An Effective Neural Approach for the Automatic ... - Semantic Scholar

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 12, DECEMBER 2008

4277

An Effective Neural Approach for the Automatic Location of Stator Interturn Faults in Induction Motor Monia Ben Khader Bouzid, Gérard Champenois, Najiba Mrabet Bellaaj, Laurent Signac, and Khaled Jelassi

Abstract—This paper presents a neural approach to detect and locate automatically an interturn short-circuit fault in the stator windings of the induction machine. The fault detection and location are achieved by a feedforward multilayer-perceptron neural network (NN) trained by back propagation. The location process is based on monitoring the three-phase shifts between the line current and the phase voltage of the machine. The required data for training and testing the NN are experimentally generated from a three-phase induction motor with different interturn short-circuit faults. Simulation, as well as experimental, results are presented in this paper to demonstrate the effectiveness of the used method. Index Terms—Diagnosis, induction machine, interturn short circuit, neural network (NN), phase shifts.

I. I NTRODUCTION

W

ITH the increase in production capabilities of modern manufacturing systems, plants have to run continuously for extended hours. Unexpected downtime due to machinery failures has become more costly than before. Therefore, a monitoring system is becoming necessary, to increase the availability and life span of machines. In fact, the monitoring of machine operating is useful to warn of impending failures, prevent further damage, and reduce maintenance costs. Induction motor (IM) is the basic component in the majority of industry process; it is mostly subject to hazardous environments such as severe shocks, vibration, heat, friction, dust, etc. Consequently, the issues of preventive maintenance and noninvasive diagnosis of the IM are of great concern in the last decades. It is required to detect, locate, and then identify different kinds of failure modes that can occur within a machine. Therefore, an efficient and permanent condition monitoring system offers powerful means to providing warning and predicting the faults at an early stage of appearance.

Manuscript received March 2, 2008; revised August 7, 2008. First published October 31, 2008; current version published December 2, 2008. M. B. K. Bouzid is with the Laboratoire des Systèmes Electriques, Ecole Nationale d’Ingénieurs de Tunis, and the High Technical Institute of Tunis, 1002 Tunis, Tunisia (e-mail: [email protected]). G. Champenois and L. Signac are with the Laboratoire d’Automatique et d’Informatique Industrielle, Ecole Supérieure d’Ingénieurs de Poitiers, University of Poitiers, 86022 Poitiers, France (e-mail: [email protected]). N. M. Bellaaj is with the Laboratoire des Systèmes Electriques, Ecole Nationale d’Ingénieurs de Tunis, and the Higher Computing Institute in Tunis, 1002 Tunis, Tunisia. K. Jelassi is with the Laboratoire des Systèmes Electriques, Ecole Nationale d’Ingénieurs de Tunis, 1002 Tunis, Tunisia (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TIE.2008.2004667

The IM is considered as a robust machine, but similar to other rotating electrical machines, it is subjected to both electrical and mechanical faults. These faults can be classified according to their location: stator and rotor. According to an IEEE and Electric Power Research Institute motor reliability study [1], stator faults are responsible for 37% of the IM failures. Stator winding faults are considered as the main types of faults in the stator, which are about 30%–40% [2]. These faults are due to turn to turn, phase to phase, or winding to earth short circuit [3]. The aim of this paper is the early detection and the location of an interturn short circuit in the stator windings of the IM. This type of fault is caused by the combination of various stresses acting on the stator, such as thermal, electrical, mechanical, and environmental stresses. This stressful condition can lead to the deterioration and eventual breakdown of the winding insulation. Insulation failure will, in most cases, lead to interturn and phase to phase faults and a destructive effect on the stator coils [4]. The detection of stator interturn short-circuit fault during the normal operation of IM is a rather difficult task. The main problem is connected with their destructive character and a tendency to a rapid transition. Therefore, early detection of interturn shorts during motor operation would eliminate subsequent damage to adjacent coils and the stator core, reducing repair costs and, consequently, motor downtime. During the last two decades, the fault detection and diagnosis of rotary systems have become a vigorous area of work. Substantial research has been carried out for the development of various techniques and methods for fault detection and diagnosis. Our survey of some relevant literature reveals three general solution approaches for supporting the task of fault monitoring [5], [6]. The first approach is based on signal analysis [7]–[15], which uses the techniques of time domain, frequency domain, time–frequency domain, and high order spectra. The second class is based on analytical approach [16]–[19], which involves detailed mathematical models that use some measured input and output and generate feature such as residuals (defined as the difference between the nominal and the estimated model parameters), parameter estimation, and state estimation. The third approach is the knowledge-based approach, which is used to automate the analysis of the measured signals by incorporating the artificial intelligence (AI) tools into the online monitoring schemes [20]. These AI tools are of great practical significance in engineering to solve various complex problems normally requiring human intelligence [2], [21], [22]. Among these tools, the powerful ones used in the fault diagnosis are [23]: expert system, fuzzy logic system, artificial neural networks (NNs), and support vector machines.

0278-0046/$25.00 © 2008 IEEE

Authorized licensed use limited to: CNUDST. Downloaded on December 5, 2008 at 13:37 from IEEE Xplore. Restrictions apply.

