You are on page 1of 8

430 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 18, NO.

2, MARCH 2010
Brief Papers
Fault Detection and Isolation of Induction Motors Using Recurrent Neural
Networks and Dynamic Bayesian Modeling
Hyun Cheol Cho, Jeremy Knowles, M. Sami Fadali, and Kwon Soon Lee
AbstractDynamic neural models provide an attractive means
of fault detection and isolation in industrial process. One approach
is to create a neural model to emulate normal system behavior and
additional models to emulate various fault conditions. The neural
models are then placed in parallel with the system to be moni-
tored, and fault detection is achieved by comparing the outputs of
the neural models with the real system outputs. Neural network
training is achieved using simultaneous perturbation stochastic
approximation (SPSA). Fault classication is based on a simple
threshold test of the residuals formed by subtracting each neural
model output from the corresponding output of the real system.
We present a new approach based on this well known scheme
where a Bayesian network is used to evaluate the residuals. The
approach is applied to fault detection in a three-phase induction
motor.
Index TermsDynamic Bayesian model, fault detection/isola-
tion, induction machines, recurrent neural networks, stochastic
approximation.
I. INTRODUCTION
I
NDUSTRIAL processes must be monitored in real time
based on input-output data observed during their operation.
Common failure modes of such systems must be classied and
detected in order to ensure safe and productive system opera-
tion, prevent damage to other connected systems, and facilitate
timely repair of failing/failed components.
Induction motors are an important part of many industrial
applications and their failure can result in signicant economic
losses. Recently, the scale of industrial processes involving
induction motors has grown considerably and fault detection
and diagnosis for such systems has become more complex. As
a result, research has focused on nding new techniques for
timely and reliable detection and diagnosis of induction motor
faults.
Manuscript received May 03, 2008; revised September 23, 2008. Manuscript
received in nal form January 12, 2009. First published June 30, 2009; current
version published February 24, 2010. Recommended by Associate Editor
A. T. Vemuri. This work was supported by research funds from Dong-A
University.
H. C. Cho is with the School of Electrical and Electronic Engineering, Ulsan
College, Ulsan 680-749, South Korea.
J. Knowles and M. S. Fadali are with the Department of Electrical and
Biomedical Engineering, University of Nevada-Reno, Reno, NV 89557-0260
USA.
K. S. Lee is with the Department of Electrical Engineering, Dong-A Univer-
sity, Busan 604-714, South Korea.
Color versions of one or more of the gures in this brief are available online
at http://ieeexplore.ieee.org.
Digital Object Identier 10.1109/TCST.2009.2020863
Several approaches have been proposed for detecting faults
in induction motor systems. One traditional method of induc-
tion motor fault detection is motor current signature analysis
(MCSA) in which signal processing technique such as the fast
Fourier transform (FFT) is used to obtain the frequency spec-
trum [1][3]. Other signal spectrum methods based on wavelet
transformation [4], time-frequency domain analysis [5], higher-
order spectra [6], etc., have also been proposed.
In [7], Lee et al. studied integrating the FFT and wavelets to
classify fault modes in induction motors. Combastel et al. [8]
investigated an online model-based wavelet algorithm for time-
varying parameters. The authors dened a hierarchical fault tree
to reach a correct fault diagnosis. In [9], Jimenez et al. used the
Hilbert transform to extract the envelope of the signal spectrum.
The envelope was premultiplied by a window to: 1) overcome
transient distortion in wavelet based fault detection and 2) im-
prove reliability for a very fast system response. Blodt et al. ap-
plied the Wigner distribution to represent a motor signal in both
the time and frequency domains for its online condition moni-
toring in [10].
High-order statistics is used for detecting fault condition
in induction motor systems including non-stationary and
non-Gaussian random systems. In [11], Arthur and Penman de-
rived a fault detection scheme for induction machines based on
high-order spectra. They require prior data describing machine
fault conditions for the implementation of their method. The
steps required to statistically estimate the high-order spectra
is complex and requires large data sets that may not be easily
available in practice.
