation 关16兴.
Mathematically, it is required that M ⱖ N in order for ⌬x៝ to be 3.1 Ambient and Operating Conditions. These measure-
uniquely estimated 关17兴. When M = N and square matrix H is in- ments, such as ambient pressure, ambient temperature, and engine
vertible and H−1 is referred as fault coefficient matrix 共FCM兲. In handle parameter, do not change with component degradations.
this case, the degradations can be predicted via They can be used for data preprocessing so that all measurement
data are corrected to the same ambient and operating conditions.
⌬x៝ = H−1 · ⌬z៝ 共3兲 As these parameters have no direct relationship with degradable
When M ⬎ N, Eq. 共3兲 is overdetermined and a pseudo-inverse of components they have no unique signature to any component deg-
H 共represented with H#兲 can be used to replace H−1 in Eq. 共3兲 radation. Therefore, these measurement parameters are essential
using least square estimation method, where but can be excluded from the measurement selection process.
H# = 共HT · H兲−1 · HT 共4兲 3.2 Derived Parameters. There are engine parameters that
could not be physically measured but can be derived from other
To determine the GPA accuracy, the errors between prediction and directly measured parameters. These are usually performance re-
actual measurement parameters can be evaluated with root mean lated parameters, such as specific fuel consumption 共SFC兲, ther-
square 共RMS兲 method

冑 兺冉
mal efficiency, etc., and should be deselected from the list of

冊
M 2 potential measurements.
1 zi,predicted − zi,measured
RMS = · 共5兲 3.3 Advanced Measurements. Measurement parameters
N i=1 zi,measured
such as turbine entry temperature or intake air flow rate are cur-
To take into account the nonlinearity of gas turbine performance rently impractical to be measured accurately due to technology
behavior and to improve the accuracy of diagnostic predictions, limitations. However, if a particular study is to evaluate the fea-
Newton–Rhapson approach using iterative application of the lin- sibility of installing revolutionary sensor technologies, measure-
ear GPA above is applied as described in Ref. 关15兴. ments of this type can be included as provisional candidates.

111701-2 / Vol. 133, NOVEMBER 2011 Transactions of the ASME

 3.4 Standard Measurements. Measurement parameters, each sensitivity coefficient in Eq. 共6兲 is divided by the calculated
which are required for gas turbine performance controls and sensitivity norm in Eq. 共7兲 to produce normalized matrix P that
safety protections are usually installed as standard sensors by the consists of normalized coefficients n with
original equipment manufacturer 共OEM兲. As far as possible, these
measurements should be utilized for engine diagnostics to avoid ni,j = si,j/储si储 共8兲
unnecessary engine modifications. However, if inclusion of any of The normalized matrix is then multiplied by its transpose to pro-
these standard measurements jeopardizes the diagnostic accuracy, duce correlation square matrix Q which consists of correlation
it can be ignored in the gas path diagnostics algorithm and left to coefficients c as the matrix elements, where
serve the intended engine control functions.
Q = P · PT 共9兲
3.5 Provisional Measurements. Measurements which do not
fall into any of the previous categories are classified as provi- Each correlation coefficient is actually the cosine of the angle
sional. These are basically the major candidates to be further between a pair of measurements, which indicates the level of vec-
tor parallelism in component change space 关11兴. Very high corre-

Downloaded from http://asmedigitalcollection.asme.org/gasturbinespower/article-pdf/133/11/111701/5685391/111701_1.pdf by Universiti Teknologi Petronas user on 05 July 2024

evaluated in the selection process, alongside standard measure-
ments or maybe some advanced sensors depending on the scope lations can be identified if the value is in between 0.9 and 1.0 or in
of study. between ⫺1.0 and ⫺0.9. Note that all diagonal members in square
matrix Q must have the value of 1 to signify perfect correlations
3.5.1 Selection Criterion No. 1: Measurement Sensitivity. of the same measurement parameter.
From Urban 关18兴, “The presence of excessively large coefficients Similar to sensitivity norm, the correlation norm 储c储 of each
in the inverted matrix would be a mathematical announcement measurement is suggested as the indicator of overall correlation
that the combination was a poor one.” Low sensitivity measure- magnitude, which is defined as