4278


NN technique has received considerable attention over the last few years [24]. It is intensively studied for fault diagnosis, and it had been successfully applied in diverse industrial applications such as pneumatic [25], chemical [26], [27], aerospace [28], and renewable energy [29]. The fault diagnosis for electrical equipment and systems has received the greatest part of the research work interest [30]–[32]. NNs had gained popularity over other techniques due to, first, their generalization capability during real-time inferences, which means that they are able to perform satisfactorily even for unseen fault. Second is their ability to learn from the examples of the relationships between the fault signatures and the corresponding operating conditions of the system, while ignoring the state’s internal process. Thus, the NN technique is robust and is an approach less model dependent for fault detection when the fault signature can be obtained directly from the available data. On the other hand, the use of the NN technique presents the advantage to be easily integrated in a comprehensive monitoring system, for complex process, using already the AI technique. The fundamental requirement for successful implementation of a fault diagnosis technique based upon an NN is the availability of relevant rich information, which is set of input data for each of the faults in question. That is why the inputs of the NN had to be meaningful indicators of fault. Selecting such a data set from a seemingly infinite of information is a difficult task. However, the best choice of suitable fault indicator is to find the parameters that give the most information about the condition of the system and discard the rest. Different kinds of fault indicators are used as NN inputs in extensive research works based on NN approach for IM stator fault diagnosis. We can find the use of the current and speed [33], [34], the three currents and three voltages [35], the negative and positive sequence stator currents, the slip and the rated slip [21], power spectral density of the residual currents [36], noisy residual current [37], current and vibrations signals [38], stator current Park’s vector patterns [39], and vibration spectra [40]. In our approach, we have selected, as the inputs of the NN, the three-phase shifts between the current and the phase voltage of the IM. In fact, in [41], the authors had also denoted that the phase shift is more preferable than the others, as interturn short-circuit fault feature signal. However, it is regretful that, in [41], the study is limited only to the detection of fault by the simple appearance of unbalance of the three-phase shifts. Thus, this method cannot give us any information about the location of fault or about the load conditions of the machine. Therefore, in this paper, a rigorous study has shown that the behavior of the three-phase shifts under fault conditions presents novel and powerful indexes, which allow the detection of the stator fault and the location of the faulty phase. These results mark the originality of this paper. Thus, the three-phase shifts are considered as robust and efficient indicators of stator fault. Consequently, monitoring these three-phase shifts by an NN allows one to detect and locate automatically an interturn short-circuit fault overcoming the problem of the noise and under different load conditions.

Fig. 1.

Block diagram of the fault location procedure.

This paper is organized as follows. A description of diagnostic system based on the NN is presented in the second section. In the third section, we describe the model including the stator fault that is used to get the simulated database for the training and the test procedure of the NN. The simulation results of the faulty phase location by the NN are detailed in the fourth section. For the experimental validation of the proposed method, the accomplishment of experimental tests on an experimental bench is indispensable. At this way, several experimental tests were done at Laboratoire d’Automatique et d’Informatique Industrielle (LAII), University of Poitiers, Poitiers, France. The results of the experimental validation of the proposed method are given in section five. Finally, conclusions and future works are provided in section six. II. F AULT D IAGNOSIS S YSTEM The proposed work consists in automating the detection and the location of an interturn short-circuit fault on the stator windings of a three-phase IM by using a feedforward multilayer-perceptron (MLP) NN. The intent of this procedure is to orientate the maintenance task, in order to avoid wasted time and the complete deterioration of the motor. Therefore, in Fig. 1, we illustrate the block diagram of the fault location procedure. The first step of this procedure is the acquisition of the three currents and the three-phase voltages from the machine in order to extract the three-phase shifts (Phia, Phib, and Phic) between the line current and the phase voltage. For this reason, we have developed an efficient algorithm on Matlab, which computes these three-phase shifts with a sampling step of 0.5 ms. The training procedure is performed offline, using the backpropagation algorithm, where the NN had to learn the relationships between the fault signatures (NN inputs) and the corresponding operating conditions (NN outputs) to be able to locate correctly the faulty phase. However, as shown in Fig. 1, the NN has three inputs, which are the three-phase shifts and three outputs corresponding to the three phases of the IM, where the fault can occur. If a short circuit is detected and located on one of the three phases, the corresponding NN output is set to “one”; otherwise, it is “zero.” III. IM F AULTY M ODEL The basis of any reliable diagnostic method is a good understanding of the machine behavior in healthy state and under fault conditions; it is also important to find the signature, related to the fault. The signature must be an electrical variable that is easily measured and very sensitive to the fault.


BOUZID et al.: NEURAL APPROACH FOR THE AUTOMATIC LOCATION OF FAULTS IN INDUCTION MOTOR

4279

with X = [ ids

Fig. 2.

Stator faulty model in dq frame.

As a first study, we have used simulated data to perform the NN and assess its performances. For this reason, a suited model which must take into account the presence of interturn shortcircuit fault in the stator winding of an IM is required. The used model is an original one, taking into account a short circuit in the stator. It was proposed first by [42]. It can explain the fault in the stator only when the fault occurs on only one phase. Hence, in [43], the author had extended this model to the general case, where he added, at each phase, the same element of fault to explain the short circuit in each phase. In faulty case, the model can be characterized by two modes. The common mode corresponds to the dynamic model in healthy operation of the machine (Park’s model), and the differential mode model explains the faults. This model, which is very simple to implement because it is expressed in Park’s frame, offers the advantage to explain the defect through a short-circuit element Qcck dedicated to each phase of the stator (k = 1, 2, 3). Fig. 2 shows the global stator faulty model of a squirrel cage IM dedicated to interturn short-circuit windings in dq Park’s axis with global leakage referred to the stator. The model of the differential mode introduces two parameters defining the faults in the stator. 1) The location parameter θcck , which is a real angle between the interturn short-circuit stator winding and the first stator phase axis. This parameter can take only the three values 0, 2π/3, and 4π/3, corresponding to the short circuit on the phase as , bs , or cs , respectively. 2) The detection parameter ηcck , which is equal to the ratio between the number of interturn short-circuit windings and the total number of turns in the healthy phase. This parameter permits one to quantify the unbalance. ηcck =

Number of shorted turns . Number of turns on healty winding

(1)

The short-circuit currents in the model of the differential mode can be expressed as follows: idqcck =

2 ηcck · P (−θ)Q(θcck )P (θ) · Udqs . 3 Rs

(2)