More recently, soft computation approaches such as neural
networks and fuzzy logics were utilized in induction motor fault
detection. A fuzzy rule base or adequately trained neural net-
work was used to represent the behavior of a healthy machine.
Fuzzy logic was then used for decision making for fault de-
tection and diagnosis in induction machine [12][15]. Alterna-
tively, the output of a trained neural network was compared to
the output of the induction motor for fault detection and di-
agnosis [16]. Several neural network types were utilized in-
cluding: radial basis networks [17], recurrent dynamic networks
[18], self organizing maps [19], and modied back-propagation
neural model [20].
The literature review shows that there are many available ap-
proaches for fault detection in induction motors. However, these
methods often fail to provide the desired detection and diag-
nosis performance on practical implementation, which is an-
1063-6536/$26.00 2009 IEEE

CHO et al.: FAULT DETECTION AND ISOLATION OF INDUCTION MOTORS 431
Fig. 1. Dynamic neuron structure.
alytically obtained from ofine design procedure, since signal
processing techniques usually require costly hardware with the
speed necessary to match their simulation results in real-time
implementation. Neural network approaches, as well as some
signal processing approaches, are often based on deterministic
models and do not take into account the randomnature of signals
in faulty induction motors. This results in performance degra-
dation in practice especially for nonstationary random signals.
Techniques based on higher-order spectra require large amounts
of data that is often unavailable in practice, and involve complex
computations.
In this brief, we investigate a neural network approach
to induction motor fault detection integrated with dynamic
Bayesian network used to model random residuals. A set of
articial neural networks (ANNs) is trained to model various
known failure modes of the system, in addition to modeling
its normal operation. The ANN model is composed of a single
layer perceptron (SLP) in cascade with an innite impulse
response (IIR) lter.
Many optimization and estimation algorithms have been used
for neural network training [21] including: back propagation,
simulated annealing, and various other forms of stochastic ap-
proximation. In this brief, we use the simultaneous perturbation
stochastic approximation (SPSA) method [22]. SPSA is an op-
timization method in which an approximation of the gradient
is made from a single pair of objective function measurements.
Unlike other gradient-based optimization methods, where eval-
uation of the gradient requires that each parameter of the objec-
tive function be varied individually, SPSA obtains an estimate
of the gradient by simultaneously varying all of the parameters.
Once a set of networks have been trained, fault detection is
achieved by arranging them in parallel with the real system, and
comparing the outputs of each network to measured system out-
puts to obtain the residual signal. The residual is random be-
cause the system outputs include measurement noise in prac-
tice. We model the residual signal using a dynamics Bayesian
network (DBN) in which a discrete Markov chain represents the
systems random behaviors.
We apply our fault detection and isolation approach to a real-
time induction motor control system. Three induction motors
are used: The rst is free of faults, the second has stator winding
fault, and the third has a bearing fault. Experimental results
demonstrate the effectiveness and reliability of the implemen-
tation of our methodology.
This brief is organized as follows. Section II describes
the dynamic neural network designed for system modeling.
Section III derives the neural network learning rule using
SPSA. We present fault modeling with dynamic neural network
and random residual modeling with a DBN, respectively, in
Sections IV and V. Section VI presents a real-time fault detec-
tion experiment for induction motors. Conclusions and future
work are given in Section VII.
II. FEED-FORWARD DYNAMIC NEURAL NETWORK
A feed-forward dynamic neural network consists of one or
more layers of dynamic neurons. The structure of the dynamic
neuron constructed in this brief is shown in Fig. 1. As in static
neurons, the dynamic neuron rst calculates the weighted sum
of its inputs
(1)
where is the number of inputs, is a
vector of input weights, and is the
vector of neuron inputs. The weighted sum is fed to an IIR lter
of order described by the linear difference equation
(2)
where , and are
the feedback and feed-forward lter weights, respectively. The
neuron output is calculated from the IIR lter output
using the following formula:
(3)
where is a slope parameter, is the output bias factor, and
is a nonlinear activation function. A multi-layer feed-for-
ward neural network is constructed by connecting neurons in
layers such that signals are allowed to travel only in a forward
direction through the network. Neuron connections are allowed
to transmit signals only fromone layer to the next until the signal
reaches the output layer of the network. No feedback loops or
interconnections between neurons in the same layer are allowed.