冑兺
ments are found to be one of the factors that cause high interfer-
L
ences in the GPA influence coefficient matrices due to large matrix
condition number. It is known that this condition number is in- 储ci储 = 共ci,h兲2 = 冑共ci,1兲2 + . . . + 共ci,L兲2 共10兲
dicative of the “magnification” of disturbances through a linear h=1
system of equations and represents the “condition” of the system where number of column L = M. Correlation analysis is performed
关14兴. In addition, measurements that register very small deviations only to assist the decision to deselect any measurement from the
when subjected to component degradations can be easily over- candidate list, where measurement parameter with low sensitivity
whelmed by measurement noise at similar magnitude. Hence, the norm and high correlations can be conveniently excluded from the
lowest sensitivity measurements shall be identified and deselected. diagnostic analysis. However, priority shall be given to measure-
The thresholds of “high” or “low” sensitivities are subjective and ment sensitivity as it has direct impacts toward diagnostics accu-
relative to other parameter deviations. racy. These selection criteria are demonstrated later in the case
In order to define the overall sensitivity of a measurement to- study.
ward component degradations, calculation of sensitivity norm
from Stamatis et al. 关10兴 is being improvised and used as sensi- 3.5.3 Selection Criterion No. 3: Measurement Subsets. As
tivity indicators in this study. Norm 共level-2兲 is chosen as it regu- mentioned before, the total number of measurements to be chosen
lates the sensitivity coefficients to produce a single positive num- M is based on the number of component health parameters N with
ber that indicates overall deviation magnitude. Note that the ideally M ⱖ N, i.e., there are equal or redundant measurements as
relative rather than absolute values of the deviations are promi- compared with health parameter indices. With redundancies, Eq.
nent in causing disturbances in GPA linear equations. 共2兲 is overdetermined and the convergence of the iteration and
Sensitivity coefficients s are actually individual elements of the prediction accuracy of the nonlinear GPA can be potentially good.
ICM in Eq. 共2兲, which is defined as the deviations of each mea- Since there are multiple diagnosable components in a gas tur-
surement parameter toward predefined changes in each compo- bine, the process of determining each and every combination of
nent health parameter as follows: faulty component is again a tedious combinatorial problem. The
proposed measurement subset analysis is intended to simplify this
⳵ z៝i/z៝i process by identifying and grouping measurement parameters that
si,j = 共6兲
⳵ x៝ j/x៝ j are relevant to individual diagnosable component. Once achieved,
multiple component fault diagnosis can be performed by using
where i = 1 , 2 , . . . , M and j = 1 , 2 , . . . , N. From Eq. 共6兲, the sensi- measurement parameters merged from relevant subsets.
tivity norm 储s储 for each measurement parameter can be calculated In this concept, each measurement set for a diagnosable com-
as follows: ponent in a gas turbine can be considered as a measurement subset

冑兺 N of the whole engine. By this definition, the universal set of the gas
共si,j兲2 = 冑共si,1兲2 + . . . + 共si,N兲2
turbine measurements SETGT is assumed as the global set of the
储si储 = 共7兲 union of all measurement subsets SETC. For R number of compo-
j=1
nents or measurement subsets
which is basically the square root of the sum of the squares of
coefficients in each row of the ICM. The norm can then be sorted SETGT = SETC1 艛 SETC2 ¯ 艛 SETCR 共11兲
from highest to lowest value to effectively indicate the order of Once the total number of measurement parameters is determined
overall sensitivity for each potential measurement parameter can- based on their sensitivities and correlations, each measurement
didate. parameter may now be assigned to a set for a particular compo-
3.5.2 Selection Criterion No. 2: Measurement Correlations. nent and become an element of a particular measurement subset
Two or more measurement parameters may indicate similar sen- SETCp = 兵z1,z2 . . . ,z M 其Cp 共12兲
sitivity profiles and magnitude when subjected to the same com-
ponent faults, hence, creating some level of correlations. These where p = 1 , 2 , . . . , R.
correlations should be identified to ensure the selected measure- The elements z in Eq. 共12兲 shall consist of selective measure-
ment parameters can uniquely isolate faults of individual degrad- ment parameters that are physically and thermodynamically re-
able component. The measurement correlation analysis in this lated to the particular component C p. As mentioned by Urban 关1兴,
study is in accordance to the method proposed by Provost 关11兴 “The measurement taken must be meaningful related to the prob-
with some reinterpretations and improvisations. lem sought and detectably respond to change therein.” This propo-
To determine the level of correlations between measurements, sition is reasonable since aerothermocalculations of component

Journal of Engineering for Gas Turbines and Power NOVEMBER 2011, Vol. 133 / 111701-3
 Fig. 4 PYHTIA model and station numberings

