Available online at www.sciencedirect.com Mathematics and Computers in Simulation 79 (2008) 318–338 Bearing condition monitoring based on shock pulse method and improved redundant lifting scheme Li Zhen a,∗ , He Zhengjia b , Zi Yanyang a , Chen Xuefeng a b a School of Mechanical Engineering, Xi’an Jiaotong University, 710049 Xi’an, China State Key Lab for Manufacturing Systems Engineering, Xi’an Jiaotong University, 710049 Xi’an, China Received 17 January 2007; received in revised form 28 December 2007; accepted 28 December 2007 Available online 17 January 2008 Abstract Due to the widespread application of rolling element bearings, it is necessary to effectively monitor their health status. The shock pulse method (SPM) has been widely used as a quantitative method for bearing condition monitoring. However, the shock value indicating the bearing condition may be mistakenly estimated by direct demodulation in the SPM. To overcome this deficiency, a new approach based on improved redundant lifting scheme (IRLS) is proposed. The classical redundant lifting scheme is improved by adding the normalization factors to avoid error propagation of decomposition results, and the IRLS is applied to preprocess the bearing vibration signals. Then the maximum normalized shock value of detail signals in decomposition results is used as a measure of the bearing condition. The effectiveness of the proposed method is demonstrated by applying it to both simulated signals and practical bearing vibration signals under different conditions. The results show that the proposed method is effective for bearing condition monitoring. © 2008 IMACS. Published by Elsevier B.V. All rights reserved. Keywords: Bearing condition monitoring; Redundant lifting scheme; Shock pulse method 1. Introduction Rolling element bearings play a critical role in industrial applications, and unexpected bearing failures may result in significant economic losses. Condition monitoring is an effective way to prevent the bearing failures. A detailed review of vibration and acoustic methods has been introduced by Tandon and Choudhury [25] for bearing condition monitoring. The shock pulse method (SPM) has achieved wide acceptance as a quantitative method for detecting the defects of rolling element bearings. The maximum normalized shock value is used as a measure of bearing condition in the SPM [1]. It gives a single value indicating the bearing condition, without the need for elaborate data interpretation as required in some other methods (such as traditional spectral analysis, cepstrum analysis [30], cyclic statistics [10], empirical mode decomposition (EMD) [3], wavelet transform [15], etc). However, direct demodulation may mistakenly estimate the shock value in the SPM. To make up this deficiency, wavelet transform based on lifting scheme [7] is used to preprocess the vibration signals of rolling bearing. ∗ Corresponding author. Tel.: +86 29 82667963; fax: +86 29 82663689. E-mail address: xjtu lizhen@163.com (L. Zhen). 0378-4754/$32.00 © 2008 IMACS. Published by Elsevier B.V. All rights reserved. doi:10.1016/j.matcom.2007.12.004 L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 319 Fig. 1. The forward transform of the lifting scheme. Wavelet transform is well known for its ability to focus on localized structures in time–frequency domain [11]. Unlike spectral analysis that describes a signal as the sum of sinusoidal functions, wavelet transform decomposes the signal into wavelet coefficients of various scales in time domain. The advantages are twofold [14]: (1) it is more effective to extract transient features and (2) it extracts signal features over the entire spectrum, without requiring a dominant frequency band. Many researchers have investigated the application of wavelet transform to vibration signals analysis for condition monitoring and fault diagnosis of rolling element bearings [2,9,16,19,21,27]. The lifting scheme [22,23] is proposed by Sweldens as a new method of wavelet construction. The main feature of the lifting scheme is that the construction is derived in the time domain. Wavelets with different time–frequency structures are obtained by the design of prediction operator and update operator. Compared with the classical wavelet transform, the lifting scheme requires less computation and memory, and can produce integer-to-integer wavelet transform. It is always perfectly reconstructed no matter how the prediction operator and update operator are designed. The application of lifting scheme for rotating machinery fault diagnosis has been reported in [5,8]. However, the lifting scheme does not have time invariant property, (i.e. using the lifting scheme, the decomposition results of a delayed signal are not the time-shifted version of those of the original signal). This may result in the loss of useful information for bearing condition monitoring. The time invariant property is particularly important in signal processing such as features detection and noise reduction [17]. The redundant lifting scheme (RLS) [4] possesses time invariant property and overcomes the disadvantage of lifting scheme by getting rid of the split step and zero padding of prediction operator and update operator. The approximation and detail signals at all levels are the same length as the original signal in the redundant lifting scheme. However, the RLS may lead to error propagation of decomposition results. In the previous paper, the authors have demonstrated the high potential of the RLS to detect bearing fault [29]. In this paper, we bring this idea and improve the algorithm of RLS. A new method based on SPM and IRLS is proposed to quantitatively indicate the bearing condition. To make up the deficiency of classical RLS, the cause of error propagation using classical redundant lifting scheme is analyzed. The RLS is improved by adding the normalization factors. To overcome the misestimate of the shock value with direct demodulation in the SPM, IRLS is applied to preprocess the bearing vibration signals. And the maximum normalized shock value of detail signals is used as a measure of the bearing condition. We find that the proposed method is more effective for bearing condition monitoring. The structure of the paper is organized as follows. In Section 2, the theories of lifting scheme and redundant lifting scheme are reviewed briefly. The deficiency of redundant lifting scheme is discussed. In Section 3, the algorithm of IRLS is introduced. In Section 4, the SPM, Hilbert transform and the proposed method are explained. Section 5 describes the disadvantage of SPM based on direct demodulation, and demonstrates the performance of the proposed method by simulation experiment. In Section 6, the proposed method is used for bearing condition monitoring. Conclusions are given in Section 7. Fig. 2. The scaling function and wavelet function with N = 6, Ñ = 6. 320 L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 Fig. 3. The detail signal of original signal using the lifting scheme decomposition. Fig. 4. The detail signal of delayed signal using the lifting scheme decomposition. Fig. 5. The forward transform of redundant lifting scheme. Fig. 6. The distortion occurs in decomposition results using redundant lifting scheme. L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 321 Fig. 7. The inverse lifting scheme with d(1) = 0. Fig. 8. The inverse lifting scheme with s(1) = 0. Fig. 9. The forward and inverse transform of IRLS. 2. The lifting scheme and redundant lifting scheme 2.1. The lifting scheme The lifting scheme is a powerful tool to construct biorthogonal wavelets in the spatial domain [22,23]. The implementation of lifting scheme is simple, fast and efficient. The lifting scheme consists in three main steps. Fig. 10. The decomposition results of signals shown in Fig. 6 using IRLS. 322 L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 Fig. 11. The flow chart of the proposed method for bearing condition monitoring. (0) In the split step, the original signal x = (xi )iz is split into the even samples s(0) = (si )i ∈ Z , and the odd samples (0) (0) d = (d i )i ∈ Z , si (0) = x2i , d i (0) = x2i+1 (1) In the prediction step, we apply an operator P on s(0) to predict d(0) . The prediction error d(1) is regarded as the detail signal of x, d (1) = d (0) − P(s(0) ) (2) where P = [p(1), p(2), . . ., p(N)] is the prediction operator, and N is the number of prediction coefficients. In the update step, an update of even samples s(0) is accomplished by using an update operator U to detail signal (1) d and adding the result to s(0) , the update sequence s(1) can be regarded as the approximation signal of x, s(1) = s(0) + U(d (1) ) (3) where U = [u(1), u(2), . . . , u(Ñ)] is the update operator, and Ñ is the number of update coefficients. Let s(1) be the input signal for lifting scheme, the detail and approximation signals at the lower resolution level can be obtained. The inverse lifting scheme can be performed by reversing the prediction and update operators and changing each ‘+’ into ‘−’ and vice versa. Fig. 1 shows the forward transform of the lifting scheme. The prediction operator and update operator can be designed by the interpolation subdivision method introduced in [24]. The scaling function and wavelet function with N = 6, Ñ = 6 are shown in Fig. 2. They are symmetrical and compactly supported. The shape of the wavelet function is very similar to an impulse. Therefore, it is desirable to extract the transient components from vibration signals. L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 323 Fig. 12. The periodic impulse signal with a period of 0.0167 s and its Hilbert spectrum. 2.2. The redundant lifting scheme The time invariance [11] of an operator L means that if the input f(t) is delayed by τ, fτ (t) = f(t − τ), then the output is also delayed by τ: g(t) = Lf (t) ⇒ g(t − τ) = Lfτ (t) (4) In the lifting scheme, the time invariant property is not ensured because there exists the split step and the length of approximation signal and detail signal decreases. Suppose the original signal is x = {. . . , x2j−4 , x2j−3 , x2j−2 , x2j−1 , x2j , x2j+1 , x2j+2 , x2j+3 , x2j+4 , . . .