Extracting sensitive indicator for early diagnosis of rotating machinery

Ishak TOUAFEK touafekishak@yahoo.com Ahmed FELKAOUI a_felkaoui@yahoo.fr

mahafida006@yahoo.fr

Hafida Mahgoun

Laboratoire de mcanique de prcision applique (LMPA) Optics and precision Mechanics institute University Ferhat Abbas Setif - (Algeria) Abstract Indicators measure the vibration or describe the condition of a machine. They shed light on trends that may lead to damage, they describe the deviation of the health status relative to a reference state. Indicators must satisfy a number of qualities that can sometimes be contradictory, such as: relevance, simplicity, objectivity, sensitivity, accuracy, reliability, robustness, etc.. The sensitivity of the indicators is the ability of these to detect any presence of a defect in the mechanical parts. It plays a very important role in the early detection and diagnosis of faults or failures in rotating machinery. Several researchers have focused on improving the sensitivity, we find, among others, in the case of filtering [1][2], de-noising [3][4][5][6], the time synchronous averaging [7], the correlation and envelope spectrum [8][9][10], etc.. This work aims to study and compare the sensitivity of statistical scalar indicators of mechanical shock (kurtosis, crest factor, RMS) applied to signals which have been decomposed by the wavelet transform (WT) [11][12] and the empirical mode decomposition (EMD) [13][14][15] to increase their sensitivity. Key words: signal processing, faults diagnosis, Emprical Mode Decomposition (EMD) , Wavelet transform (WT), Kutosis, crest factor, RMS. 1- Introduction : Vibration monitoring using scalar indicators such as kurtosis, crest factor and root mean square (RMS) based on their sensitivity to detect a malicious purpose machines defects during operation. For this, the application of some methods can help to increase the sensitivity by eliminating some non significant signals mixed with in the original signal recorded using accelerometers sensor. Because these signals include the signals of gears, rolling bearings, etc ... ... and to monitor these elements, one must know the special characteristics of each element that makes up the machine. In our case, the diagnosis is made on the gears. We used the gear signals of CETIM laboratory in France to implement the methods we used. In this work, we will give the signal characteristics of gear, simulating a signal with his meshing frequency and rotating frequency. The methods we used are methods that have given perfect analysis for non-stationary and non-linear signals using the wavelet discrete transform (WDT) and the empirical mode decomposition (EMD). 2- Gears 2-1 Definition The gear is one of the power transmission largest used in construction of rotating machines. Gear consists of the set of two gears meshing with each other to transmit power between two shafts close together with a constant gear ratio. When a gear is defective, he generates shocks which can be seen on the spectrum of the signal that represents and its, in general, with the form of non-stationary. Also the spectrums measured on the machines that are equipped are rich in harmonic frequencies that we will be defining below. 2-2 Meshing frequency We have a gear with two pinions 1 and 2, with Z1 and Z2 teeth and turning frequencies F1 and F2. Each time a teeth of the wheel engages in leading a driven wheel 2, there is a charging jack periodic rate of engagement of the teeth meshing with a frequency Fe equal to the frequency of rotation of the wheel multiplied by the number of teeth. [16] = 1 1 = 2 2 (1)

If the teeth are correct, the spectrum consists of components whose frequencies correspond to the meshing frequency and its harmonics.

Fig (1) : Spectrum of correct gear 2-3 Modeling of a gear: 2-3-1 Mathematical model of gear faults If we consider a gear vibration signal which is averaged synchronously with respect to the frequency fr of rotation of a wheel of teeth Z, we will have: (2 + ) () = cos
=0

M : order signal analysis of meshing. Xm the amplitude of the harmonic m.

(2)

If the gear now has a defect, such as a crack at the base of teeth, this will result in an amplitude modulation and phase signal meshing. The modulated signal is given by (2 + + ()) () = 1 + () cos () = . cos (2 + )
=0 =0 =0

a m (t) and b m (t) represent respectively the amplitude modulation and frequency signal meshing due to the presence of a defect of teeth. These functions are defined by their Fourier series. In the spectrum, the amplitude modulation will create lines around the side peaks meshing. The spacing of the satellite lines is equal to the frequency of rotation of the wheel bearing failure. The extent of the satellite lines depends strongly on the shape of the modulating signal. It can therefore be based on the assumptions above, to a demodulation of amplitude and phase of the signal ()
x 10
7

() = . cos (2 + )

(4)

(3)

