Home

Search

Collections

Journals

About

Contact us

My IOPscience

Gear Fault Signal Detection based on an Adaptive Fractional Fourier Transform Filter

This article has been downloaded from IOPscience. Please scroll down to see the full text article. 2011 J. Phys.: Conf. Ser. 305 012022 (http://iopscience.iop.org/1742-6596/305/1/012022) View the table of contents for this issue, or go to the journal homepage for more

Download details: IP Address: 117.211.91.202 The article was downloaded on 08/09/2011 at 08:26

Please note that terms and conditions apply.

9th International Conference on Damage Assessment of Structures (DAMAS 2011) IOP Publishing Journal of Physics: Conference Series 305 (2011) 012022 doi:10.1088/1742-6596/305/1/012022

Gear Fault Signal Detection based on an Adaptive Fractional Fourier Transform Filter
Xiaojun Zhou[1*], Yimin Shao[1,3], Dong Zhen[2], Fengshou Gu[2], Andrew. Ball[2] State Key Laboratory of Mechanical Transmission, Chongqing University, 400044, PR China; 2 School of Computing and Engineering, University of Huddersfield, HD1 3DH, UK; 3 College of Automobile and Mechanical Engineering, Changsha University of Science and Technology, Changsha, 410004, PR China.
1

Abstract: Vibration-based fault diagnosis is widely used for gearbox monitoring. However, it often needs considerable effort to extract effective diagnostic feature signal from noisy vibration signals because of rich signal components contained in a complex gear transmission system. In this paper, an adaptive fractional Fourier transform filter is proposed to suppress noise in gear vibration signals and hence to highlight signal components originated from gear fault dynamic characteristics. The approach relies on the use of adaptive filters in the fractional Fourier transform domain with the optimised fractional transform order and the filter parameters, while the transform orders are selected when the signal have the highest energy gathering and the filter parameters are determined by evolutionary rules. The results from the simulation and experiments have verified the performance of the proposed algorithm in extracting the gear failure signal components from the noisy signals based on a multistage gearbox system.

1. Introduction Gearbox health monitoring problems have received increased interest for the last two decades due to maximising safety and minimising cost in operations. Vibration signal are sensitive quantities with respect to a fault and it is often used for gearbox fault detection diagnosis. However, gear fault feature signal are always buried in the heavy noise which are from the healthy gear-pair meshing vibration and other machine vibration sources [1-2]. How to extract the gear fault feature signal efficiently or how to improve the signal-noise-ratio (S/N) actively is the main research problems of gearbox monitoring. Many signal detection methods such as wavelet transform (WT) [3], wavelet package transform (WPT) [4], genetic algorithm (GA) [5], ensemble empirical mode decomposition (EEMD) [6] and artificial neural network (ANN) [7] have been proposed to suppress noise in gear vibration signals and hence to highlight components originated from gear dynamic characteristics. Adaptive noise cancelling (ANC) is an effective method for estimating signal when it is corrupted by additive noise [8]. The effectiveness of the signal enhancement depends on the correlation degrees between the reference signal and the noise contained in the primary signal. For better effectiveness, many adaptive filter parameters search algorithms have been proposed, e.g. least mean square (LMS)
*Authorcorrespondence,Phone:(+8623)65112520,Email:xiaojun@cqu.edu.cn.

Published under licence by IOP Publishing Ltd

9th International Conference on Damage Assessment of Structures (DAMAS 2011) IOP Publishing Journal of Physics: Conference Series 305 (2011) 012022 doi:10.1088/1742-6596/305/1/012022

algorithm, recursive least squares (RLS) algorithm, LMS/RLS algorithms in the frequency domain and so on. However, there are little researches on the variation of output signal when the adaptive noise cancelling realized from the time domain to frequency domain. In the mean time, the fractional Fourier transform (FrFT) belongs to the class of time-frequency representations which have been extensively used in the field of signal processing for extracting useful signal components from complicated noise and interferences data [9-12]. FrFTs response to the ordinary time domain and ordinary Fourier transform when the transform order are equal to 0 and ! / 2 , respectively. An optimal transform fractional order will lead to an optimal energy density and the signal intensity simultaneously in the time and frequency plane. So in this work, a novel algorithm of adaptive fractional Fourier transform filter (AFTF) is proposed for the gear fault signal extraction.. The approach relies on the use of adaptive noise cancelling filters in the fractional Fourier transform domain with the optimised fractional order and optimised filter parameters, while the transform order is selected when the transform signal have the highest energy gathering and the filter parameters are determined by evolutionary rules, because the filters based on the evolutionary rules can converge to a global minimum when deal with the multiple peak problem [13]. The content of this paper is organized into 5 sections. In section 2, the algorithm of adaptive fractional Fourier transform filter is described. The main body of the paper is in section 3 and 4 where both the simulation study and the test evaluations are addressed to show the performance of the method. Finally, section 5 summaries the conclusions. 2. Adaptive fractional Fourier transform filter 2.1. Review of fractional Fourier transform The Fractional Fourier transform (FrFT) is a time-frequency distribution extended from the classical Fourier transform (FT). There are several applications of FrFT in the areas of signal processing, especially in noise reduction. The FrFT of signal x(t ) with order denoted by X ! (u ) is defined as [9]

