ARTICLE IN PRESS

Mechanical Systems and Signal Processing

Mechanical Systems and Signal Processing 22 (2008) 12161236 www.elsevier.com/locate/jnlabr/ymssp

A simplied scheme for induction motor condition monitoring

Pedro Vicente Jover Rodr guez, Marian Negrea, Antero Arkkio
Laboratory of Electromechanics, Department of Electrical and Communication Engineering, P.O. Box 3000, FIN-02015 HUT, Helsinki University of Technology, Finland Received 14 July 2006; received in revised form 13 November 2007; accepted 17 November 2007 Available online 28 November 2007

Abstract This work proposes a general scheme to detect induction motor fault by monitoring the motor current. The scheme is based on signal processing (predictive lters) and soft computing technique (fuzzy logic). The predictive lter is used in order to separate the fundamental component from the harmonic components. Fuzzy logic is used to identify the motor state. Finite element method (FEM) is utilised to generate virtual data that allows to test the proposed technique and foresee the change in the current under different motor conditions. A simple and reliable method for the detection of stator winding failures based on the phase current amplitudes is implemented and tested. The layout has been proved in MATLAB/SIMULINK, with both data from FEM motor simulation program and real measurements. The proposed method has the ability to work with variable speed drives and avoids the detailed spectral analysis of the motor current. This work shows the feasibility of spotting broken rotor bars, eccentricities and inter-turn short-circuit by monitoring the motor currents. r 2007 Elsevier Ltd. All rights reserved.
Keywords: Induction motor; Fault detection; Fuzzy logic; Predictive lter; Condition monitoring

1. Introduction Three-phase induction motors are the most widely used machines. They are simple and have high reliability. However, owing to the thermal, electrical and mechanical stresses, mechanical failures are unavoidable in induction motors. Early detection of abnormalities in the motor will help to avoid expensive failures. Operators of electric drive systems are under continual pressure to reduce maintenance costs and prevent unscheduled downtimes that result in lost production and loss of nancial income. The modern industry has widely used reliability-based maintenance and condition-based maintenance strategies to reduce unexpected failures and downtime. These techniques can increase the time between planned shutdown for standard maintenance and reduce maintenance and operational costs. The operation of the machine in unsafe condition must also be avoided. Nevertheless, the failures are unavoidable and failure statistics [1] has reported that the percentage of failures in induction motor components is as follow: 1. Bearing related: 40%. 2. Stator related: 38%.
Corresponding author. Tel.: +358 9 451 2391; fax: +358 9 451 2991.

E-mail address: vicent@cc.hut. (P.V.J. Rodr guez). 0888-3270/$ - see front matter r 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.ymssp.2007.11.018

ARTICLE IN PRESS
guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr 1217

3. Rotor related: 10%. 4. Others: 12%. There have been some integrated motor protection systems for monitoring electrical fault in induction motor by analysing the motor current [2,3]. On the other hand, bearing faults are quite difcult to spot from the motor current. Stack [4] showed from experimental results that faults in bearings produce unpredictable and broadband changes in the motor current. The authors arrived to the conclusion that for this failure the motor current is a very poor indicator. However, bearing failures because their precise nature, have a clear route to spot through vibration monitoring [5,6]. Schoen [2] developed an unsupervised, online system for induction motor fault detection based on articial neural network. Firstly, the system utilised a selective lter in order to reduce the amount of harmonics to a manageable number. After sufcient training period, the neural network is able to sign a potential failure condition. While this technique has demonstrated success in identifying an incipient failure, a prerequisite for its operation is a priori fault data, which is clearly no available. This impedes the practical operation of such methods. FEM (nite element method) simulations can remove this requirement by predicting the machine behaviour under various operational conditions. FEM is used in our investigation to foresee the changes of motor performance due to the changes in the internal parameters when the motor is working under fault conditions. Numerical simulations generate useful data, which are used to test the diagnostic techniques. Acosta [3] presented an online monitoring system that uses the combination of motor current signature analysis (MCSA) and Parks vector approach. The authors based on experimental observation and on knowledge of the machine constructed a knowledge-based system. The motor condition identication task requires the interpretation of data and makes a decision from this data. From the point of view that sees induction motor condition as a fuzzy concept, there has been some fuzzy logic approaches for diagnosis. The lack of proper processing of fuzzy input data and the construction of the membership functions are presented as the major difculties [7]. This problem is tackled in this work by using FEM in order to generate reliable virtual data, which allows the construction of the membership functions in all fault and load conditions. Fuzzy logic approach is used to make decisions about the motor condition. Fuzzy logic can describe the characteristics of an industrial process with linguistic terms. Fuzzy logic was chosen because the motor condition constitutes a fuzzy set. In practice, the users are concerned about condition of the motor in terms of a linguistic variable that can be expressed as good, damaged or seriously damaged. The task of the diagnostic system presented in this work is to detect an upcoming machine fault as early as possible, in order to save expensive manufacturing processes or to replace faulty parts. The proposed monitoring system can monitor eccentricity, rotor and stator related faults by analysing the motor current. This work focuses on the application MCSA to diagnose faults in three phase induction motor drives, establishing a general scheme that permits to spot failures in variable frequency. A traditional MCSA utilises results of spectral analysis of the supply current of an induction motor to detect an existing or incipient failure of the motor in the drive system. The spectral analysis is rather complicated and knowledge of the slip of the motor as well as motor data are needed [6,8,9]. The scheme developed in this work suggests new ideas with respect to the traditional scheme given by Nandi and Toliyat [9]. The proposed scheme permits the identication of faults in variable frequency and avoids the detailed analysis of the current spectrum, thus reducing the computation task. Further, in many applications it is desirable to detect the presence of the fault with minimal computation and cost. The motor current amplitudes are also used in order to spot failures in the stator winding. Motor current amplitudes contain potential fault information and constitute the most suitable indicator for diagnosing stator winding fault, in term of easy accessibility, reliability and sensitivity. 2. Methods 2.1. Motor current signature analysis A traditional MCSA is a non-invasive, on line monitoring technique to diagnose problems in induction motor. A large amount of research has been directed toward using the stator current spectrum to sense motor faults. The monitored spectral components can result from a number of sources, including those related to

