Induction Motor Broken Rotor Bars Diagnosis using Stator Current Time-Frequency Analysis

A.F. Aimer, A. Bendiabdellah, A.H. Boudinar, C. Mokhtar and S. Laribi L.D.E.E Laboratory, Department of Electrical Engineering, USTO-Oran, Algeria e-mail: ameurfethi@yahoo.fr
Abstract--Recent advances in the field of power electronics and control circuits, have contributed to the increasing use of induction machines in electrical systems. The use of induction machines is mainly due to their robustness, their power/weight ratio and their low cost of manufacture. Still, various faults may appear at such machines. The Power Spectral Density (PSD) based on the Fourier Transform (FT), is used as a method of analysis for many years for its simplicity and computing time relatively low. However, it is ineffective in faults detection in the case of a small slip (harmonics too near to the fundamental). In addition, the fact that this method is based on the calculation of the FT, implicitly implies that the spectral properties of the signal are stationary. With the development of variable speed applications, the spectral characteristics of the stator current become non-stationary and the spectra are much richer in harmonics. To resolve these problems, we used in this paper, a time-frequency representation called Short Time Fourier Transform, giving therefore, an additional information on changes of the frequencies over time in the stator current signal. We conduct several simulations and experimental tests in the aim of validating our approach. Index TermsInduction motor, fault diagnosis, timefrequency analysis, broken rotor bars.

on almost all the faults that can appear on the induction motor [4], [5], [6]. Signal processing, spectral analysis in particular, is used for many years in detecting and locating faults in induction machines [7]. This type of failures is well suited to this approach since many phenomena are reflected by the appearance of sideband frequencies directly related to the motor speed. However, in most cases, induction motor operates directly from the mains and running in the steady state, with known loads. The development of technology and the advances in variable speed drives have given another dimension to various problems including non stationary condition of measurable signals, disruptions and distortions caused by power converters, etc. In addition, the requirements in terms of industrial maintenance guide the research to a more effective fault diagnosis. The power spectral density or PSD of the stator current, calculated based on the Fourier transform (FT) is the most currently tool used by researchers and industrials [8]. This is justified by its simplicity and the low cost of current sensors and the harmonic content of stator current. However, this technique has several disadvantages which are due to the problem of frequency resolution. Indeed, the calculation of the FT introduces a smoothing effect and a side effect. These effects are reflected by the appearance of sideband lobes in the stator current spectrum and reduce the analysis efficiency [9]. To analyze a signal, it has an interest in having a main lobe as narrow as possible and sideband lobes of very low amplitude when it is not possible to have both properties simultaneously. However, it is possible to reduce the amplitudes of sideband lobes by replacing the rectangular window by a smoother and seamless window. Following this resolution problem, the PSD finds difficulties in detecting faults with a small motor slip (harmonics close to the fundamental). Also, the fact that the PSD is based on the FT, implicitly implies that the spectral properties of the signal are stationary. In addition, the modulus of the FT of a signal provides only a time average of the spectral content of the signal without giving details on possible

1. INTRODUCTION With a simple technology, the induction motor is widely used in most electric drives, especially for constant speed applications. Advances in power electronics associated with modern controls techniques have led to consider variable speed applications efficiently, which was previously reserved exclusively for DC motor and synchronous motor. Thus, given this growing interest, a general reflection was engaged for fault detection and condition monitoring of induction machine. Fault detection in electrical machinery has been the subject of research and industrial achievements in recent years. There are several types of techniques; the vibration analysis is the most used method [1], [2], [3]. This method is mainly used for the detection of mechanical faults. Another technique is also used increasingly in recent years based on analysis of stator current also called Motor Current Signature Analysis or MCSA, its particularity is that the stator current contains information