X ! (u ) = $ x(t ) K! (t , u )dt
#"

"

(1)

Where the kernel function K! (t , u ) is shown as

% & " (t $ u ) & & K! (t , u ) = ' " (t + u ) & u 2 +t 2 & 1 $ j cot ! e $ j ( 2 ) cot ! $ jut cos ec (! ) & 2# ( if is a multiple of 2# if +# is a multiple of 2# if is not a multiple of #

(2)

It is found that the FrFT of the signal x(t) exists under the same conditions in which its Fourier transform exists. And the FrFT of order =0 is the input signal itself, while the FrFT of order =/2 corresponds to the Fourier transform of the input signal. if a signal x(t) is a periodic signal and can be expressed as [10]

x(t ) =
where X n =

n =#"

$X

"

exp( j 2! nt )

(3)

x(t ) exp(" j 2! nt )dt , the FrFT of the signal can be shown as

9th International Conference on Damage Assessment of Structures (DAMAS 2011) IOP Publishing Journal of Physics: Conference Series 305 (2011) 012022 doi:10.1088/1742-6596/305/1/012022

X ! (u ) =
where " (u , # ) =

! j ( $" u 2 tan ! ) 1 e 2 # (u , ! ) cos !

(4)

n =%$

&X

exp(

j 2! nu ) exp(% j! n 2 tan # ) . cos #

Figure 1 gives the chirp signal and its time-frequency distribution with the transform order in the fractional Fourier domain. Figure 1(b) shows the chirp signal can convergence to an appropriate transform order in the fractional Fourier transform domain.
1

e d u t i l p m A

0.5

-0.5

-1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Time (s)

a) Waveform of the signal b) Energy distribution with the transform order Figure 1. Waveform of the signals 2.2. Algorithm adaptive fractional Fourier transform filter (AFTF) Figure 2 shows the diagram of the classic adaptive noise cancelling algorithm. Based on the reference signal N2, which is related to the noise signal N1 and unrelated to the feature signal S1, using the adaptive algorithms, we can extract the feature signal S 1 from sensor signal S.
Signal Source Signal Path

S1
Sensor Primary Input

Primary Input Noise Path Noise Source Refenence Input Noise Path

S = S1 + N1

"

Output

S$

N1

N2

Reference Input Sensor

N2

Adaptive Algorithm

Figure 2. Diagram of adaptive noise cancelling. The diagram of the adaptive fractional Fourier transform filter (AFTF) is shown in Figure 3. It contains the fractional Fourier transform algorithm and the adaptive noise cancelling algorithms which is based on the evolutionary algorithm. And the proposed adaptive fractional Fourier transform filter contains the following steps: 1) The primary input signal and the reference signal are transformed into the fractional Fourier domain by the FrFT with a fractional order; 2) The primary signal in the fractional transform domain is filtered by a series of filter with different parameters, while the filter parameters are controlled and refreshed based on the strategy of the evolutionary rules;
3

9th International Conference on Damage Assessment of Structures (DAMAS 2011) IOP Publishing Journal of Physics: Conference Series 305 (2011) 012022 doi:10.1088/1742-6596/305/1/012022

3) The error signals among the reference signal and the outputs from the filters are calculated; 4) The rules are used to evaluated the performance of the filters and for selection of the optimum filter parameters and the fractional orders, and go to step 2 if the output signal dont satisfy the requirements; 5) The optimum output signal is transformed to the time domain by the inverse fractional Fourier transform with the inverse of the optimum fractional order, and the gear fault diagnosis based on the noise cancelled signal can be much more accuracy.
Reference signal FrFT S 1(k) FrFT Y1(k) d(k) Input signal s(k) Filter (1)

+" +

D(k)

Filter (2)

y2(k)

"

! ! !