(5)

spectrum of modeled gear with a defect Fe 2 Fe

3 Fe

15 10 5 0

200

400

600 800 Frequency (Hz)

1000

1200

1400

Fig (2): simulated spectrum of a gear We see in the figures (2), the spectrum of a gear modeled with a defect. This spectrum shows the peaks of the meshing frequency and its harmonics, in addition to the side stripes representing the frequency of rotation around the meshing frequency. For the presence a defect, the signal failure of the gear change at the side stripes that will increase in amplitude. 3Empirical Mode Decomposition (EMD) :

The technique of Empirical Mode Decomposition (EMD) is a signal analysis method that was introduced by Norden Huang. Applied since (1998) to decompose any signal into a sum of oscillating components derived directly from it. [17].

These components (or IMF for Intrinsic Mode Functions) are interpreted as non-stationary waveforms, (modulated in amplitude and frequency) may optionally be associated with non-linear oscillations. IMF satisfying the conditions [18] [13] 1- In the whole data set, the number of extrema and the number of zero-crossing must either be equal or differ at most by one. 2- At any point, the mean value of the envelope defined by local maxima and the envelope defined by the local minima is zero. The empirical mode decomposition is a method of deflation is an iteration in which each estimate is subtracted from the signal mode and analyzed or estimate the next mode is driven from the signal residue. 3-1 The EMD method. An IMF represents a simple oscillatory mode as a counterpart to the simple harmonic function of the Fourier series method. To obtain an IMF from the original signal, we do the following steps:[19][20] Pick out all of the maxima of the series signal X(t) and calculate the upper envelope with cubic spline function. Pick out all of the minima of the series signal X(t) and calculate the lower envelope with cubic spline function.
upper envelope

Signal

lower envelope

The mean envelope m1(t) of the signal X(t) is the mean value of the upper envelope and lower envelope.

mean envelope

A new series h 1 with low frequency removed is calculated by subtracting the mean envelope from the serie X(t): 1 () = () 1 () (6)

3Generally speaking, h1 is still a non stationary series, so the above procedure must be repeated k times until the mean envelope is approximate to zero, so the first IMF component C (t) is obtained: 1(1) () 1( ) () = 1( ) () 1 () = 1( ) () (7)

The first IMF component represents the highest frequency component of the original series. The second IMF component C2 (t) can be obtained from the margin series r 1 (t) which is calculated by subtracting the first IMF component from series X(t). such procedure needs to be repeated until the last margin series r n(t) cannot be decomposed, here rn(t) represents the mean value or trend of original series. () 2 () = 2 () . (1) () () = () So, the original series can be presented as: () = () + ()
=1

(8)

(9) (10)

4- The discret wavelet transform: From the mother wavelet (t) we constructed by translation and dilation a family of functions ( , ) (). These functions ( , ) () are given by the relationship [4] [21]:

Where a is the scaling parameter and b is the translation parameter

( , ) =

(11)

The discrete wavelet transform is derived from the continuous version, unlike the latter; the DWT uses a scale factor and a translation discrete. Called discrete wavelet transform any dyadic wavelet basis working with a scale factor a = 2n. From the functional point of view, we replace a continuous representation by a discrete representation. This is done by sampling the wavelet coefficients. So In this case the parameters a and b become:[22]
= 0

et

0 : is a scale parameter.

= 0 0

Among the methods used to calculate the discrete wavelet transform, we find the multi resolution. This technique consists of passing the original signal through two filters one that presents the high- Pass for details and the other is the low- pass for approximations.

0 : is a translation parameter

Fig(3): discrete wavelet transform with multi resolution The idea then is to separate the signal into several frequency bands "low frequencies and high frequencies." The approximation coefficients , at resolution 2 1 are obtained from the approximation coefficients at the resolution 2 by filtering using a filter impulse response h[n], followed by a decimator of order 2, and the detail coefficients , are obtained from the coefficients approximation , by filtering using a filter impulse response g[n] followed by a decimator of order 2.[22] The approximations were calculated with (12), and details with (13):
4T 2T 2T

, = 2 [] 1,2 +
= +

(12) (13)

, = 2 [] 1,2 +
=

So, the signal will be writing with this form: () = 1 + 2 + + + (14)

5- Statistical indicators : 5-1 Kurtosis : This quantity introduced by the mathematician Pearson is deduced by the ratio: [23][24][25] : () = 4 () [2 ()]2 (15)

M4 (sig) and M 2 (sig) are respectively the moments of order 4 and 2 signal (sig). than 3.