Downloaded from http://asmedigitalcollection.asme.org/gasturbinespower/article-pdf/133/11/111701/5685391/111701_1.pdf by Universiti Teknologi Petronas user on 05 July 2024

power output of 26.0 MW, thermal efficiency of 35.8%, total pres-
sure ratio of 20.0, total air mass flow rate of 91.2 kg/s, and turbine
entry temperature of 1193° C.
Fig. 2 Venn diagram for measurement subset concept Based on Fig. 3, there are five major degradable components
being identified as follows:

health parameters 共such as efficiency index and flow capacity in- 1. LP compressor 共COMP1兲
dex兲 are derived from the measurement parameter values 共such as 2. HP compressor 共COMP2兲
pressure and temperature兲 of that component. If the measurement 3. HP turbine 共TURB1兲
parameters do not satisfy the above conditions they will not be 4. LP turbine 共TURB2兲
included in any of the measurement subset and should be ex- 5. power turbine 共TURB3兲
cluded from the diagnostic analysis.
To evaluate any single or multiple component fault cases using For every component 共compressor or turbine兲, it needs two
a global set of measurements, these measurement subsets should component health indices, i.e., efficiency index and flow capacity
hypothetically be combined in accordance to set rules as unions. index, to represent its degradation mechanism. These indices rep-
There also exist overlaps or sharing of measurement parameters resent the deviations of speed lines on compressor or turbine char-
between components, which are basically the intersections of the acteristic maps due to degradation. For the compressors, blade
component subsets. From this set analysis, it can be deduced that fouling is selected as the test case with typical deviation values of
the measurement parameters that fall in the intersection areas are ⫺1% in efficiency index and ⫺4% in flow capacity index. For the
more critical toward GPA as they can detect faults for more than turbines, blade erosion is used as the test case with assumed de-
one component. To illustrate the concept, a Venn diagram for a viations of ⫺1% in efficiency index and +2% in flow capacity
gas turbine with three diagnosable components is shown in Fig. 2. index as shown in Fig. 5. These component degradations are
Even though a diagram for three components can be conve- seeded into engine performance model to simulate measurements
niently illustrated, the complexity is increasing with higher num- of degraded engine for diagnostic analysis.
ber of components. However, the basic set rules are valid for any The individual or combinations of potential component faults
number of components and alternative diagrams can be used. This are referred as component fault cases 共CFCs兲 关15兴. Since there are
will be demonstrated in case studies later. many possible combination of CFC in a multicomponent gas tur-
bine, only few cases could be tested based on common degrada-
4 Case Study tion mechanisms as described by Ogaji et al. 关7兴. The test cases
are performed without any simulated measurement noise in order
In order to test the effectiveness of the proposed measurement to demonstrate the methodology without any interference.
selection approach, a three-shaft aeroderivative gas turbine model As there are five degradable components and two component
is used in the case study. This model engine is similar to an in- health parameters for each component, there are ten health param-
dustrial Rolls-Royce RB211-24G units installed onboard Angsi eters for the whole engine, as tabulated in Table 1.
offshore production platform in South China Sea, Malaysia oper-
ated by PETRONAS Carigali Sdn. Bhd. 关19兴. The engine layout 4.2 Measurement Classifications. Initially, 30 possible mea-
and existing sensors schematics are shown in Fig. 3. surements are chosen and classified as shown in Table 2. Based on
the classifications, 13 potential measurements are selected for sub-
4.1 Engine Modeling. The engine performance is modeled sequent analysis, as shown in Table 3.
using PYTHIA, a gas turbine performance and diagnostics software
关15兴 developed by Cranfield University, UK. The engine model 4.3 Sensitivity Analysis. The ICM or sensitivity matrix of the
and station numbers is shown in Fig. 4. selected measurements is determined by individually implanting
The major engine performance specifications are as follows: +1% deviation toward each component health index. From the
ICM, the sensitivity bar chart is then plotted in Fig. 6 and the

Fig. 3 Gas turbine components and existing sensor layout Fig. 5 Component degradation test cases

111701-4 / Vol. 133, NOVEMBER 2011 Transactions of the ASME

 Table 1 Selected component health parameters Table 3 Potential measurements

No. Component health parameters Symbol No. Potential measurement parameters Symbol

1 LP compressor efficiency CE1 1 LP compressor exit pressure P3

2 LP compressor flow capacity CFC1 2 LP compressor exit temperature T3
3 HP compressor efficiency CE2 3 HP compressor exit pressure P5
4 HP compressor flow capacity CFC2 4 HP compressor exit temperature T5
5 HP turbine efficiency TE1 5 HP turbine exit pressure P9
6 HP turbine flow capacity TFC1 6 HP turbine exit temperature T9
7 LP turbine efficiency TE2 7 LP turbine exit pressure P11
8 LP turbine flow capacity TFC2 8 LP turbine exit temperature T11
9 Power turbine efficiency TE3 9 Power turbine exit pressure P12
10 Power turbine flow capacity TFC3 10 Power turbine exit temperature T12

Downloaded from http://asmedigitalcollection.asme.org/gasturbinespower/article-pdf/133/11/111701/5685391/111701_1.pdf by Universiti Teknologi Petronas user on 05 July 2024

11 Fuel flow FF
12 LP compressor speed PCN1
13 HP compressor speed PCN2
Table 2 Measurement classifications

No. Measurement parameters Symbol Category Include?

