U.S. Government work not protected by U.S.
copyright
Gas Turbine Bearing and Vibration Classification
of Using Multi-layer Neural Network
Moneer Ali Lilo 1, Latiff, L.A*1, Aminudin Bin Haji Abu 2, Yousif I. Al Mashhadany3, Abidulkarim K. Ilijan 1
1
Razak School of Engineering and Advanced Technology, Universiti Teknologi Malaysia, Kuala Lumpur Malaysia.
2
Malaysia - Japan International Institute of Technology, Universiti Teknologi Malaysia. Kuala Lumpur, Malaysia.
3
Engineering Department, Engineering College, University of Anbar, Iraq.
E-mail:moneerlilo@yahoo.com; *Liza.kl@utm.my
Abstract Gas Turbine (GT) is a vital component to a power
plant. This system contains many signals that are used to control
and protect the GT from damage or accidents caused by
vibration, speed, and temperature. Moreover, the vibrations of
GT at dangerous levels might lead to damages to the system. In
this paper, a concerted effort is made to identify the number of
the bearing and vibration levels during operations. We designed
and compared two types of the Neural Networks (NNs); series
and parallel NNs. They are based on the two stages from NN's
employed by MATLAB. The results indicated that the parallel
NN is better, depending on the time training and the produced
error. Moreover, the two stages of NNs can identify the bearing
number and vibration situations. The structure of the NNs puts
the system in sleep mode until the vibration is in high level,
however, sleeping system leads to the reduction of power
consumption when designing the hardware system.
Keywords Neural Network; vibration identification;
vibration classification; gas turbine.
I. INTRODUCTION
GT is a system that produces a massive torque with a
flexible speed adjustment that rotates the coupled electrical
energy generator. It is also a distinguished technology due to
numerous concerns such as reliability, availability, operating
costs, maintenance costs, and low CO2 emissions, principally
when used in a combined cycle, which is why they are
commonly used in the industry[1][2] [3]. Analyses of vibration
in gas turbines are considered as one of the most challenging
preventive maintenance. Methods of vibration analysis and
systems for condition monitoring of this equipment are under
rapid development since the previous decade. Mostly, two
kinds of the vibration measurement were used in GT; peak-to-
peak and absolute vibration, based on the Root Mean
Square (RMS) value. Indeed, vibration monitoring data can be
collected from one of the following probes; velocity pickup
detectors and accelerometers, both of which are typically fixed
to the casing of the turbine [4].
Sensor fault detection has garnered extensive recent
attention in numerous engineering areas, due to the profits of
decreasing downtime, productivity loss, increasing the
guarantee of safety, reliability, and quality of the system.
Artificial Neural Networks (ANNs) have been extensively
reported in the context of fault detection. Specifically,
multilayer perception nervous networks were trained using
*Corresponding author
U.S. Government work not protected by U.S. copyright
back- propagation BP for classification purposes. ANN-BP
realized very high success rates for the identification and
classification machine faults. Other methods
include probabilistic neural network (PNNs), which detect
faults on gas turbines, were also taken into account. Principal
component analysis detects sensor faults based on the
utilization of the Squared Prediction Error SPE[5]. NNs are
classified based on the supervision and unsupervised
techniques, moreover, NNs use a different strategy to collect
error feedback from the output to other stages. Clearly, BP
used the error output as feedback to all of the stages. Other type
are Dynamic neural networks, which are a type of neural
network, where one employs an internal feedback between the
input and output [6]. The intentions of this research are
categorized into three parts; the first is a comparison between
the series and parallel NNs, the second is an identification of
the sensor number and classify the vibration level for that
sensor, and the last is a design of multi-stage NNs with sleep
mode for classification stage.
This paper will illustrate in Section II related work for
other investigators , followed by methodology of identify the
vibration data based on the NN and architectures of the NN
in section III, and section IV will present and discuss the
results utilizing MATLAB program based on the neural
algorithm ,eventually, Section V will show conclusion on the
investigate done.
II. RELATED WORK
In recent years, the investigators began augmenting the
accuracy of classifying and identifying a default on a machine.
They utilized many methods to realize those objectives, such as
fuzzy, neural, and wavelets [7] [8]. Moreover, the researchers
start to identify the problem, and decide to minimize damage,
culminating in them utilizing a hybrid method to realize the
aforementioned objectives. The hybrid are termed fuzzy-neural
and fuzzy-wavelet [9] [10]. This part will involve researchers
employing NN, with one of them being (Zhang et al, 2008),
who discussed the classification and identification vibration of
thermal turbine based on the probability of a neural network
PNN. The power of the spectrum, singular spectrum, and
spectrums based on wavelet energy were inputted into the
PNN. The accuracy in these networks is 100% in the context of
classification [11].
20
 U.S. Government work not protected by U.S. copyright
(McKee et al 2015) proposed utilizing analysis statistical
metric and vibration cavitation to discover the vibration
cavitation. The values associated with ISO 10816 were
employed in this work. Subsequently, the vibration data in FFT
mode has been identified and classified based on ISO
10816[12].
(Fei et al, 2014) have augmented the accuracy of detecting
the vibration on a steam turbine based on three techniques.
Those techniques consist of back propagation NN, radial basis,
