The Journal of Engineering Research Vol. 8 No.

1 (2011) 53-60

Neuro-Fuzzy Sensor Fault Diagnosis of an Induction

Motor
M. L. Benloucif

Automatic Control Laboratory, LAS, Department of Electrical Engineering,

Faculty of Engineering, University of Skikda, Algeria

Received 20 January 2010; accepted 2 March 2010

¤j²*”xp˜•Dz~zpg´*&b€³*Žb~{gC°¤+bc~9Á¡c~|<o3¡¿
£~6¡D¡›£+—š

¢•<]˜g†-b€³*Žb~{gC*fJx:*¤j²*”xp˜•Dz~zpg´*¯Ÿ~|pAH&b€³*Žb~{gC°¤+bc~9Á¡c~|<o3¡¿kpcD**wGš]J f~8ɳ*
i°b²*Hf£€1ÉD*f£’£Eb›J]D* i°b²* ‡E $¡ŒC –’~{+–Eb†gD*bž›’¤gD*H¤+bc~¦D*€›´*ib£›-Hf£c~|†D*bJɳ*Í+bEf›£mGib£›-
š]sg~z-f£FbjD*Hb€s•Df£cg´*fD*]D*]£D¡gDf£c~|†D*bJɳ*¶H°*f•0x´*š]sg~z-Íg•0xEœEfF¡’gEfJx:¢•<š]´*o3¡˜›D*]˜g†J,ÌÏ*
šbƒ›D*,$bŒCf~6*4]DkpcD*¯ŸE]Eo3¡˜›•D,bCbÏ*n)bgFb€s•Df£cg´*fD*]D*͘sgDf£+bc~9f£F¡c~|<f›£mGfJx:

ŸDy<Hz~z˜pg˜•Db€³*Žb~{gC*fc£~|†D*bJɳ*¤j²*”xÏ*¤+bc~¦D*€›´*f£0bgŒ´*ib˜•’D*
Abstract: In this paper, a neuro-fuzzy fault diagnosis scheme is presented and its ability to detect and
isolate sensor faults in an induction motor is assessed. This fault detection and isolation (FDI) approach
relies on a combination of neural modelling and fuzzy logic techniques which can deal effectively with
nonlinear dynamics and uncertainties. It is based on a two step neural network procedure: a first neural
network is used for residual generation and a second fuzzy neural network performs residual evaluation.
Simulation results are given to demonstrate the efficiency of this FDI approach.

Keywords: Fuzzy logic, Induction motor, Neural networks, Sensor fault detection and isolation

1. Introduction

The problem of fault detection and isolation (FDI) fault indicators, and residual evaluation which
is a crucial issue for the safety, reliability and perform- involves decision making. The model-based FDI
ance of industrial processes. approach which has received intensive attention uses
The usual approach to fault diagnosis is based on mainly state and parameter estimation techniques
hardware redundancy (multiple sensors, actuators and (Frank 1990). Model based FDI performance is
components) and uses a voting technique to decide if a directly related to the accuracy of the mathematical
fault has occurred and to locate it among the redundant model of the monitored system. The effect of model
system elements (Frank 1990). Instead, the analyti- uncertainties, disturbances and noise is therefore a key
cal redundancy FDI approach, also referred to as the issue in model based fault diagnosis.
model-based FDI approach, makes use of a mathemat- The main design requirements of model based fault
ical model of the monitored system [(Frank 1990). diagnosis procedures are thus concerned with the
The task of model based diagnosis methods consists of problems of robustness with respect to model uncer-
detecting faults that may occur in the system and tainties and enhancement of sensitivity to faults. These
which can be additive or multiplicative in nature. requirements are contradictory so a trade off is needed
Basically the FDI procedure consists of two main to cope with sources of false alarms and missed detec-
steps: generation of residuals which should be useful tions. Two strategies may be used: an active strategy
___________________________________________ consisting in robust residual generation and a passive
*Corresponding author’s e-mail: m.l.benloucif@gmail.com one through robust residual evaluation. Most of the
54