ARTICLE IN PRESS
1218 guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr

normal operating condition. It is necessary to use some degree of expertise in order to distinguish a normal operating condition from a potential failure mode. This requirement is even more acute when analysing the current spectrum of an induction motor since a multitude of harmonics exists due to both the design and the load condition. In fact, MCSA utilises results of spectral analysis of the stator current, precisely the supply current of an induction motor to spot an existing or incipient failure of the motor or the drive system. Each induction motor fault has its own effect in the stator current. A symmetrical three-phase stator winding fed from a symmetrical supply will produce a resultant forward rotating magnetic eld at synchronous speed, and if an exact symmetry exists there will not be a backward rotating eld. Any asymmetry of the supply or stator winding impedance will cause a resultant backward rotating eld from the stator winding. Ref. [6] made a reviewed of the frequencies components given by motor faults. 2.1.1. Eccentricity related faults The presence of static and dynamic airgap eccentricity can be detected using MCSA. One of the equations describing the frequencies components is given by [10]   1 s f ecc f s kR nd u , (1) p where nd 0 in case of static eccentricity, and nd 1,2,3,y in case of dynamic eccentricity, fs is the fundamental frequency, R is the number of rotor slots, s is the slip, p is the number of pole pairs and u is the order of time harmonic that is presented in the power supply. As can be seen from Eq. (1), important motor information is needed to spot the related eccentricity components. However, in practice both types of eccentricity co-exist together and when this happens low frequencies related with the rotation frequency do appear [11] given by f e f s f r, (2) where fr is the rotational speed. Our proposed system is based on detecting these low frequency components. 2.1.2. Broken rotor bars During broken bars events the sideband currents around the fundamental are well known [6] f b 1 2ksf s . Nevertheless, there are other spectral components [6,12], given by    k fb 1 s s f s . p (3)

(4)

The components given by Eq. (4) are more separated from the main frequency and will be the ones to spot in our work. 2.1.3. Stator fault Stator winding faults are difcult to identify from the current spectrum. Ref. [13] showed that from the nature of the cage induction motor, all ux density waves from the stator side will be reected, and occur only at the following frequencies in the stator spectrum:   kR f rf f s 1 1 s . (5) p frf components reected from the stator side and R is the number of rotor slots. Ref. [13] arrived to the conclusion that under stator short circuit conditions there are MMF and ux density waves at all numbers of pole pairs and in both directions of rotation. Therefore, no new frequency components appear in the stator current spectrum as a result of fault in the stator winding, only a rise in the rotor slot harmonics can be expected. Ref. [13] showed that these components cannot be used as specic signs of the occurrence of inter-turn short circuit in the stator winding. An important fact also is that any stator short circuit produces a negative sequence component in the input currents. When a short circuit occurs, the

