Applied Thermal Engineering 75 (2015) 648e657

Contents lists available at ScienceDirect

Applied Thermal Engineering

journal homepage: www.elsevier.com/locate/apthermeng

Optimization of operating conditions for steam turbine using an

articial neural network inverse
ndez b, Victor Salazar b
Y.El. Hamzaoui a, J.A. Rodrguez b, *, J.A. Herna
a
rez (UACJ),
n, CU, Universidad Auto
noma de Ciudad Jua
Instituto de Ingeniera y Tecnologa, Dpto. Ingeniera El
ectrica y Computacio
rez, Chihuahua, Mexico
Av. Del Charro # 450 Norte, CP 32310, AP 1594-D Ciudad Jua
b
n en Ingeniera y Ciencias Aplicadas (CIICAp-UAEM), Av. Universidad #1001, Col Chamilpa, CP 62209 Cuernavaca, Morelos, Mexico
Centro de Investigacio

h i g h l i g h t s

The failure assessment in blades is optimized using articial neural network inverse (ANNi).

(ANNi) is a very effective modeling the useful life in blades of steam turbines.

Failure assessment in blades is optimized using articial neural network inverse.

a r t i c l e i n f o

a b s t r a c t

Article history:
Received 30 June 2014
Accepted 23 September 2014
Available online 2 October 2014

The useful life (UL) of the failure assessment in blades of steam turbines is optimized using the articial
intelligence. The objective of this paper is to develop an integrated approach using articial neural
network inverse (ANNi) coupling with a Nelder Mead optimization method to estimate the resonance
stress when the UL of the blades is required. The proposed method ANNi is a new tool which inverts the
articial neural network (ANN). Firstly, It is necessary to build the articial neural network (ANN) that
simulates the output parameter (UL). ANN's model is constituted of feedforward network with one
hidden layer to calculate the output of the process when input parameters are well known, then
inverting ANN. The ANNi could be used as a tool to estimate the optimal unknown parameter required
(resonance stress). Very low percentage of error and short computing time are precise and efcient,
make this methodology (ANNi) attractive to be applied for control on line the UL of the system and
constitutes a very promising framework for nding set of good solutions.
2014 Elsevier Ltd. All rights reserved.

Keywords:
Inverse neural network
Optimal parameters
Optimization
Steam turbine failure
Life cycle assessment in blades

1. Introduction
Steam turbines have many applications in various industrial
sectors. However, by common experience blade failures are the
main origin of operational breakdowns in these machines, causing
great economic lost in turbo machinery industry. The turbines are
designed to work in stable operation condition [1e4]. Nevertheless,
failure in blades has been present after a short time period of work.
These failures commonly attributed to resonance stress of the
blades at different stages to certain excitation frequencies. The
expense of downtime and repair is about the millions of dollars [5].
The useful life (UL) is a very important variable for determining the

* Corresponding author. Tel.: 52 7772677638.

E-mail address: [email protected] (J.A. Rodrguez).
http://dx.doi.org/10.1016/j.applthermaleng.2014.09.065
1359-4311/ 2014 Elsevier Ltd. All rights reserved.

performance of steam turbines [6]. Therefore, the critical components which determine the useful life of the turbine should be
evaluated to determine the rehabilitation or replacement of them.
The critical components are the blades of steam turbines [7]. Most
of the existing analytical models used to predict the useful life of
the failure assessment in blades of steam turbines are based on
analysis using analysis of vibrations for the construction of the diagram of Campbell, which shows the natural frequencies of the
blades like a function of the speed of the rotor (RPM) [8]. These
models do not provide reliable predictions for useful life (UL). This
is caused by the complexity of solving the equations that involve
the radiant energy balance, the spatial distribution of the absorbed
radiation, mass transfer, and the mechanisms of steam turbines [5].
Moreover, in the light of the rapid development witnessed by
the modern world in different elds of knowledge, science and
technology, due to the increased speed of complexity of the system,
in response to the issues requiring urgent attention of the people, in