changes in frequency over time. Therefore, information regarding the location of the frequencies over time can not be inferred from the Fourier transform. To overcome these constraints, we use a timefrequency representation. Indeed, the Gabor works in the 40s have conducted to the foundations of a new type of analysis called Short Time Fourier Transform or STFT. He was the first to imagine a local Fourier transform based on a windowing signal analysis to observe changes in frequency over time. This transformation requires the division of the signal in consecutive short segments and then calculates the Fourier transform of each segment. The idea is to introduce the local frequency parameter so that the Fourier transform is applied to the signal through a sliding window on which the signal is approximately stationary. This method represents the results into three dimensions; the description of the signal is carried out in the time-frequency composed of spectral characteristics as a function of time [10]. The main disadvantage of this transformation is its limited resolution. Indeed, it is impossible to have a perfect localization in both time and frequency. In other words, a good resolution in time (small time window) necessarily implies poor resolution in frequency (large frequency window). Conversely, good resolution in frequency implies poor resolution in time. The window width in each domain remains fixed as it is translated. This results in a fixed resolution across the entire timefrequency plane [11], [12]. In this context, this paper focuses on the application of the Short Time Fourier Transform in detecting and locating induction motor faults. To this end, the STFT is evaluated and used for the analysis of stator current in the presence of broken rotor bars faults. The results of simulations and experimental tests will support and validate our approach. 2. STATOR CURRENT SIGNATURE ANALYSIS The stator current spectral analysis is the most commonly method used in recent years, because the resulting spectrum contains a source of information on most faults that may appear on an induction machine. It is important to note that, as in vibration analysis case, the more the fault progresses, its characteristic spectral components continue to increase with time [6]. Induction motor broken rotor bars is considered among the most common fault studied because of its simplicity of implementation. This fault induced changes in the stator current spectral component and thus the appearance of sideband frequencies in the current spectrum produced by the magnetic field anomaly of the broken rotor bars [7]. Indeed, broken rotor bars give rise to a sequence of sidebands given by:

Where: fs is the supply frequency and fc the sideband frequencies associated with the broken rotor bars, s is the motor slip and k = 1, 2, 3 3. TIME-FREQUENCY ANALYSIS The Fourier transform is expressed by the following equation [11]:

FTx ( f ) =

x(t )e

j 2ft

dt

(02)

We define the power spectral density or PSD as the square modulus of the Fourier transform. The PSD is thus independent of the signal phase. Therefore, any information on the frequency changes over time is lost in the PSD. It is really not necessary to present the problem of the Fourier analysis as the overall weaknesses of this type of representation in the non stationary cases were identified [8], [10], [13]. Gabor's work in the 40s have laid the foundations of the Short Time Fourier Transform. It was the first to imagine a local Fourier transform based on a windowing analyzed signal to observe changes in frequency over time. The time-frequency distribution, known as Short Time Fourier Transform or STFT is defined by [11]:

STFTx (t , f ) =

STFTx (t , f ) = X ( + f )H ( )e j 2t d

x( )h( t ) e

j 2f

(03) (04)

The second expression of the STFTx is inferred from the classical properties of the FT: conservation of the scalar product, shift properties and transformation of a normal product into a convolution one. obtained by weighting x( ) by the window witch is a short time analysis window localized around t and that shifts by varying the parameter t.

The STFT is constituted by the FT of x( )h

( t ) h ( t )

Join to h( ) , the family of functions depending on two parameters t and f, defined by [14]:

ht , f ( ) = h( t )e j 2f ,
The numbers

(t , f ) 2

(05)

f c = (1 2ks ) f s

(01)

projections of x( ) on the functions system ht,f. If h is the rectangle window of T support, the STFT consists in taking the FT of a sequence of signals equal to x on the support and zero elsewhere.

STFTx (t , f ) are commonly called

We begin by the discrete-time signal [xn = x(nT)], T>0. Let hn = h(nT) and we mean by N the number of samples in the analysis window. Finally, we introduce a discretization of the frequency variable f. The STFT is then defined by the entire numbers Xk,n calculated as follows [15]:

This issue requires the search for the right timefrequency compromise suitable to the case considered in determining the correct window width. Gaussian window has the best time-frequency localization. Indeed, it verifies the following equality:

X k ,n = xl+ k hle
l =0

N 1

j 2lT

n N

t f =
, k Z, n = 1,2, (06)

1 4

(11)

As for the FT, the zero-padding technique allows the improvement of the frequency resolution. The principle of this method is to complete by M zeros a set of N samples so that M + N is a power of 2 and thereafter can perform calculations using the Fast FT algorithm using the N + M points. When M = N, the method use the Discrete FT algorithms that are being made to calculate 2N points from the spectrum, from only N points of the signal. 3. HEISENBERG-GABOR UNCERTAINTY PRINCIPLE The uncertainty principle, also called time-frequency inequality, is based on the uncertainty relationships established by Werner Heisenberg in quantum mechanics. The analogy with the work of Heisenberg for the Fourier transform was made by Dennis Gabor in 1946. Let us consider the finite energy signal x(t), centred in time and frequency around zero. Gabor defines the duration t and the spectral band f as follows [14]:

Finally, the choice of the window is important because it represents another compromise (comparable to the time-frequency compromise) between the main lobe width and the amplitude of the sideband in the frequency domain. 4. SIMULATION RESULTS AND DISCUSSION To simulate the presence of a rotor fault in a squirrel cage induction machine with a motor slip s = 5%, we construct the signal to be analyzed as follows:

x(t ) = 5 sin (2f s t ) + 0,066 sin (2f c1t )