1 Ambient pres. P1 Constant No norm is calculated and sorted in Table 4. Referring to the results,
2 Ambient temp. T1 Constant No it can be observed that T9 and T11 have much lower overall
3 LP comp. inlet pres. P2 Constant No sensitivity 共i.e., 0.2908 and 0.3506, respectively兲 as compared
4 LP comp. inlet temp. T2 Constant No
5 LP comp. exit pres. P3 Standard 共PS25兲 Yes
with the next lowest ones, i.e., P12 共0.5458兲, T12 共0.5909兲, and T5
6 LP comp. exit temp. T3 Standard 共T25兲 Yes 共0.6052兲. Hence, these two measurements can be considered for
7 HP comp. inlet pres. P4 Constant No exclusion.
8 HP comp. inlet temp. T4 Constant No
9 HP comp. exit pres. P5 Standard 共PS31兲 Yes
4.4 Correlations Analysis. Measurement correlations analy-
10 HP comp. exit temp. T5 Provisional Yes sis is conducted next as shown in Table 5. T5 and T11 can be
11 HP turb. inlet pres. P8 Advance No identified to have highest correlations with the factor of ⫺0.9936.
12 HP turb. inlet temp. T8 Advance No Then, normalized absolute deviations from the ICM are plotted in
13 HP turb. exit pres. P9 Provisional Yes Fig. 7 to validate this finding. Note that in this chart, a pair of high
14 HP turb. exit temp. T9 Provisional Yes correlation measurements will show similar profile and magni-
15 LP turb. inlet pres. P10 Constant No tudes. When Fig. 7 is carefully analyzed, T5 and T11 both show
16 LP turb. inlet temp. T10 Constant No very similar deviation characteristics. Hence, one of the measure-
17 LP turb. exit pres. P11 Standard 共PS455兲 Yes ment parameters can be excluded.
18 LP turb. exit temp. T11 Standard 共T455兲 Yes
19 Power turb. exit pres. P12 Provisional Yes
In this case study, minimum condition of M = N need to be
20 Power turb. exit temp. T12 Provisional Yes fulfilled since we are evaluating all component health parameters.
21 Exhaust pres. P13 Constant No Hence, ten component health parameters from Table 1 require ten
22 Exhaust temp. T13 Constant No measurement parameters to be chosen from the 13 potential mea-
23 Fuel flow FF Standard 共FF兲 Yes surement parameter candidates in Table 3.
24 Shaft power SHP Derived No However, less measurement parameters may be applicable de-
25 Thermal efficiency TE Derived No pending on the component fault case as the number of health
26 LP comp. speed PCN1 Standard 共N1兲 Yes parameters may be less in individual cases. This can be done by
27 HP comp. speed PCN2 Standard 共N2兲 Yes using measurement subset concept, which is demonstrated later. In
28 Power turbine speed PTN Constant No
29 Inlet mass flow MF1 Derived No
this concept, the total number of measurement parameters re-
30 Exhaust mass flow MF8 Derived No quired should be the natural outcome of the component health
parameters in the union of the subsets.

Fig. 6 Sensitivity analysis bar chart

Journal of Engineering for Gas Turbines and Power NOVEMBER 2011, Vol. 133 / 111701-5
 Table 4 Overall sensitivity norm

Measurement P3 P9 P5 FF P11 PCN1 T3 PCN2 T5 T12 P12 T11 T9

Sensitivity norm 2.2627 2.0301 1.8033 1.3185 1.2294 1.1742 0.9020 0.8499 0.6052 0.5909 0.5458 0.3506 0.2908

Table 4 indicates T9 have very low sensitivity norms and T11 The subsets for all degradable components in this case study are
have low sensitivity norm plus highly correlated with T5. Thus, selected as follows:
both measurement parameters can be conveniently deselected
SETCOMP1 = 兵P3,T3,PCN1,FF其
from the candidates. However, sensitivity norms of P12, T12 and
T5 are quite close together. Among these parameters, P12 has the
SETCOMP2 = 兵P5,T5,PCN2,FF其
highest correlation norm 共i.e., 2.2462兲 in Table 6 and, hence, be-

Downloaded from http://asmedigitalcollection.asme.org/gasturbinespower/article-pdf/133/11/111701/5685391/111701_1.pdf by Universiti Teknologi Petronas user on 05 July 2024

ing omitted. Finally, ten measurements for GPA diagnostics for SETTURB1 = 兵P9,PCN2,FF其
the whole engine are shown in Table 7.
4.5 Measurement Subset Analysis. For measurement subset SETTURB2 = 兵P11,PCN1,FF其
analysis, five subsets may be selected with each referring to a
degradable component: SETTURB3 = 兵T12,FF其

SETGT 傶 SETCOMP1 艛 SETCOMP2 艛 SETTURB1 艛 SETTURB2 As there are more than three components, conventional Venn dia-
gram per Fig. 2 could not be easily drawn. Thus, an alternative
艛 SETTURB3 diagram is used to similarly illustrate the measurement subset
Based on the final measurements in Table 7, component measure- concept in Fig. 8.
ment subsets are determined from component related parameters
that have direct influence toward the component health indices.
For example, the element for LP compressor 共COMP1兲 must in- 5 Result Verifications
clude the LP compressor exit pressure 共P3兲, exit temperature 共T3兲, In order to compare the GPA diagnostics accuracies using the
rotational speed 共PCN1兲, and fuel flow 共FF兲, which are all related standard and selected measurements, the following are three dif-
to the efficiency and flow capacity of the particular component. ferent measurement groups being evaluated in the case study.

