Energy and Power En gi neering, 2011, 3, 513-516
doi:10.4236/epe.2011.34062 Published Online September 2011 (http://www.SciRP.org/journal/epe)
Copyright © 2011 SciRes. EPE
Fault Detection and Isolation Based on Neural Networks
Case Study: Steam Turbine
Djamel Benazzouz, Samir Benammar, Smail Adjerid
Solid Mechanic and Systems Laboratory (LMSS), M’Hamed Bougara University, Boumerdès, Algeria
Received February 9, 2011; revised March 20, 2011; accepted April 8, 2011
The real-time fault diagnosis system is very important for steam turbine generator set due serious fault re-
sults in a reduced amount of electricity supply in power plant. A novel real-time fault diagnosis system is
proposed by using Levenberg-Marquardt algorithm related to tuning parameters of Artificial Neural Network
(ANN). The model of novel fault diagnosis system by using ANN are built and analyzed. Cases of the diag-
nosis are simulated. The results show that the real-time fault diagnosis system is of high accuracy and quick
convergence. It is also found that this model is feasible in real-time fault diagnosis. The steam turbine is used
as a power generator by SONELGAZ, an Algerian company located at Cap Djinet town in Boumerdes dis-
trict. We used this turbine as our main target for the purpose of this analysis. After deep investigation, while
keeping our focus on the most sensitive parts within the turbine, the weakest and the strongest points of the
system were identified. Those are the points mostly adequate for failure simulations and at which the de-
signed system will be better positioned for irregularities detection during the production process.
Keywords: Failure, Diagnosis, Artificial Neural Networks, Isolation, Steam Turbine
The first role of the industrial diagnosis is to increase the
availability of the industrial installations to reduce the
direct and indirect maintenance costs of the production
equipments. The direct costs of this maintenance are
mainly those related to the various spare parts. On the
other hand, the indirect costs are essentially due to the
off line production [1,2]. The increase repair time influ-
ences negatively on the indirect maintenance costs.
The objective of this paper is to minimize this waiting
time for detecting the failure in the industrial installa-
tions. The proposed model will supervise the system,
detect and localize any faulty in real time. An important
characteristic of the proposed model is that it has the
possibility of detecting and locating several failing points
at the same time. For example: an increase in the vibra-
tion level in the four landings of the turbine. The data
vectors for the training in the Artificial Neural Network
(ANN) model are intervals limited by two values, mini-
mum and maximum. The used symbol “1” represents a
normal functioning and the symbol “–1” represents a
failure situation. The training algorithm used for the
network is the Levenberg-Marquardt algorithm, the choice
of this algorithm is that it gives a fast training of the
ANN compared to the other algorithms of decent of gra-
dient [3,4]. The programming was completely developed
under MATLAB 7.5.
2. Steam Turbine Presentation
The study case concerns a steam turbine of an Algerian
electrical production thermal plant SONELGAZ located
at Cap-Djinet, Boumerdes. The turbine transforms the
thermal energy contained in the vapor coming from the
boiler into a rotation movement of the tree. Mechanical
work obtained is used to actuate the alternator. It is
composed of three bodies, HP body (High Pressure), MP
body (Average Pressure) and BP body (Low Pressure). It
has a power and a nominal number of revolutions of 176
MW and 3000 rpm respectively. The line of tree rests on
four landings, each one of these landings thus carries two
relative vibration sensors, it is the total of eight sensors
on all the line of tree, but for model simplification we
consider only four sensors. The maximum value of rela-
tive vibrations that can be supported by the system is 120
μm. Figure 1 represents the supervision and placement
sensors site in the landings turbine.
D. BENAZZOUZ ET AL.
Figure 1. Sensor vibrations site in the turbine landings C1,
C2, C3 and C4 are respectively the supervision sensors of the
tree relative vibrations compared to landings 1, 2, 3 and 4.
3. Calculation of the Optimal Neural
There exist many applications of ANN in industry par-
ticularly in data analysis, model and command identifi-
cation [4,5]. Among various types of ANN, the most
used is the multi-layer perceptron (MLP) which is re-
tained in our application as a powerful tool. A network
MLP is generally composed by one neural entry layer,
one or more hidden in the intermediate layers and one
output layer. Figure 2 shows the proposed architecture
model where the input vector is ,
with M = 951, the hidden layer neurons varies from j to
N with N = 15 and then the output vector
,,,Yyyyy. The general laws for the calculation
of the hidden layer neurons and the output layer are re-
f1, f2 are respectively the sigmoid activation (3) and
linear functions (4).
The network training MLP implies to find connection
weights values w1i,j and w
2l,k, which reduce to the
minimum the average error function (MSE) between the
measured value and the theoretical (desired) value cor-
responding in the training step. The network training by
the Levenberg-Marquardt algorithm is constructed di-
rectly using Matlab neural network toolbox [6,7].
The optimal architecture, after several trials, was found
by this configuration “951-12-04” which gives the small-
est error of 1.90756e−005, after 111 iterations, during
Figure 2. Proposed network structure.
n: number of examples.
1) Flowchart for the Calculation of Optimal Archi-
W, b: are the weight matrix and the bias vector (the
initialization of the weight and bias values is taken ran-
domly between [–1, +1]). The following symbols are
“it”: iterations number,
Ei:(internal excitation) desired error,
µ: training rate (between 0 and 1),
f1: activation function of the hidden layer,
f2: activation function of the output layer,
Aap: training algorithm,
NNS: neurons number in the output layer,
p: the step,
Ni: desired number,
R: correlation coefficient,
RG: generalized network coefficient,
MSE: mean square error,
Tap: training time,
Ti: training desired time.