More detailed description regarding the faulty model is presented in reference [44]. The state’s representation of the faulty model is as follows: X˙ = A(ω).X + B.U (3) Y = C.X + D(ηcc , θcc ).U

iqs

Φdr

Φqr ]T

U = [ Uds Uqs ]T ⎡ Rs +Rr ⎤ Rr ω − Lf ω Lm .Lf Lf ⎢ −ω ⎥ Rr r − RsL+R − Lωf ⎢ Lm .Lf ⎥ f A(ω) = ⎢ ⎥ ⎣ Rr 0 − LRr 0 ⎦ m 0 Rr 0 − LRmr

T 1/Lf 0 0 0 B= 0 1/Lf 0 0

1 0 0 0 C= 0 1 0 0

3 2 D(ηcc , θcc ) = ηcck .P (−θ).Q(θcck )P (θ) 3.Rs k=1

cos(θcck )2 cos(θcck ) sin(θcck ) Qcck = sin(θcck )2 cos(θcck ) sin(θcck )

cos(θ) − sin(θ) P (θ) = sin(θ) cos(θ) ids , iqs dq stator current components; Φdr , Φqr dq rotor flux linkages; Uds , Uqs dq stator voltages; θ electrical angle; ω = dθ/dt. Rs , Lf , Rr , and Lm are the stator resistance, global leakage inductance referred to the stator, rotor resistance, and magnetizing inductance, respectively. IV. S IMULATION R ESULTS A. Characteristics of the Simulated Phase Shifts Features extracted from one or more characteristics of the machine were utilized as indicators of faults [38] to effectively discern the fault. The choice of these indicators is an important initial step in any monitoring and fault diagnosis system. Its accuracy directly affects the final monitoring results. In this section, we will present the study of the behavior of the three-phase shifts in the presence of interturn short-circuit fault and under different load conditions. According to their good features, we have selected the three-phase shifts (Phia, Phib, and Phic) as the best suitable inputs of the NN. Under normal operation and balanced conditions, machine performance gives phase voltages and line’s currents equal in magnitude and shifted by 120◦ electrical, but under faulty operation, the currents are altered and, consequently, the phase shifts. To investigate the currents of the IM under interturn shortcircuit fault, we have simulated the model shown in Fig. 2 by “Simulink” under the environment of Matlab. Steady-state operation under constant speed is assumed with a fixed sampling step of 0.5 ms. The resolution of differential equations of the model was made by the order four Runge–Kutta method. The simulated machine is a 1.1-kW IM


4280


Fig. 5.

Phase shift characteristics for fault on phase bs .

Fig. 6.

Phase shift characteristics for fault on phase cs .

Fig. 3. Fault effect on the three line’s currents.

Fig. 4. Phase shift characteristics for fault on phase as .

having 464 turns per phase winding on the stator and 28 bars in the rotor. Fig. 3 shows the profiles of the simulated three line’s currents with no load torque and under a stator fault of 48-shorted turns on one of the three phases (as , bs , and cs ) introduced at 0.5 s. The stator fault is expressed by an unbalance of the three currents. We can observe that, when a fault occurred on one of the three phases, an important arising of the current appears particularly in the corresponding phase. Thus, it is clear that an interturn short circuit principally affects the stator current of the faulty phase in peak value. The other stator phase currents suffer smaller influences. Furthermore, we can notice that the alteration of the current values is achieved according to a circular permutation in direct sense (ias→ ibs→ ics ). However, in this figure, we cannot see the significant change of the value of the phase shift between the phase voltage (which experiences little changes due to the unbalanced supply) and the line’s current of the machine. To study the fault effect on the phase shift, we have plotted, in Figs. 4–6, the characteristics of the three-phase shifts under a load torque (T ) of 3 N · m, as function of the faulty turn number

(n), in the case of fault on the phases as (Fig. 4), bs (Fig. 5), and cs (Fig. 6). It is very important to note that, for any number n of faulty turns, the simultaneous three values of the phase shifts are quite distinct. This difference of values is linked to the importance of the fault, which is expressed by the number n of shorted turns. It can be also noted that, in the case of a stator fault on one of the three phases, the smallest value of the three-phase shifts is usually upon the phase where the fault has occurred. With this index, we can localize the faulty phase. Furthermore, for a short circuit on phase as and for a power supply in direct sense (Fig. 4), the phase shifts are in the following increasing order: (Phia→ Phic→ Phib). We had also studied the load effect on the phase shifts. For this purpose, we have shown, in Figs. 7–9, the three-phase shifts in the case of stator fault in phases as , bs , and cs under two load conditions (T = 3 and 7 N · m). These figures show that the phase shifts do not present any overlap between the values, even with different torques and in faulty conditions. We have also studied the behavior of the phase shift with noisy conditions, and we have obtained the same curves of the phase shifts; therefore, we deduce that the phase shift values are not sensitive to the noise.



Fig. 7. Phase shifts corresponding to fault on phase as and under two different load conditions.

4281

Fig. 10. Simulated training input data set of NN.

After this detailed analysis, we can conclude that the novel distinctive features extracted from the behavior of the threephase shifts under fault conditions are efficient indicators to detect a stator fault, locate the phase where this fault has occurred, and also provide the information about the fault severity (the number of the faulty turns). Thus, the three-phase shifts constitute an ideally noninvasive sensor, which provides the adequate data for the neural diagnosis system in order to ensure effective monitoring. B. Database Selection

Fig. 8. Phase shifts corresponding to fault on phase bs and under two different load conditions.

Fig. 9. Phase shifts corresponding to fault on phase cs under two different load conditions.