Table 5 Measurement correlations matrix

P3 T3 P5 T5 P9 T9 P11 T11 P12 T12 FF PCN1 PCN2 Norm

P3 1 0.9295 0.5920 0.4816 0.8221 0.4940 0.3761 ⫺0.4570 0.6782 ⫺0.3915 0.6406 0.7072 ⫺0.2336 2.3010
T3 0.9295 1 0.4653 0.5341 0.7059 0.4987 0.1789 ⫺0.5242 0.5267 ⫺0.3642 0.4400 0.6226 ⫺0.2241 2.1121
P5 0.5920 0.4653 1 0.7496 0.7803 ⫺0.2385 0.6211 ⫺0.7063 0.7303 ⫺0.6414 0.7316 0.6952 0.3863 2.4066
T5 0.4816 0.5341 0.7496 1 0.4816 ⫺0.4600 0.1797 ⫺0.9936 0.3874 ⫺0.6615 0.2472 0.4576 0.3463 2.1261
P9 0.8221 0.7059 0.7803 0.4816 1 0.2957 0.7218 ⫺0.4251 0.8101 ⫺0.5065 0.8648 0.8503 ⫺0.0392 2.4878
T9 0.4940 0.4987 ⫺0.2385 ⫺0.4600 0.2957 1 0.0929 0.4767 0.1619 0.2313 0.2597 0.1991 ⫺0.5957 1.6206
P11 0.3761 0.1789 0.6211 0.1797 0.7218 0.0929 1 ⫺0.0901 0.5769 ⫺0.4370 0.8186 0.5104 0.1050 1.8960
T11 ⫺0.4570 ⫺0.5242 ⫺0.7063 ⫺0.9936 ⫺0.4251 0.4767 ⫺0.0901 1 ⫺0.3747 0.5970 ⫺0.1908 ⫺0.4359 ⫺0.3612 2.0590
P12 0.6782 0.5267 0.7303 0.3874 0.8101 0.1619 0.5769 ⫺0.3747 1 ⫺0.2205 0.9011 0.7732 0.1037 2.2462
T12 ⫺0.3915 ⫺0.3642 ⫺0.6414 ⫺0.6615 ⫺0.5065 0.2313 ⫺0.4370 0.5970 ⫺0.2205 1 ⫺0.4026 ⫺0.4203 ⫺0.2537 1.8558
FF 0.6406 0.4400 0.7316 0.2472 0.8648 0.2597 0.8186 ⫺0.1908 0.9011 ⫺0.4026 1 0.7310 0.0573 2.2878
PCN1 0.7072 0.6226 0.6952 0.4576 0.8503 0.1991 0.5104 ⫺0.4359 0.7732 ⫺0.4203 0.7310 1 0.0039 2.2589
PCN2 ⫺0.2336 ⫺0.2241 0.3863 0.3463 ⫺0.0392 ⫺0.5957 0.1050 ⫺0.3612 0.1037 ⫺0.2537 0.0573 0.0039 1 1.3965

Fig. 7 Correlations analysis bar chart

111701-6 / Vol. 133, NOVEMBER 2011 Transactions of the ASME

 Table 6 Overall correlations norm

Measurement P9 P5 P3 FF PCN1 P12 T5 T3 T11 P11 T12 T9 PCN2

Correlations norm 2.4878 2.4066 2.3010 2.2878 2.2589 2.2462 2.1261 2.1121 2.0590 1.8960 1.8558 1.6206 1.3965

Table 7 Selected diagnostics set 1. Standard measurements—all standard measurements in-

stalled on the engine 共Table 8兲.
No. Final measurement parameter Symbol Category 2. Selected measurements—all selected measurements for the
whole engine 共Table 8兲.
1 LP compressor exit pressure P3 Standard 3. Measurement subsets—specifically selected measurements
2 LP compressor exit temperature T3 Standard
for each CFC with corresponding M = N 共Table 9兲.