Y.El. Hamzaoui et al. / Applied Thermal Engineering 75 (2015) 648e657

order to check for access with high performance and reliability. In

recent years, computer science and articial intelligence have
received increasing attention proving effectiveness in solving many
issues and outstanding challenges [6]. Articial intelligence has
been used in a wide range of elds including industrial application
system has already proved successful, both on the technical level
and as reservoirs of expertise [7,8].
Therefore in practice an articial intelligence tools such as an
articial neural network (ANN) can provide a new approach to
process without take into account any previous assumptions [5].
ANN is a collection of interconnecting computational elements
which function like neurons in biological brain. It has the ability to
model processes by learning from input and output data, without
mathematical knowledge of the process. The (UL) could be
calculated on-line, when the input variables are well known in
blades of steam turbines. Indeed, the problem is that this (UL)
computed by ANN is not ideal in the system, and therefore it is
necessary that its input variables are well known when a given
(UL) is required, that means Find the effect of this cause. However, the term of neural network inverse can be misleading
sometimes. However, within the scope of this paper is to show the
new strategy which is proposed in our current paper based on
ANN model about the use of articial neural network (ANN) for
modeling the useful life of the failure assessment in blades of
steam turbines. The main idea is coupling the neural network
model with optimization algorithms to estimate the adequate
value of a selected input to obtain the desired output. On the other
hand, to understand very well the difference between inverse and
optimization problem, Marcelo J. Colao and others [9] presented
basic concepts of inverse and optimization problems, and illustrated the inverse and optimization problems are conceptually
different, despite their similarities. According to Marcelo J. Colao
et al., inverse problems are concerned with the identication of
unknown quantities appearing in the mathematical formulation of
physical problems, by using measurements of the system
response. On the other hand, optimization problems generally deal
with the minimization or maximization of a certain objective or
cost function, in order to nd design variables that will result on
desired state variables [9e15].
The problem was developed through inverting an articial
neural network (ANNi) to estimate the optimum input variables on
a required (UL) in the system. However, articial neural networks
inverse (ANNi) have been used successfully in different applications
in which an output desired parameter is selected and then the input
parameters values are determined [6e8,16e29].
The proposed method ANNi is a new tool which inverts articial
neural network (ANN) and it uses an optimization method to nd
the optimum parameter value (or unknown parameter) for given
required conditions in the process. In order to do so, rst, it is
necessary to build the articial neural network (ANN) model that
simulates the output parameters of the failure assessment in blades
of steam turbines is constituted of a feedforward network with one
hidden layer to simulate output, considering one or more wellknown input parameters of the process. LevenbergeMarquardt

Outputk yfkg purelin

S
X
s1

"
LWfk;sg $ tansig

K
X
k1

649

Fig. 1. General neural network inverse model.

learning algorithm, hyperbolic tangent sigmoid transfer-function,

linear transfer-function and several neurons in the hidden layer
(due to the complexity of the process) are considered in the built
model. As soon as the model was validated, the second step was to
invert the model. With the required output and some input parameters it is possible to calculate the unknown input parameters.
However, it is important to note that the analytical solution with one
neuron in the hidden layer neural model exists, and it is described
in Section 4. Nevertheless, in the case that a proposed ANN model
has more than one neuron in the hidden layer it is necessary to use
an optimization method. On the other hand, in many cases, when an
optimal output is required, the optimal input parameters are unknown, that's why, for this reason, we found that the inverse articial neural network (ANNi) is a fundamental strategy to estimate
the optimal operation condition. Rodriguez et al., have performed
the sensitivity analysis to show which parameters have the most
inuence on (UL) [4] in order to optimize them by means of inverse
neural network (ANNi).
The paper is organized as follows second section will give an
overview about articial neural network inverse (ANNi), third
section will discuss the Nelder Mead method, fourth section is
devoted to the optimization approach applied on ANNi, then the
fth section is assigned to the results and discussions. Whilst, the
comparative study is showed in Section 6. Finally, the conclusions
on this work are drawn.

2. Articial neural network inverse (ANNi)

A general neural network is shown in Fig. 1 which is constituted
by hyperbolic tangent (tanh) or sigmoid function (tansig) in the
hidden layer and linear transfer functions in the output layer. Then
the output is given by,

!!#
IWs;k $Ink b1s

!
b2k

(1)

650

Y.El. Hamzaoui et al. / Applied Thermal Engineering 75 (2015) 648e657

According to the denition of purelin and tansig functions are

used, y(k2) is given by,

y2

13

B
S 6
X
B
6
6LW2;s $B
B
6
@
s1 4

2
1 exp

2

PR

r1 IWs;r $pr b1s

C7
C7
7
!!  1C
C7 b22
A5

(2)

Let p(r3) be the input to be estimated and k 2 the required

output y(2) then

2
y2

S
S 6
X
X
6
6
b22 
LW2;s
6
s1
s1 4

3
2LW2;s
1 exp

 2 IWs;3 $p3

PR

rs3 IWs;rs3 $pfrs3g b1s

where subscripts s is the number of neurons in the hidden layer;

subscripts r is the number of neurons in the input layer; l is the
number of neurons in output layer; S is the number of neurons in
the hidden layer; R is the number of input; Tansig is the hyperbolic
tangent sigmoid transfer function; Purelin is linear transfer function; and IW, LW and b1s, b2l are the input and output weights and
the biases, respectively.
The Eq. (4) is to be minimized to zero to nd the optimal input(s)
parameter(s) in a general ANN, in this case, x is the p3 value to be
computed to zero by an optimization method.

7
7
!!7
7
5

method approximately nds a local optimal solution with N variables when the objective function varies smoothly. NeldereMead
generates a new test position by extrapolating the behavior of the
objective function measured at each test point arranged as a simplex. Then, the algorithm chooses to replace one of these test points
with the new test point. Thereby, a new simplex is generated with a
single evaluation of the objective. The numerical algorithm of the
NeldereMead simplex method has been described in detail by
Nelder and Mead [30].

2
f x b22 

S
X

LW2;s  y2

s1

S 6
X
6
6
6
s1 4

(3)

3
2$LW2;s
1 exp

 2 IWs;3 $x

Therefore, optimization can be done using the NeldereMead

method.