+ 0,068 sin (2f c 2 t ) + b(t )

(12)

With: b(t) a signal that represents the white noise introduced by using the concept of the SNR. The SNR being the Signal to Noise Ratio given by the following equation:

SNR = 10 log10

Ps Pb

(13)

t =

1 + 2 2 t x(t ) dt Ex 1 + 2 2 f = f X ( f ) df Ex

(07) (08)

Where Ps and Pb are, respectively, the signal and noise powers. The supply frequency used is equal to fs = 50Hz. The sideband frequencies of broken rotor bars are calculated by equation (1). Therefore, the sideband frequencies of one broken bar are given by: fc2 = 55Hz. fc1 = 45Hz et Figure 01 shows the signal of stator current in time domain simulated with and without broken bars. We can observe the change in the form of the stator current after the introduction of the broken rotor bar fault. In Fig. 02, the STFT of the stator current is used for the detection of harmonics corresponding to one broken bar. This simulation is done with a motor slip of 5% and SNR = 50 dB corresponding to a moderately noisy signal. The sideband frequencies of the faults are highlighted in the STFT using the three dimensional grid technique. Fig. 03 represents the of stator current signal analysis simulation achieved by the STFT algorithm. The simulation parameters used correspond to one broken rotor bar with a motor slip of 5% and a SNR of 10 dB. The sideband frequencies corresponding to the simulated fault are easily located for this value of the SNR which corresponds to a highly noisy signal. This verifies the robustness of the STFT against noise.

Where Ex is the energy of the signal given by the Parceval relationship:

E x = x(t ) dt =
2

X ( f ) df
2

(09)

Therefore, the time-frequency inequality is defined by [12]:

t f

1 4

(10)

It expresses the fact that the duration-band product of a signal is lower bounded for a t duration and a f spectral band. In other words, the greater the accuracy in frequency localization, the lower the accuracy in time localization and vice versa. The STFT is subject to the uncertainty principle due to the use of Fourier transform.

signal du courant statorique 10 10

signal du courant statorique

5 Amplitude (A) Amplitude (A) 4 4.2 4.4 4.6 Temps (s) 4.8 5

-5

-5

-10

-10

4.2

Time (s)

4.4 4.6 Temps (s)

4.8

Healthy motor

Time (s) 1 broken bar

Fig. 01: Stator current signal in time-domain for healthy and faulty cases.

Spectre 3D 100 100

STFT

STFT Spectre 3D

80 P uis s anc e (dB )

80

P uis s anc e (dB )

Power (dB)

Power (dB)

60

48,11dB 45Hz 47,18dB 55Hz

60

48,11dB 45Hz

47,18dB 55Hz

40

40

20

20

0 0 2 4

0 0 2 4 0 50 100 150 200

Time (s) Temps (s)

Frequency (Hz) Frquence (Hz)

Time (s) Temps (s)

30

35

40

45

50

55

60

65

70

Frquence (Hz) Frequency (Hz)

(a)

(b)

Fig. 02: Stator current analysis by STFT for one broken bar, a motor slip of 5 % and an SNR=50 dB: (a): without zoom (b): with zoom.
STFT Spectre 3D
signal du courant statorique 10
80 100

48,11dB 45Hz

P uis s anc e (dB )

5 Amplitude (A)

47,18dB 55Hz

Power (dB)

60

40

20

-5
0 0

-10

4.2

4.4 4.6 Temps (s) Time (s)

4.8

2 4

Time (s) Temps (s)

30

35

40

45

50

55

60

65

70

Frequency (Hz) Frquence (Hz)

Fig. 03: Time-domain representation and STFT analysis of stator current for one broken bar, a motor slip of 5 % and an SNR=10 dB.

Signal du courant statorique

5. EXPERIMENTAL RESULTS AND DISCUSSION 5.1 Experimental Setup description In order to validate the results obtained by the simulation tests, we apply the time-frequency analysis in the processing of experimental signals obtained from a test conducted at the Laboratory of Automatics and Industrial Informatics from the University of Poitiers (France). Indeed, this test creates an artificial fault of one broken bar and two broken bars. The block diagram of the experimental setup is given in Fig. 04. AC Grid 380 V Electrical sensors
Butterworth Filters
3-phase Induction Motor
A m (dB) Powerplitude

2 A m p litude (A )

-2

-4

2.1

2.2 2.3 Temps (s) (s) Time

2.4

2.5

Fig. 05: stator current signal for the healthy motor test.
STFT

Spectre temps-frqeunce 80

50 Hz 74.08 dB 150 Hz 250 Hz 24.64 dB 45.03 dB 350 Hz 33.20 dB 621 Hz 29.40 dB

60

isa , isb , isc vsa , vsb , vsc Data Acquisition dSPACE 1104

40

20

Shaft encoder 1024 pts/r

-20 0 2 4 6 Temps (s) Time (s) 0 100 200 500 400 300 Frquence (Hz) 600 700 800

Frequency (Hz)

Fig. 04: Experimental setup bloc diagram.