Downloaded from http://asmedigitalcollection.asme.org/gasturbinespower/article-pdf/133/11/111701/5685391/111701_1.pdf by Universiti Teknologi Petronas user on 05 July 2024

3 HP compressor exit pressure P5 Standard
4 HP compressor exit temperature T5 Provisional
5 HP turbine exit pressure P9 Provisional For demonstration purposes, the 15 CFCs chosen in Table 9 are
6 LP turbine exit pressure P11 Standard sufficient to cover most common component fault occurrences in
7 Power turbine exit temperature T12 Provisional industrial gas turbines. Any combination of the measurement pa-
8 Fuel flow FF Standard rameters in the same subset can be chosen in each CFC and the
9 LP compressor speed PCN1 Standard ones used in Table 9 are arbitrarily chosen to showcase this con-
10 HP compressor speed PCN2 Standard cept. Finally, nonlinear GPA 共NLGPA兲 simulations are conducted
for each CFC to compare the accuracy of the three measurement
groups. The accuracies are gauged based on RMS of the errors
between the predicted and measured measurement parameters per
Eq. 共5兲. The simulation results are shown in Figs. 9–11.

6 Discussions
As demonstrated via the case study, the methodology managed
to deselect from 30 possible measurements to 13 potential mea-
surements and finally to ten diagnostics measurements. In order to
establish the baselines of the effectiveness of the selection meth-
odology, the performance of the standard measurements are sum-
marized in Table 10.
From this table, the RMS error soars in each CFC that includes
power turbine 共TURB3兲 component. These escalations are ex-
pected since the standard engine does not have sufficient gas path
measurement to detect power turbine degradations, hence, reduc-
ing the GPA accuracies.
In addition, all of the multiple component fault cases that in-
Fig. 8 Measurement subset diagram
clude the HP turbine 共TURB2兲 also produce high RMS error 共ex-
cept single fault case in CFC4, where the number of MP is highly
redundant兲. This suggests that the use of gas generator 共GG兲 exit
Table 8 Standard and selected measurements
temperature 共T11兲 in the standard measurements would jeopardize
No. of GPA diagnostics accuracy since this measurement is shown to
Measurement Measurement parameters parameters have very low sensitivity norm. Instead, the use of gas generator
exit pressure 共P11兲 included in the selected measurements is more
Standard P3, T3, P5, P11, T11, PCN1, PCN2, FF 8 suitable for GPA.
Selected P3, T3, P5, T5, P9, P11, T12, PCN1, PCN2, FF 10 A comparison of diagnostic accuracies for different fault cases
using standard, selected and subset measurements is shown in

Table 9 CFC and component related measurement subsets

Degraded No. of Component related No. of

CFC components comp. measurements MP

1 COMP1 1 P3, PCN1 2

2 COMP2 1 P5, PCN2 2
3 TURB1 1 P9, PCN2 2
4 TURB2 1 P11, PCN1 2
5 TURB3 1 T12, FF 2
6 COMP1 + COMP2 2 P3, P5, PCN1, PCN2 4
7 COMP1 + TURB1 2 P3, P9, PCN1, PCN2 4
8 TURB1 + TURB2 2 P9, P11, PCN1, PCN2 4
9 COMP2 + TURB3 2 P5, T12, PCN2, FF 4
10 COMP1 + COMP2 + TURB1 3 P3, P5, P9, PCN1, PCN2, FF 6
11 COMP1 + TURB1 + TURB2 3 P3, P9, P11, PCN1, PCN2, FF 6
12 TURB1 + TURB2 + TURB3 3 P9, P11, T12, PCN1, PCN2, FF 6
13 COMP1 + COMP2 + TURB1 + TURB2 4 P3, T3, P5, P9, P11, PCN1, PCN2, FF 8
14 COMP1 + TURB1 + TURB2 + TURB3 4 P3, T3, P9, P11, T12, PCN1, PCN2, FF 8
15 All components 5 P3, T3, P5, T5, P9, P11, T12, PCN1, PCN2, FF 10

Journal of Engineering for Gas Turbines and Power NOVEMBER 2011, Vol. 133 / 111701-7
 Downloaded from http://asmedigitalcollection.asme.org/gasturbinespower/article-pdf/133/11/111701/5685391/111701_1.pdf by Universiti Teknologi Petronas user on 05 July 2024
Fig. 9 NLGPA RMS errors „single component fault cases…

Fig. 10 NLGPA RMS errors „two component fault cases…

Fig. 11 NLGPA RMS errors „three or more component fault cases…

111701-8 / Vol. 133, NOVEMBER 2011 Transactions of the ASME

 Table 10 GPA accuracy of standard measurement set

CFC No. COMP1 COMP2 TURB1 TURB2 TURB3 RMS⬎ 0.02

1 X
2 X
3 X
4 X
5 X X
6 X X
7 X X
8 X X X
9 X X X
10 X X X

Downloaded from http://asmedigitalcollection.asme.org/gasturbinespower/article-pdf/133/11/111701/5685391/111701_1.pdf by Universiti Teknologi Petronas user on 05 July 2024