For a Gaussian distribution, the kurtosis has a value of 3, then for the presence of a default, it takes a value greater

5-2 Crest factor : The crest factor is defined as the ratio between the peak (max) in signal and RMS (RMS) [23][24] = (16)

This indicator is used to detect specific pulse or shock present in the vibration signal.

For a Gaussian distribution, the crest factor of 3 to 6, but for the presence of default, this flag is set higher than 6. 5-3 Root Mean Square (RMS) It is a value characteristic of the signal, since it has a direct relationship with the energy contained in it: 1 = 2 () 0 (17)

sig (t) is the representative function of the signal et T analysis time. 6- Gear signals of CETIM

The system consists of a motor, a gear ratio of 40/42 and closure of the gear tested, report 20/21. Figure (4) shows the system.

Fig (4) : System of CETIM The table (1) gives the characteristic of the system tested with CETIM Sampling frequency Number of points recorded Meshing frequency (Hz) Number of teeth of the pinion 1 Number of teeth of the pinion 2 Frequency of rotation of the pinion 1 (Hz) Frequency of rotation of the pinion 2 (Hz) Fs N fe z1 z2 f1 f2 20000 60170 333.48 21 20 16.67 15.88

Table (1) : Characteristic of the chassis CETIM .

Jour 2 3 4 5 6 7 8 9 10 11 12 13

Observation First day of acquisition, no anomaly no anomaly // // // // // // Scaling in the teeth 1/2 No change tooth 1 / 2 no change Evolution of tooth scaling 15/16 // // // // Scaling across the width of the tooth 15/16 Table (2): Expert report

The analysis of time signals allows us to visualize the occurrence of a fault very marked the twelfth day. However, as can be seen in figure (5), it is not possible to detect early the occurrence of a fault using the time signals.
x 10
4

day 2 2 0

2 0 -2

x 10

day 3 2 0

x 10

day 4 2 0

x 10

day 5

0 x 10
4

0.5 time (s) day 6

-2

0 x 10
4

0.5 time (s) day 7

-2

0 x 10
4

0.5 time (s) day 8

-2

0 x 10
4

0.5 time (s) day 9

2 0 -2

2 0

2 0

2 0

0.5 time (s)

-2

0.5 time (s)

-2

0.5 time (s)

-2

0.5 time (s)

2 0 -2

4 x 10 day 10

2 0

4 x 10 day 11

5 0

4 x 10 day 12

5 0 -5

4 x 10 day 13

0.5 time (s)

-2

-5 0.5 1 0 0.5 1 time (s) time (s) Fig(5): Time signals of the gear train CETIM

0.5 time (s)

The calculation of statistical indicators for the 12 days of the experiment back to draw the curves in the following figures.

kurtosis
20 10 0 jour 2 jour 3 jour 4 jour 5 jour 6 jour 7 jour 8 jour 9 jour 10 jour 11 jour 12 jour 13 kurtosis

Fig(6): kortosis in 12 days

Crest factor
10 5 0 jour 2 jour 3 jour 4 jour 5 jour 6 jour 7 jour 8 jour 9 jour 10jour 11jour 12jour 13 Fig(7): Crest factor in 12 days Crest factor

RMS
10000 0 jour 2 jour 3 jour 4 jour 5 jour 6 jour 7 jour 8 jour 9 jour 10 jour 11 jour 12 jour 13 Fig(8): Root Mean Square (RMS) in 12 days RMS

7- Application of EMD and wavelets on the signals CETIM The EMD method decomposes the signal into several signals in mono mode known IMF. And each IMF covers a frequency band. The decomposition process of high frequency to lower frequencies. In what, monitoring indicators are calculated by taking the total size of the signal and not around the frequencies selected such as meshing frequencies. Empirical mode decomposition decomposes signal into several signals called intrinsic mode function IMF.Each IMF covers a frequency band; it acts as a filter band (strip). The table (3) shows the frequency band of each IMF around the maximum frequency for signals of CETIM. IMF Bande frquentielle [Hz] 1 2950 - 3150 2 1270 - 1440 3 610 - 780 4 230 - 400 5 200 - 400 6 200 - 400 7 50 - 150 8 20 - 120 9 0 - 50 10 0 - 35 11 010 12 0 - 10 Table 3: Frequency bands of each IMF IMF 5 shows a spectrum represents the meshing frequency IMF 4 and IMF 3 present a spectrum which shows the second harmonic of the meshing frequency. IMF 9 et IMF 10 show the spectrum of rotating frequency IMF 8 shows the spectrum of the third harmonic of rotating frequency. Le signal reconstruit partir de ces IMFs est de la forme suivante : Defects found in the gears are kind of shock, then most of the monitoring indicators used for this type is the Kurtosis and crest factor.

