DIAGNOSIS OF COMBINED CYCLE GAS TURBINE USING FEED FORWARD
NEURAL NETWORKS
5. Camporeate, L. Dambrosio, A, Milella, M. Mastrovit, B. Fortunato
Distec.
‘Sea Macchine el ners
Politecnico di Ba
Via Re Davia 200, 70125 BARI- ITALY
Emil camporele@polbait
ABSTRACT
A diagnostic tool based on Feed Forward Neural Networks (FFNN) is proposed to detect the
‘rin of performance degradation in 8 Combined Cycle Gas Tusbine (CCGT) power plant
[In such a plant, due the connection of the steam cycle tothe gas turbine, any deterioration of
28 turbine components affects the gas turbine itself but also the steam eyele. At the same
time, fouling of the heat recovery boiler may cause the increase ofthe turbine back: pressure,
reducing the gas turbine performance. Therefore, the fault variable set used for detecting
Faults in the gas turbine can include measurements taken from the steam cycte
The diagnostic tool is composed of wo FFNN stages the frst network stage pre-processes
fault data in order to evaluate the influence of the single fault variable on the single fault
condition. The second FFNN stage detects the fault condition. The influence of the set of
‘measurements chosen to constitute the neural network training is analysed, taking into
account also the effect of uncertainties in the instrumentation, Tests with simulated data
show thatthe diagnostic tool i able to recognize single faults onthe CCGT with a high rate
of success, in case of fll fault intensity, even in presence of uncertainties in measurements,
In case of partial fault intensity, faults concerning gas turbine components. and the
superbeater are well recognized, while false alarms occut for the other steam plant
‘component faults, in presence of uncertainties in daa,
Finally, some combinations of faults, belonging citer to the gas turbine or the steam plan,
have been examined. In this case, the network is applied twice: in the frst step the amount
ofthe faut parameters that originate the primary fault ae estimated; in the second step, the
diagnostic tool curls the contribution of the main fault to the fault parameters, and the
Aiagnostic process is reiterated. In the examined fault combinations, the diagnostic tool vas
able to detect at leat one of the two faults, even in presence of uncertainty in measurements
and partial Fault intensity.
Norartoy
G mtfowrac : [
N Fano ed
Pome Abbrevitions
pre ANN” Aiea! Ne Nework
}petomane fr CCT Conined ye Gs Trine
1 Ener FANN Feo Fv Neal Nes
Urea raster cot Gh Gos tain
oreo HRSG Hen Recon Stam Gente
Por meer IG Int Cade Vane
SF CORED URALEAT-Gmenzot ISTLHV Low Heating Value sup supetheater
RH Relative Humidity evap evaporator
ST Steam Turbine econ economizer
‘TET Turbine Entry Temperature d— design point
exh exhaust gas
Subscripts is isentropic
0. ambient condition st steam mass Mow rate
b fuel stack stack gas
compressor T turbine
CC combined cycle
ced compressor discharge
1 INTRODUCTION
‘hn efficient diagnostic tool, able to detect fault conditions at the early stages of thei
‘panifestaion. assumes an essential role for condition monitoring in gas turbine power
Plants. This kind of application allows t0 increase the plant components reliability and
Frailabity because of the deterioration process results in loss of power and thermal
efficiency which together contribute 10 lower revenue and to reduce profitability. Im
Combined cycle and cogeneration power plants the degradation of gas turbine components
involves not only the gas turbine engine itsel? but also the heat recovery system and the
‘ottoming steam eycle (Boccaeiti eta, 2000: Zovebek and Pilidis, 2001-pat.).
‘Various techniques have been proposed and are curently applied in onder to determine the
health sate of stationary gas turbines on the basis of the data collected from the engine
jnstrumentation, Some authors investigated the application of the “trend analysis" together
‘with “inverse design” techniques (Bettocchi and Spina, 1999 Bettocchi eta, 2001; Pinel
‘and Venturini, 2002: Benvenuti, 2001)
‘The atfiial ural networks are also very attractive for gas turbine diagnostic, due to their
ability to recognize fault paterns, despite the presence of noise. In particular, due to its
flexible and straightforward structure, the Feed Forward Neural Networks (FFNN) seem to
bre very stitale for fault detection, DePold and Glass (DePold et al, 1999) investigated
oural network methods for automated trend detection, Gluch proposed a neural network
‘with back-propagation earning algorithm for thermal turbines diagnosis (Gluch et al, 19%;
Gluch etal, 1997) and Lu (La et al, 2001) investigated the isolation of single fault for an
‘ero engine using Artificial Neural Networks (ANN). Probabilistic Neural Network (PNN}
‘vith a feed! forward structure has been proposed for their capability (0 integrate statistical
Jrformation about the engine faults (Romessis et al, 2001). A condition monitoring by
jeans ofan ato associative back propagation neural network has heen also proposed (Pong
erat, 2000), Volponi (Valpon eta, 2000) compared the Kalman filter (KF) method with a
puck propagation neural nctwork for engine diagnosis, The simplicity and the high level of
ffeodom led the authors to propose a diagnostic tool, based on FTN theory to a heavy-duty
gas turbine (Dambrosio era, 2001).