11 X X X X
12 X X X X
13 X X X X X
14 X X X X X
15 X X X X X X

Figs. 9–11. In most CFCs, the RMS errors for measurement subset 7 Conclusions
are less than 0.008, which indicate very good accuracy. However,
The main objective of this research is accomplished, which is to
slight inaccuracy can be detected in CFC4 with RMS value of
provide better understanding on the effect of each measurement
0.015 in Fig. 9. This confirmed the hypothesis that the measure-
parameter toward GPA accuracy and simplify the measurement
ment subset method does not always produce the best measure-
selection process. The effectiveness of this method was demon-
ment set. However, this concept would suggest sufficiently good
strated in an engine case study, where minimal measurements are
measurements with the benefits of simplified selection process at
selected from 30 possible measurements depending on various
component-level diagnostics. Most importantly, the GPA diagnos-
component fault cases.
tics using all selected measurements generate much superior ac-
In addition to measurement sensitivity and correlations, mea-
curacies in each and every component fault case compared with
surement subset concept being introduced is a useful tool to sim-
the same fault case using the standard measurement set in Figs.
plify the measurement selection process for multiple component
9–11. This proves that the sensitivity, correlation and redundant
fault diagnosis. This concept also managed to identify the critical-
number of measurements are important for GPA diagnostics.
ity of each measurement parameter and the measurement subset
Based on the findings, design modifications to the existing en-
diagram provides a useful graphical overview of the required
gine for GPA application are inevitable. There are three provi-
measurement parameters.
sional measurements included in the selected diagnostics mea-
One of the drawbacks of measurement subset concept is the
surement set. The compressor exit temperatures 共T5兲 and power
resulting GPA predictions may not be as accurate as compared
turbine exit temperatures 共T12兲, which both operate at around
with using redundant measurements. However, measurement re-
450° C can be easily retrofitted using conventional resistive tem-
dundancies are not always available at site and the operators can
perature detectors 共RTD兲 or thermocouples.
actually determine alternative combinations of measurements via
However, the sensor for HP turbine exit pressure 共P9兲, which
the subset analysis. In addition, the inaccuracy is within accept-
operates at high temperature environment of about 800° C is tech-
able range and proved to be better than using standard measure-
nically challenging. The emergence of new sensor technologies
ments only. Based on the results from the selection process, asser-
can be explored or alternatively, these parameters can be approxi-
tive recommendations can then be made with regards to
mated using maximum likelihood technique as proposed in Ref.
diagnostics sensor requirements prior to applications of GPA in
关5兴 or combinatorial approach in Ref. 关20兴. In addition, HP turbine
the industry.
is the first component downstream of the combustor, which is very
likely to experience deteriorations such as hot corrosion, erosions
and creep. Since this measurement is crucial for GPA, extensive Acknowledgment
technology research for high temperature sensor is hereby recom- The authors would like to thank PETRONAS for sponsoring
mended. the first author to carry out the research works at Cranfield Uni-
With regards to the measurement subset concept, it is in fact a versity.
supplementary tool being introduced to complement the outcomes
from the sensitivity and correlation analyses. First, this concept
provides visualization of the measurement parameters to the cor- Nomenclature
responding gas turbine components. We could also determine CFC ⫽ component fault case
from the set analysis the components that could not be accurately CHP ⫽ component health parameter
diagnosed using GPA due to insufficient measurements. COMP ⫽ compressor
Second, any combination of measurement parameters that falls E ⫽ efficiency
within a subset can be used for accurate diagnostics of that par- FC ⫽ flow capacity
ticular component. It is an advantage for GPA practitioners to FF ⫽ fuel flow
know that many options exist if they want to focus on diagnostics FCM ⫽ fault coefficient matrix
for particular components. GPA ⫽ gas path analysis
Lastly, the operators or even the engine designers would appre- HP ⫽ high pressure
ciate which measurement parameters are more crucial for the GPA ICM ⫽ influence coefficient matrix
diagnostics. Measurement parameters that cover more than one LP ⫽ low pressure
component are obviously regarded as “critical,” hence, the avail- M ⫽ number of measurement parameters
ability, accuracy and reliability of that particular measurements MP ⫽ measurement parameter
are of utmost importance. N ⫽ number of component health parameters