day 2

day3

5 0 imf1 imf2 imf3 imf4 imf5 imf6 imf7 imf8 imf9 imf10 imf11 imf12

10 5 0 imf1 imf2 imf3 imf4 imf5 imf6 imf7 imf8 imf9 imf10 imf11 imf12
day5

day4

5 0

day 6

10 5 0 imf1 imf2 imf3 imf4 imf5 imf6 imf7 imf8 imf9 imf10 imf11 imf12

60 40 20 0 imf1 imf2 imf3 imf4 imf5 imf6 imf7 imf8 imf9 imf10 imf11 imf12
day9

day8

100 50 0 imf1 imf2 imf3 imf4 imf5 imf6 imf7 imf8 imf9 imf10 imf11 imf12

5 0 imf1 imf2 imf3 imf4 imf5 imf6 imf7 imf8 imf9 imf10 imf11 imf12
day11

day10

4 2 0 imf1 imf2 imf3 imf4 imf5 imf6 imf7 imf8 imf9 imf10 imf11 imf12

60 40 20 0 imf1 imf2 imf3 imf4 imf5 imf6 imf7 imf8 imf9 imf10 imf11 imf12
day13

day12

40 20 0 imf1 imf2 imf3 imf4 imf5 imf6 imf7 imf8 imf9 imf10 imf11 imf12

40 20 0 imf1 imf2 imf3 imf4 imf5 imf6 imf7 imf8 imf9 imf10 imf11 imf12

Fig (9) : Variation of kortosis for IMF of 12 days Figure (9) shows the distribution and variation of the kurtosis of each IMF during the 12 days of testing. We see that this indicator reach the value (54) on day 7 and (96) for the IMF 7 and (47) for the IMF 8 and IMF 9 on the day 8.. We can say that the fault has occurred from day 7. But, before the use of EMD, the variation of kurtosis shows that the fault has occurred in the 12 days. Figure (10) shows a comparison of the variation of kurtosis in the 12 days before and after the application of EMD.

imf1 imf2 imf3 imf4 imf5 imf6 imf7 imf8 imf9 imf imf imf
day 7

5 0

10

8 6 4 2 0

20 15 10 5 0 0 5 imf1 imf2 imf3 imf4 imf5

day6 day4 day2

10 5 0

imf1 imf2 imf3 imf4 imf5 imf6 imf7 imf8 imf9 imf10 imf11 imf12

imf1

imf2

imf3

imf4 imf6 imf7 imf8 imf9 imf10 imf11 imf12

imf5

day8

imf6

imf7

imf8

imf9

imf10

imf11

imf12

Fig(10) : kurtosis in 12 days with EMD and without EMD

We clearly see the increased sensitivity of this indicator on the days 7, 8, 11, 12, and 13 after application of EMD.

9
6 4 2 0 6 4 2 0 20 10 0 imf1 imf2 imf3 imf4 imf5
day7

8 6 4 2 0 imf1 imf2 imf3 imf4 imf5

day5

imf1 imf2 imf3 imf4 imf5

imf1

imf2

imf3

imf4 imf6 imf7 imf8 imf9 imf10 imf11 imf12 imf9 imf10 imf11 imf12 imf8 imf7 imf6

imf5

day9

day3

imf6

imf6 imf7 imf8 imf9 imf10 imf11 imf12

imf7

imf8

imf9

imf10

imf11

imf12

day10

day11

10 5 0 imf10 imf11 imf12 imf7 imf1 imf2 imf3 imf4 imf5 imf6 imf8 imf9

20 10 0 imf10 imf10 imf11 imf11 imf12 imf12 imf7 imf7 imf1 imf2 imf3 imf4 imf5 imf6 imf8 imf8 imf9 imf9

day12

day13

15 10 5 0 imf10 imf11 imf12 imf7 imf1 imf2 imf3 imf4 imf5 imf6 imf8 imf9

15 10 5 0 imf1 imf2 imf3 imf4 imf5 imf6

Fig(11) : Variation of crest factor of IMF in 12 days The application of the method of empirical mode decomposition (EMD) has detected the presence of a fault on the day 7 and not on the day 12. The value of crest factor is close to 10 for the IMF 8 and IMF 9 on the day 7, and on the day 8, this value is increased to 16 for the IMF 7. Figure (12) shows a comparison of the change in this indicator before and after the application of EMD.

