Documente Academic
Documente Profesional
Documente Cultură
www.springerlink.com/content/1738-494x
DOI 10.1007/s12206-010-1207-5
(Manuscript Received April 28, 2010; Revised July 23, 2010; Accepted October 24, 2010)
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Abstract
In the present paper, limitations concerning three implementations of gas path analysis (GPA) methods are investigated and their diag-
nostic effectiveness is evaluated. The methods were tested for different sets of faults on a twin shaft gas turbine with an instrumentation
set typical of today’s engines. Test results revealed that classical GPA is not sufficient. Correct diagnosis is provided only when one al-
ready knows a subset of components containing the fault; otherwise, the fault may be attributed to other component (s). The effectiveness
of a second method that implements Multi Operating Point Analysis (MOPA) is related with the assumption of non varying health pa-
rameters with deviations along with the operating point. Cases of wrong diagnosis were detected when the above assumption was vio-
lated. Improvement on the diagnostic effectiveness of the MOPA method has been verified through careful selection of the parameters
defining the operating point. Further improvement on diagnostic ability was achieved when a third, recently proposed method was ap-
plied. The method uses information produced from existing sensors when artificial operating points are defined close to the initial operat-
ing point. It was found that the third method can detect and correctly identify faults even in cases where the multipoint method fails.
Keywords: Diagnostics; Gas path analysis; Gas turbine; Fault identification; Multipoint analysis
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
∆y = C ⋅ ∆f + v (3a) (a)
∆fˆ = S -1 ⋅ CT ⋅ R -1 ⋅ ∆y (3b)
-1 T -1
S = M + C ⋅R ⋅C (3c)
∆y k = Ck ⋅ ∆f k = 1, NOP (4a) discussed by Doel [4]. The algorithm tends to ‘‘smear’’ the
Ck = ⎡⎣cij , k ⎤⎦ (4b) fault over many components.
Regarding the MOPA approach, the so-called information
cij , k = ( ∂∆yi / ∂∆f j ) (4c) matrix P is crucial in the sense that its condition determines
k
NOP the diagnostic effectiveness. The condition of the matrix is
∆fˆ = P −1 ⋅ ∑ (C
k =1
T
k ⋅ R −k 1 ⋅ ∆y k ) (4d) represented by its condition number. Investigations concern-
ing effects of both the number of operating points used and the
NOP ‘distance’ of the operating points on information matrices
P= ∑ (C
k =1
T
k ⋅ R −k 1 ⋅ Ck ) (4e) have been reported by Mathioudakis and Kamboukos [15] and
Skölde et al. [16].
∆z k = G k ⋅ ∆ f k = 1, NAOP (5a) Additional details on assessing identifiability in multipoint
G k = ⎡⎣ gij , k ⎤⎦ (5b) gas turbine estimation problems are given by Gronstedt [17].
Although all the works implementing the multipoint approach
gij , k = ( ∂∆zi / ∂∆f j ) (5c) agree that the idea more or less improves the diagnostic effec-
k
NAOP tiveness, there are also results (see Henriksson et al. [18]),
∆fˆ = H −1 ⋅ ∑ (G T
k ⋅ R −k 1 ⋅ ∆z k ) (5d) indicating that the theoretically attainable multi-point im-
k =1 provements are difficult to realize in practical engine applica-
NAOP
tions. In order to understand the reasons for potential problems
H= ∑ (G
k =1
T
k ⋅ R −k 1 ⋅ G k ) (5e)
concerning diagnosis using a multipoint approach, it is neces-
sary to examine the underlying assumptions of the method.
The main assumption of the method is that the ‘deltas’ con-
Although the formulation for classical GPA has proven to cerning the health parameters remain constant with regard to
be successful for practical purposes and existing commercial change in operating conditions (see Fig. 3(a)). This assump-
systems are based on it (e.g. Doel [13], Barwell [14]), identifi- tion is obviously true for some parameters (for example, the
ability problems exist due to limited instrumentation. In this parameter expressing the effective turbine area or the area of
case, inversion of matrix S is only possible when it is domi- non variable nozzle jet engine), but there are indications that
nated by M. The main drawback of the method is the effect for other parameters this is a week assumption. Several works
472 A. G. Stamatis / Journal of Mechanical Science and Technology 25 (2) (2011) 469~477
u OPV y MV
1 Load 1 NGG
2 NPT 2 EGT
3 Tamb 3 WF
4 Pamb 4 CDP
5 CDT
f HPV 1 2 3 4 5 6 7 8
1 SWc + + + - + + - -
Fig. 4. Variability of compressor flow capacity delta with operating
point. 2 SEc + + + - + + - -
3 SWt + + + + - - + -
(Diakunchak [19], Aker and Saravanamuttoo [20]), have pro- 4 SEt + + + + - - + -
vided evidence that when deterioration is present, the devia- 5 SWpt + + - + + - - +
tions of parameters such as flow compressor capacity and
6 SEpt + + - + + - - +
efficiency change with the operating point. In fact, a different
working-point means different aerodynamic conditions and, in 7 SEb + - - - - - - -
this sense, efficiency and flow capacity deltas can significantly
vary with the operating condition (Fig. 3(b)).