e1(k)

e 2(k)

Evaluation of the output signal Inverse fractional Fourier transform with the optimum order

Filter (i)

y i(k)

" + "

y*(k )
ei(k)
Updating order of fractional Fourier transform and the filter parameters

Output signal S_ m(k)

Filter (n)

yn (k)

+ ! e (k)
n

Figure 3. Diagram of the AFTF algorithm. In time domain, the output of the filter can be given by [13]

y (k ) = " an (k ) y (k ! n) + " bm (k ) x(k ! m)

n =1 m=0

(5)

Where, M and N are the orders of the moving averaging part and the autoregressive part, respectively. So in the fractional Fourier domain with fractional order , the output of the ith sub-filter can be given by

Yi (k , ! ) = # an ,i (k )Yi (k " n, ! ) + # bm ,i (k ) X i (k " m, ! )

n =1 m=0

(6)

where Y(k,) X(k,) stand for the fractional Fourier transform of y(k) and x(k) with the order . And the coefficients an ,i and bm ,i are correspond to the feature of an individual in the evolutionary algorithm. In this paper, the filter coefficients of each sub-filter at time k are given as a column vector

Wi (k ) = [an ,i (k ), bm ,i (k )]

(7)

The evolutionary algorithm is constructed to refresh the coefficients of the AFTF. The cloning and mating method are used according to the parents objective function fitness value, while cloning is for the higher fitness, and the rules are described as follows:

Wi ,cc , j (k ) = Wi ,cp (k ) + ! ini , j Wi ,mc , j (k ) = (Wi ,mpa (k ) + Wi ,mpb (k )) / 2 + ! ini , j

(8) (9)

9th International Conference on Damage Assessment of Structures (DAMAS 2011) IOP Publishing Journal of Physics: Conference Series 305 (2011) 012022 doi:10.1088/1742-6596/305/1/012022

where Wi ,cc , j (k ) Wi , mc , j (k ) are the children from the cloning method and mating method for the ith filter, respectively. Wi ,cp (k ) and Wi , mp (k ) are the parents for the cloning method and mating method, respectively. , are fluctuation coefficients. ni , j is a Gaussian random variable vector with zero mean and unit variance. The choice of the objective function is crucial for the algorithm definition since it can affect several characteristics of the overall algorithm including computational complexity, speed of convergence, robustness, and most importantly for the IIR adaptive filtering case, the occurrence of biased and multiple solutions. In this paper, the objective function for the ith filter is given by mean square of the error signal:
2 2 {! , i } = min! ,i E[ ei (k ) ] = min! ,i E[ Yi (k ) " D(k ) ] (10)

And the optimum FrFT order is estimated according to the following rule:

{! } = max! (kurtosis (ei (k )))

(11)

where ei(k) is the residual signal from the ith sub-filter; kurtosis stands for the kurtosis calculation, and it is sensitive to the shape of the signal, hence is well suited to the impulsive nature of the stimulating forces generated by component damage. 3. Simulation Validations According to the references [1,6], there will be periodic impulse signal which is induced by the gear failure, when the gearbox works under the constant speed condition. Supposed the noisy signal, x(t), contains the feature signal, s(t), the additive random noise, n(t), and the periodic noise signal, r(t). It expresses as x(t ) = s (t ) + n(t ) + r (t ) (12) where the feature signal s(t) is a periodic impulse signal, the noise n(t) is the random signal, and the noise signal r(t) is a periodic signal which has a different period with the feature signal. To validate the proposed algorithm under this condition, the periodic pulse signal with the noise is simulated. The feature signal is simulated using a periodic Gaussian pulse, with 50Hz signal frequency and 10Hz repetitive frequency, as shown in Figure 4(a), and the noise signal combines the 15 Hz sinusoidal signal and the rand noise with zero mean is shown in Figure 4(b), while the feature signal and the noise signal is formed into the noised signal (the primary signal) which have the -5dB signalnoise-ratio, as shown in Figure 4(d). The reference signal is shown as in Figure 4(c), which is relative to the noise contained in the primary signal but with different amplitude and phase, and the sampling frequency is 2 kHz.
e d u t i l p m A
1

0.5

-0.5

e 0.35 ed d 0.3 u u t 0.25 t i i l 0.20 l p p0.15 mm 0.1 AA

0.05

-1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

-5 0 0 0

0.1 100

0.2 200

0.3 300

0.4 400

0.5 500

0.6 600

0.7 700

0.8 900 0.9 1000 1 800

Time (s)
1

Time (Hz) Frequency(s)