PR

rs3 IWs;rs3 $prs3 b1s

7
7
!!7
7
5

(4)

4. Optimization approach
4.1. Neural network learning

3. NeldereMead method
The NeldereMead method is a generally used nonlinear optimization algorithm. This method is a numerical method to minimize to zero an objective function in a multi-dimensional space.
This algorithm is a direct search method that does not use numerical or analytic gradient [27]. It attends to minimize a scalarvalued nonlinear function of n real variables using only function
values, without any derivative information. The method uses the
concept of simplex, which is a polyhedron of N 1 in N dimensions.
Simplices are a line, a triangle and tetrahedron in one-, two-, and
three-dimensional space, respectively, and so forth [28]. The

A learning (or training) algorithm is dened as a procedure that

consists of adjusting the coefcients (weights and biases) of a
network, to minimize an error function (usually a quadratic one)
between the network outputs, for a given set of inputs, and the
correct (already known) outputs as shown in Fig. 2. If smooth nonlinearities are used, the gradient of the error function can be
computed by the classical backpropagation procedure. To determine the best backpropagation training algorithm, ten backpropagation algorithms were studied. In addition, three neurons
were used in the hidden layer for all backpropagation algorithms.
Table 1 shows a comparison of different backpropagation training

Y.El. Hamzaoui et al. / Applied Thermal Engineering 75 (2015) 648e657

651

Fig. 2. Numerical procedure used for the ANN learning process, and the iterative architecture used by the model to predict the UL of the failure assessment in blades of steam
turbines (S is the number of the neuron in the hidden layer).

Consequently, RMSE and R2 were used as the error function

which measures the performance of the network. Therefore, the
network having minimum RMSE and maximum R2 was selected the
best ANN model. More detail about neural network training was
already described by Rodriguez et al. [4].

algorithms. LevenbergeMarquardt backpropagation training algorithm could have smaller mean square error (RMSE), on the other
hand, we found training with Levenberg Marquardt algorithm can
run smoothly in computer with lower expanded memory specication (EMS), and the training time is quickly, than the other
backpropagation algorithms. Because, the LevenbergeMarquardt
algorithm was designed to approach second order training speed
without having to compute the Hessian matrix.
However, the performance of the ANN model was statistically
measured by the root mean square error (RMSE) and regression
coefcient (R2), which are calculated with the experimental values
and network predictions as illustrated in Figs. 3 and 4. These calculations are used as a criterion for model adequacy, obtained as
follows:

RMSE

v
!
u

2
u P
N
u
u
n1 yn;pred  yn;exp
t

In the case of one neuron in the hidden layer. The analytical

solution is performed as follow:

If tansig and purelin are considered as the hyperbolic tangent
sigmoid and linear transfer function in the hidden layer and
output layer, and k 1.

  1A b2


Out1 LW1;1 $@
1 exp  2$ IW1;r $pr b1

(5)

(7)

2
yn;pred  yn;exp
n1
R2 1 
2
PN 
n1 yn;exp  ym
PN

4.2. Inverse neural network considering one neuron in the hidden

layer in ANN model

This can be transformed into:

(6)

1
2
@
  1A


Out1  b2 LW1;1 $
1 exp  2$ IW1;r $pr b1
0

where N is the number of data points, yn,pred is the network prediction, yn,exp is the experimental response, ym is the average of
actual values and n is an index of data.

(8)

Table 1
Comparison of 10 backpropagation algorithms with ve neurons in the hidden layer.
Backpropagation algorithm

Function

Root mean square

error (RMSE)

Epoch

Correlation
coefcient (R2)

Best linear equation

LevenbergeMarquardt backpropagation
Batch gradient descent
Batch gradient descent with momentum
PolakeRibiere conjugate gradient backpropagation
Scaled conjugate gradient backpropagation
BFGS quasi-Newton backpropagation
PowelleBeale conjugate gradient backpropagation
One step secant backpropagation
FletchereReeves conjugate gradient backpropagation
Variable learning rate backpropagation

trainlm
traingd
traingdm
traincgp
trainscg
trainbfg
traincgb
trainoss
traincgf
traingdx

0.00235005
0.01657932
0.01982303
0.03267017
0.44944913
0.48619630
0.50820237
0.02753381
0.01756329
0.02039637

1000
2000
2000
2000
2000
2000
2000
2000
2000
2000

0.990
0.988
0.987
0.979
0.974
0.971
0.965
0.782
0.725
0.718

Y
Y
Y
Y
Y
Y
Y
Y
Y
Y

0.990X 0.306
0.986X 0.927
0.988X 0.837
0.957X 2.53
1.020X  0.783
0.982X 1.23
0.960X 2.03
0.617X45.3
0.425X 34.8
0.386X 38

652

Y.El. Hamzaoui et al. / Applied Thermal Engineering 75 (2015) 648e657

Fig. 3. Schematic methodology.