ARTICLE IN PRESS
guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr 1219

phase winding has less impedance, less turns and therefore less magnetomotive force. This gives a possibility of diagnosing stator short circuit by monitoring only the amplitude of the phase currents [14]. 2.2. Finite element method This software package was designed for the transient magnetic eld analysis of electrical machines coupled with the circuit equations of the machine windings. This method allows the simulation of an electrical machine fed from measured voltages used in experiments. The simulation of the induction machine is based on the time stepping, nite element analysis. The magnetic eld in the cross-section area of the test machine is computed by using an in-house 2-D FE program. The software uses the time-stepping method, which takes into account the motion of the rotor and the induced voltage due to this motion. The motion of the rotor is achieved by changing the mesh in the airgap and the time dependence is modelled by the CrankNicholson method. Some of the 3-D effects like ux fringing and end windings are modelled with analytical and electric circuit approaches. The magnetic eld, the current and the potential differences of the winding are obtained in the solution of the coupled eld and circuit equations. A full description of the software and its accuracy is given in [15]. A detailed description about fault implementation can be found in [16]. A diagram of the FEM modelling is given in Fig. 1. The FEM program permits the generation of data with the motor working in different condition of load as well as changing the severity of the fault. The simulations are carried out with a xed time step of 25 ms and a total of 40 000 steps, assuming a constant speed in the steady state. As examples, Fig. 2 shows the generated data for a healthy motor at full load (3% slip) and fed at 100 Hz. Fig. 3 shows the generated data for a motor working with 33% of dynamic eccentricity at the same load condition supply with a sinusoidal source at 100 Hz. From this last gure can clearly be seen the harmonic components due to eccentricity given by formula (2) and the main component at 100 Hz. Our motor has two parallel paths in the stator winding. The analysed current corresponds to one of these branch currents. The aim of FEM here is to foresee the changes of motor performance due to the changes in the internal parameters when the motor is working under fault conditions. FEM permits the evaluation of the inuence of different motor faults in an inexpensive and accurate manner. 2.2.1. Eccentricity harmonics with loading and eccentricity degree Fig. 4 shows the generated data for a motor working with a mixed eccentricity of 37% dynamic and 10% static at half load condition (1.9% slip). From Fig. 4, can be clearly seen the harmonic components due to eccentricity at fs7fr and the main component at 100 Hz. The amplitude of the sideband currents increases proportionally to the level of dynamic eccentricity. Fig. 5 shows the variation of the sideband currents with the degree of dynamic eccentricity when the motor is working at full load (3% slip).

Inputs Supply voltage Shaft torque Dimensions Material Data Winding Data

FEM model

Outputs Currents

The electromagnetic field couples the inputs and outputs

Rotational speed Fluxes in coils

Fig. 1. Electrical machine modelling by FEM.

A +

A = t

u l

Forces Losses Other..

ARTICLE IN PRESS
1220 guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr

Fig. 2. Motor current (a) and spectrum (b). Data generated by FEM. Healthy motor. Rated load (3 % slip).

Fig. 3. Motor current and spectrum. Data generated by FEM. Dynamic eccentricity.

It is important for condition monitoring strategies to make clear the dependence between the amplitude of the sideband currents and the machine loading. The load dependence is studied for the case of a pure dynamic eccentricity, see Fig. 6. A non-linear relationship between harmonic amplitudes and loading is obtained. The

ARTICLE IN PRESS
guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr 1221

lowest values for the sideband current is obtained at full load, while the highest is obtained at no-load condition. During the loading, the currents, which create the MMF increase theirs amplitudes, however, the amplitude of the induced current harmonics do not change linearly with loading as can be seen in Fig. 6. This is explained by the fact that when the machine is slightly loaded, the generated asymmetric ux due to the eccentricity can easily ow through the airgap producing electromotive force. When the load increases, the

Fig. 4. Motor current (a) and spectrum (b). Data generated by FEM. Mixed eccentricity, 10% static and 37% dynamic. Half load (1.9% slip).

Upper Harmonic (fs+fr) 2 1.8 1.6 1.4 Current [A] 1.2 1 0.8 0.6 0.4 0.2 0 0 0.1 0.2 0.3

Lower Harmonic (fs-fr)

0.4

0.5

0.6

0.7

Eccentricity degree
Fig. 5. Sideband currents as a function of the eccentricity, full load (3% slip).

ARTICLE IN PRESS
1222 guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr

Upper harmonic (fs+fr) 4.5 4 3.5 Current [A] 3 2.5 2 1.5 1 0.5 0 0 0.005 0.01 0.015

Lower harmonic (fs-fr)

0.02

0.025

0.03

0.035

Slip [s]
Fig. 6. Load dependence of the sidebands currents. Motor working with 33% pure dynamic eccentricity.