Fig. 06: stator current analysis by STFT for the healthy motor test.

The motor used in the experimental investigation is a three phases, 1.1 kW, 50Hz, 2 pole pairs, squirrel cage induction machine. All acquisitions were done in steady state with a sampling frequency equal to 1428Hz. Each measured signal is acquired on 13347 samples. 5.2 Healthy motor test The healthy motor test is established in order to provide a reference which will be taken in consideration in the next tests to distinguish the normal operation of induction motor from the faulty ones. Fig. 05 represents the time-domain signal of stator current for the healthy motor case with a full load. The stator current time-frequency analysis by STFT is given by the following figure 06.

In Fig. 06, we can observe through the time-frequency analysis of the stator current different harmonics on the spectrum. However, the most imposing harmonics are the multiples of the fundamental (150 Hz, 250 Hz, 350 Hz) and the principal rotor slot harmonic observed at 621 Hz. Figure 07 represents a zoom of the time-frequency spectrum produced around the fundamental in the frequency band [30Hz-70Hz]. This band concerns us insofar as the broken bars fault, studied in this paper, is located in this frequency band. 5.3 One broken bar test The motor slip calculated in this test is equal to 4.06%. This value is obtained through the mechanical speed which is measured at 1432 rpm.

100

STFT Spectre temps-frqeunce

5.4 Two broken bars test The sideband frequencies are calculated in the same way regardless of the number of broken rotor bars. Indeed, stator current analysis in case of broken bars gives only an indication of the presence of the fault. Nevertheless, we can differentiate between the case of one broken bar and the case of two broken bars by monitoring the evolution of the power of the faults component. Indeed, several experimental studies have shown that increasing the number of rotor bars increases the power of the faults component [7]. Figure 09 shows the stator current signal in the timedomain for two broken rotor bars.

80

Power (dB) P uis s anc e (dB )

60

40

20

0 0 2

Signal du courant statorique

4

Time (s) Temps (s)

30

35

40

45

50

55

60

65

70

Frquence (Hz) Frequency (Hz)

A m plitude (A )

Fig. 07: stator current analysis by STFT for the healthy motor test with a zoom around the fundamental.

The theoretical sideband frequencies of one broken rotor bar fault calculated according to the motor slip and depending on the value of k=1,2 are given by: [41.88Hz;45.94Hz;54.04Hz;58.12Hz]. These frequencies are located around the fundamental. Through the zoom done on the time-frequency spectrum shown in fig. 08, we can observe the sideband frequencies of the fault of one broken bar around the fundamental. Also, it should be noted that the detected frequencies are equal to the sideband frequencies calculated theoretically based on the motor slip.
100 Spectre STFT temps-frqeunce

-2

-4

2.2

2.4 2.6 Temps (s) Time (s)

2.8

Fig. 09: stator current signal for the two broken bars test.

80

45.81 Hz 42.92 dB 41.85 Hz 30.81 dB

54.16 Hz 44.55 dB 58.44 Hz 33.00 dB

The zoom done on the time-frequency spectrum around the fundamental is represented in fig. 10. We can observe the sideband frequencies of the fault of two broken bars. STFT Spectre temps-frqeunce
100 80

Power e (dB P uis s anc (dB) )

60

41.21 Hz 33.26 dB

45,70 Hz 54.48 Hz 46,36 dB 45.79 dB

40
P uis s anc e (dB )

58.87 Hz 36.43 dB

20

Power (dB)
2 4 30 35 40

60

40

0 0

20

0 0

Temps (s) Time (s)

Frequency (Hz)

55 50 45 Frquence (Hz)

60

65

70

2 4 Temps (s) Time (s) 30 35 40

Frequency (Hz)

55 50 45 Frquence (Hz)

60

65

70

Fig. 08: stator current analysis by STFT for the one broken bar test with a zoom around the fundamental.

Fig. 10: stator current analysis by STFT for the two broken bar test with a zoom around the fundamental.