1
2$LW1;1
  LW1;1 A


Out1  b2 @
1 exp  2$ IW1;r $pr b1

 LW


1;1  Out1 b2

exp  2 IW1;1 $p1 b1

Out1 LW1;1  b2

(13)

(9)
Out1  b2 LW1;1

2$LW1;1




1 exp  2$ IW1;r $pr b1



LW1;1  Out1 b2
2 IW1;1 $p1 b1 ln
Out1 LW1;1  b2

!
(14)

(10)



1 exp  2$ IW1;r $pr b1

LW1;1  Out1 b2
1
IW1;1 $p1 b1  ln
2
Out1 LW1;1  b2

2LW1;1
Out1  b2 LW1;1
(11)

 LW


1;1  Out1 b2

exp  2 IW1;r $pr b1

Out1 LW1;1  b2
As k 1, R 1, So

(12)

LW1;1  Out1 b2
1
IW1;1 $p1  ln
2
Out1 LW1;1  b2

!
(15)

!
 b1

(16)

Let p{r1} would be the input parameter to be calculated when

one output parameter is required. Then:

Y.El. Hamzaoui et al. / Applied Thermal Engineering 75 (2015) 648e657

653



exp  IW1;1 $p1 b1

IW1;1 p1 ln


LW1;1
 1  b1
Out1  b2


b1  ln
p1

Fig. 4. Performance plots of ANN during training, validating and testing of the
network.

LW1;1  Out1 b2
1
p1 
ln
2$IW1;1
Out1 LW1;1  b2

!


b1
IW1;1

(17)

LW1;1
1
Out1  b2

LW1;1
Out1b2

(21)

(22)


1

IW1;1

(23)

Therefore, according to Eq. (1), that it is possible to simulate the

outputs values, when input parameters are well known. However,
in many cases, the problem is that ANN predicted output values
which are not satisfactory in the system, and therefore, it is
necessary that its inputs variables are well known when giving a
required or satisfactory output. Consequently, the new control
strategy which is proposed here using ANN model applied to energy systems. The proposed strategy uses an inverse of neural
network and the NeldereMead optimization algorithm to nd the
optimal input values for the required output value. Then in this
ANNi methodology, as mentioned above, the required output value
is well-known.

5. Results and discussion

If logsig and purelin are considered as the logistic tangent sigmoid and linear transfer function in the hidden layer and output
layer, and k 1

1
1

A b2


Out1 LW1;1 @
1 exp  IW1;1 p1 b1
Out1  b2

LW1;1


1 exp  IW1;1 $p1 b1



1 exp  IW1;1 p1 b1

LW1;1
Out1  b2

Fig. 5. The cracked leading edge of the L-0 blade.

(18)

(19)

(20)

The experimental set up used in this work has been previously

described in detail elsewhere [4]. Failures of turbine blade usually
initiate at the zone of high stress concentration which occur in
metallurgical discontinuities or where corrosion is present or even
in regions of excessive wear. During operation of the turbine, the
cracks are frequently caused by erosion, corrosion or small imperfections and then propagate into the fracture. However, these imperfections increase the fatigue stress concentration factor and of
course the stresses themselves. Turbine inspection revealed that
sets of 10 blades failure of the L-0 in low pressure stage of a 110 MW
steam turbine were illustrated within 15 cm from the root. The
blades are signicantly affected by a crack practically is a fracture
observed in a blade root is shown in Fig. 5. The L-0 stage had 110
blades of 0.6 m in length with groups of 10 blades. The blades of
each group are connected in their top end by a shroud and two
wires as shown in Fig. 6. In addition, a visual inspection combined
with a revision of the turbine operation history was carried out into
system description and experimental data. Furthermore, a turbine

Fig. 6. Discrete models of blade group for the calculation of natural frequencies.

654

Y.El. Hamzaoui et al. / Applied Thermal Engineering 75 (2015) 648e657

ow analysis was also conducted together with verication of the

blade dimensions and tolerances. The natural frequencies and vibration modes analysis of the 10 blades group and the stress
analysis of blades were realized in experimental mode and a nite
element program called ANSYS [2,31e34].
The operation history of the turbine was carefully studied and
revised. The turbine was operated by approximately 1800 h in
intermittent mode, with a record of 650 start-ups, during a period
of 5 years with at least ve replicates of information acquisition to
ensure the verication of measurements a database about 2500
samples was obtained According to the work developed by Garcia
[31], Kubiak [32] and Rodrguez [34].
Rodriguez et al. [5] proposed a neural network model which
demonstrating an efciency of 99% in predicting useful life (UL) of
the failure assessment in blades of steam turbines. This developed
ANN model has three neurons in the hidden layer (21 weights and 4
biases) and considering 6 inputs parameters (resonance stress,
frequency ratio, dynamic stress, damping, fatigue strength, mean
stress). The proposed equation developed by Rodriguez et al. [5] is

2
UL

(25)

where


X1 2 IW1;1 V1 IW1;2 V2 IW1;3 V3 IW1;4 V4

IW1;5 V5 IW1;6 V6 b11

(26)


X2 2 IW2;1 V1 IW2;2 V2 IW2;3 V3 IW2;4 V4

IW2;5 V5 IW2;6 V6 b12

(27)