asymmetric ux induces current in the motor cage, which opposes the asymmetric ux, damping the magnitude of the electromotive forces. Furthermore, the asymmetric ux induces circulating currents in the parallel branches (the motor studied has two parallel branches) of the stator winding. These currents tend to equalise the ux distribution reducing the radial ux. From this analysis is deduced that a no-load test is the most informative for the identication of dynamic eccentricity. In the no-load condition, there are no rotor currents to damp the asymmetric ux, thus the induced harmonics in the motor current have the highest amplitudes, see Fig. 6, it shows that the worst case scenario to be detected is at full load when the harmonics have the lowest amplitude. The nonlinear relationship implies that eccentricity detection is more reliable at no-load condition. 2.2.2. Behaviour of the broken bars harmonics with loading and number of broken bars Fig. 7 shows the current spectrum for a motor working at full load with three broken rotor bars and fed from a sinusoidal supply of 100 Hz. The main harmonics (from 1 to 6) given by formula (4) can be clearly seen. The sidebands around the fundamental (100 Hz) correspond to the third and fourth harmonics. The amplitudes of the harmonics due to broken bars are proportional to the number of broken bars as it is shown in Fig. 8. This gure shows the cases of healthy, one broken bar, three broken bars and ve broken bars at full load condition (3% slip). The amplitude of harmonics due to broken bars is proportional to loading as can be seen in Fig. 9. This is the case of simulated data for a motor working with three broken bars. The fourth harmonic hardly varies with the loading; this harmonic is called the upper sideband and is due to consequent speed oscillations [6] during broken bars events. 2.3. The proposed motor protection system The proposed detection system in this work is a consequence of the analytical results. The scheme of the system can be seen in Fig. 10. The system consists of two main blocks. The left-hand block to spot electromechanical faults from the rotor (broken rotor bars and eccentricity) is based on monitoring the content of the spectrum of the current. The right-hand block to spot faults from the stator is based on monitoring the amplitudes of the motor currents. The left-hand block has a new idea with respect to the traditional scheme given by Nandi and Toliyat [9], in order to make the system able to work in variable frequency and avoid the detailed spectral analysis carry out

ARTICLE IN PRESS
guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr 1223

Fig. 7. Spectrum of the motor current for a motor with three broken rotor bars at full load (3% slip).

Fig. 8. Current spectrum in different rotor conditions. Motor working at full load (3% slip).

by a traditional MCSA. After the A/D block, a predictive lter is used to remove the fundamental component. This lter is able to work in variable frequency keeping the ltering properties [17]. The role of this lter is very important. This lter does not produce delay between the incoming signal and the ltered signal [17]. This quality is useful for detecting harmonic components by subtracting the input signal

ARTICLE IN PRESS
1224 guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr

First harmonic 2.00 1.80 1.60 1.40 Amperes [A] 1.20 1.00 0.80 0.60 0.40 0.20 0.00 0.00 Second harmonic Third harmonic Fourth harmonic Fifth harmonic Sixth harmonic

0.01

0.02 Slip [s]

0.03

0.04

Fig. 9. Harmonics during broken bars with loading, three broken bars.

Three-phase source

Ia

Ib M Ic

Low Pass Filter

A/D

A/D

STATOR FAILURE

Subtraction

Predictive filter

Calculate rms of current

Preprocessor FFT and averaging

S/rms

Fuzzy inference system

BROKEN ROTOR BARS AND ECCENTRICITY FAILURES

Rotor & Stator

Fig. 10. Block diagram of the proposed integrated motor protection system.

ARTICLE IN PRESS
guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr 1225

from the output of the ltered signal without delay, yielding a measurement of the harmonic components rather than the fundamental sinusoidal. The multistage lter is shown in Fig. 11. The rst block is a three point median lter, which removes the disturbing impulses. The median lter is a non-linear lter that operates by sorting the samples inside the moving lter window by magnitude, choosing the middle value, and removing completely isolated impulses. The median lter causes one sample delay and does not restore a sinusoidal signal after removing the impulses. The adaptive predictor compensates these drawbacks. The predictor predicts two steps ahead; one to compensate the delay in median lter and the other to allow interpolation in the last stage. This lter behaves as a band pass lter with a narrow pass-band centred at the main frequency. Since adaptive predictor is used, the prediction step remains accurate, even if the line frequency changes. Thus, the lter is able to work in variable frequency, only by changing some control parameters correspondingly [17]. The frequency range is limited only by the Nyquist frequency of the discrete time system. Fig. 12 shows an example of the lter performance in steady state when the motor is working with 20% dynamic eccentricity, half load (1.9% slip) and fed at 100 Hz. The lter structure is shown in Fig. 13, there are two nite impulse response (FIR) lters. The FIR inside the loop produces the update coefcients for the FIR that produces the ltered output. The coefcients of the adaptive lter are updated using the least mean square (LMS) algorithm, which is chosen due to its robustness and computational simplicity [17]. The coefcient update formula is H n 1 H n 2 u en X n 2, where en xn H T n X n 2 H n h0; . . . ; hn 1T is the error; (6)

is the coefficient vector; is the data vector:

X n 2 xn 2; . . . ; xn N 1T

The constant u is xed small enough to guarantee the stability of the LMS algorithm. The adaptive lter conguration was implemented in MATLAB/SIMULINK, with 22 coefcients. The lter shows good convergence properties and a signicant reduction of the distortion of the primary sinusoidal component. Another advantage from the use of the predictive lter is that digital processing offer higher accuracy than analog processing. The lter was tested at 50, 60 and 100 Hz with current data from FEM motor simulation program. When the motor is working under fault conditions relevant changes occur in the motor current, e.g. in Fig. 14 can be seen the harmonics components due to broken rotor bars. After the subtraction cancels the main component, just the harmonics components can be seen (see Figs. 15 and 16). The subtraction result of the ltered signal from the input signal constitutes the inputs for the FFT and pre-processor block. By removing the main component, is obtained in the output the harmonic components, which constitute the useful information for identifying the motor condition. Once we do not have the main component, the sum of the harmonics with higher amplitude than a pre-set level is calculated. Several references indicate [79] the drawback of the noisy environment and also suggest the idea of reducing the fundamental component in order to improve the results in the fault identication system. But by removing the main component we obtain in the output the harmonic components, which constitute the useful information for detecting the motor condition. Once we do not have the main component we just have to calculate the sum of the harmonics with higher amplitude than a preset level with a simple algorithm for detecting the faulty condition, as it is illustrated in Fig. 15 (healthy case) and Fig. 16 (faulty case). The preset level is determined by the harmonic contents during healthy situation. Consequently, avoiding the detailed spectral analysis of the line current for detecting the

Median Filter

Adaptive Predictor

Enhancen and Interpolation

Fig. 11. Block diagram of the adaptive multistage lter.

ARTICLE IN PRESS
1226 guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr

Fig. 12. Input current when the motor is working with 20% dynamic eccentricity and ltered current.

x (n) z-2 LMS FIR

e (n)

Y (n) FIR

Fig. 13. Filter block diagram.

presence of the fault. This means, minimal amount of data memory requirements and minimal computation and cost. Further, very precise information it is not necessary neither about the motor slip nor about motor data, which determine the exact position of the spectral components, related to a motor fault. Further, the multistage adaptive lter is a truly adaptive algorithm. Thus, the lter automatically follows any variation in the supply frequency, while maintaining their ltering property [17]. The system is also suitable to work in variable frequency, which means that the same lter can work at 50, 60 Hz or any frequency, only changing some control parameters, correspondingly. It was tested at 50, 60 and 100 Hz with motor data from FEM motor simulation program and real measurements. The faulty condition is determined by the total sum of harmonics with amplitude higher than the threshold level, see Figs. 15 and 16. A block diagram of the pre-processor and the averaging block is shown in Fig. 17. In this ow chart, integer N is the number of harmonic components determined by the wide of the useful spectrum after the low pass ltering. The chosen value in this work was 500, which means from 1 to 500 Hz, with a frequency resolution of 1 Hz. I integer controls the harmonic order. The sampling frequency used was 40 kHz and the length of data set 40 000 samples for both the simulated and measured data.

ARTICLE IN PRESS
guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr 1227

Fig. 14. Stator current and its spectrum. Motor working with two broken bars at half load.

Fig. 15. Current after subtracting the fundamental component and its harmonic contents, healthy condition.

ARTICLE IN PRESS
1228 guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr

Fig. 16. Current after subtracting the fundamental component and its harmonic contents, faulty condition.

The value of S is load dependent. Then, it is divided by the rms values of the total current. Later on, this value from the division that we call R is one of the inputs for the fuzzy logic engine. If the value of R is large means that the variable rotor condition RC is damage. If the value of R is normal, we have a healthy rotor. Fig. 18 shows the membership function for the variable R. In a membership function, the xaxis indicates the normalised value of the universe of discourse of the linguistic variable (in this case R), while the y-axis indicates the probability of belonging to one of the classes (in this case large or normal). The fuzzy system allows the transformation of heuristics and linguistics terms into numerical values via fuzzy rules and membership functions and it is able to approximate the complex relationship related with the identication task. A no-load test must decide if the harmonics are due to broken bar or eccentricity events because the harmonics due to broken bars are approximately zero at no load (see Fig. 9), while by the contrary, the harmonics due to eccentricity have maximum amplitude at no load (see Fig. 6). For the detection of stator winding faults (following the right-hand block in Fig. 10), the amplitudes of the currents (rms) are categorised using four linguistics values. Fig. 19 shows the membership functions for these categories. These categories are Very Small (VS), Small(S), Medium (M) and Large (L). The membership functions for the input and output variables are constructed by the analysis of data generated by FEM. Thus, the FEM program is run from no-load to full-load in the healthy situation and the rms values of the phase currents are calculated, obtaining values from 25 to 65 A. At every operational point, the loading is constant. These values are normalised between [0,1], dening the membership function which corresponds with the linguistic term M, for a healthy motor. Thus, the load information is explicitly in the membership functions. A similar process is repeated for the faulty conditions, dening the VS, S, and L. The VS condition is obtained when the motor has a fully open phase, L corresponds to a short-circuit in one phase and S corresponds to a unloaded motor. All these situations are simulated in FEM, allowing the denition of the trapezoidal membership functions. The system was tested with triangular, trapezoidal and Gaussian membership functions. It was found that the combination of triangular and trapezoidal membership function

ARTICLE IN PRESS
guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr 1229

FFT of difference current

Set Threshold (T), S=0, I=0

If Harmonic (I)>T S=S + Amplitude (I)

I=I + 1

I<N

Save the value of S

Fig. 17. Block diagram of pre-processor and averaging block.

