Journal of Software Engineering and Applications, 2013, 6, 87-90
doi:10.4236/jsea.2013.63b019 Published Online March 2013 (
The Research on the Methods of Diagnosing the Steam
Turbine Based on the Elman Neural Network
Junru Gao, Yuqing Wang
Hebei University of Engineering, Handan China.
Received 2013
This paper introduces a kind of diagnosis principle and learning algorithm of steam turbine fault diagnosis which based
on Elman neural network. Comparing the results of the Elman neural network and the traditional BP neural network
diagnosis, the results shows that Elman neural network is an effective way to improve the learning speed , effectively
suppress the minimum defects that the traditional neural network easily trapped in, and shorten the autonomous learning
time. All these proves that the Elman neural network is an effective way to diagnose the steam turbine.
Keywords: Steam Turbine Fault Diagnosis; Elman Neural Network; BP Neural Network
1. Foreword
The fault of steam turbine is a complicated system which
is nonl near and multivariable identification. If using
traditional mathematical model to diagnose its fault, be-
cause traditional mathematical model can only reflect
linear law of steam turbine’s fault. When the construction
of system is unknown and nonlinear, the accuracy rate of
diagnosing by traditional athematical model is low,
which may lead to misdiagnose. As for the complicated
relation between the sign and type of steam turbine fault,
Elman neural network provide a new way to deal with
steam turbine’s fault. This paper discusses the method of
diagnosing steam turbine’s fault based on Elman neural
network and BP neural network.. And proved the training
and result of these two neural network by doing simu-
lation example, which result in compared with BP neural
network, if using Elman neural network , it will be easier
to change structural parameter, take shorter training time
and have reliable performance, which is better than BP
neural network to apply to diagnosing fault.
2. BP Neural Network Structure and
Due to steam turbine’s generator set has a complicated
construction system, multiple failure mode and many
sign of fault signature. We can diagnose the fault of
steam turbine’s generator set by analyzing these sign of
fault signature, and by means of formerly experience,
more than 90% of all kinds of fault of steam turbine can
be found by diagnosing vibration fault of steam turbine.
This paper refers to the related research, making ten
regular fault as analysis object. The form 1 provides fre-
quency spectrum’s value distribution of fault signature
on 0.01/0.39f0.40/0.49f0.5f0.51/0.99ff2f3/5f
odd times f and more than 5f, which were given after
normalization all ten regular fault information. We want
to make it as a training sample for the method of diag-
nosing neural network model’s fault.
Using the data of Table 1 as input sample and Table 2
as output sample to have training. This paper takes 3
layer BP neural network model which has 9 input layer
neuron numbers: x1, x2,x3,…,x9, Inputting the vibration
fault information after pretreatment. The number of neu-
rons in the output layer is 10, they are Rotor unbal-
ance(F1), Aerodynamic coupling(F2),shaft misalign-
ment(F3), oil whirl(F4), Rotor adia lub(F4),Symbiotic
Loosening(F6), Thrust bearing damage(F7),surge (F8),
Bearing pedestal looseness (F9),Range bearing stiffness
(F10). According to 12
 and the adjust-
ment after practical training, The number of hidden layer
neurons is 12, Training error is 0.01, learning rate is 0.2,
excitation function of hidden layer neuron is S function,
The excitation function of the output layer neuronsis
purelin function that is linear function, the function of
training is TRAINSCG, but the disadvantage of it is
when the training is misconvergence, it will stop training,
and takes less time than other algorithm.
Foundation item: natural science funds of Hebei province (F2012
402021), Junru Gao (1969-), female, associate professor, Master In-
structor, Mainly engaged in power and electrical engineering teaching
and research work.
n order to explain the application of BP neural network
in the fault diagnosis of steam turbine is reasonable,
Copyright © 2013 SciRes. JSEA
The Research on the Methods of Diagnosing the Steam Turbine Based on the Elman Neural Network
Table 1. The sorts of fault diagnosis for steam turbine and results of spectrum analysis.
Fault samples F1 F2 F3 F4 F5 F6 F7 F8 F9
frequency range 0.01/0.39f 0.40/0.49f 0.5f 0.51/0.99ff 2f 3/5f Odd times f high frequency >5f
Rotor unbalance 0.00 0.00 0.000.00 0.900.050.050.00 0.00
Aerodynamic coupling 0.00 0.30 0.100.60 0.10
Shaft misalignment 0.00 0.00 0.000.00 0.400.500.100.00 0.00
Oil whirl 0.10 0.80 0.000.10 0.00
Rotor radial rub 0.10 0.10 0.100.10 0.10
Symbiotic loosening 0.00 0.00 0.000.00 0.25
Thrust bearing damage 0.00 0.00 0.100.90 0.00
surge 0.00 0.30 0.100.60 0.00
Bearing pedestal looseness 0.90 0.00 0.000.00 0.00
Range bearing stiffness 0.00 0.00 0.000.00 0.000.800.200.00 0.00
Table 2. The outputs of ANN fault diagnosis.