and wavelet, all of which are built parallel to each other. They
are interested in three fault singles, which are imbalance rotor,
oil whip of turbine, and misalignment to the same machine.
They proved the fault diagnosis of vibration was ameliorated,
depending on the outcome of that technique [8].
Eventually, (Marichal g. et al, 2010) used the hybrid
method to detect the vibration fault of ingoing and over the
bearing. They seem to reduce the complexity of a vibration
signal by dividing the frequency with the equilateral category.
This concept will be based on the approximated value to each
group. Thereafter, these signals were applied to a neural fuzzy
system to locate vibrations in the bearing. Finally, they applied
these signal to the many stages of the neuro-fuzzy technique,
where the results of first network was inputted into the second
stage for finding the other kind of vibration that might be
generated duration operations [13].
III. IDENTIFY THE VIBRATION BASED ON A NEURAL
Identifying the vibration data of the gas turbine and other
machines, which are related to classifying the vibration
situations, and decide on the control system to protect the
turbine from dangerous values. In the work, we will describe
the vibration signals that were collected, and compare the two
types of NN. The parallel NN represented final network which
is utilized in this work.
A. Vibration data collecting
Vibration data was collected from vibration sensors
installed in a gas turbine. These sensors were classified based
on the type of vibration being collected, similar to operation
frequency of the machine. However, sensors of the vibration
are defined by the velocity, acceleration, and displacement
sensor [14]. In this work, the vibration signal was collected
from the Al-Quds power plant, which contains a gas turbine.
Moreover, these signals are related to the startup situation of
the gas turbine, where the turbine started rotating until it
reaches the power generation situation. These signals represent
the normal situational vibration. Indeed, this work identifies the
vibration of the gas turbine in the alarm cases. Thus, we
augmented the vibration value for it to be multiplied by its
constant value. Thereafter, some of them will be in alarm
situation, and can be utilized to classify the vibration
situation[15].
B. Comparison between series and parallel neural network
The data of vibration sensors were applied on both NN; the
first one is a parallel- input parallel- output (PIPO), while the
second type is related to seriesinput-series-output (SISO). The
comparisons between the outcomes of those two kinds are
based on time in learning and running situations. Moreover, it
depends on the maximum error produced from those networks,
U.S. Government work not protected by U.S. copyright
the results of which are illustrated in Table 1. Eventually, we
need the same comparison to be applied to the data related to
classifying the alarm situations of each sensor. Table 2
depicted the outcome of the comparison between two types of
NN when applied to the alarm data in different situations. The
results showed that PIPO is better than SISO, based on the time
and maximum errors. Similarly, this work will be based on the
applied signal on the PIPO NN.
Table 1 compression result for the main data input
No. Type Time of Max. error Sample Test
learning number number
1 Parallel 2.9225 0.0538 5*380 10
2 Series 29.84 0.055 5*380 10
Table 2 comparison result of the alarm data input
No. Type Time of Max. error sample Test
learning number number
1 Parallel 2.1727 0.0533 3*380 10
2 Series 10.45 0.06 3*380 10
C. Architectures of the neural network in this work
The main intention of using forward back propagation NN
is to identify and classify the vibrations of the gas turbine.
Thus, the vibration sensors will categorize and discriminate the
vibration values at its previous levels [16] [12]. The
architectures of NN's of this work are based on two main parts;
the first one will identify the vibration values where a bearing
might occur in an alarm level. Moreover, the first part will be
switched to wake up sleeping NN's, which were stilled in the
second part. The first part of the NN will lead to the second
part by sending signals to wake up the NN related to one of the
input data, or to one bearing data. The second part consist of
the four NN's on representing and learning different situation to
one bearing when it is in alarm or high levels of vibration data.
Moreover, the result of first stage is represented where the
bearing vibration occurred. However, every outcome of the
first stage will wake up one NN, which is related to that
bearing to classify a vibration level of one input only.
Eventually, if we have four inputs on the main NN, it will
consist of five NN's, where only one network will be in
continues operation. Similarly, four NNs will be in sleep
mode, until the first stage wakes up one of four NNs to classify
the vibration level, as illustrated in Figure 1. If the algorithm of
the first stage is applied to smart wireless sensor this leads to
minimization of power consumption [17]. The main parameter
of the NN is 4-inputs, 45 neurons in hidden layer, three layers,
and 4-outputs.
IV. RESULT AND ANALYSIS OF THE MULTI- NEURAL
NETWORK
The results of learning NN to identify the sensors in the
alarm situation are illustrated in the Figures 2, 3, 4 and
5.Moreover; these figures characterize the typical behavior of
the vibration values related to the four bearing, which is
starting from 0 to steady state speed value. Indeed, this part
represented the first stage from the work. We learned NN's
based on trial and error in catching better data to minimize the
error rate and the time needed to learn a data. Learning rate ,
momentum mu, ratio, and the neurons number nm in the
hidden layer are the main parameters, which are changed to
produce the minimum error with smallest time [16].
21
 U.S. Government work not protected by U.S. copyright