These frequencies are equal to the sideband frequencies calculated theoretically based on the motor slip. There is an increasing in the power of the faults component compared to the case of one broken bar.

[2] [3]

6. CONCLUSION Through this paper, accent is placed on the presentation of an effective diagnostic method capable of detecting and locating the sideband frequencies of rotor faults particularly near the fundamental. Indeed, it was used in this paper, a signal corresponding to the stator current signal in the case of a squirrel cage induction motor with one and two broken rotor bars. Therefore, an algorithm based on Short Time Fourier Transform is developed to perform a correct diagnosis and relatively precise location of the sideband frequencies of the considered fault. Several tests have been established to verify and to validate the robustness and effectiveness of our approach. Indeed, The STFT is robust to variations of the noise and allows a clear and easy location of the faults frequencies for even low values of motor slip. Also, we have experimentally validated our approach using experimental signals acquired from a test conducted at the Laboratory of Automatics and Industrial Informatics at the University of Poitiers (France). The time-frequency analysis of stator current from the different tests led to the detection of fault sideband frequencies. Finally, it should be noted that this work constitutes a first step in analyzing non-stationary signals, through which, we have validated the effectiveness of our approach in the case of stationary signals. In this context, two perspectives should be considered for this work: extending the diagnosis to the transient state (case of load changes) to solve the non-stationarity problem observed in the variation of load of the induction machine and the experimental validation of this approach in the case of load changes.

[4]

[5] [6]

[7]

[8] [9] [10]

[11] [12] [13]

[14] [15]

W. Li, C.K. Meshefske, Detection of induction motor faults: a comparison of stator current, vibration and acoustic methods , Journal of Vibration and Control, vol. 12, pp. 165-188. 2006 B. Liang, S.D. Iwnicki, A.D. Ball, Asymmetrical stator and rotor faulty detection using vibration, phase current and transient speed analysis , Mechanical Systems and Signal Processing, Elsevier, vol. 17, pp. 857-869. 2003 A. Bendiabdellah, N. Benouzza, D. Toumi, Cage rotor faults detection by speed estimation and spectral current analysis , in proc. of the 3rd IET International Conference on Power Electronics, Machines and Drives (PEMD06), Dublin, Ireland. April 4-6, 2006. S. Nandi, H.A. Toliyat, Condition monitoring and fault diagnosis of electrical machinesA review , IEEE Trans. on Energy Conversion, vol. 20, pp. 719-729, 2005. A.H. Boudinar, A. Bendiabdellah, N. Benouzza, N. Boughanmi, Three phase induction motor incipient rotors faults detection based on improved Root-MUSIC approach , International Review of Electrical Engineering, vol. 02, N 03, pp. 406-413. May - June 2007 M.E.H. Benbouzid, M. Viera, C. Theys, Induction motors faults detection and localization using stator current advanced signal processing techniques , IEEE Trans. on Power Electronics, vol. 14, pp. 14-22, January 1999 M.E.H. Benbouzid, A review of induction motors signature analysis as a medium for faults detection , IEEE Trans. on Industry Electronics, vol. 47, pp. 984-993, October 2000 M. Benidir Thorie et traitement du signal 2 : Mthodes de base pour lanalyse et le traitement du signal , dition Dunod. 2004 B. Yazici, G.B. Kliman, An adaptive statistical timefrequency method for detection of broken bars and bearing faults in motors using stator current , IEEE Trans. on Industry Applications, vol. 35, pp. 442-452, Mar./Apr. 1999 P. Wallisch, Matlab for neuroscientists : An introduction to scientific computing in Matlab , Elsevier Inc. 2009 P. Flandrin, Time-Frequency , Hermes. 1998 R. Burnett, J.F. Watson, S. Elder, The application of modern signal processing techniques for use in rotor fault detection and location within three-phase induction motors , Signal Processing, Elsevier, vol. 49, pp. 57-70. 1996 G. Blanchet, M. Charbit, Digital signal and image processing using MATLAB , ISTE Inc., USA. 2006 V.C. Chen, H. Ling, Time-frequency transforms for radar imaging and signal processing , Artech House Edition. 2002

ACKNOWLEDGEMENTS The authors wish to thank all the team of the Laboratory of Automatic and Industrial Informatics of the University of Poitiers (FRANCE) and special thanks to Mr. Gerard Champenois for the aid he afforded for our experimental work validation. REFERENCES
[1] D.F. Shi, F. Tsung, P.J. Unsworth, Adaptive timefrequency decomposition for transient vibration monitoring of rotating machinery , Mechanical Systems and Signal Processing, Elsevier, vol.18, pp. 127141. 2004