The output of the network layer
Fault samples
F1 F2 F3 F4 F5 F6 F7 F8 F9
Rotor unbalance 1 0 0 0 0 0 0 0 0
Aerodynamic coupling 0 1 0 0 0 0 0 0 0
Shaft misalignment 0 0 1 0 0 0 0 0 0
Oil whirl 0 0 0 1 0 0 0 0 0
Rotor radial rub 0 0 0 0 1 0 0 0 0
Symbiotic loosening 0 0 0 0 0 1 0 0 0
Thrust bearing damage 0 0 0 0 0 0 1 0 0
surge 0 0 0 0 0 0 0 1 0
Bearing pedestal looseness 0 0 0 0 0 0 0 0 1
Range bearing stiffness 0 0 0 0 0 0 0 0 1
making turbine vibration fault data of document 3 and 4
as the BP neural network calibration sample, that is The
input of the neural network is: [0.39 0.07 0.00 0.06 0.00
0.13 0.00 0.00 0.35]. On this data set by the establish-
ment of the BP neural network model for fault diagnosis
we can get the output layer of the network output is:
[0.78 0.91 0.64 0.49 0.00 0.89 0.47 0.01 1.00 0.30].
Therefore we can determine the fault type of steam tur-
bine is the ninth kind of fault, which has the same con-
clusion with document 3 and 4. BP neural network train-
ing error curve as shown in Figure 1.
Figure 1. The training error curve of BP neural network.
Copyright © 2013 SciRes. JSEA
The Research on the Methods of Diagnosing the Steam Turbine Based on the Elman Neural Network 89
Table 3. He result of initial diagnosis of BP net and Elman net.
Fault type
Fault diagnosis method
F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 m(α)
BP net 0.11 0.13 0.09 0.07 0 0.13 0.07 0 0.14 0.04 0.22
Elman net 0.09 0.10 0.09 0.06 0 0.11 0.07 0.05 0.12 0.08 0.23
Figure 2. The training error curve of Elman neural net-
3. Elman Neural Network Model Structure
and Algorithm
Elman neural network is a kind of feedback neural net-
works, which has better nonlinear fitting ability than ratio
of forward neural network. Elman neural network intro-
duced the following layer on the basis of BP neural net-
work. As the network storage unit, it is used to store last
output function. The structure of Elman neural network
consisted of input layer, hidden layer, undertake layer
and output layer. Compared with BP neural network, On
the one hand Elman neural network’s hidden layer output
is delivered to the output layer neurons, on the other,
transferred to undertake layer neurons to storage, during
the next output in the hidden layer neurons prior to join-
ing a hidden layer output effect. As a result of following
the joining layer, Elman neural network Improve the
overall stability compared these two neural network.
On the other hand, Elman neural network learning al-
gorithm based on the improved BP algorithm, that is
adaptive learning rate momentum gradient descent back-
propagation algorithm. Due to it adopt this kind of algo-
rithm, when training on the same training sample, it has
better astringency than BP neural network and can also
avoid network training being into local minimum .
4. The Elman Neural Network Structure and
Using the same network settings in BP neural network
model, taking the data of form 1 and 2 as sample to have
a training on Elman neural network. Similarly, making
turbine vibration fault data of document 3 and 4 as the
Elman neural network calibration sample, On this data
set by the establishment of the BP neural network model
for fault diagnosis we can get the output layer of the
network output as shown in form Table 3. Compared the
result we can find that two fault diagnosis methods are
equally effective in Steam turbine fault diagnosis. Elman
neural network training error curve as shown in Figure
Compared Figures 1 and 2, we can see the training
curve of Elman neural network model is more smooth
than BP neural network. And training time is greatly re-
duced, reduced from 82 to 41, significantly speeding up
the convergence rate .
5. Conclusion
This paper uses Elman neural network learning algorithm
to diagnose fault of steam turbine, compared with BP
neural network, which enhance the training speed and the
stability of the whole system. Through the examples of
simulation validated the training and result of these two
neural network, also proved that Elman neural network
can achieve the anticipated target on recognizing typical
type fault of Steam turbine, overcome static feedforward
neural network convergence and the disadvantage of eas-
ily to fall into local minimum.
[1] Sichuan Province Electric Power Industry Bureau,
Sichuan Province Electric Power Education Association.
Vibration of steam turbine generator units[M]. Beijing:
China Electric Power Press, 1998.
[2] Jianhua Zhang, Guolian Hou, Xiaogang Sun, Guili Yuan,
Using a probabilistic neural network fault diagnosis
method [J], Power engineering, 2005, 25 (5):698~701
[3] Zhihong Yao. Kohonen Network in the steam turbine
vibration fault diagnosis using [J]. Turbine technology,
2004, 46 (1):67~68
[4] Heji Yu, Changzheng Chen, etc. Intelligent diagnosis
based on neural network [M]. Beijing : Metallurgical
Industry Press, 2000
[5] Akbari A A, et al. Induction motor identification using
Elman neural network[J].WSEAS Transactions on
Copyright © 2013 SciRes. JSEA
The Research on the Methods of Diagnosing the Steam Turbine Based on the Elman Neural Network
[6] WangL, ShiX, LiangY.An improved Elman neural network
with profit factors and its applications[A]. International
Con-ferenceon Intelligent Computing, CIC2006[C],
Aug16-19 2006, Springer Verlag, Heidelberg, D-69121,
Germany, Kun-ming, China, 2007, 317~322
[7] Huang Jinying, Pan Hongxia.The Research on Condition
Detection and Fault Characteristic in one Vehicle Wheel
Box. Proceedings of the 5th International Symposium on
Test and Measurement,2009.
Copyright © 2013 SciRes. JSEA