13

B
S 6
X
B
6
6LW1;s $B
B
6
@
s1 4
1 exp



2LW1;1 2LW1;2 2LW1;3
UL

1 eX1
1 eX2
1 eX3


 LW1;1 LW1;2 LW1;3 b21

 2$

PR

2


r1

IWs;r $pr b1s

C7
C7
7
!!  1C
C7 b2

A5

where the number of neurons in the input layer is 6. According to

the model given by Eq. ( 24), it is possible to simulate the useful life
of the failure assessment in blades of steam turbines, while the
input parameters are well known. Since we found, that the resonance stress is the most inuential parameter [5]. Therefore, it is
important to know in this process, what optimal resonance stress is
needed for a required UL (input number 1). Consequently, we have
developed a strategy to estimate the optimal resonance stress in
the failure assessment process about blades of steam turbines from
the inverse articial neural network (ANNi). In the meantime,
ndez et al. [23] and El-Hamzaoui et al. [18] have been applied
Herna
ANNi with ANN model prediction in order to different processes of
R2 0.915 and R2 0.986, respectively. The authors mentioned that
it is possible to use ANN model with predictions of R2 > 0.985 for
implementing ANNi. In our case the ANN model is of R 0.99.
Consequently, we believe that it is possible to developed ANNi in
this process. The results have been showed that the UL's error
evaluation between the experimental and simulated by ANNi is
0.7%. As we can see very small. In addition, the neural network
model developed by Rodriguez et al. has an efciency of 99%.
Therefore, with this motivation mentioned above, we can use this
model to perform ANNi model.
The proposed method (ANNi) inverts the articial neural
network. Then we have the following equation that calculates UL
during the failure assessment process in blades of steam turbines.
The key information (optimal performance) for useful life of the
failure assessment in blades of steam turbines, when controlling
the required output is to know the optimal input parameters. An
inverted ANN could be considered as a model based method of
supervisory control, the control action in which the unknown input
parameters are obtained by solving an on line optimization problem for the desired output. The inverse of the articial neural
network deduced from Eq. (24) is the following Eq. (25) that calculates UL in the system.

(24)


X3 2 IW3;1 V1 IW3;2 V2 IW3;3 V3 IW3;4 V4

IW3;5 V5 IW3;6 V6 b13

(28)

V1 Resonance stress [MPa]

V2 Frequency ratio
V3 Dynamic stress [MPa]
V4 Damping
V5 Fatigue strength [MPa]
V6 Mean stress [MPa]
UL Useful Life [Min]
At this step, we have obtained the function which has to be
optimized to get the optimal input parameter:

f V1 A

2LW1;1
1 eX11 3:58V1

2LW1;2
1 eX22 34:1V1

2LW1;3
1 eX33 5:5V3
(29)

Table 2
Adjustable parameters obtained (weights and bias) in the proposed model with
S 3, K 6.
IW(s,k)

Wo(s)
b1(s)
b2

Wi(1,1)
1.79
Wi(2,1)
17.05
Wi(3,1)
2.75
Wo(1)
0.19
b1(1)
123.17
0.29

Wi(1,2)
1.14
Wi(2,2)
7.45
Wi(3,2)
0.17
Wo(2)
0.13
b1(2)
17.15

Wi(1,3)
2.53
Wi(2,3)
43.37
Wi(3,3)
3.58
Wo(3)
0.81
b1(3)
6.39

Wi(1,4)
1.21
Wi(2,4)
5.38
Wi(3,4)
0.59

Wi(1,5)
2.37
Wi(2,5)
27.57
Wi(3,5)
0.69

Wi(1,6)
1.17
Wi(2,6)
10.29
Wi(3,6)
12.76

Y.El. Hamzaoui et al. / Applied Thermal Engineering 75 (2015) 648e657

655

Table 3
Some samples of the experimental and simulated information of the system.
Test number
Input
Mean stress
Frequency ratio
Dynamic stress
Damping
Fatigue strength
Resonance stress_Exp
Output
[Useful Life]EXP
[Useful Life]SIMANN

500

1000

2000

3000

4000

5000

6000

7000

8000

9000

557.12
0.91
24.12
0.018
1295.36
244

655.97
1.052
35.41
0.020
967.85
221

483.77
1.017
101.75
0.0201
1186.33
208

562.34
0.902
31.66
0.023
977.33
208

602.87
0.916
30.184
0.0188
923.95
211.5

500.71
1.029
85.985
0.0201
1070.33
216

634.83
0.966
60.516
0.0185
971.45
219

492.83
0.889
21.95
0.0201
1128.079
221

673.60
1.036
60.59
0.0207
883.3121
221.5

566.39
0.937
32.159
0.0198
1038.347
220

4.67E11
4.67E9

4.27E17
4.27E15

4.7855E10
4.7855E8

5.38E14
5.38E12

6.87E14
6.87E12

3.19E14
3.19E12

1.54E14
1.54E12

4.18E14
4.18E12

6.03E12
6.03E10

3.04E13
3.04E11

where

A UL  b21 LW1;1 LW1;2 LW1;3

(30)