The Journal of Engineering Research Vol. 8 No. 1 (2011) 53-60

existing model based FDI techniques rely on the use of motor.

linear system models (Benloucif and Staroswiecki
2002 and Frank 1990). Often, nonlinear systems are 2. Model of the Induction Motor
described by linear models with additive disturbances.
Robust residual generation based on unknown input Assuming linear magnetic circuits and a balanced
observers to achieve disturbance decoupling may pro- three-phase system in the (a, b, c) frame, the electrical
vide an efficient solution to fault detection and isola- equations of the induction motor expressed in the
tion problems. As far as linear systems are concerned two-phase stationary (d, q) reference frame
the problem of robust residual generation may be con- (Benloucif and Balaska 2006) are:
sidered to be mature ((Benloucif and Staroswiecki
2002, Frank 1990 and Patton and Chen 1997) where-
as the FDI problem for nonlinear dynamic systems has
been investigated to a lesser extent (Benloucif and
Balaska 2006; Garcia and Frank 1997 and Jiang et al. (1)
2001).
Alternatively, FDI can be performed using qualita-
tive techniques such as expert systems, fuzzy logic,
neural networks (Al;exandru et al. 2000; Benloucif
and Mehennaoui 2002; Benloucif and Mehennaoui
2005; Chen and Lee 2002; Evsukoff et al. 1999; Frank
1990; Isermann 1998; Schneider and Frank 1996; (2)
Simani and Fantuzzi 2002; Takagi and Sugeno 1985,
Theilliol et al. 1997 and Uppal et al. 2002). To over-
come the limitations of the analytical FDI approach,
the actual trend integrates model based (analytical)
and knowledge based (non analytical) methods in where ), I, V are the stator/rotor fluxes, currents and
order to take advantage of their respective performanc- voltages expressed in the (d, q) reference frame. Ts is
es. Residual generation and residual evaluation for
the angle between the stator reference frames (a, b, c)
decision making may be achieved by using appropri-
and (d, q), and Tr is the angle between the rotor ref-
ate combinations of different techniques such as state
estimation, parameter estimation, neural networks, erence frames (a, b, c) and (d, q). Rs, Rr, Ls, Lr are the
fuzzy logic inference. stator/rotor resistances and inductances, respectively,
In (Benloucif and Mehennaoui 2002) a fault diag- and Lm is the magnetizing inductance. For a squirrel-
nosis procedure for linear systems used a combination cage IM the rotor voltages are zero. The mechanical
of an analytical residual generator based on Kalman equation is:
filtering and a fuzzy neural network for residual eval-
uation. In this work, an extension of the neuro-fuzzy (3)
FDI scheme given in (Benloucif and Balaska 2006) is
proposed. It is based on a two step neural network
procedure: The first network which has the ability to and dthe electromagnetic torque Te is given by:
model a wide class of nonlinear dynamic systems acts
as an on-line residual generator. The second network (4)
performs the decision making which consists in detect-
ing and isolating a fault when it occurs. This neural
network coupled to a fuzzy inference block acts as an
on-line fault classifier.
The paper is organized as follows. In section 2 the
model of the induction motor is presented, starting
from the classical Park transformation. The architec-
ture of the neuro-fuzzy scheme used for residual gen-
eration and evaluation is discussed in section 3.
Simulation results are given in section 4 to illustrate
the performance of the proposed neuro-fuzzy FDI
scheme for sensor fault diagnosis of the induction
 55

The Journal of Engineering Research Vol. 8 No. 1 (2011) 53-60

 57

The Journal of Engineering Research Vol. 8 No. 1 (2011) 53-60

A neuro-fuzzy network is based on the association et al. 2000; Benloucif and Mehennaoui, 2005 and
of fuzzy logic inference and the learning ability of Chen and Lee 2002).
neural networks.
The neuro-fuzzy approach is a powerful tool for
solving important problems encountered in the design
of fuzzy systems such as: determining and learning
membership functions, determining fuzzy rules, adapt-
ing to the system environment.
The main points of the residual evaluation proce-
dure are described below.