Normal 1

Large

Universe of discourse

Fig. 18. Membership function of the variable R.

VS 1

Universe of discourse

Fig. 19. Fuzzy membership functions for current the normalised current Ia.

ARTICLE IN PRESS
1230 guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr

is the most appropriated for fault diagnosis of induction motors. The trapezoidal function has well dened corners as the motor has well dened rated parameters. The linguistic variable Stator winding condition (SC), interpreting the stator condition, may have values Good (G), Damaged (D) or Seriously damaged (SD). G refers to a stator with no faults. D might be a stator with voltage unbalance, and SD a stator with an open phase or coil short circuit. Fig. 20 shows the membership functions for these categories, which are also dened by the analysis of data generated by FEM. The design of rules is based on the expert understanding. The simplied number of rules between the three-rms inputs and the classication of every current is 14. It is also considered a rule that takes into account the rms variance of the phase currents. The last is done to improve the system sensitivity and reliability. The membership function for the amplitude variance of current is shown in Fig. 21. The variance is calculated as follow: j2 jI b I j2 jI c I j2 jI a I , (7) 3 is the mean value of Ia,Ib,Ic. where Ia,Ib,Ic are the rms of the input currents I There are two membership output functions, one to evaluate the stator condition and other to evaluate the rotor condition. There are totally 17 rules in the inference engine, 15 rules for the stator condition and two rules for the rotor condition. The set of rules is as follows: v Rule Rule Rule Rule Rule 1: 2: 3: 4: 5: If If If If If Ia is VS then SC is SD. Ib is VS then SC is SD. Ic is VS then SC is SD. Ia is L then SC is SD. Ib is L then SC is SD.

Good 1

Damaged

Seriously Damaged

SC

Universe of discourse

Fig. 20. Fuzzy membership function for the stator condition.

Normal 1

Large

Universe of discourse

Fig. 21. Fuzzy membership functions of the variable variance.

ARTICLE IN PRESS
guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr 1231

Rule Rule Rule Rule Rule Rule Rule Rule Rule Rule Rule Rule

6: If Ic is L then SC is SD. 7: If Ia is S and Ib is S and Ic is M then SC is D. 8: If Ia is S and Ib is M and Ic is M then SC is D. 9: If Ia is M and Ib is S and Ic is M then SC is D. 10: If Ia is M and Ib is M and Ic is M then SC is G. 11: If Ia is S and Ib is S and Ic is S then SC is G. 12: If Ia is S and Ib is M and Ic is S then SC is D. 13: If Ia is M and Ib is S and Ic is S then SC is D. 14: If Ia is M and Ib is M and Ic is S then SC is D. 15: If v is L then SC is D. 16: If R is large then RC is D. 17: If R is normal then RC is G.

2.4. Simulation results A SIMULINK/MATLAB model was implemented to test the proposed system. The model was tested with data from FEM motor simulation program. The rated parameters of the motors are given in Appendix A. The sampling frequency in the FEM simulation program was 40 kHz and number of samples was 40 000. The data was introduced from MATLAB workspace, in such a way that the model is working online with the data. The FEM program was run under different stator winding conditions, healthy, open phase, coil short-circuited and inter-turn short circuit. The data were collected under three load conditions, no load, half load and full load. The simulation model was able to identify the fault with excellent accuracy. During every data set, the fuzzy lter executes 25 validations of the stator condition. In order to prove the performance of the SIMULINK model under noise condition, a source of noise to each phase was added. Table 1 shows the percentage of correct detection of stator condition under noise. The FEM program was also run under different rotor conditions: one, two, three and ve broken rotor bars, 10%, 22%, 33%, 37% and 50% dynamic eccentricity and a mixed eccentricity of 37% dynamic and 10% static. This was done at different load conditions. The model was able to identify the rotor condition in all the data sets corresponding with broken bars, dynamic and mixed eccentricity with the simulated data. 2.5. Measurement results A measuring setup was arranged to get data from a working motor. The motor used in the measurements was the same as used in the FEM motor simulation program. The data was recorded through a transient recorder. The sampling frequency used was 40 kHz. The tests were carried out with the motor in healthy condition and with a real inter-turn short circuit, mixed eccentricity and broken rotor bars. The real short circuit was done between two adjacent turns. The insulation of winding wires were scratched and two wires were soldered to them. These were long enough to be closed from outside the motor through a switch. The short circuit was made active during a short time, just enough to take the 1 s of data. Different rotor faults were prepared, from one to ve broken rotor bars and mixed eccentricity. The mixed eccentricity was obtained