Journal of Engineering for Gas Turbines and Power NOVEMBER 2011, Vol. 133 / 111701-9
 NLGPA ⫽ nonlinear gas path analysis 关8兴 Arora, J. S., 2006, Introduction to Optimum Design, McGraw-Hill, New York.
关9兴 Sowers, T. S., Kopasakis, G., and Simon, D. L., 2008, “Application of the
P ⫽ total pressure Systematic Sensor Selection Strategy for Turbofan Engine Diagnostics,”
PCN ⫽ shaft rotational speed ASME Paper No. GT2008-50525.
RMS ⫽ root mean square 关10兴 Stamatis, A., Mathioudakis, K., and Papailiou, K., 1992, “Optimal Measure-
SHP ⫽ shaft horse power ment and Health Index Selection for Gas Turbine Performance Status and
Fault Diagnosis,” ASME J. Eng. Gas Turbines Power, 114共2兲, pp. 209–216.
T ⫽ total temperature 关11兴 Provost, M. J., 1994, “The Use of Optimal Estimation Techniques in the
TET ⫽ turbine entry temperature Analysis of Gas Turbines,” Ph.D. thesis, Cranfield University, UK.
TURB ⫽ turbine 关12兴 Cotae, P., Yalamnchili, S., Philip Chen, C. L., and Ayon, A., 2008, “Optimi-
x ⫽ component health parameter zation of Sensor Locations and Sensitivity Analysis for Engine Health Moni-
toring Using Minimum Interference Algorithms,” EURASIP Journal on Ad-
z ⫽ measurement parameter vances in Signal Processing, 2008, pp. 1–9.
关13兴 Borguet, S., and Leonard, O., 2008, “The Fisher Information Matrix as a
Relevant Tool for Sensor Selection in Engine Health Monitoring,” Int. J. Ro-
References tating Mach., 2008, pp. 1–10.

Downloaded from http://asmedigitalcollection.asme.org/gasturbinespower/article-pdf/133/11/111701/5685391/111701_1.pdf by Universiti Teknologi Petronas user on 05 July 2024

关1兴 Urban, L. A., 1969, Gas Turbine Engine Parameter Interrelationships, 2nd 关14兴 Mathioudakis, K., and Kamboukos, P., 2006, “Assessment of the Effectiveness
ed., HSD UTC, Windsor Locks, CT. of Gas Path Diagnostic Schemes,” ASME J. Eng. Gas Turbines Power,
关2兴 Urban, L. A., 1980, Gas Path Analysis—A Tool for Engine Condition Moni- 128共1兲, pp. 57–63.
toring, Flight Safety Foundation, Christchurch, New Zealand. 关15兴 Li, Y. G., and Singh, R., 2005, “An Advanced Gas Turbine Gas Path Diagnos-
关3兴 Volponi, A. J., 1982, “Gas Path Analysis: An Approach to Engine Diagnos- tic System—PYTHIA,” 17th International Symposium on Airbreathing En-
tics,” 35th Symposium Mechanical Failure Prevention Group, Gaithersbury, gines, Munich, Germany, Paper No. ISABE-2005-1284.
MD. 关16兴 Li Y. G., 2010, “Gas Turbine Performance and Health Status Estimation Using
关4兴 Li, Y. G., 2002, “Performance-Analysis-Based Gas Turbine Diagnostic: A Re- Adaptive Gas Path Analysis,” ASME J. Eng. Gas Turbines Power, 132共4兲, p.
view,” Proc. Inst. Mech. Eng., Part A: Journal of Power and Energy, 216, pp. 041701.
363–377. 关17兴 Golub, H. G., and Van Loan, C. F., 1996, Matrix Computations, 3rd ed., John
关5兴 Mathioudakis, K., Kamboukos, P., and Stamatis, A., 2004, “Gas Turbine Com- Hopkins University Press, Baltimore.
ponent Fault Detection From a Limited Number of Measurements,” Proc. Inst. 关18兴 Urban L. A., 1975, “Parameter Selection for Multiple Fault Diagnostics of Gas
Mech. Eng., Part A: Journal of Power and Energy, 218共8兲, pp. 609–618. Turbine Engines,” ASME J. Eng. Power, 97共2兲, pp. 225–230.
关6兴 Mushini, R., and Simon, D., 2005, “On Optimization of Sensor Selection for 关19兴 Jasmani, M. S., 2009, “Systematic Measurement Selections for Gas Path Di-
Aircraft Gas Turbine Engines,” 18th International Conference on Systems En- agnostics of Industrial Gas Turbines, MSc. thesis, Cranfield University, UK.
gineering 共ICSEng 2005兲, pp. 9–14. 关20兴 Aretakis, N., Mathioudakis, K., and Stamatis, A., 2003, “Nonlinear Engine
关7兴 Ogaji, S. O. T., Sampath, S., Singh, R., and Probert, S. D., 2002, “Parameter Component Fault Diagnosis From a Limited Number of Measurements Using
Selection for Diagnosing a Gas Turbine Performance-Deterioration,” Appl. a Combinatorial Approach,” ASME J. Eng. Gas Turbines Power, 125共3兲, pp.
Energy, 73, pp. 25–46. 642–650.

111701-10 / Vol. 133, NOVEMBER 2011 Transactions of the ASME