Fig (12): Crest factor before and after the application of EMD The sensitivity of this indicator is also increased with the use of EMD. 8- Application of discrete wavelet transforms db7 on signals of CETIM

The wavelet Daubchies "db7" is used, the number of levels for decomposition is N = 10. This number is calculated using a following expression (18)[26]:
= log ( ) log (2)

(18)

10

Table (4) shows different frequency bands for the details and approximations after decomposition. Niveau 1 2 3 4 5 6 7 8 9 10 A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 Bande frquentielle Bande frquentielle 0 5000 D1 5000 - 10000 0 1250 D2 2500 - 5000 0 625 D3 1250 - 2500 0 312.5 D4 625 - 1250 0 156.25 D5 312.5 - 625 0 78.125 D6 156.25 - 312.5 0 39.062 D7 78.25 - 156.25 0 19.531 D8 39.062 - 78.25 0 9.765 D9 19.531 - 39.0625 0 4.882 D10 9.765 - 19.531 Table(4) : Bands for the details and approximations

Our characteristic frequencies are: Meshing frequency Fs= 333.48 Hz. The frequency of rotation is Fr= 16.67 Hz. The decomposition is made on 10 levels, the detail D10 covers the band included the rotational speed of the machine. For the meshing frequency, the details D5
day2 day3

150 100 50 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10

day4

10 5 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day5

20 10 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day6

10 5 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day7

4 3 2 1 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day8

30 20 10 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day9

6 4 2 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10

6 4 2 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10

11

day10

day11

60 40 20 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day12

30 20 10 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day13

300 200 100 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10

200 150 100 50 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10

Fig (13): kurtosis of the details and approximations in 12 days After the decomposition of each signal of each day using the discrete wavelet transform (DB7), we see that this technique could detect the presence of the same shock for days, and it can be hard to assemble the system because the teeth were in good condition in these days and that from the expert report, but it has also say that the of the teeths cracks started from the day 7 wheel, and we can see that after the decomposition fault appeared from the seventh day, and it caused by increased values of the Kurtosis in detail d6, d7, d8, and d9 for harmonic frequency of rotation.
day2 day3

15 10 5 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day4

6 4 2 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day5

10 5 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day6

6 4 2 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day7

6 4 2 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10

10 5 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10

12

day8

day9

10 5 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day10

6 4 2 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day11

10 5 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day12

15 10 5 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10
day13

30 20 10 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10

30 20 10 0 d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 a10

Fig (14): Crest factor of details and approximations in 12 days 9- Conclusion In this work, we applied methods for non-stationary signals and non-linear in order to see their effect on the sensitivity of statistical indicators for monitoring such as kurtosis and crest factor. The method of empirical mode decomposition (EMD) is applied to the test signals for broken down into several IMF to represent the frequency bands exist in the original signal. As for the discrete wavelet transform (WDT). With this techniques, we have seen that the shock-like defects, occurred in the seventh day and not the twelfth day, which shows that the sensitivity of the indicators used for this purpose given to increased and improved monitoring and diagnostic reading. The method of the EMD, and as it is based on a specified algorithm is not a well defined theory such as wavelets, is simple compared to the wavelet which requires a good choice of the mother wavelet and good for the calculation number of decomposition level. 10- Reference [1] Jrome Antoni, R.B. Randall : The spectral kurtosis: application to the vibratory surveillance and diagnostics of rotating machines .Mechanical Systems and Signal Processing 20 pp 308331. 2006 [2] Adriano O. Andrade , Slawomir Nasuto, Peter Kyberd, Catherine M. Sweeney-Reed, and F.R. Van Kanijn: (EMG signal filtering based on Empirical Mode Decomposition) Biomedical Signal Processing and Control Volume 1,pp 44-55. January 2006. [3] David L. Donoho De-Noising by Soft-Thresholding IEEE Transactions on informations theory.Vol. 41, NO. 3. MAY 1995. [4] Michel Misiti, Yves Misiti : Les ondelettes et leurs applications . Edition Lavoisier, Paris. 2003. [5] Quan Pan, Lei Zhang, Guanzhong Dai, and Hongcai Zhang (Two Denoising Methods by Wavelet Transform) . IEEE Transactions on signal processing. VOL. 47, NO. 12. Dcembre 1999. [6] J.-C. Cexus, A.-O. Boudraa, S. Benramdane Dbruitage des signaux par approche EMD : Multi-EMDSG . Colloque GRETSI, 11-14 septembre 2007, Troyes. [7] Frdric Bonnardot : Comparaison entre les analyses angulaire et temporelle des signaux vibratoires de machines tournantes. Etude du concept de cyclo stationnarit floue . Institut national polytechnique de grenoble. Thse 2004.