Table 1 Percentage of correct detection of stator condition under noise Motor condition Healthy motor Open phase Inter-turn short Coil short circuited Data sets 3 9 3 3 Accuracy (%) 96 100 92 100

ARTICLE IN PRESS
1232 guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr

by tting non-concentric support parts between the shaft and bearing, see Fig. 22. It was achieved (measured) 37% dynamic eccentricity and about 10% static eccentricity. Data were collected at 50, 60 and 100 Hz and in three load conditions for the cases of broken rotor bars and mixed eccentricity. The model was fed with this data in such a way is working online. The model was able to identify two, three and four broken rotor bars with total accuracy. Fig. 23 shows the motor current when the motor was working with three broken bars, full load, inverter supply at 100 Hz measured data. Fig. 24 shows the measured data when the motor was working with the mixed eccentricity. Fig. 25 shows the three phase

Fig. 22. Elements of the articially created dynamic eccentricity.

Fig. 23. Measured current and its spectrum from the motor working with three broken bars. Inverter supply at 100 Hz.

ARTICLE IN PRESS
guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr 1233

currents in a healthy machine. Fig. 26 shows the three phase currents during the real inter-turn short circuit, from this gure can be seen the current unbalance due to the short circuit. The SIMULINK model was fed with motor data in healthy condition and with a real inter-turn short circuit. Data were collected in three load conditions, no load, half load, and full load, at different frequencies. Table 2 shows the detection rate results for the case of inverter supply. The model was also tested with motor data, when the motor was fed with sinusoidal supply. The results were also with high accuracy.

Fig. 24. Motor current when the motor was working with a mixed eccentricity. Measured data.

Fig. 25. Terminal phase currents in a healthy motor. Inverter fed at 100 Hz.

ARTICLE IN PRESS
1234 guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr

Fig. 26. Terminal phase currents in a faulty motor during an inter-turn short circuit. Inverter fed at 100 Hz. Measured data.

Table 2 Percentage of correct detection. Measured data with Inverter supply Frequency (Hz) Load Detection accuracy (%) Healthy 25 No-load Half load Full load No-load Half load Full load No-load Half load Full load No-load Half load Full load 94.7 97.2 96.3 94.7 96 97.3 100 100 100 100 100 100 Faulty 100 94.4 100 100 100 100 100 100

50

75

100

3. Discussions The main objective of this work was to establish a layout capable of detecting the motor condition by monitoring the motor currents. The data analysis from a FEM motor simulation program showed the same feature in the motor current as it was predicted in the literature. Thus, new harmonics in the current spectrum appear when there are broken rotor bars and eccentricity faults. On the other hand, as was expected from the analytical results, no new prominent current harmonics are generated when the motor is working under stator winding fault. From the analytical and simulation results, a novel fuzzy logic and predictive lter layout was designed. In this layout, a predictive lter to cancel the main harmonic component was introduced. The lter was tested at different frequencies with simulation and measured data. It showed good performance and it was able to work with both sinusoidal and inverter supplies. The lter was able to enhance the spectrum and improve the

ARTICLE IN PRESS
guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr 1235

identication system. As a comparison, a classical MCSA requires a frequency resolution better than 0.1 Hz and in practice is unable to distinguish the side bands for so1.5 [9]. Our proposed system with a frequency resolution of 1 Hz obtained a good accuracy detection rate. This means minimal amount of data memory requirements and minimal computation and cost. Further, the very precise information about the motor slip is not needed. The detection rate was accurate for the cases of broken rotor bars and dynamic eccentricity with both, simulated and measured data. In this layout, fuzzy logic was used to analyse the data and make decisions after the cancellation of the main component. It was able to detect the motor condition with high accuracy. However, it is also possible to use other soft-computing or traditional techniques to carry out the identication task. The rotor condition can be identied even from simple rms values of the harmonic components, once, that has been cancelled the main component. The SIMULINK model was able to identify the stator winding condition with simulation and real data. It was also tested with noisy simulation data. The model was able to identify the stator condition with good accuracy even under noisy condition. The model was fed with measured data. It was tested at different frequencies with both inverter supply and sinusoidal supply. In both cases, the SIMULINK model was able to identify the stator condition. Two conditions were tested with measured data: healthy condition and inter-turn short circuit. The model showed that comparing the rms current of the phases reveals changes in the internal electrical balance of the machine. It was sensitive enough to reveal one shorted turn in the stator winding, where there were 11 turns per coil. It is important to sign that in the case of stator winding faults in induction machines the most important fault to be detected is a primary inter-turn short circuit (the fault in an early stage). Because by detecting it at the early stage would avoid the total destruction of the stator winding. Our proposed system showed good performance in detecting such a condition. 4. Conclusions This work showed the feasibility of spotting stator winding faults and broken rotor bars by monitoring the motor current with appropriate use of the existing techniques for signal processing and soft computing. The proposed system is able to detect different induction motor faults with high accuracy and it is remarkable that the detailed spectral analysis is avoided. This work is an example of fusion between the soft-computing technique (fuzzy logic) and hard-computing technique (FEM) in order to make a reliable identication system. A strait forward application of the system is in variable speed drives. Appendix A. Motor rated parameters Parameter Rated power Rated frequency Rated voltage Rated current Connection Number of pole pairs Number of stator slots Number of rotor bars 35 kW 100 Hz 400 V 64 A Star 2 48 40