We may formulate this variability (Fig. 4) as nosis since the assumption of non-variability of health pa-
rameter deltas is ‘insured’ by the method definition. On the
∆fl = ∆f r + δ∆fl l = 1, NOP (6a) other hand, the way the operating points are defined seems to
improve diagnostic accuracy, as will be shown in the follow-
where ∆f r represents the deltas in a reference operating ing application.
point r among the NOP operating points and δ∆fl the devia-
tion from this base situation in operating point l. 4. Application test case
It can be seen (after some matrices manipulation) that when
the deltas are estimated with the multipoint approach, a bias is Results from application to a case of a twin shaft gas turbine
introduced in the estimation given in the linear case by the engine are used to demonstrate limitations and capabilities of
relation the variants of GPA discussed previously. A computerized
engine model based on the methodology introduced by Stama-
NOP tis et al. [21], resembling the ABB GT10 engine, was used to
∆fˆ − ∆fl = P −1 ⋅ ∑ (Q k ⋅ δ∆fk ) − δ∆fl (6b) simulate engine performance.
k =1 The engine operating point vector (OPV) is defined initially
Q k = CTk ⋅ R -k1 ⋅ Ck (6c) through ambient pressure and temperature, power turbine
NOP speed (NPT) and Load (vector u in Table 1). The measure-
P= ∑Q k . (6d) ments Vector (MV) provided (Table 1, vector y) includes
k =1 compressor delivery pressure (CDP) and temperature (CDT),
exhaust gas temperature (EGT), fuel flow rate (WF) and gas
The resulting diagnosis risk is not only imprecise calcula- generator speed (NGG).
tion of the engine’s state after some deterioration but, more- The engine main component faults such as erosion, fouling,
over, indicating the wrong component (s) as responsible for turbine nozzle area change, etc. are represented by changing
the fault. mass flow capacity and efficiency with respect to their values
The question emerging through the previous analysis is the at the healthy component condition (see Aker and Saravana-
following: is it possible to overcome the diagnosis risk using muttoo [20], Urban and Volponi [10]).
multipoint analysis based on the ‘questionable’ assumption of It must be stressed at this point that changes for each com-
non-varying deltas of health parameters, without abandoning ponent have to be independent from the operating point shift
at the same time the basic idea of the method? We recall here due to change in another component. In order to make it pos-
that this idea is to exploit information supplied from existing sible, the changes have to be defined not between final (faulty)
instrumentation when more operating points are considered in and initial (healthy) values but between final values and the
the analysis. values at the initial map corresponding to the same final val-
The AMOPA method, although implementing a multipoint ues of the independent parameters of the map (i.e. pressure
analysis, at the same time minimizes the risk for wrong diag- ratio and corrected speed). Here, the health performance vec-
A. G. Stamatis / Journal of Mechanical Science and Technology 25 (2) (2011) 469~477 473
(c)
tor (HPV) of the main engine components (compressor, tur- Fig. 5. GPA and MOPA fault diagnosis.
bines and burner), is expressed through a flow factor SW
and/or an efficiency factor SE (Table 1, vector f), defined as: data could be generated, as would happen in an actual situa-
tion, when measurements are periodically recorded.
x An indicative set of engine faults considered is shown in
Sx = (7)
xref Table 2. Cases Cx represent a compressor fault, Tx a turbine
fault and PTx a power turbine fault. Two variants, x, of each
where x is the corrected airflow or isentropic efficiency for fault are examined. Variant 1 represents invariant health pa-
compressor and turbines, combustion efficiency for the burner, rameters deltas along with different operating points and vari-
subscript ref represents the healthy engine and x, xref corre- ant 2 assumes that some deltas are decreasing and increasing
spond to the same values of the independent parameters of the correspondingly.