3.2.2 Residual Fuzzification

It consists in converting the numerical values of
residuals into linguistic variables. Each input (resid-
ual) may be described by three linguistic variables
(Negative, Zero, Positive). Each linguistic variable is
represented by a membership function which has gen-
erally a triangular or trapezoidal shape. The linguistic
Figure 5. Example of a RNN used for residual
variable Zero defines the range where the residual may
evaluation
be considered to be unaffected by a fault. The linguis-
tic variables Negative and Positive define the residual 3.2.4 Training
amplitude ranges indicating the presence of a fault. Prior to on-line use, network training is performed
The corresponding membership functions give the for all possible fault scenarios. During training a resid-
extent to which a residual is or is not affected by a ual pattern corresponding, eg. to fault f1, is applied to
fault. the network input and a one is assigned to the corre-
sponding output. The network weights are then adjust-
3.2.3 Neural Network Structure ed by an appropriate algorithm thus enabling the neu-
For fault diagnosis it is desirable to use a neural net- ral network to learn the imposed input-output pattern.
work to model the nonlinear relationship between the The use of the backpropagation algorithm is recom-
fuzzified residuals and the fault decision functions. A mended (Benloucif and Mehennaoui 2005). The ulti-
multilayer perceptron network is therefore a good can- mate goal of the training is to achieve the extraction
didate. Moreover, to account for memory in the deci- and selection of the necessary parameters defining the
sion process it is necessary to use a recurrent neural inference rules
network (RNN). The RNN may be implemented as a
neural model described by: 4. Numerical Results
(11)
Results using MATLAB simulation are next pre-
where Dk ( fi ), i = 1...nf , are the fault decision func- sented to assess the ability of this diagnosis approach
tions also referred to as fault indicators and fi are the based on neural and fuzzy techniques to detect and iso-
faults acting on the process. The regression vector late sensor faults in an induction motor. Its model
contains the fuzzy residuals Ri (k), i = 1...nr , and the expressed in the two-phase reference frame (d, q) is
delayed decisions Dk-1 (fi), i = 1...nf . Because of the given by the nonlinear state space Eq. (5).
feedback introduced, the recurrent neural model may The squirrel-cage induction motor considered here
be realized by a three-layer MLP. has power rating of 1 kW and its electrical and
This is illustrated by the example given in Fig. 5 mechanical parameters are as follows:
which shows a residual evaluation scheme processing
three residuals (r1, r2, r3) to diagnose three faults (f1,
f2, f3).
The corresponding neural network has the follow-
ing architecture: an input layer with 12 units represent-
ing all possible states of the fuzzy residuals together
Simulation is carried out with a sampling period of
with the past decisions, a hidden layer having 4 units,
1 msec, with 400 V and 50 Hz sinusoidal inputs. In
and an output layer with 3 units each assigned to a
normal operation, the outputs (Isd, Isq, :) and the elec-
decision function. The use of this RNN architecture
ensures reliable dynamic decision making (Alexandru tromagnetic torque Te are shown in Fig. 6.
58

The Journal of Engineering Research Vol. 8 No. 1 (2011) 53-60

Table 1. Inference table

Motor
Various simulation tests have been performed in
Figure 6. Stator currents (Isd, Isq), rotor speed : ,
scheme and the results are quite conclusive. Bias and
torque Te (normal operation) drift type sensor faults are introduced during steady
4.1 Residual Generation state conditions of the system. For illustrative purpos-
A NNARX model having the architecture shown in es only a few fault scenarios summarized in Tables 2
Fig. 3 has been used with the following parameters: n1 to 4 are discussed.
= n2 = n3 = n1 = m1 = m2 = 1, d = 1. Training of this Table 2. Case 1
MLP network was achieved by the Levenberg-
Marquardt algorithm for different numbers of hidden
neurons. For nh = 4, the output error cost reached at 36
iterations is E = 1.528e-002. After validation this Table 3. Case 2
NNARX model is used to generate the residuals:

(12)
4.2 Residual Evaluation
Table 4. Case 3

After many tests on residuals for different fault sen- 4.3.1 Case 1
sor situations to achieve a good trade off between A bias type fault is injected on sensor 1 as described
missed detections and false alarms, the following in Table 2.
membership functions for each residual were selected: The corresponding residuals are shown in Fig. 7.
Although a single fault may induce changes in several
Residual 1: N1= [-1,-1,-0.005,-0.002] residuals ( here a fault on sensor 1 affects positively
Z1= [-0.0025,0 , 0.0045] , P1= [0.0035, 0.006, 1, 1]. the first residual and negatively the second residual at
Residual 2: N2= [-1, -1, -0.04, -0.015] time t=2.5 sec) the decision functions ensure success-
Z2= [-0.02, 0, 0.005], P2= [0.004, 0.009, 1, 1]. ful detection and isolation of the fault on sensor 1 as
Residual 3: N3= [-1, -1, -0.018, -0.015] shown in Fig. 7. The neuro-fuzzy classifier has been
Z3=[-0.016,-0.0135,-0.012],P3=[-0.0125,-0.0115,1,1]. trained to recognize the faulty situations from the
fuzzified residual patterns according to the rule base
The RNN used in this simulation study is shown in given in Table 1.
Fig. 5. Its training is based on the rules summarized in
Table 1 which have been obtained after many simula- 4.3.2 Case 2
tion tests. The learning operation realized by the back- This fault scenario of bias faults on sensors 2 and 3
propagation algorithm converged after 3600 epochs is described in Table 3.
with a sum of squared error E=0.025. The residuals and the corresponding decision func-
Each row of the Inference table represents a rule. tions are shown in Fig. 8. The faulty sensors are
For example, rule 2 is expressed as: promptly detected and correctly isolated.
IF {residual 1 is positive and residual 2 is negative
and residual 3 is zero} THEN sensor 1 is faulty. 4.3.3 Case 3
This fault scenario uses drift faults on sensors 2 and
 59

The Journal of Engineering Research Vol. 8 No. 1 (2011) 53-60

3 as described in Table 4. Drift faults are modelled as a two step procedure: a neural NNARX model is used
ramp functions with given slopes. for residual generation and a recurrent fuzzy neural
The diagnosis effectiveness in the presence of sen- network performs the residual evaluation task. Fault
sor drift faults is illustrated in Fig. 9. We notice a diagnosis is achieved by training the network to recog-
detection delay for fault sensor 2. This delay, which is nize the fault signatures from the patterns of the fuzzi-
dependent on the slope of the drift, gives rise to a tem- fied residuals. The successful results obtained in sim-
ulation demonstrate the efficiency of this neuro-fuzzy
diagnosis scheme to detect and isolate bias and drift
sensor faults in an induction motor.