X11 2 IW1;2 V2 IW1;3 V3 IW1;4 V4 IW1;5 V5

IW1;6 V6 b11

(31)


X22 2 IW2;2 V2 IW2;3 V3 IW2;4 V4 IW2;5 V5

IW2;6 V6 b12

(32)


X33 2 IW3;2 V2 IW3;3 V3 IW3;4 V4 IW3;5 V5

IW3;6 V6 b13

(33)

The weights (IW and LW) and biases (b1 and b2) of ANN's model
are showed in Table 2 and the input parameters are reported by
Rodriguez et al. [4] in order to minimize to zero the Eq. (29), an
optimization method (NeldereMead Simplex Method) is used to
calculate the resonance stress (V1). Where, the Input of ANN are
frequency ratio; dynamic stress; damping; fatigue strength; mean

stress. Whilst, the required output is the useful life. However, the
resonance stress will be estimated by ANNi.
The optimization method nds the minimum of a scalar function of several variables, starting at an initial estimate. However,
this is generally referred to as unconstrained nonlinear optimization. However, the optimization method starts at the initial value
and nds a local minimum V1 described in Eq. (29). On the other
hand, Table 3, shows some samples of the experimental and
simulated information of the system as shown in Fig. 7.
An example of this application is shown to calculate the
required resonance stress (V1) considering the experimental data
reported by Rodriguez et al. [34,35], which values are obtained
from experimental test database showing in test number 1000,
thus only, we want to calculate the resonance stress value (V1):
Mean stress (V6) 655.97 [MPa];
Frequency ratio (V2) 1.052;
Dynamic stress (V3) 35.41 [MPa];
Damping (V4) 0.020;
Fatigue strength (V5) 967.85 [MPa];
Resonance stress (V1) ?
And an output value, UL 4.27E 17 [Min]
According to the weights and biases of Table 1 and optimization
method of the NeldereMead, it is possible to calculate the optimum
Resonance
Stress
(V1),
which
is:
Resonance
Stress
ANNi V1ANNi 224 [MPa].

Fig. 7. Architecture of the articial neural network inverse for determining the optimum resonance stress.

656

Y.El. Hamzaoui et al. / Applied Thermal Engineering 75 (2015) 648e657

Fig. 8. Experimental resonance stress vs resonance stress by ANNi.

However, to validate this value, a test was performed with

different data to optimize the resonance stress in different conditions to demonstrate the feasibility of this method about ANNi.
However, the simulation outcomes were then compared with
experimental data in order to check the accuracy of ANNi. This error
is given by:

Err 100

jExp  Simj
Exp

(34)

That means Resonance Stress estimated by ANNi is compared to

experimental resonance stress of the experimental data using Eq
(34). In Table 2, within the test number 1000, the experimental
value of Resonance Stress is V1EXP 221 [MPa]. Whilst, the error is
given by Eq (34). Therefore, in this case, an error of 1.5% is obtained
which is very acceptable. On the other hand, the elapsed time to
calculate this Resonance Stress from this methodology (ANNi with
Nelder Mead Simplex) is only 35.17 s. It's seem that this time is good
enough to control the process. In addition, Fig. 8, illustrates that
there is a good agreement between the experimental resonance
stress and resonance stress estimated by ANNi. In the meantime,
the tting quality is so good. It has been an outstandingly successful
models in estimating experimental results by ANNi.

6. Comparative results
The remarkable thing is that, according to Fig. 9, there is good
agreement between the predicted values for useful life of the failure assessment in blades of steam turbines by ANN and ANNi
models with experimental data. Indeed, it has been an outstandingly successful models in predicting the experimental results.
Consequently, the UL error between the experimental and simulated by ANNi is 0.7%.
These models: articial neural network (ANN) and articial
neural network inverse (ANNi) prove to be very effective in
modeling the useful life of the failure assessment in blades of steam
turbines. The smaller RMSE and larger R2 mean better performance
[5]. However, the performance of the ANN and ANNi on modeling
UL of the failure assessment in blades of steam turbines is presented in Fig. 9, where the two models are trained using the same
training datasets and validated by the same datasets (fresh data). In
practice, however, the calculation required for system are so
complicated, that's why, all the calculations were carried out on
LINUX system, Intel D CPU 2.80 Ghz, 2.99 GB of RAM. According
to Fig. 9, we can distinguish the following results: The ANN model
has smaller error for datasets than the ANNi (about 0.85%). In this

Fig. 9. UL versus number of test patterns for failure assessment of steam turbines.