A training database constituted by input and output data sets has been applied to train the NN. The input data are collected through simulations by Matlab, using the model in Fig. 2. This possibility is of special interest for collecting data under any different faulty conditions, since generally, those data are not available for the real machine. To achieve a good location of the IM faulty phase, the training data should represent the complete range of the operating conditions, which contain all possible fault occurrences and even the healthy cases. For this purpose, the input data set, which is shown in Fig. 10, is composed by a successive range of several examples in different operating conditions of the IM. Each example is composed by the three-phase shifts. All these examples are presented to the NN under three load conditions (T = 7, 5, and 3 N · m) and represent the following different operating cases of the IM: healthy (three points) and fault of an odd number n of shorted turns (n = 1, 3, 5, 7, 9, 11, 13, and 15) on each stator phase [24 (8 × 3) points]. Thus, a total of 75(24 × 3 + 3) training samples is taught to the NN (Fig. 10). We have reserved the fault examples corresponding to an even number of shorted turns (n = 2, 4, 6, 8, 10, 12, and 14) to the test. The output data set is formed by the following desired outputs (Ti ) which indicate the state of each phase: T1 = 1 for a short circuit at phase as ; otherwise, T1 = 0; T2 = 1 for a short circuit at phase bs ; otherwise, T2 = 0;


4282


Fig. 11. NN architecture.

T3 = 1 for a short circuit at phase cs ; otherwise, T3 = 0. Therefore, the output states of the NN are set to the following: [0; 0; 0] no fault (healthy mode); [1; 0; 0] fault occurred on phase as ; [0; 1; 0] fault occurred on phase bs ; [0; 0; 1] fault occurred on phase cs .

Fig. 12.

Training performances of the NN.

Fig. 13.

Training outputs and errors of NN.

C. Structure of the NN In this paper, a feedforward MLP network performed by back-propagation training algorithm is found adequate [44]. The number of inputs and outputs is fixed by the number of the fault indicators, which are the three-phase shifts, and the number of fault classes, which are the three phases of the IM, respectively, but the optimal number of neurones in the hidden layer is not known. Indeed, if the number of the neurones in the hidden layer is too few, the NN cannot learn well, and if this number is too large, the NN may simply memorize the training set (overtraining). However, in the two cases, the NN does not present a good generalization. To avoid the overtraining problem, for each assumed number of neurones in the hidden layer, the NN must be trained and then tested successively with separate data not previously introduced during the training in order to check both the training and test errors. First, we start with a few neurones (i.e., two neurones); then, we add other ones until an appropriate number that provides us a low training and test mean square error (mse) is reached and no further improvement in classification could be obtained. In our case, the best training and test performances of the NN are obtained with five neurones in the hidden layer. Thus, the optimal structure of the MLP network adopted for the location of the faulty phase of IM is shown in Fig. 11. This network has three inputs (Phia, Phib, and Phic), three outputs (phases as , bs , and cs ), and a hidden layer of five neurons. The activation functions of the hidden and output layers are “tansig” and “logsig,” respectively. D. Training Results The performance of the NN is indicated by its mse shown in Fig. 12. After learning for 5000 epochs, the NN reaches a low training mse that is equal to 7.28 × 10−21 . The training outputs and errors of NN are shown in Fig. 13. The training errors are very low, proving that the NN has well learned the input data and has correctly reproduced the desired outputs.

E. Test Results The test procedure was made by a test data set that is different from the training data set to assess the generalization capacity of the adopted network. The NN is tested for faults on each of the three phases; satisfactory results are achieved. The test data set is composed by three examples of healthy operating (three points) and 20 examples of interturn short-circuit faults on one phase with a number of shorted turns (n = 2-4-6-8-10-12-14-16-18-20-2224) and with load torques different from those of the learning cases (2-4-6 N · m; 20 points). Figs. 14–16 show the NN test outputs and their errors for faults on phases as , bs , and cs , respectively. As can be easily seen in Fig. 14, the NN outputs (O1 , O2 , O3 ) are set to (1, 0, 0) to indicate faults on phase as with good accuracy. This means that the NN is able to locate correctly the faults on phase as . In Fig. 15, the NN outputs are set to (0, 1, 0), indicating the presence of faults on phase bs with low errors.



4283

Fig. 17. Description of the experimental test bed. TABLE I CHARACTERISTICS OF MOTOR IM1 STAR CONNECTED Fig. 14. Test outputs and errors for faults on phase as .

In the same way, Fig. 16 shows that the NN has recognize faults on phase cs with satisfactory exactitude. In this case, the NN outputs are set to (0, 0, 1). Thus, according to these test results, we can conclude that the NN is able to recognize correctly the healthy operating and locate any faults on one of the three phases of the machine. It also presents a good generalization capacity. It should be noted that the NN can locate correctly a fault of 24 shorted turns on each of the three phases, knowing that the NN had learned up to 15 turns. V. E XPERIMENTAL R ESULTS A. Test Bench Description Fig. 15. Test outputs and errors for faults on phase bs .

Fig. 16. Test outputs and errors for faults on phase cs .

In order to validate the method of location of an interturn short-circuit fault proposed in this paper, experimental data are required. For this purpose, many tests are carried out on two similar test beds. Each of the two test beds is assembled in the LAII and designed to study the behavior of IM under different stator fault conditions through various numbers of shorted turns on the stator windings. Each test bed is composed, as shown in Fig. 17, by a threephase squirrel cage IM of 1.1 kW, two pole pairs, star connected, three-phase squirrel cage IM, and a dc generator loaded by a variable resistance which is used as load system. The variation of the motor load of the machine is achieved through the variation of the resistance connected to the dc machine. The two motors (IM1 and IM2 ) have the same characteristics reported in Table I. They are supplied directly by a balanced three-phase source of sinusoidal voltages. Each of them has a rotor cage of 28 bars and 464 turns per phase winding in the stator. The dc generator is connected to the electric IM through flexible coupling and a torque meter for the load torque measurement. The torque value is displayed through a personal


4284


Fig. 18. Configuration of the stator windings of the motor IM1 with additional access points.