References
[1] Motor Reliability Working Group, Report of large motor reliability survey of industrial commercial installations, part I, IEEE Transactions on Industry Applications IA-1 (1985) 853872.

ARTICLE IN PRESS
1236 guez et al. / Mechanical Systems and Signal Processing 22 (2008) 12161236 P.V.J. Rodr [2] R. Schoen, B. Lin, T.G. Habetler, J. Schalag, S. Farag, An un-supervised online system for induction motor fault detection using stator current monitoring, IEEE Transactions on Industry Applications 31 (6) (1995) 12801286. [3] G.G. Acosta, C.J. Veruchy, E.R. Celso, A current diagnosis system for diagnosing electrical failures in Induction Motors, Mechanical Systems and Signal Processing 20 (2006) 953965. [4] J.R. Stack, T.G. Hableter, R.G. Harley, Fault classication and fault signature production for rolling element bearings in electrical machines, IEEE Transactions on Industry Applications 40 (3) (2004). [5] P. Tavner, J. Penman, Condition Monitoring of Electrical Machines, Research Studies Press, Letchworth, Herefordshire, England, 1987, 302pp. [6] S. Nandi, H. Toliyat, Condition monitoring and fault diagnosis of electrical motorsa review, IEEE Transactions on Energy Conversion 20 (4) (2005). [7] F. Fillipetti, C. Tassoni, P. Vas, Recent developments of induction motor drives fault diagnosis using AI techniques, IEEE Transactions on Industrial Electronics 47 (5) (2000) 9941000. [8] W.T. Thomson, M. Fenger, Current signature analysis to detect induction motor faults, IEEE Transactions on Industry Applications July/August (2001) 2634. [9] M.E.H. Benbouzid, M. Viera, C. Theys, Induction motors fault detection and localization using stator current advanced signal processing techniques, IEEE Transactions on Power Electronics 14 (1) (1999) 1422. [10] J.R. Cameron, W.T. Thomson, C. Eng, A.B. Dow, Vibration and current monitoring for detecting airgap eccentricity in large induction motors, IEE Proceedings 133 (3) (1986). [11] D.G. Dorrell, W.T. Thomson, S. Roach, Analysis of airgap ux, current, and vibration signals as s function of a combination of static and dynamic eccentricity in 3-phase induction motors, IEEE Transactions on Industry Applications 33 (1997) 2434. [12] G.B. Kliman, R.A. Koelg, J. Stein, R.D. Endicott, M.W. Madden, Noninvasive detection of broken bars in operating induction motors, IEEE Transactions on Energy Conversion 3 (4) (1988) 873879. [13] G.M. Joksimovic, J. Penman, The detection of inter-turn short circuits in the stator winding of operating motors, IEEE Transactions on Industry Applications 47 (5) (2000) 10781084. [14] M.E.H. Benbouzid, A Simple Fuzzy Logic Approach for Induction Motor Stator Condition Monitoring, IEMDC 2001, IEEE International, January 2001, pp. 634639. [15] A. Arkkio, Analysis of induction motor based on numerical solution of the magnetic eld and circuits equations, Acta Polytechnica ScandinavicaElectrical Engineering Series 59 (1987) 97 Available at /http://lib.hut./Diss/list.html#1980S. [16] S. Po yho nen, M. Negrea, P. Jover, A. Arkkio, H. Hyo tyniemi, Numerical magnetic eld analysis and signal processing for fault diagnostic of electrical machines. COMPEL, The International Journal for Computation and Mathematics in Electrical Engineering 22 (4) (2003) 969981. [17] O. Vainio, S. Ovaska, Multistage adaptive lters for in-phase processing of line frequency signal, IEEE Transactions on Industrial Electronics 44 (2) (1997) 258264.