References

Alexandru, M., Combastel, C. and Gentil, S., 2000,

"Diagnostic Decision using Recurrent Neural
Networks," Proc. IFAC Safeprocess 2000,
Budapest, Hungary.
Babuska, R., 1998, “Fuzzy Modeling for Control,”
Kluwer Academic Publishers.
Figure 7. Faculty residuals and corresponding Benloucif, M.L. and Staroswiecki, M., 2002, "Fault
decisions (Case 1) Diagnosis using a Robust Estimation Method,"
Proc. International Conference CIFA-2002,
Nantes, France.
Benloucif, M.L. and Mehennaoui, L., 2002, "A Mixed
Analytical Neuro-Fuzzy Approach for Fault
Diagnosis," Proc. 2ND Instrumentation and
Measurement in Petroleum Applications
Conference IMPAC-2002, Boumerdès, Algeria.
Benloucif, M.L. and Mehennaoui, L., 2005, "A Fuzzy
Neural Scheme for Fault Diagnosis," Proc.
International Computer Systems and Information
Technology Conference ICSIT'05, Algiers,
Algeria.
Figure 8. Faculty and corresponding decisions
Benloucif, M.L. and Balaska, H., 2006, "Robust Fault
(Case 2)
Detection for an Induction Machine," 7th World
Automation Congress- WAC 2006, Budapest,
Hungary.
Chen, Y.M. and Lee, M.L., 2002, "Neural Networks
Based Scheme for System Failure Detection and
Diagnosis," Mathematics and Computers in
Simulation, 58, pp. 101-109.
Evsukoff, A., Combastel, C. and Gentil, S., 1999,
"Qualitative Reasoning and Neural Network
Decision Procedures for Fault Detection and
Isolation," Proc. 14th IFAC World Congress,
Beijing, China.
Figure 9. Faculty residuals and corresponding Frank, P.M., 1990, "Fault Diagnosis in Dynamic
decisions (Case 3) Systems using Analytical and Knowledge Based
Redundancy - A Survey and Some New Results,"
5. Conclusions Automatica, 26, pp. 459-474.
Frank, P.M., 1994, "Application of Fuzzy Logic to
In this paper, a neuro-fuzzy scheme for on-line fault Process Supervision and Fault Diagnosis," Proc.
diagnosis was applied to the induction machine. This IFAC Safeprocess 94, Espoo, Finland.
FDI approach relies on combinations of neural model- Garcia, E. and Frank, P.M., 1997, "Deterministic
ling and fuzzy logic which can deal effectively with Nonlinear Observer Based Approaches to Fault
nonlinear dynamics and uncertainties. Diagnosis: A survey," Control Eng. Practice, vol.
The proposed neuro-fuzzy FDI scheme is based on 5(5), pp. 663-670.
60

The Journal of Engineering Research Vol. 8 No. 1 (2011) 53-60

Isermann, R., 1998, "On Fuzzy Logic Applications for Supervision and Fault Detection in Robots using
Automatic Control, Supervision and Fault Nonlinear and Fuzzy Logic Residual Evaluation,"
Diagnosis," IEEE Trans. on Systems, Man and IEEE Trans. on Control System Technology, Vol.
Cybernetics, Vol. 28(2). 4(3), pp. 274-282.
Jiang, B., Staroswiecki, M. and Cocquempot, V., 2001, Simani, S. and Fantuzzi, C., 2002, "Neural Networks
"Robust Observer Based Fault Diagnosis for a for Fault Diagnosis and Identification of Industrial
Class of Nonlinear Systems with Uncertainty," Processes," Proc. of the 10th European
Proc. 40th IEEE Conference on Decision and Symposium on Artificial Neural Networks,
Control, CDC'01, Orlando, USA. Bruges , Belgium.
Mamdani, E. and Assilian, S., 1995, "An Experiment Takagi, T. and Sugeno, M., 1985 "Fuzzy Identification
in linguistic Synthesis with Fuzy Logic of Systems and Applications to Modelling and
Controller," Int. J. Man-Machine Studies, 7(1). Control," IEEE Trans. On system, Man and
Norgaard, M., Ravn, O., Poulsen, N.K.. and Hansen, Cybernetics, Vol. 15(1).
L.K., 2000, "Neural Network for Modelling and Theilliol, D., Sauter, D. and Vela Valdes, L.G., 1997,
Control of Dynamic Systems," Springer Verlag, "Integration of Qualitative and Quantitative
London. Methods for Fault Detection and Isolation,"
Patton, R. and Chen, J., 1997, "Observer-Based Fault Proc. IFAC Safeprocess 97, Hull, U.K.
Detection and Isolation: Robustness and Uppal, F.J., Patton, R.J. and Palade, V., 2002, "Neuro-
Applications," Control Eng. Practice, Vol. 5(5), Fuzzy Based Fault Diagnosis Applied to an
pp. 671-682. Electro- Pneumatic Valve," 15th IFAC World
Schneider, H. and Frank, P.M., 1996, "Observer Based Congress, Barcelona, Spain.