Fig. 19. Configuration of the stator windings of the motor IM2 with additional access points.

computer (PC) in order to guarantee the same operation conditions in all the accomplished tests. The two motors are equipped by specific access points to diverse turns of stator windings, in order to emulate different cases of interturn short-circuit fault on one phase. The IMs IM1 and IM2 are designed to perform with faults through low and high number of shorted turns, respectively. The access points inserted in IM1 are at the level of (1-4-13) turns on phase as and (2-8-29) turns on phase bs , from the middle point of the stator winding, as shown in Fig. 18. On the other hand, the turns which can be accessed on the IM2 are (18-58116) on the phase as and (29-58-116) on the phase bs , from the beginning of the stator windings, as shown in Fig. 19. It is clear that, on the two motors, we cannot made tests for faults on phase cs , since we have no intermediate point access. Thus, to acquire the correspondent signals, we had to make circular permutations of the three currents, as explained previously in Section IV-A. B. Test Description To verify and validate the simulated results of the faulty phase location by NN, we had to extract the experimental threephase shifts from the stator currents and the phase voltages of the machine. Therefore, to acquire these signals, we had carried out several experimental tests, where three voltage sensors and three noninvasive current sensors with galvanic insulation are used. The six analogy sensors are connected to the inputs of a data acquisition board (six channels). The latter is connected (see Fig. 17) to a PC, where the Matlab software was used for the computation of the three-phase shifts by a specific developed algorithm and the achievement of the fault diagnosis task by the NN.

Tests had been conducted for several operating conditions. For each test, three samples of the six signals (three currents and three voltages) are recorded by the data acquisition board and then submitted to the PC. The acquisition time of each sample is about 1 s with a sampling step of 0.5 ms. To emulate an interturn short circuit on the stator winding of the motor, an external adjustable resistor Rs was used to restrict the circulating current in the shorted portion of the stator winding to a safe level to avoid a permanent motor winding damages. In these tests, the loop current in the shorted turns was not allowed to exceed 10 A. In this paper, we have carried out two different sequences of tests. Tests N◦ 1 and N◦ 2 are performed on motors IM1 and IM2 , respectively. The tests N◦ 1 and N◦ 2 are used to collect the useful data set for the training and the test procedure of the NN, in order to verify and validate experimentally the performance of the NN. These data are also used to evaluate the capacity of the generalization of the NN, when the data are provided from another machine with the same characteristics and faults with other number of shorted turns. 1) Test N◦ 1: In this case, the tests were conducted on the motor IM1 turning in the direct sense and after heating it at the nominal temperature. According to the configuration of the access points inserted in the stator windings of IM1 shown in Fig. 18, we have achieved the following tests for different operation conditions of the motor: 1) healthy operation; 2) an interturn short-circuit faulty operation across (3-9-12) shorted turns on phase as ; 3) an interturn short-circuit faulty operation across (6-2127) shorted turns on phase bs . We have obtained seven cases of operation conditions. Each operating case was performed under eight different load torques, ranging from unload to full load (0 to 7 N · m). Therefore, 56 (7 × 8) tests are carried out, and a total of 168 [56 × 3(by circular permutations)] samples of the six signals is collected to obtain the different phase shifts for the training and test data sets. 2) Test N◦ 2: These tests are achieved on the motor IM2 turning in the inverse sense and, also, after heating it at the nominal temperature. The configuration of the IM2 , shown in Fig. 19, leads to accomplish the following tests for different operation states of the motor: 1) healthy operation; 2) three cases of an interturn short-circuit faulty operation corresponding to faults of (18-40-58) shorted turns on phase as ; 3) two cases of an interturn short-circuit faulty operation corresponding to fault of (29 and 58) shorted turns on phase bs . We obtained six cases of operating conditions. Each operation case was performed under eight different load torques, ranging from unload to full load (T = 0 to 7 N · m). Therefore, 48 (6 × 8) tests are made, and a total of 144 [48 × 3(by circular permutations)] samples is formed only for the test data set.



Fig. 20. Experimental phase shifts in the case of fault on phase as .

4285

Fig. 22. Experimental phase shifts in the case of fault on phase cs .

Fig. 21. Experimental phase shifts in the case of fault on phase bs .

C. Characteristics of Experimental Phase Shifts

Fig. 23. Training input data set (from IM1 ).

We have plotted, in Figs. 20–22, the three characteristics of the experimental phase shifts for a fault on phases as , bs , and cs , respectively, with a load torque of 3 N · m. This is in order to compare them with the simulated ones (Figs. 4–6). It is important to take into account that the computed phase shifts, from experimental signals, may have values different from those of the simulation, even without fault. This is because the motor may have little abnormalities in construction (obviously, depending on its quality control) and also because of unbalanced supply network, internal errors in sensors, quantification errors, and others. Despite this difference in values, the curves of the simulated and experimental phase shifts have similar evolution as long as there is an increase of fault. It is also important to note that the curves of the experimental phase shifts are not strictly monotonic. This is due to the permutation of the experimental values. Moreover, in practice, the machine presents small asymmetries of construction (magnetic unbalance, poor concentricity of the stator and rotor,. . .); thus, the phase shifts obtained experimentally can be slightly different for the same fault on two different phases.

Ideally, it would take an experimental machine with the same exit of turns on the three phases, make the three tests, and then take the average of the measurements to minimize these effects, mentioned earlier. D. Training Results The training input data set, which is shown in Fig. 23, is composed by a successive range of examples, acquired from the motor IM1 . Each example is presented under three cases of load conditions (T = 7, 5, and 3 N · m). The examples corresponding to different states of machine are as follows: 1) healthy operating: one case for each load condition (three points); 2) fault on phase as : five cases of (n = 3-6-9-12-21) shorted turns under the three load conditions [15 points (5 × 3)]; 3) fault on phase bs : five cases of (n = 3-6-9-12-21) shorted turns under the three load conditions [15 (5 × 3) points]; 4) fault on phase cs : five cases of (n = 3-6-9-12-21) shorted turns under the three load conditions [15 (5 × 3) points].