13

[8] Frdric Bonnardot, Robert Bob Randall, Jrome Antoni, Franois Guillet : Enhanced unsupervised noise cancellation (E-SANC) using angular resampling application for planetary bearing fault diagnosis. Surveillance 5 CETIM Senlis 11- 13 octobre 2004. [9] R. B. RANDALL, J. ANTONI, S. CHOBSAARD : the relationship between spectral correlation and envelope analyzing the diagnostics of bearing faults and other cyclostationary machines signal Mechanical Systems and Signal Processing vol (15) N(5),pp 945-962. 2001. [10] Hlne Laurent, Menad Sid Ahmed, Christian Doncarli : Dtection prcoce dendommagement de denture sur des trains dengrenages droits par analyse vibratoire Dix-septime colloque GERTSI, Vannes, 13 17 septembre 1999. [11] Jie Liu,WilsonWang, Farid Golnaraghi, Kefu Liu ( Wavelet spectrum analysis for bearing fault diagnostics). Meas. Sci. Technol. 19 (2008). Measurement science and technology [12] Truchetet F. Ondelettes pour le signal numrique Editions Hermes, Paris, 1998 [13] Q.Gao, C.Duan, H.Fan, Q.Meng Rotating machine fault diagnosis using empirical mode decomposition .Mechanical Systems and Signal Processing. Vol (22), pp1072-1081. 2008. [14] Flandrin P. Emprical Mode.Decomposition as a Filtre Bank . IEEE signal processing letters.Vol 11, pp.112-114. 2004. [15] Rilling G. Dcompositions Modales Empiriques : Contributions la thorie, l'algorithme et l'analyse de performances . Thse : Ecole Normale Suprieure de Lyon2007. pp 1-278. [16] (schenk, 1994) vibrations quilibrage sur le site : Application la maintenace industrielle.Edition Schenk S A. 1994. [17] Patrick Flandrin, Fellow,IEEE: Emprical Mode.Decomposition as a Filtre Bank. IEEE signal processing letters. Vol 11. pp.112-114, 2004. [18] Qiuhua Du, Shunian Yang : Application of the EMD methode in the vibration analysis of ball bearings . MSSP. Vol 21, pp 2634-2644. 2007 [19] Hui Li : Bearing fault detection based on instantaneous energy spectrum . Seventh international conference on Fuzzy systems and knowledge discovery (FSKD 2010). 2010. [20] Deng Yongjun, Wang Wei et al.. : Boundary- processing technique in EMD method and Hilbert transform . Chinese Science Bulletin. Vol 46, no 1, January 2001. [21] Fred Truchetet : Ondelettes pour le signal numrique . Edition Hermes, Paris. 1998. [22] Hafida Mahgoun : Analyse et caractrisation dfes signaux vibratoires cyclo-stationnaires . Thse de magister, Universit de Stif. Algrie, 2000. [23] Alain Boulenger et Christian Pachaud : Analyse vibratoire en maintenance : Surveillance et diagnostic des machines . Edition DUNOD, Paris. 2003. [24] Baptiste TRAJIN : Analyse et traitement de grandeurs lectriques pour la dtection et le diagnostic de dfauts mcaniques dans les entranements asynchrones : Application la surveillance des roulements billes . thse de doctorat, Institut National Polytechnique de Toulouse. 2009. [25] Philipp Arqus : Diagnostic prdictif de ltat des machines . Edition MASSON, Paris. 1996. [26] BOUZIDA Ahcene : DIAGNOSTIC DE DEFAUTS DE LAMACHINE ASYNCHRONE A CAGE PAR LA TECHNIQUE DES ONDELETTES . Thse de magister, Ecole Nationale Polytechnique dAlger, Algerie, 2007.

14

15