Figure 3 learning result of the bearing 2 of parallel NN

Figure 1 Architecture of the neural network was designed

Therefore, the first stage modification of those parameters

are = 0.043, mu=0.0325, and nm= 45, to produce the results
shown in Figures 2, 3, 4 and 5 with errors depicted in Figure 6.
Eventually, based on the principle of the trial and error, we test
another four NN's based on selected and mu to get better
results through NN training, as shown in Figure 7. The
vibration behavior to three situations to the same bearing is
represents in Figure 7, applied to train the NN to classify the
vibration level. This NN in online mode has ability to classify
the vibration level into three categories: A, B, and C that are
related to alarm zone of the vibration scheme.
After training NNs many times, we store the weight and Figure 4 learning result of the bearing 3 of parallel NN
basis of every NN. Thereafter, we apply a vibration signal to
the online network to check the actuality situation of these NN.
The outcome of these NN's are zero when the bearing are in a
normal situation, and produce the number of the bearing when
applied to high level vibration values, which will lead to the
second stage wake up and classification of the value of the
vibration. Finally, the design of the NN's was based on the
MATLAB code only, and Table 3 presented how many times
we tested the network with different situation and the average
results of those experiments.
Table 3 accuracy result of online NN's
No. Stage number Training number Average accuracy
1 First 50 100%
2 second 50 98%