4286


Fig. 24. Training performance of the NN. Fig. 26.

Test outputs and errors of the NN (using IM1 ).

Fig. 27.

Input data for the test (from IM2 ).

Fig. 25. Training outputs and errors of the NN (using IM1 ).

Using the same structure of the NN, which had performed with the simulated data, shown in Fig. 11, and with the same parameters (activation function, epoch number, and training algorithm), we have trained the NN with an experimental data set shown in Fig. 23. Satisfactory training results are obtained. In fact, in Fig. 24, we can see that after 5000 epochs, the reached mse is 1.23 × 10−22 . The training outputs and its corresponding errors are shown in Fig. 25. It is clear that the classification of healthy and faulty examples is achieved successfully with low errors. E. Test Results 1) Test Results Using Data From the Same Motor IM1 : To test the performances of the network with data acquired from the same machine, we have introduced to the NN successively a range of 21 fault examples on each of the three phases of the IM1 , which are different from those of the training examples concerning the fault (number of turns in short circuit) and the load conditions (different torques). It can be observed from the test outputs and its corresponding errors shown in Fig. 26

that the network presents a good generalization since it had recognized the total examples with very low errors. 2) Test Results Using Data From Other Identical Motor: The objective of this section is to evaluate the capacity of generalization of the NN, when the input data are acquired from another identical machine having the same power (1.1 kW) and with more important faults. However, for this purpose, the database of the test procedure is composed by successive examples of four cases of interturn short-circuit faults with (n = 18-29-40-58) shorted turns on each phase of the motor IM2 . Each fault is presented to the NN under three load conditions (T = 0, 2, and 4 N · m). Thus, for each phase as , bs , and cs , a total of 12 (4 × 3) examples of fault is presented to the NN (Fig. 27). Obviously, this data set is quite distinct from the training data set, since it is acquired from the motor IM2 while the data set of the training procedure is obtained from the motor IM1 . The test outputs and its errors are shown in Fig. 28. It is clear that, according to the neglected errors of the test (about 10−11 ), the NN is able to recognize perfectly faults with a big number of shorted turns, produced



Fig. 28. Test outputs and errors of the NN (using IM2 ). TABLE II PHASE SHIFTS OF 18 SHORTED TURN FAULT ON PHASE as UNDER FULL LOAD

4287

The novel features of phase shifts including information about the detection and the location of fault let them to be reliable indicators of interturn short-circuit fault in the stator windings of the IM. The simulation results are verified and experimentally validate on two IMs. The successful obtained results prove that this approach is effective to ensure a reliable and precise fault diagnosis process. Furthermore, in this paper, it is important to mention an interesting benefit of this diagnosis method, which is resumed in the fact that, once the NN is trained by data from a machine and performed correctly, it can find its application with other machine having the same power without training it again. As further scope of this paper, we can extend our research to identify the number of the shorted turns on the faulty phase and, in this fact, achieve a comprehensive diagnosis procedure. It is also worthwhile to find out supplementary suitable fault indicators to take into account the effect of the rotating sense on the phase shifts.

ACKNOWLEDGMENT This work was performed in the frame of the Comité Mixte Franco-Tunisien de Coopération Universitaire (CMCU) Project (CMCU 04/S 1122). The authors would like to thank LAII, University of Poitiers, for the experimental data sets for this paper.

on another similar machine. Knowing that, it was learned faults with small number of shorted turns. The inverse case is also investigated. The test results do not give us good accuracy for all test examples. That is why we proposed for the application of this method to teach the NN faults with small number of shorted turns. The synthesis of this experimental study let us conclude that the good obtained results confirm the efficiency and the feasibility of this suggested procedure and how accurate this method is in locating the faulty phase of an induction machine by a simple NN. It can also be mentioned that the robustness of the suggested method is demonstrated by the fact that, even if we teach the NN faults from a motor, the same NN structure can be applied on other identical motor having the same power to recognize faults more important than the learned ones. We also analyzed the effect of the rotation sense of the motor on the phase shifts. It is very important to note that the change of the IM rotation sense permutes only two phase shifts. This is shown in Table II. In this table, we reported the values of the three-phase shifts for an interturn short-circuit fault of 18 shorted turns on phase as , performed in the direct and inverse senses. VI. C ONCLUSION AND F UTURE W ORK This paper presents an effective method to detect and locate automatically an interturn short circuit on the stator windings of the IM. The diagnostic process is automated through monitoring simultaneously the values of the three-phase shifts between the line current and the phase voltage by a simple MLP NN.