To obtain optimal architecture we must firstly vary the
neurons number of the hidden layer (N) to 1 until reach-
ing the desired number (Ni) to satisfy condition MSE ≤
(Quadratic average error ≤ desired error), if there is no
value of N which satisfies this condition, we must return
to the network parameters to fix another initials values.
In the second step we observe the behavior of the net-
work and identify all the correlation factors which should
be at “1”. If it is not we should increase the desired error
(Ei). The last step is the generalization phase. It consists
of calculating the correlation coefficient of the general-
ized network RG which must be close to “1” as shown
by the developed flowchart in Figure 3.
Copyright © 2011 SciRes. EPE
D. BENAZZOUZ ET AL.515
Figure 3. General flowchart for the choice of an optimal
2) Network Training
The training graph (Figure 4) converges towards an error
of the order of 10–4 after 111 iterations during 113.86 sec.
3) Output Network Simulation
We observe in Figure 5, the network graphs in con-
tinuous line, superpose with the desired functions graphs
in dashed line. This allows us to say that it is an accept-
able architecture since the error is of the order of 10–4.
Vrp1: relative vibrations for landing 1,
Vrp2: relative vibrations for landing 2,
Vrp3: relative vibrations for landing 3,
Vrp4: relative vibrations for landing 4,
Svp1: desired output of the relative vibrations of land-
Svp2: desired output of the relative vibrations of land-
Svp3: desired output of the relative vibrations of land-
Svp4: desired output of the relative vibrations of land-
Figure 4. Training of the founded neur al ne twork.
Figure 5. Output network simulation.
4) Fault Detection and Localization (Test of the
To be sure that the proposed network will detect any
faulty we will try to inject a known faulty and see the
behavior of the network. This is presented in the follow-
Example 1: We consider 3 correct values and 1 incor-
rect value, for landing 1 is 160 μm, for landing 2 is 110
μm, for landing 3 is 100 μm and 90 μm for landing 4.
Figure 6 shows the unacceptable value test, where the
curve does not superpose the learned one.
Example 2: In this example we will generalize our test
to give random intervals which do not belong to the ac-
ceptable interval. Instead of considering value, now we
consider a set of values. We test by set values [120 μm -
239 μm] for Vrp1, [300 μm - 419 μm] for Vrp2, [120 μm
- 239 μm] for Vrp3 and by [400 μm - 519 μm] for Vrp4.
The network output detects four alarms in the four points
of measurement as shown in Figure 7 due to an unac-
ceptable vibrations values given to the network.
Copyright © 2011 SciRes. EPE
D. BENAZZOUZ ET AL.
Copyright © 2011 SciRes. EPE
Figure 8. Performance of the proposed network.
Figure 6. Unacceptable value test.
work maintenance operations by reducing both the time
and cost of troubleshooting.
 M. Basseville and M.-O. Cordier, “Surveillance et Diag-
Nostic de Systèmes Dynamiques: Approche Complémen-
taire du Traitement de Signal et de l’Intelligence Artifi-
Cielle,” Rapport INRIA, Thèmes 3-4, No. 2861, 1996, pp.
 N. Palluat, D. Racoceanu and N. Zerhouni, “Utilisation
des Réseaux de Neurones Temporels pour le Pronostic et
la Surveillance Dynamique,” Etude Comparative de Trois
Réseaux de Neurones Récurrents, RSTI-RIA, Vol. 19, No.
6, 2005, pp. 911-948.
 M. Bouamar and M. Ladjal, “Système Multicapteur
Utilisant les Réseaux de Neurones Artificiels pour la
Surveillance des Eaux Potables,” 4th International Con-
ference: Sciences of Electronic, Technologies of Infor-
mation and Telecommunications, LASS, Laboratoire
d’Analyse des Signaux et Systèmes, Université de M’sila,
Algérie, 25-29 March 2007.
Figure 7. Unacceptable vibrations under a set of values test.
5) Performance Evaluation
In Figure 8 we show the correlation coefficient “R”
which tends towards “1”, this fulfill the condition corre-
lation requirement as mentioned before and confirm that
the output values are compared to the desired output
values in the network. This is represented in Figure 8 by
superposing the continuous line which is the best linear
adjustment fit between the targets and the network out-
puts (dashed line).
 M. I. A. Lourakis, “A Brief Description of the Leven-
berg-Marquardt Algorithm,” Implemented Institute of
Computer Science Foundation for Research and Tech-
nology—Hellas (FORTH) Vassilika Vouton, Heraklion,
 J. Karim, “Surveillance, Diagnostic et Pronostic en
Temps Réel de Systèmes Hybrides: Application à des
Bancs d’Essais CERTIA,” LAAS-CNRS Groupe DISCO
7, Toulouse, 2007.
In this investigation, we have proposed MLP architecture
for detecting and locating any faulty in the monitored
system using the Levenberg-Marquardt algorithm. That
is due to its speed and training availability.
 T. Alani, “Réseaux de Neurones Tutorial en Matlab,”
Département Informatique ESIEE-Paris, Paris, 2008.
 M. S. Patil, Jose Mathew and P. K. R. Kumar, “Bearing
Signature Analysis as a Medium For Fault Detection: A
Review,” Journal of Tribology, Vol. 130, No. 1, 2008, pp.
The suggested network system will help the mainte-
nance team to better localize and automatically identify
the sources of failures. This way, all they will have to do
is to fix the defective parts, which will improve the net-