0.35 0.3 0.25 0.2 0.15 0.1 0.05

a) Waveform of the feature signal

e d u t i l p m A

b) Waveform of the noise signal

e d u t i l p m A

0.5

5
-0.5

-1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

100

200

300

400

500

600

700

800

900

1000

9th International Conference on Damage Assessment of Structures (DAMAS 2011) IOP Publishing Journal of Physics: Conference Series 305 (2011) 012022 doi:10.1088/1742-6596/305/1/012022 -5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

d u t i l p m A

d u t i l p m A

0.8

0.6

0.4

0.2

100

200

Time (s)
5 13

e d u t i l p m A

ee dd 0.82 uu tt 1 i i 0.6 ll 0 pp 0.4 m -1 m AA 0.2

-2
0 -3 0 0

0.5

e d u t i l p m A

0.4

0.3

0.2

0.1

-5

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Time (s)
3

100 0.1

200 0.2

300 0.3

Frequency(s) Time (Hz)

400 0.4

500 0.5

600 700 0.6 0.7

800 900 1000 0.8 0.9 1

0 0 100 200

c) Waveform of the reference signal

e e Figure 4. Waveform of the signals d 2 d 0.4 u u t 1 t According to the Figure 5(a), the optimum transform 0.3 order is /2 for this simulation, and in the i i l 0 l evolutionary algorithm, the number of parents and children are 16 and 4, and the generation is 60. p p 0.2 Fromm -1 Figure 5(b), we can find the indicator calculated from the output signal converged quickly, that m A is to A say, the proposed algorithm can converge quickly and not time-consuming. And the extracted

0.5

d) Waveform of the noised signal

signal with the proposed algorithm and frequency domain adaptive filter (FDAF) using the LMS -3 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 100 200 300 400 500 600 700 800 900 1000 algorithm [14] are shown in Figure 6, 0.8 0.9 1 to the Figure 4(a), the noise signal is eliminated and compared Time (s) Frequency (Hz) the feature signal is highlighted by the proposed algorithm, while the output signal from the FDAF is still heavy noised. The simulation results verified the performance of the proposed algorithm is much better than the classic frequency domain adaptive filter algorithm.
20

-2

0.1

4.5

s i s e o d t u r t u i K l p m A

15

1
10

out-lms
4 3 2 1 0 -1
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

e d 0.5 5 u t i0 0 l p -5 m 0 -0.5 A
-1

-2 -3

s i s e o d t u r t u iK l p m A

4 0.35

out-dft
4

Transform order

e 0.3 d3.5 3 0.25 u 2 t 0.2 i 31 l 0.15 0 p -1 2.5 m0.10 A -2

10

20

30

40

50

60

0 -4

a) 0.1 0.2 0.3value vs. transform order 1 Kurtosis 0.4 0.5 0.6 0.7 0.8 0.9 b) Indicators evolutions vs. generations 100 200 300 400 500 600 700 800 900 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Time (s) Frequency (Hz) Figure 5. Kurtosis value vs. transform order and Indicators evolutions vs. generations. Time (s) Time (s)
error-lms
3 2 1 0 -1 -2 -3

0.05 -3 0 0 -4

Generation
1000 1

e d u t i l p m A

e d 0.5 u t i l 0 p m -0.5 A

e d u t i l p m A

0.35

e 0.3 d 2 u 0.25 t 1 i 0.2 l 0 0.15 p m0.1-1 A

0.05-2 0 -3 00

-1 -4 0 0 0.1 0.1 0.2 0.2 0.30.3 0.40.4 0.50.5

0.6 0.6 0.7 0.7

0.8 0.8

0.9 0.9

11

100 0.1

200 0.2

300 0.3

400 0.4

500 0.5

600 0.6

700 0.7

800 0.8

900 0.9

1000 1

Time (s) Time (s)

Frequency (Hz) Time (s)

a) Output signal from propose algorithm b) Output signal of FDAF Figure 6. Waveform of the output signal using different adaptive algorithms. 4. Experiment Validations 4.1. Experiment Set-up To confirm the performance of the proposed algorithm, an experimental study was carried out based on a gearbox transmission system. The gear system under experiment is designed by Gear fault diagnosis group, University of Huddersfield (UK). As shown in Figure 7(a), the test rig comprises a 3-