Y.El. Hamzaoui et al. / Applied Thermal Engineering 75 (2015) 648e657

way, the ANN achieves better performances than the ANNi model.
Therefore, ANN is a good choice for modeling the useful life of the
failure assessment in blades of steam turbines. It is believed also,
that ANN and ANNi could be used to handle many other types of
problems about the failure assessment in blades of steam turbines.
7. Conclusions
The useful life (UL) of the failure assessment in blades of steam
turbines is optimized using articial neural network inverse in order
to calculate an ideal input value from an ideal (UL) and taking into
account the above well known input values excepting required input
value as resonance stress. Then, NeldereMead method is applied in
the inverse problem to optimize the optimal operating condition is
tested for a single parameter. Thanks to this method, it is possible to
nd any unknown input variable on line in the engineering failure
analysis in blades of steam turbines. Indeed, it is very important to
note that the elapsed time to calculate the optimum input parameter is only a few seconds (<40 s), thus it is feasible to get optimal
parameters on line and is sufciently suitable to direct control of
steam turbines. Briey, ANNi integrated with NeldereMead method
signicantly reduced the computational time with better convergence for optimal solution for useful life of the failure assessment in
blades of steam turbines. Despite its successes, ANNi, is still in its
infancy. It's part of the future. In a way it's amazing we have done so
much with so little, and we have barely begun. However, if there are
many input parameters to be found (solution to multi-parameter
problems) then NeldereMead method couldn't be able to solve
the optimization problem. It would be recommended to use another
advanced techniques for solving optimization problem, such as genetic algorithms (GAs) and particle swarm optimization (PSO).

[10] M.J. Colao, G.S. Dulikravich, T.J. Martin, Optimization of wall electrodes for
electro-hydrodynamic control of natural convection effects during solidication, Mater. Manuf. Process 19 (4) (2004) 719e736.
[11] M.J. Colao, G.S. Dulikravich, T.J. Martin, Optimization of wall electrodes for
electro-hydrodynamic control of natural convection effects during solidication, in: Proceeding ASME International Mechanical Engineering Congress &
Exposition, 2003.
[12] M.J. Colao, G.S. Dulikravich, T.J. Martin, Reducing convection effects in solidication by applying magnetic elds having optimized intensity distribution, in: Proceeding ASME Summer Heat Transfer Conference, 2003.
[13] M.J. Colao, H.R.B. Orlande, G.S. Dulikravich, F.A. Rodrigues, A comparison of
two solution techniques for the inverse problem of simultaneously estimating
the spatial variations of diffusion, in: Proceeding ASME International Mechanical Engineering Congress & Exposition, 2003.
[14] M.J. Colao, H.R.B. Orlande, Inverse forced convection problem of simultaneous estimation of two boundary heat uxes in irregularly shaped channels,
Numer. Heat Transf. A Appl. 39 (2001) 737e760.
[15] J. Marcelo, M.J. Colao, H.R.B. Orlande, Inverse natural convection problem of
simultaneous estimation of two boundary heat uxes in irregular cavities, Int.
J. Heat Mass Transf. 47 (2004) 1201e1215.
s, G. Urquiza, J.A. Hern
[16] O. Corte
andez, Optimization of operating conditions for
compressor performance by means of neural network inverse, Appl. Energy
86 (11) (2009) 2487e2493.
[17] J.A. Hernandez, A. Bassam, J. Siqueiros, D. Juarez-Romero, Optimum operating
conditions for a water purication process integrated to a heat transformer
with energy recycling using neural networks inverse, Renew. Energy 34 (4)
(2009) 1084e1091.

[18] Y. El Hamzaoui, J.A. Hernandez, S. Silva-Martnez, A. Bassam, A. Alvarez,

[19]

[20]

[21]

Acknowledgements
[22]

J.A. Rodrguez, expresses his gratitude to CONACYT for the

project with title is: Experimental and numerical study for evaluation of reliability and life estimation of turbine blades under
resonance conditions [In Spanish] whit reference number: 156757.
Y.El. Hamzaoui, expresses his gratitude to PROMEP for nancial
support (Project PROMEP/103.5/13/7073.UACJ-PTC-289).

[23]

[24]

References
[25]
[1] C.R.F. Azevedo, A. Sin
atora, Erosion-fatigue of steam turbine blades, Eng. Fail.
Anal. 16 (2009) 2293e2303.
lez, J.C. Garca, F. Sierra, J.G. Nebradt,
[2] J. Kubiak, J.A. Segura, G. Gonza
J.A. Rodriguez, Failure analysis of the 350 MW steam turbine blade root, Eng.
Fail. Anal. 16 (2009) 1270e1281.
[3] Y. Jung, P. Papadopoulos, R.O. Ritchie, Constitutive modelling and numerical
simulation of multivariant phase transformation in superelastic shapememory
alloy, Int. J. Numer. Meth Eng. 60 (2004) 429e460.
[4] D. Taylor, Crack modelling: a technique for the fatigue design of components,
Eng. Fail. Anal. 3 (1996) 129e136.
ndez, J.C. Garca, J.E. Flores,
[5] J.A. Rodrguez, Y.El. Hamzaoui, J.A. Herna
A.L. Tejeda, The use of articial neural network (ANN) for modeling the useful
life of the failure assessment in blades of steam turbines, Eng. Fail. Anal. 35
(2013) 562e575.
lisation et Optimisation en Mise en Forme, Une approche
[6] A. Chamekh, Mode
e sur la Me
thode des Ele
ments Finis et les Re
seaux de Neurone
hybride base
enne Ed,
Articiels: Application 
a l'emboutissage et 
a l'hydrofomage, Europe
2010.
[7] A. Chamekh, H. BelHadjSalah, R. Hambli, A. Gahbiche, Inverse identication
using the bulge test and articial neural networks, J. Mater. Process. Technol.
177 (1e3) (2006) 307e310.
[8] Abdessalem
Chamekh,
Souad
BenRhaiem,
Houda
Khaterchi,
Hedi BelHadjSalah, Ridha Hambli, An optimization strategy based on a metamodel applied for the prediction of the initial blank shape in a deep drawing
process, Int. J. Adv. Manuf. Technol. 50 (1e4) (2010) 93e100.
[9] M.J. Colao, H.R.B. Orlande, G.S. Dulikravich, Inverse and optimization problems in heat transfer, in: Proceedings the 10th Brazilian Congress of Thermal
Sciences and Engineering, 2004.