However, the network includes feedback in the IIR lter of each
neuron.
III. NEURAL NETWORK TRAINING VIA SPSA
The SPSA algorithm is used to estimate the optimal value of
a vector of unknown parameters, such that some loss function
is minimized. Given a set of constraints dening the feasible
range of , this minimization can be expressed as follows, where
is the theoretical optimum:
(4)

432 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 18, NO. 2, MARCH 2010
Fig. 2. FDI scheme using DNN models.
SPSA is a special case of stochastic approximation (SA) and
the general form of the SA recursive algorithm is given by [22]
(5)
where is an estimate of the gradient based on
noisy measurements of the loss function , each
is a positive scalar gain discussed as follows, and is the step
number. The components of the gradient approximation are cal-
culated as follows (two-sided/symmetric simultaneous pertur-
bation) [22]:
(6)
where is a -dimensional random perturbation vector.
The elements of this vector must be independent and sym-
metrically distributed around 0, with nite inverse moments
. The symmetric Bernoulli distribution
satises these properties [23], while the uniform and normal
distributions do not. The symmetric Bernoulli distribution
is used in this brief. The gain sequences and are
calculated as follows [23]:
(7)
where , , , and are constants chosen by the user. Spall pro-
vides conditions for almost sure convergence of to [22].
Among these conditions are the following constraints on the
gain sequences:
i) , , where ;
ii) , as ;
iii) ;
iv) .
Constraints i) and ii) impose the following conditions on the
user selected constants: , , and , . Under
conditions (iiv) and the integral test for convergence of an in-
nite series, we can show that constraints iii) and iv) imply the
following additional conditions on the user selected exponents
and as and .
IV. MODELING OF FAULT SYSTEMS USING DNNS
For fault detection and isolation, a set of dynamic neural net-
works is trained to mimic the behavior of the system using ex-
perimental data or simulation data. One network is trained to
approximate normal system behavior, and additional DNN
models are trained to simulate faulty system behavior, one for
each of fault conditions to be monitored. After training, the
DNN models are placed in parallel with the system to be mon-
itored. The output of each model is compared to the output of
the real system for the same input to estimate the fault state of
the real system.
We construct a DNN bank that is composed of DNN models
trained to represent healthy and faulty systems as shown in
Fig. 2. Here, DNN 0 is a representation of the healthy system,
and each DNN , represents the system with
fault . This is realized for fault isolation strategy. In other
words, each DNN learns one fault condition. Each residual
, , is the difference between the actual system
output and the output of the th DNN model output. If the
magnitude of the residual is larger than a threshold , a
fault is detected, otherwise the system is deemed sound. This
concept is expressed with the binary hypothesis
no fault
fault
(8)
where and indicates the 1-norm. For fault isolation,
fault occurs when . Fault isolation is simi-
larly given as the binary hypothesis
fault occurred
no fault occurred.
(9)
The hypothesis allows us to identify the location of the fault
within the system.

CHO et al.: FAULT DETECTION AND ISOLATION OF INDUCTION MOTORS 433
Fig. 3. DBN structure for modeling of a random residual.
V. DBN-BASED RESIDUAL MODELING
In practice, the residual is stochastic since the system output
measurement includes noise. Thus, a deterministic approach for
fault detection and isolation against such random signal is often
unsatisfactory for practical implementations. We use DBNmod-
eling to sequentially represent random residuals in Fig. 2. We
ignore about DBN and refer [24] for details. Of several DBN
models, we adopt a simple DBNmodel, a discrete Markov chain
in this brief, shown in Fig. 3. In order to construct a DBNmodel,
we rst discretize the continuous absolute residuals:
(10)
where is positive constant. Its probability is given by
(11)
From the probability axioms, we have the condition
(12)
A. Online DBN Modeling
The network parameter indicates the conditional proba-
bility of each variable between and , dened as
(13)
where for simplicity. Similarly, we have
(14)
The optimal network parameters are estimated online based
on the observation sequence. We adopt the estimation algorithm
of [25] for our DBN modeling. The parameters in (13) are alter-
natively expressed as
(15)
where is the average likelihood and is a normalizing factor
to satisfy the probability constraints (14). This variable is se-
quentially updated based on observation data. We dene its up-
date rule in the recursive from
(16)
where is involved with observation data and is selected one
of both rules dened by
if
otherwise
(17)
where is state of the random variable and .