component map. Sign + in the table means that the corre- Classical GPA cannot be used for diagnosis when health pa-
sponding health parameter is assumed to vary and is included rameter vectors (HPV1 or HPV2) have dimensions larger than
in the estimation process while sign – means that the corre- the engine measurement vector (MV). Still, a solution is ob-
sponding health parameter is assumed to be constant. tained when health parameter vectors with smaller dimensions
For the purposes of the present work, we produced the re- than the measurement vector are estimated (HPV3- HPV8).
quired information from the engine when different faults oc- However, as it can be seen from Fig. 5, correct diagnosis using
cur using the computerized engine model. The model provides HPV6- HPV8 is provided only when one already knows the
simulated data (the values of the 5 measured quantities of the faulty component, or the pair of components containing the
engine) for a given operating point, defined through the 4 fault when using HPV3- HPV5; otherwise, the fault may be
relevant parameters, when the condition of the engine compo- attributed to other component (s). GPA results in these figures
nents, namely a set of values of the 7 health parameters, is are obtained at 100% load.
known. The data produced can be contaminated with noise to No a priori knowledge is needed when MOPA is applied.
resemble actual measured data. Sets containing series of noisy Fig. 5 shows that estimation of health parameters deltas of the
474 A. G. Stamatis / Journal of Mechanical Science and Technology 25 (2) (2011) 469~477
AOP 1 2 3 4 5
x z x z x z x z x z
Tamb + + + + +
Pamb + + + + +
NPT + + + + +
Load + + + + +
NGG + + + + + Fig. 6. MOPA wrong fault diagnosis.
EGT + + + + +
WF + + + + +
CDP + + + + +
CTD + + + + +
∆ (CDP) NGG 1.2 -0.9 -0.6 -0.7 -0.56 0 Fig. 7. Faults detected using AMOPA.
∆ (CDT) NGG 0.7 -1.9 -0.3 -0.4 -0.3 0
6. Conclusions
The scope of the present paper was to investigate limitations
concerning various implementations of gas path analysis
methods in order to evaluate their diagnostic effectiveness and
suggest possible improvement.
The methods were tested on different sets of main compo-
nent faults, with an instrumentation set typical of today’s en-
(a) gines. Main limitations of the examined implementations
evaluated on a twin shaft gas turbine have been identified as
follows: Classical GPA is not sufficient when the number of
health parameters is bigger than the number of measurements.
Correct diagnosis is provided only when one already knows a
subset of components containing the fault; otherwise the fault
may be attributed to other component (s). MOPA method
effectiveness is directly related with the assumption of non-
varying health parameter deviations along with the operating
point. The weaker the assumption is, the greater the risk for
wrong diagnosis. The AMOPA method, though more efficient
(b) than the previous method, has limited success when combined
faults are considered.
Fig. 9. Effect of the variable defining operating point on MOPA diag-
nosis. Improvement of the diagnostic effectiveness of MOPA
methods has been verified when selection of the parameters
senting the condition of engine components, the model is ca- defining the operating point was done after careful analysis.
pable of producing realistic simulated fault data. Diagnosis results confirm that success is influenced strongly
A key aspect of the AMOPA method is the assumption of by this selection. Further improvement on diagnostic ability
non varying health parameters deviations with the operating resulted when the AMOPA method was implemented. It was
point. This assumption is quite reasonable as can be seen in found that AMOPA can detect and correctly identify faults,
compressor map (Fig. 8) where all the artificial operating even in cases where existing multipoint methods may produce
points are quite close with respect to usual MOPA while the an incorrect diagnosis.
extra two points are far enough to justify the assumption. Still,
some of the artificial points are discrete as observed in the
Nomenclature------------------------------------------------------------------------
zooming area. Therefore, the AMOPA method can be applied
to data coming from neighboring operating conditions for CDP : Compressor Discharge Pressure
which existing MOPA methods are not effective. That the CDT : Compressor Discharge Temperature
method was found to be less successful in accurately predict- EGT : Exhaust Gas Temperature
ing health parameters when combined faults affecting both N : Shaft rpm
compressor and turbines are considered is probably due to NGG : Gas Generator rpm
linearization errors. Ongoing research is focused on the non- NPT : Power turbine rpm
476 A. G. Stamatis / Journal of Mechanical Science and Technology 25 (2) (2011) 469~477