657

[26]

[27]
[28]

[29]
[30]
[31]
[32]

[33]

[34]

[35]

C. Lizama-Bahena, Optimal performance of COD removal during aqueous

treatment of alazine and gesaprim commercial herbicides by direct and inverse neural network, Desalination 37 (2011) 325e337.
Maamar Laidi, Salah Hanini, Optimal solar COP prediction of a solar-assisted
adsorption refrigeration system working with activated carbon/methanol as
working pairs using direct and inverse articial neural network, Int. J. Refrig.
36 (1) (2013) 247e257.
Nour Hattab, Mikael Motelica-Heino, Application of an inverse neural network
model for the identication of optimal amendment to reduce copper toxicity
in phytoremediated contaminated soils, J. Geochem. Explor. 136 (2014)
14e23.
Yasaman Sanayei, Naz Chaibakhsh, Ali Chaibakhsh, Ali Reza Pendashteh,
Norli Ismail, Tjoon Tow Teng, Long-term prediction of biological wastewater
treatment process behavior via wiener-laguerre network model, Int. J. Chem.
Eng 2014 (2014) 1e7.
Mohammad Rostamizadeh, Mashallah Rezakazemi, Kazem Shahidi,
Toraj Mohammadi, Gas permeation through H2-selective mixed matrix
membranes: experimental and neural network modeling, Int. J. Hydrogen
Energy 38 (2) (2013) 1128e1135.
ndez, W. Rivera, D. Colorado, G. Moreno-Quintanar, Optimal COP
J.A. Herna
prediction of a solar intermittent refrigeration system for ice production by
means of direct and inverse articial neural networks, Sol. Energy 86 (4)
(2012) 1108e1117.
M. Behloul, H. Grib, N. Drouiche, N. Abdi, H. Lounici, N. Mameri, Removal of
malathion pesticide from polluted solutions by electrocoagulation: modeling
of experimental results using response surface methodology, Sep. Sci. Technol. 48 (4) (2013) 664e672.
Ian Buchanan, H.C. Liang, Vahid Razaviarani, Pesticides and herbicides, Water
Environ. Res. 35 (2012) 1533e1567.
Shahrokh Asadi, Jamal Shahrabi, Peyman Abbaszadeh, Shabnam Tabanmehr,
A new hybrid articial neural networks for rainfallerunoff process modeling,
Neurocomputing 121 (2013) 470e480.
P.C. Wang, T.E. Shoup, Parameter sensitivity study of the NeldereMead simplex method, Adv. Eng. Soft. 42 (2011) 529e533.
G. Lepoittevin, G. Kress, Optimization of segmented constrained layer
damping with mathematical programming using strain energy analysis and
modal data, Mater. Des. 31 (2010) 14e24.
Xin-She Yang, Engineering Optimization. An Introduction with Metaheuristic
Applications, 2010, ISBN 978-0-470-58246-6.
J.A. Nelder, R.A. Mead, Simplex method for function minimization, Comput. J.
7 (1965) 308e313.
J.C. Garcia, Determination of Induced Forces by Steam Flow in Turbines
(Doctoral thesis), CIICAp, Cuernavaca, Morelos, 2008 (in Spanish).
J. Kubiak, Diagnostic of Fracture of the Blade of the Steam Turbine, Internal
report of EM-CIICAp, No 6P/DM CIICAp, University of the State of Morelos,
2002.
J. Kubiak, J.A. Rodrguez, J.A. Segura, J.C. Garca, The effect of cracks in the
steam turbine L-0 blades on their natural frequencies and their remaining life,
in: Proceeding of the International Conference on Electric Power, Rosemont,
2007.
J.A. Rodrguez, Experimentation and Simulation of Fracture in Blades of the
Last Stage in Steam Turbines (Doctoral thesis), CIICAp, Cuernavaca, Morelos,
2009 (in Spanish).
C. Cuevas Arteaga, J.A. Rodrguez, C.M. Clemente, J.A. Segura, G. Urquiza, Y.
El. Hamzaoui, Estimation of useful life in turbines blades with cracks in corrosive environment, Eng. Fail. Anal. 35 (2013) 576e589.