The reader is referred to [25] for details of this algorithm and
the convergence property of this estimator. Applying a sliding
windowto adopt relatively recent data sequence, we rewrite (16)
as
(18)
where window size . If is large, then somewhat
older data is chosen, but a small corresponds to a short data
sequence such that only recent data are considered. The update
rule of (18) requires the current time to be larger than the
window size , i.e., . At an initial data point, it is
possible for the data set to be shorter than the window size
. Thus, windowing is not available until .
B. Decision Making for FDI by DBN
The posterior probability vector of the model in Fig. 3 is lin-
early expressed as
(19)
where time-varying stochastic matrix ,
is updated through the previous estimation procedure. The state
probability is recursively computed from multiplying the sto-
chastic matrix by the prior probability vector. Using this esti-
mation, we obtain probability density for each random residual
, where and determine a variable related to a
maximum posterior probability in (10), i.e.,
(20)
Using this selection, we alternatively express the hypothesis
for decision making in fault detection as
no fault
fault
(21)
where a reference threshold . This rule indicates
that if a variable with maximum posterior probability is smaller
than , the decision making is applied by , otherwise, .
This procedure is sequentially accomplished through online es-
timation of DBN based on the residual. Similarly, for fault iso-
lation we dene a hypothesis as
fault occurred
no fault occurred
(22)
where and . Unfortunately, there is no an-
alytical guideline for selecting the threshold in decision making
rule, but we must determine its proper value through iterative
real-time experiments.
VI. EXPERIMENTAL RESULTS: APPLICATION
TO INDUCTION MOTORS
We apply our FDI algorithmto a three-phase induction motor.
The specications of the motor are given in Table I.

434 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 18, NO. 2, MARCH 2010
TABLE I
PARAMETERS OF THE INDUCTION MOTOR
Fig. 4. Experimental conguration of the induction motor system.
The experimental setup of the induction motor is shown in
Fig. 4. The motor is controlled by an inverter system in which a
vector control strategy [26] is embedded to manage the desired
speed by leveling the input voltage. We measure each phase cur-
rent of the motor using a Hall sensor. The cutoff frequency of
the anti-aliasing lter is 5 kHz and the current is discretized at
13 kHz sampling frequency. We select the Hanning window to
overcome Gibbs phenomenon and use a 16 bit high-precision
AD converter. The measurement is fed to a PC interface board
where the proposed FDI algorithmis programmed using Matlab.
Three induction motors with the same specications are utilized
of which one is healthy, one has a faulty stator where the stator
winding is short-circuited, and one has a faulty bearing where
some balls inside the bearing are cracked. We run each induction
motor under the same experimental environment so that each in-
herent current value is acquired. Fig. 5 gives plots of the current
measurements for each of the three motors. Here, we observe a
regular sinusoidal waveform for the healthy motor in Fig. 5(a)
and irregular sinusoidal waveforms for the two faulty motors in
Fig. 5(b) and (c).
Each current is used to train one neural network to repre-
sent a motor, one network to represent a healthy motor, and
one to represent each of the two faulty motors. The three net-
works are identical in structure with three input nodes and the
same sigmoid activation function as shown in Fig. 1. We select
a second-order IIR lter in Fig. 1, which includes two poles and
one zero because of easily constructing a stable lter. We iter-
atively train each neural network until its output is close to the
actual current measurement fromwhich its weights are properly
Fig. 5. Current waveform of the induction motors.
determined. We estimate the probability density of the current
sequence using the proposed DBN approach of Section V. To
construct each DBN model, the continuous current signal (de-
noted by ) is rst discretized as
(23)
with

CHO et al.: FAULT DETECTION AND ISOLATION OF INDUCTION MOTORS 435
Fig. 6. Probability estimation of the current sequence.