R EFERENCES [1] A. Siddique and G. S. Yadava, “A review of stator fault monitoring techniques of induction motors,” IEEE Trans. Energy Convers., vol. 20, no. 1, pp. 106–114, Mar. 2005. [2] M. Awadallah and M. Morcos, “Application of AI tools in fault diagnosis of electrical machines and drives—An overview,” IEEE Trans. Energy Convers., vol. 18, no. 2, pp. 245–251, Jun. 2003. [3] G. K. Singh and S. A. S. Al Kazzaz, “Induction machine drive condition monitoring and diagnostic research—A survey,” Elect. Power Syst. Res., vol. 64, no. 2, pp. 145–158, Feb. 2003. [4] C. Gerada, K. J. Bradley, M. Summer, and P. Wheeler, “Operating induction motor drives with turn-to-turn faults,” in Proc. IEEE IEMDC, San Antonio, TX, 2005, pp. 770–776. [5] A. Siddique, G. S. Yadava, and B. Singh, “A review of stator fault monitoring techniques of induction motors,” IEEE Trans. Energy Convers., vol. 20, no. 1, pp. 106–114, Mar. 2005. [6] V. Venkatasubramanian, R. Rengaswamy, N. Kavuri, and K. Yin, “A review of process fault detection and diagnosis Part III: Process history based methods,” Comput. Chem. Eng., vol. 27, no. 3, pp. 327–346, Mar. 2003. [7] M. Eltabach, J. Antoni, and M. Najjar, “Quantitative analysis of noninvasive diagnostic procedures for induction motor drives,” Mech. Syst. Signal Process., vol. 21, no. 7, pp. 2838–2856, Oct. 2007. [8] M. Benbouzid and H. El, “A review of induction motors signature analysis as a medium for faults detection,” IEEE Trans. Ind. Electron., vol. 47, no. 5, pp. 984–993, Oct. 2000. [9] M. E. H. Benbouzid et al., “Induction motors’ detection and localization using stator current advanced signal processing techniques,” IEEE Trans. Power Electron., vol. 14, no. 1, pp. 14–22, Jan. 1999. [10] T. W. S. Chow and S. Hai, “Induction machine fault diagnostic analysis with wavelet technique,” IEEE Trans. Ind. Electron., vol. 51, no. 3, pp. 558–565, Jun. 2004. [11] G. G. Acosta, C. J. Verucchi, and E. R. Gelso, “A current monitoring system for diagnosing electrical failures in induction motors,” Mech. Syst. Signal Process., vol. 20, no. 4, pp. 953–965, May 2006. [12] 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. 1842– 1852, Dec. 2006.


4288


[13] W. Thomson and M. Fenger, “Current signature analysis to detect induction motor faults,” IEEE Ind. Appl. Mag., vol. 7, no. 4, pp. 26–34, Jul./Aug. 2001. [14] S. Nandi and H. A. Tolyat, “Novel frequency domain based technique to detect incipient stator inter-turn faults in induction machines,” in Conf. Rec. IEEE IAS Annu. Meeting, 2000, pp. 367–374. [15] N. Arthur and J. Penman, “Induction machine condition monitoring with higher order spectra,” IEEE Trans. Ind. Electron., vol. 47, no. 5, pp. 1031– 1041, Oct. 2000. [16] R. Isermann, “Model-based fault-detection and diagnosis—Status and applications,” Annu. Rev. Control, vol. 29, no. 1, pp. 71–85, Feb. 2005. [17] S. Bachir, S. Tnani, J. C. Trigeassou, and G. Champenois, “Diagnosis by parameter estimation of stator and rotor faults occurring in induction machines,” in Proc. EPE, Graz, Autriche, Aug. 2001. [18] S. Bachir, S. Tnani, J.-C. Trigeassou, and G. Champenois, “Diagnosis by parameter estimation of stator and rotor faults occurring in induction machines,” IEEE Trans. Ind. Electron., vol. 53, no. 3, pp. 963–973, Jun. 2006. [19] F. Filippetti et al., “State of art of model diagnostic procedures for induction machines inter-turns short circuits,” in Proc. IEEE SDEMPED, Gijon, Spain, Sep. 1999, pp. 19–31. [20] V. Uraikul, C. W. Chan, and P. Toniwachwuthikul, “Artificial intelligence for monitoring and supervisory control of process systems,” Eng. Appl. Artif. Intell., vol. 20, no. 2, pp. 115–131, Mar. 2007. [21] F. Filippetti, G. Franceschini, C. Tassoni, and P. Vas, “Recent developments of induction motor drives fault diagnosis using AI techniques,” IEEE Trans. Ind. Electron., vol. 47, no. 5, pp. 994–1003, Oct. 2000. [22] F. Filippetti, G. Franceschini, C. Tassoni, and P. Vas, “AI techniques in induction machines diagnosis including the speed ripple effect,” IEEE Trans. Ind. Appl., vol. 34, no. 1, pp. 98–108, Jan./Feb. 2008. [23] A. Siddique, G. S. Yadava, and B. Sin, “Applications of artificial intelligence techniques for induction machine stator fault diagnostics: Review,” in Proc. SDEMPED, Atlanta, GA, Aug. 24–26, 2003, pp. 29–34. [24] K. Bimal, “Neural network applications in power electronics and motor drives—An introduction and perspective,” IEEE Trans. Ind. Electron., vol. 54, no. 1, pp. 14–33, Feb. 2007. [25] M. Karpenko and N. Sepehri, “Neural network classifiers applied to condition monitoring of a pneumatic process valve actuator,” Eng. Appl. Artif. Intell., vol. 15, no. 3/4, pp. 273–283, Jun.–Aug. 2002. [26] Y. Zhou, J. Hahn, and M. S. Mannan, “Fault detection and classification in chemical processes based on neural networks with feature extraction,” ISA Trans., vol. 42, no. 4, pp. 651–664, Oct. 2003. [27] D. L. Yu, J. B. Gomm, and D. Williams, “Sensor fault diagnosis in a chemical process via RBF neural networks,” Control Eng. Pract., vol. 7, no. 1, pp. 49–55, Jan. 1999. [28] Y. M. Chen and M. L. Lee, “Neural networks-based scheme for system failure detection and diagnosis,” Math. Comput. Simul., vol. 58, no. 2, pp. 101–109, Jan. 2002. [29] S. A. Kalogirou, “Artificial neural networks in renewable energy systems applications: A review,” Renew. Sustain. Energy Rev., vol. 5, no. 4, pp. 373–401, Dec. 2001. [30] S. Rajakarunakaran, P. Venkumar, K. Devaraj, and K. S. P. Rao, “Artificial neural network approach for fault detection in rotary system,” Appl. Soft Comput., vol. 8, no. 1, pp. 740–748, Jan. 2008. [31] S. R. Kolla, “Identifying three-phase induction motor faults using artificial neural networks,” ISA Trans., vol. 39, no. 4, pp. 433–439, Sep. 2000. [32] J. F. Martins, V. F. Pires, and A. J. Pires, “Unsupervised neural-networkbased algorithm for an on-line diagnosis of three-phase induction motor stator fault,” IEEE Trans. Ind. Electron., vol. 54, no. 1, pp. 259–264, Feb. 2007. [33] M. Y. Chow and S. O. Yee, “Using neural networks to detect incipient faults in induction motors,” J. Neural Netw. Comput., vol. 2, no. 3, pp. 26– 32, 1991. [34] M. Y. Chow, R. N. Sharpe, and J. C. Humg, “On the application and design of artificial neural networks for motor fault detection,” IEEE Trans. Ind. Electron., vol. 40, no. 2, pp. 181–196, 1993. [35] S. R. Kolla and S. D. Altman, “Artificial neural network based fault identification scheme implementation for a three-phase induction motor,” ISA Trans., vol. 46, no. 2, pp. 261–266, Apr. 2007. [36] M. Bouzid, N. Mrabet, S. Moreau, and L. Signac, “Accurate detection of stator and rotor fault by neural network in induction motor,” in Proc. IEE SSD, Hammamet, Tunisia, Mar. 21, 2007, vol. III, pp. 1–7. [37] M. Bouzid, N. Mrabet, J. Khaled, G. Champenois, and L. Signac, “Location of an inter turn short circuit fault in stator windings of induction motor by neural network,” in Proc. IEE, IET, Colloq. Rel. Electromagn. Syst., Paris, France, May 24, 2007, pp. 32–39.