‘The main aim of the present paper is to analyse the problem of fault recognition in &
combined cyele power plant composed by an heavy-duty gas turbine and a single pressure
tottoming seam eyele. In the first past ofthe paper, the effects ofthe deterioration of gas
forhine, heat exchangers and steam turbine components, on the measured parameters are
tstimated, through a mathematical rode able to simulate the power plant under deteriorated
Conditions Since the performance and measured variables depend on the operating poi
the analysis has been caried out under the hypothesis that the gas turbine is suitably185
controlled by & system able to maintain constant the turbine exhaust temperature and the
power output, Some selected faults have been modelled and the influence coefficient matix,
{that collects the fingerprints of the selected faults, is presented and discussed
In the socond part, the proposed FENN-based diagnostic tool is described and applied to
detect the considered single fault and also two diferent fauls that occur contemporstiy
‘The capability ofthe diagnostic tool of recognizing the fault itensity is also analysed. The
architecture and the complexity of the Arificial Neural Network (ANN) for the given
Aiagnostc problem are firstly examined. The authors propose a FFNN-base diagnostic tol
that requires limited computational resources to be trained, Such diagnostic tool is composed
by a set of FFNNs organized in two stages, The fits stage is specialized to evaluate the
influence of each fault condition on each fault variable. The second stage operates a post.
processing ofthe output coming from the first FENNs stage and detects the CCGT faults,
‘The described FFNN-based diagnostic too constitutes the fst tool proposed bythe authors,
A second tool, detiving from the first achieves the detection of combined fault conditions in
{wo different steps, In such a two-step strategy the proposed diagnosis algorithm firstly
«detects the mos likely fault condition, estimating also its intensity: secondly, the diagnosis
‘ool subtracts the influence of the detected fault condition from the global influence ofthe
‘combined fault condition and then the detection proces is reiterated,
2. FAULTS ANALYSIS
‘This section investigates the effects of degraded components of a CCGT povser plan. The
Plant i composed of asingle-shatt heavy-duty zas turbine and a bottoming steam plant with
a single pressure heat recovery steam generator. A sketch of the plant is
where the measured variables are also schematically indicated,
‘The gas turbine modelled in this work is the PG6S4-B, manufactured by Nuovo Pignone
(Nuovo Pignone, 1990) For simplicity, a single pressure HRSG has been ehosen, instead of
4 more efficent and often adopted double pressure scheme with deaerator. The knowledge
of GT and HRSG characteristics is necessary to evaluate the off-design behaviour under
Ueteriorated conditions. A preliminary power plant design has been carried out in onler to
82 the parameters needed forthe off-design analysis. As far as the gas turbine is concemed,
4 limited set of information is obtained from the manufacturer's technical pon (Neovo
Pignone, 199%) and from commercial simulation tools like GateCycle (Enter Software Ine
1990), The numerical simulation of the gas turbine, under both on-design and offdesign
conditions, has been earred out by means of a code, previously developed by two of the
authors (Camporeale and Fortunato, 1998). A conventional design procedure, based on
‘signed values of pinch-point and approach point temperature deltas has heen adopted in
‘order to determine the heat exchanger surfaces of the HRSG, Global performance and min
characteristics resulting from the design procedure, summarized in Tab. 1, are referred tothe
‘nominal working poi,
For the off-design analysis, the compressor is modelled using a normalized map given in
{Enter Software Ine. 1990), that provides correted mas flow rate and isentropic eficency
4 a function of corrected rotational speed and pressure ratio. To take into account the
Position of variable inlet guide vanes (IGV), the data given by the normalized map are
‘modified according to the formulas given in the same previous reference, A view of the
compressor map obtained by means of such a procedure is given in the Fig2, where the
GV position effect is shown, The off-design expansion, instead, is examined by means of a
siage-by-stage procedure hased on the velocity diagram evaluated at the piteh-line from
onstructive data of blade angles and eeosssetional are.