9th International Conference on Damage Assessment of Structures (DAMAS 2011) IOP Publishing Journal of Physics: Conference Series 305 (2011) 012022 doi:10.1088/1742-6596/305/1/012022

phase electrical induction motor coupled with a DC generator that applies load to the motor, and the motor is connected to a two stage tested helical gearbox, which contains a gear with fault, and the schematic of the gearbox is shown as Figure 7(b). The accelerometer sensors were set on the gearbox to collect the vibration on the vertical and horizontal direction, respectively, and the sampling frequency of the acquisition system is 84kHz. The input motor speed is about 1500rpm, because it varies with the current fluctuation. The experiment will take when load condition of the system is 78% of the full loaded. The numbers of the teeth and the feature frequencies are listed in the table 1. Table 1. The parameters of the gear system
Items No. of teeth Feature frequency (Hz) Mesh frequency (Hz) 1st pinion 58 25 1450 2nd gear 47 30.9 3rd pinion 13 30.9 401.1 4th gear 59 6.8

a). The gearbox transmission system. b). Schematic of the gearbox. Figure 7. Gearbox transmission system test rig. 4.2. Experiment Results The experiment was conducted in two steps. Vibration measurements were first captured for brand new gears and then with a fault which was created on one tooth of the first gear wheels with 58 teeth. The longer the data sets, the time-consuming the program running, so the original signal collected from the sensors are decimated to a lower sampling rate of 28kHz firstly, another reason is that frequency higher than 14kHz carry little information. Figure 8(a) and 8(b) show the collected signals of the vertical vibration in time domain for normal gear pairs and gear pairs with failure teeth respectively. Compared between them it is seen that the amplitude of vibration signal increases when there is a fault and there are some impulses buried in the noised signal. However, the interval period is not very clear and it is hard to diagnose which part of the gearbox is abnormal.

9th Conference on Damage Assessment of -1.5 Journal of Physics: Conference Series 305 (2011) 012022
-2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

e d 1.5 u 1 t 0.5 i 0 l p -0.5 m A -1 International

e d u 0.04 t i 0.03 l p 0.02 m A Structures

0.01 0

0.05

(DAMAS 2011) IOP Publishing doi:10.1088/1742-6596/305/1/012022

2000 4000 6000 8000 10000 12000 14000

Time (s)
2 1.5 1 0.5 0 -0.5 -1 -1.5 -2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.035

Frequency (Hz)
2 0.05

e d u t i l p m A

e d u t i l p m A

e 0.03 d 1.5 u 1 0.025 t 0.5 i 0.02 0 l 0.015 p -0.5 m 0.01 A -1

0.005 -1.5 0 -2 0 0 2000 0.1 4000 0.3 6000 0.2 0.4 8000 0.5

e d u t i l p m A

0.04

0.03

0.02

0.01

0.610000 0.7

12000 0.8

14000 0.9

2000

40

Time (s)

Frequency (Hz) Time (s)

0.035 0.03 0.025 0.02 0.015 0.01 0.005 0

e d u t i l p m A

respectively. Compared to the signals in Figure 8(b), the noise signal is eliminated significantly and p -0.01 -1.5 m the feature signal is much more clearly by the proposed algorithm. The interval for the impulse can be -0.1 A -0.02 -2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 seen is about 0.04s (0.283-0.2436=0.0364s), which is the rotation frequency of the first shaft (25Hz). Time (s) -0.15 -0.03 Based on this the gear is diagnosed to have is a local fault on the first pinion. On the other hand there -0.2 are no 0.2 such clear impulses in the output signal from -0.04 0 FDAF.0.2 0.3experimental case results verified the 0.1 The 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.4 0.6 0.8 1 1.2 1.4 Time of Time (s) the better performance (s) the proposed algorithm in extracting the gear fault feature signal.
-0.05 0.15

e d d 0.15 0.04 u u 1 e 0.03 0.1 Applying the proposed method has found that the d t 0.5transform order is also 1 and the number of t optimal i i u 0.02 0 l 0.05 parents and children are 16 and 4 while the generation is 60. The extracted two signals, shown in l t p p 0.01 -0.5 i Figure 9, are from the proposed algorithm and the frequency domain LMS adaptive filter algorithm, m 0 m 0 l A -1 A

a) Vibration from the normal gear pairs b) Vibration from the failure gear pairs 2 e Figure 8. Collected vibration signal when gearbox under normal and failure condition. out-lms out-dft 1.5