[38] B.-S. Yang and K. J. Kim, “Application of Dempster–Shafer theory in fault diagnosis of induction motors using vibration and current signals,” Mech. Syst. Signal Process., vol. 20, no. 2, pp. 403–420, Feb. 2006. [39] A. Nejjari, M. Benbouzid, and H. El, “Monitoring and diagnosis of induction motors electrical faults using a current Park’s vector pattern learning approach,” IEEE Trans. Ind. Appl., vol. 36, no. 3, pp. 730–735, May/Jun. 2000. [40] C. T. Kowalski and T. O. Kowaska, “Neural networks application for induction motor faults diagnosis,” Math. Comput. Simul., vol. 63, no. 3–5, pp. 435–448, Nov. 2003. [41] X. Boqiang, L. Heming, and S. Liling, “Feature signal extraction of interturn short circuit fault in stator windings of induction motors,” in Proc. IEEE ICIT, Bangkok, Thailand, 2002, vol. 1, pp. 97–100. [42] E. Schaeffer, “Diagnostic des machines asynchrones: Modèles et outils paramétriques dédiés à la simulation et à la détection de défauts,” Ph.D. dissertation, Université de Nantes, Nantes, France, 1999. [43] S. Bachir, “Contribution au diagnostic de la machine asynchrone par estimation paramétrique,” Ph.D. dissertation, Ecole supérieure d’ingénieurs de Poitiers, Poitiers, France, Dec. 17, 2002. [44] S. J. Hong and G. S. May, “Neural-network-based sensor fusion of optical emission and mass spectroscopy data for real-time fault detection in reactive ion etching,” IEEE Trans. Ind. Electron., vol. 52, no. 4, pp. 1063– 1072, Aug. 2005.

Monia Ben Khader Bouzid was born in Tunisia in 1965. She received the M.S. degree in electrical engineering from the Ecole Nationale d’Ingénieurs de Tunis, Tunis, Tunisia, in 1992. Since 1992, she has been an Assistant Professor with the High Technical Institute of Tunis, Tunis. Since 2004, she has also been with the Laboratoire des Systèmes Electriques, Ecole Nationale d’Ingénieurs de Tunis. Her current research activities concern the application of intelligent techniques to the fault diagnosis of electrical machines.

Gérard Champenois was born in France in 1957. He received the Ph.D. and the “habilitation” degrees from the Institut National Polytechnique de Grenoble, Grenoble, France, in 1984 and 1992, respectively. He is a Professor with the Laboratoire d’Automatique et d’Informatique Industrielle, Ecole Supérieure d’Ingénieurs de Poitiers, University of Poitiers, Poitiers, France. His major fields of interest in research are electrical machines associated with static converter, control, modeling, and diagnosis.

Najiba Mrabet Bellaaj was born in Tunisia in 1966. She received the Diploma in engineering and the Ph.D. degree in electrical engineering in 1988 and 2001, respectively. Since 2002, she has been an Assistant Professor with the Higher Computing Institute in Tunis, Tunis, Tunisia. She also does research with the Laboratoire des Systèmes Electriques, Ecole Nationale d’Ingénieurs de Tunis, Tunis. Her research interests include the application of intelligent techniques to electrical power systems as tools of diagnosis, the identification and prediction of consumption, and the production of electrical energy.



Laurent Signac was born in France in 1973. He received the Ph.D. degree from the University of Poitiers, Poitiers, France, in 2001. He is currently an Assistant Professor with the Laboratoire d’Automatique et d’Informatique Industrielle, Ecole Supérieure d’Ingénieurs de Poitiers, University of Poitiers, Poitiers. His major fields of interest in research are system identification and simulation with neural networks.

4289

Khaled Jelassi was born in Tunis, Tunisia, in 1962. He received the B.S. degree in electrical engineering from the National School of Engineering of Monastir, Monastir, Tunisia, in 1987, the Ph.D. degree in engineering from the Institut National Polytechnique de Toulouse, Toulouse, France, in 1991, and the Habilitation degree from the National School of Engineering of Tunis, Tunis, Tunisia, in 2003. He is currently a Researcher with the Laboratoire des Systèmes Electriques, Ecole Nationale d’Ingénieurs de Tunis, Tunis. His research interest is in the field of supervision and diagnosis of fault in electromechanical systems composed of ac drives coupled to a gear which is driving the load.