Fig. 6 shows time history of probability estimation for each
state. Here, the estimated probabilities are almost constant in
the steady-state, but initial vary during the transient phase. The
residual between the currents of healthy and faulty motors is de-
ned as
(24)
Fig. 7. Residual signal for healthy motors.
where symbol and 2 denote the stator and bearing fault
respectively. We calculate total residual by summing all resid-
uals, i.e.,
(25)
Fig. 7 shows the time histories of the residual for the healthy
motor. We observe the average value is about 0.05, 0.06, and
0.05 for each phase. Figs. 8 and 9 are plots of the histories of

436 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 18, NO. 2, MARCH 2010
Fig. 8. Residual signal for faulty stator.
the residual for each phase of the faulty motors. In Fig. 8 a
large transient error is followed by convergence with steady-
state error of about 0.7, 0.6, and 0.2% for phases A, B, and C,
respectively. Similarly, in Fig. 9 the residual signals have about
0.9, 0.8, and 0.7% in the steady-state for phases A, B, and C,
respectively. These six values are obviously larger than those of
the healthy motor. From this result we conclude the two fault
motors are abnormally operated and the proposed detection al-
gorithm is effectively achieved. Moreover, the latter is larger on
the average than the former, which indicates that the bearing
fault of the motor is more easily detected than the stator winding
fault.
Fig. 9. Residual signal for faulty bearing.
VII. CONCLUSION
This brief presents a stochastic fault detection and isolation
technique for induction motor systems. We analytically model
healthy and faulty motors by means of dynamic neural networks
in which the SPSA-based learning algorithm is derived without
calculating the gradient terms. The random residual signal is
represented by a DBN for stochastic decision making. We con-
ducted real-time experiments with three motors to evaluate our
FDI scheme. We estimated the probability distributions for the
motors states and observe random residues online.

CHO et al.: FAULT DETECTION AND ISOLATION OF INDUCTION MOTORS 437
Future work will include other applications with the pro-
posed FDI method such as generator systems. Such is usually
involved to complicated fault diagnosis of large-scale dynamic
systems, thus we will enhance our FDI approach suitable to the
framework.
REFERENCES
[1] S. Nandi, H. A. Toliyat, and X. Li, Condition monitoring and fault
diagnosis of electrical motorsA review, IEEE Trans. Energy Conv.,
vol. 20, no. 4, pp. 719729, Dec. 2005.
[2] W. Li and C. K. Mechefske, Detection of induction motor faults: A
comparison of stator current, vibration and acoustic methods, J. Vibr.
Control, vol. 12, no. 2, pp. 165188, 2006.
[3] J.-H. Jung, J.-J. Lee, and B.-H. Kwon, Online diagnosis of induction
motors using MCSA, IEEE Trans. Ind. Electron., vol. 53, no. 6, pp.
18421852, Nov./Dec. 2006.
[4] A. Boggess and F. J. Narcowich, A First Course in Wavelets With
Fourier Analysis. Upper Saddle River, NJ: Prentice-Hall, 2001.
[5] B. Boashash, Time Frequency Analysis. Oxford, U.K.: Elsevier Sci-
ence, 2003.
[6] C. Nikias and A. P. Petropuou, Higher-Order Spectra Analysis.
Upper Saddle River, NJ: Prentice-Hall, 1993.
[7] S.-H. Lee, S.-P. Cheon, Y. Kim, and S. Kim, Fourier and wavelet trans-
formations for thefault detectionof inductionmotor withstator current,
in Proc. Int. Conf. Intell. Comput., Kunming, China, 2006, pp. 557569.
[8] C. Combastel, S. Lesecq, S. Petropol, and S. Gentil, Model-based and
wavelet approaches to induction motor on-line fault detection, Control
Eng. Practice, vol. 10, pp. 493509, 2002.
[9] G. A. Jimenez, A. O. Munoz, and M. A. Duarte-Mermoud, Fault
detection in induction motors using Hilbert and wavelet transforms,
Elect. Eng., vol. 89, pp. 205220, 2007.