2000

40

e d u t i l p m A

e 0.1d u 0.05 t 0i l -0.05p m -0.1 A

-0.15 -0.2 0

0.04 0.03 0.02 0.01 0 -0.01 -0.02 -0.03 -0.04 0 0.2

X: 0.2426 Y : 0.03143

error-lms X: 0.283
Y : 0.0319

0.04

e d u t i l p m A

0.03 0.02 0.01 0 -0.01 -0.02 -0.03

e d u t i l p m A

2.5

x 10

-3

1.5

0.5

0.1

0.4

0.2

0.3 0.6

0.4

0.8

0.5

0.6 1

0.7 0.8 1.2

0.9 1.4

-0.04 0

0 0 0.1

0.2

2000 0.3

4000 0.4

6000 0.5 0.6

8000 0.7

10000 12000 0.8 0.9 1

14000

Time (s) Time (s)

Frequency (Hz) Time (s)

a) Output from proposed algorithm b) Output from the FDAF-LMS algorithm Figure 9. Waveform of the output signal using different adaptive algorithms. 5. Conclusions To detect the gear fault signal from the heavy noised signal, the adaptive fractional Fourier transform filter (AFFT) is proposed in the paper, which is realized the adaptive noise cancelling algorithm in the fractional transform domain. It also provides the rules for the selection of the optimal fractional transform order and the evolutionary rules for the refreshments of the filter parameters. The results from simulation confirm the algorithms quickly convergence and results from the experiment verify that the proposed algorithms are effective in cancelling the heavy noise and highlighting the gear fault feature signal. Future work: apply the algorithm to the speed variation condition (especially for the speed-up condition) and loose relationship between the reference signal and the noise contained in the primary signal. 6. Acknowledgement This research was supported by the National Natural Science Foundation of China under No.51035008 and by the Visiting Scholar Foundation of the State Key Lab. of Mechanical

9th International Conference on Damage Assessment of Structures (DAMAS 2011) IOP Publishing Journal of Physics: Conference Series 305 (2011) 012022 doi:10.1088/1742-6596/305/1/012022

Transmission in Chongqing University. 7. References [1]Shao Y M and Mechefske C K 2009 Gearbox vibration monitoring using extended Kalman filters and hypothesis tests J. Sound Vib. 325 629-48. [2] Gu F S, Yesilyurt I, Li Y H, Harris G and Ball A 2006 An Investigation of the Effects of Measurement Noise in the Use of Instantaneous Angular Speed for Machine Diagnosis Mech. Syst. Signal Process. 20 1444-60. [3] Jafarizadeha M A, Hassannejadb R, Ettefaghb M M and Chitsaz S 2008 Asynchronous input gear damage diagnosis using time averaging and wavelet filtering Mech. Syst. Signal Process.22 172-201. [4] Pan Y N, Chen J and Lin X2010 Bearing performance degradation assessment based on lifting wavelet packet decomposition and fuzzy c-means Mech. Syst. Signal Process. 24 559-566. [5] Zhang Y X, Randall RB 2009 Rolling element bearing fault diagnosis based on the combination of genetic algorithms and fast kurtogram Mech. Syst. Signal Process. 23 1509-17. [6] Lei Y G, He Z J and Zi Y Y 2009 Application of the EEMD method to rotor fault diagnosis of rotating machinery Mech. Syst. Signal Process.23 1327-38. [7] Saravanan N, Kumar Siddabattuni V N S and Ramachandran K I 2010 Fault diagnosis of spur bevel gear box using artificial neural network (ANN), and proximal support vector machine (PSVM) Appli. Soft Comp. 10 344-60. [8]Shao Y M and Nezu K 2005 Design of mixture denoising for detecting faulty bearing signals J. Sound Vib. 282 899-917. [9] Ozaktas H M, Zaevsky Z, and Kutay M A, 2001 The Fractional Fourier Transform with Application in Optics and Signal Processing, Wiley New York. [10] Kutay M A, Ozaktas H M, Arikam O and Onural L 1997 Optimal filtering in fractional Fourier domains IEEE Trans. Signal Proc. 45 1129-43. [11] Ashok Narayanan V and Prabhu K M M 2003 The fractional Fourier transform: theory, implementation and error analysis Microproc. Microsyst. 27 511-521. [12] Tao R, Deng B and Wang Y 2009 Fractional Fourier Transform and Its Applications, Tsinghua University Press. [13]Abe M, Kawamata M and Higuchi T 1996 Convergence behaviour of evolutionary digital filters on a multiple-peak surface In Proc. IEEE Inter. Symp. Circuits & Syst. 2 185-188. [14] Paulo S R D 2008 Adaptive filtering algorithm and practical implementation. The international series in Engineering and Computer Scienc. Third Edition.