Figure 5 learning result of the bearing 4 of the parallel NN

Figure 6 learning error result for bearing 1, 2, 3 and 4 of the parallel NN

Figure 2 learning result of the bearing 1 of parallel NN

U.S. Government work not protected by U.S. copyright 22

 U.S. Government work not protected by U.S. copyright
vibration in level A vibration in level B
vibration value
vibration in level C
vibration value
error result
Figure 7 learning and error result for bearing 1 for classifications vibration
value
V. CONCLUSION
A comparison was drawn between the serial and parallel
network, proving that the series NN type require more time
compared to a parallel NN. Moreover, the parallel network
produces less error than a series network. This work is
interested in two things, first is detecting and identifying which
sensor has high vibration value, while another is related to
classifying the vibration value of the sensor. Indeed, only the
main network or the first stage was still in continuous
operation, but the other NN's stayed in sleep mode until it was
switched from the main NN. However, when applied that
algorithm to the wireless sensor, it leads to decreased power
consumptions.
REFERENCES
[1] M. Yari, M. Aliyari Shoorehdeli, and I. Yousefi, V94.2 gas turbine
identification using neural network, 2013 First RSI/ISM Int. Conf.
Robot. Mechatronics, pp. 523529, Feb. 2013.
[2] N. Hino, A. Kothet, and N. Kusumi, A new concept of gas turbine
system: Motor-assisted gas turbine with high-speed motor, 2014
IEEE Energy Convers. Congr. Expo., pp. 25552560, Sep. 2014.
[3] H. Hanachi, J. Liu, S. Member, A. Banerjee, Y. Chen, and A. Koul,
A Physics-Based Modeling Approach for Performance Monitoring
in Gas Turbine Engines, pp. 19, 2014.
[4] O. A. Zargar, Vibration Analysis of Gas Turbine SIEMENS
162MW- V94 . 2 Related to Iran Power Plant Industry in Fars
Province, vol. 2, no. 1, pp. 110, 2014.
[5] Y. Zhang, C. M. Bingham, M. Gallimore, Z. Yang, and J. Chen,
Applied sensor fault detection and validation using transposed
input data PCA and ANNs, 2012 IEEE Int. Conf. Multisens. Fusion
Integr. Intell. Syst., pp. 269274, Sep. 2012.
U.S. Government work not protected by U.S. copyright
[6] FAULT DETECTION OF GAS TURBINE ENGINES USING
DYNAMIC NEURAL NETWORKS S . S . Tayarani-Bathaie , Z .
N . Sadough Vanini and K . Khorasani, 2012.
[7] M. Hashemi and M. S. Safizadeh, Design of a fuzzy model based
on vibration signal analysis to auto-detect the gear faults, Ind.
Lubr. Tribol., vol. 65, no. 3, pp. 194201, 2013.
[8] E. Background, Fault Diagnosis in Power Plant Based on Multi-
Neural Network, 2014.
[9] S. D. Nguyen, K. N. Ngo, Q. T. Tran, and S.-B. Choi, A new
method for beam-damage-diagnosis using adaptive fuzzy neural
structure and wavelet analysis, Mech. Syst. Signal Process., vol.
39, no. 12, pp. 181194, Aug. 2013.
[10] V. Muralidharan and V. Sugumaran, Rough set based rule learning
and fuzzy classification of wavelet features for fault diagnosis of
monoblock centrifugal pump, Measurement, vol. 46, no. 9, pp.
30573063, Nov. 2013.
[11] Y. Zhang, S. Huang, W. Gao, and T. Shen, Vibration Fault
Diagnosis of Steam Turbine Shafting Based on Probability Neural
Networks, 2008 Congr. Image Signal Process., pp. 582585, 2008.
[12] K. K. McKee, G. L. Forbes, I. Mazhar, R. Entwistle, M.
Hodkiewicz, and I. Howard, A vibration cavitation sensitivity
parameter based on spectral and statistical methods, Expert Syst.
Appl., vol. 42, no. 1, pp. 6778, Jan. 2015.
[13] G. Marichal, M. Artes, and J. Garcia-Prada, An intelligent system
for faulty-bearing detection based on vibration spectra, J. Vib.
Control, vol. 17, no. 6, pp. 931942, Oct. 2010.
[14] Z. Hameed, Y. S. Hong, Y. M. Cho, S. H. Ahn, and C. K. Song,
Condition monitoring and fault detection of wind turbines and
related algorithms: A review, Renew. Sustain. Energy Rev., vol. 13,
no. 1, pp. 139, Jan. 2009.
[15] M. A. Lilo, L. A. Latiff, Y. I. Al Mashhadany, B. H. Abu, and A. K.
Ilijan, Modeling and Classification a Vibration Signal of Steam
Turbine Based on Fuzzy System, wulfenia journal,Austria, vol. 21,
no. 12, pp. 234243, 2014.
[16] JACEK M. ZURADA, Introduction to Artificial Neural System.
New York -United States of America: west publishing company,
1992.
[17] J. Neuzil, O. Kreibich, and R. Smid, A Distributed Fault Detection
System Based on IWSN for Machine Condition Monitoring, vol.
10, no. 2, pp. 11181123, 2014.
23