[10] M. Blodt, D. Bonacci, J. Regnier, M. Chabert, and J. Faucher, On-
line monitoring of mechanical faults in variable-speed induction motor
drives using the Wigner distribution, IEEE Trans. Ind. Electron., vol.
55, no. 2, pp. 522533, Mar./Apr. 2008.
[11] N. Arthur and J. Penman, Induction machine condition monitoring
with higher order spectra, IEEE Trans. Ind. Electron., vol. 47, no. 5,
pp. 10311041, Sep./Oct. 2000.
[12] D. F. Leite, M. B. Hell, P. H. Diez, B. S. Ganglio, L. O. Nascimento,
and P. Costa, Jr, Real-time model-based fault detection and diagnosis
for alternators and induction motors, in Proc. IEEEInt. Electric Mach.
Drives Conf., Antalya, Turkey, 2007, pp. 202207.
[13] P. V. J. Rodriguez and A. Arkkio, Induction motor stator fault detec-
tion using fuzzy logic, in Proc. IASTED Int. Conf. Circuits, Signals,
Syst., Cancun, Mexico, 2003, pp. 96100.
[14] F. Zidani, D. Diallo, M. E. H. Benbouzid, and R. Nait-Said, A fuzzy
based approach for the diagnosis of fault modes in a voltage-fed PWM
inverter induction motor drive, IEEE Trans. Ind. Electron., vol. 55, no.
2, pp. 586593, Mar./Apr. 2008.
[15] P. V. J. Rodriguez and A. Arkkio, Detection of stator winding fault
in induction motor using fuzzy logic, Appl. Soft Comput., vol. 8, pp.
11121120, 2008.
[16] H. Su and K. T. Chong, Induction machine condition monitoring using
neural network modeling, IEEE Trans. Ind. Electron., vol. 54, no. 1,
pp. 241249, Jan./Feb. 2007.
[17] L. Eren, A. Karahoca, and M. J. Devaney, Neural network based motor
bearing fault detection, in Proc. IEEE Instrumentation Meas. Tech.
Conf., Como, Italy, 2004, pp. 16571660.
[18] A. G. Parlos, K. Kim, and R. Bharadwaj, Detection of induction
motor faultsCombining signal-based and model-based techniques,
in Proc. Amer. Control Conf., Anchorage, AK, 2002, pp. 45314536.
[19] N. Li and C. K. Mechefske, Induction motor fault detection and diag-
nosis using articial neural networks, Int. J. COMADEM, vol. 9, no.
3, pp. 1523, 2006.
[20] V. Duraisamy, N. Devarajan, D. Somesundareswary, and S. N.
Sivanandam, Modied back propagation algorithm for incipient fault
detection in three-phase induction motor, Model., Meas. Control A,
vol. 79, no. 12, pp. 1525, 2006.
[21] S. Haykin, Neural Networks: A Comprehensive Foundation. Upper
Saddle River, NJ: Prentice-Hall, 1999.
[22] J. C. Spall, Multivariate stochastic approximation using a simulta-
neous perturbation gradient approximation, IEEE Trans. Autom. Con-
trol, vol. 37, no. 3, pp. 332340, Mar. 1992.
[23] K. Patan and T. Parisini, Identication of neural dynamic models
for fault detection and isolation: The case of a real sugar evaporation
process, J. Process Control, vol. 15, pp. 6779, 2005.
[24] K. Murphy, Dynamic Bayesian networks: Representation, inference
and learning, Ph.D. dissertation, Dept. Comput. Sci., Univ. California,
Berkeley, 2002.
[25] H. C. Cho, M. S. Fadali, and K. S. Lee, Design of time-varying sto-
chastic process with dynamic Bayesian networks, KIEEJ. Electr. Eng.
Technol., vol. 2, no. 4, pp. 543548, Dec. 2007.
[26] B. Singh, V. Garg, and G. Bhuvaneshwari, A new waveshaper for
harmonic mitigation in vector controlled induction motor drives, KIEE
J. Electr. Eng. Technol., vol. 3, no. 2, pp. 152161, Jun. 2008.

You might also like