} The prediction operator is P = {p1 , p2 }, and the update operator is U = {u1 , u2 }. The decomposition procedure of original signal is given in Fig. 3. The detail signal d of original signal can be calculated by d[j] = x2j+1 − (p1 x2j + p2 x2j+2 ) (5) when delaying the original signal by one time step, the detail signal d of the delayed signal can be computed by d[j] = x2j − (p1 x2j−1 + p2 x2j+1 ). Fig. 13. The periodic impulse signal with a period of 0.0125 s and its Hilbert spectrum. (6) 324 L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 Fig. 14. The simulated signal with two periodic impulse signals and its Hilbert spectrum. The decomposition procedure of the delayed signal is shown in Fig. 4. Obviously, the detail signal d of delayed signal is not equal to the detail signal d of original signal, so the time invariant property is not ensured in the lifting scheme. In the redundant lifting scheme, the split step is discarded. The redundant prediction operator P(l) and the redundant update operator U(l) are computed by padding the prediction operator P and update operator U with zeros at the corresponding level l. The redundant lifting scheme possesses time invariant property and well keeps the signal information. The forward transform of redundant lifting scheme is shown in Fig. 5. The decomposition results of an approximation signal ṡ(l) at level l with redundant lifting scheme are expressed by following equations, ḋ (l+1) = ṡ(l) − P (l) ṡ(l) ṡ(l+1) = ṡ(l) + U (l) ḋ where ḋ (l+1) (7) (l+1) (8) and ṡ(l+1) are detail signal and approximation signal at level l + 1. Fig. 15. The decomposition results of simulation signal using IRLS. L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 325 Fig. 16. The Hilbert spectrums of detail signals r(1) and r(2) . 2.3. The deficiency of redundant lifting scheme Since the redundant lifting scheme does not control the error propagation, the decomposition results using redundant lifting scheme always bring distortion. The prediction operator and update operator with N = 6, Ñ = 6 are used to demonstrate the deficiency of redundant lifting scheme. The inspected signal is a practical vibration signal of rolling bearing with heavy outer-race defects. The waveform contains periodic impulses. The range of original signal is [−0.8, 0.8]. The decomposition results are shown in Fig. 6. From Fig. 6, the range of the detail sig(1) nal (ḋ ) is [−1.5, 1.5], which far exceeds the range of original signal. Moreover, the energy of the signal is given by the classical Euclidian norm. The total energy of the original signal is 98.26, but the total energy of the decomposition results is 360.64. It shows that the distortion occurs in decomposition results using redundant lifting scheme. Fig. 17. The decomposition results of simulation signal using redundant lifting scheme. 326 L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 Fig. 18. The Hilbert spectrums of detail signals ḋ (1) (2) and ḋ . 3. The improved redundant lifting scheme (IRLS) 3.1. The cause of the error propagation in redundant lifting scheme The error propagation can be controlled by inverse transform in the lifting scheme [7,20]. We analyze the cause of the error propagation in the redundant lifting scheme by comparing the decomposition results of redundant lifting scheme with the reconstruction results of lifting scheme. The signal is performed 1-level decomposition by using lifting scheme then reconstructed by oneself. The total energy of reconstruction results is approximately equal to that of original signal. From the forward transform of lifting scheme, the detail signal d(1) and the approximation signal s(1) can be expressed as follows: (1) d i = x2i+1 − N pr x2i−N+2r r=1 Fig. 19. The decomposition results of the simulation signal using the Jiang’s method. (9) L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 Fig. 20. The Hilbert spectrums of detail signals d̂ (1) si = x2i + Ñ (1) uj d i−Ñ/2+j−1 (1) 327 (2) and d̂ . (10) j=1 when d(l) = 0, the inverse transform of the lifting scheme is shown in Fig. 7. The reconstructed signal ŝ(1) of approx(1) (1) (1) imation signal s(1) is calculated. The even samples ŝ(1) e = {ŝ2i } and the odd samples ŝo = {ŝ2i+1 } are given by the following equations: (1) (1) ŝ2i = si (1) ŝ2i+1 = N (11) (1) pr si+r−N/2 (12) r=1 Fig. 21. The vibration signals of bearing under different conditions: (a) healthy bearing, (b) weak outer-race defect and (c) heavy outer-race defect. 328 L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 Fig. 22. The decomposition results using IRLS of the vibration signal of healthy bearing. Fig. 23. The Hilbert spectrum of detail signal r(1) in Fig. 22. Fig. 24. The decomposition results using IRLS of the vibration signal with weak outer-race defect. L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 329 Fig. 25. The Hilbert spectrum of detail signal r(1) in Fig. 24. Fig. 26. The decomposition results using IRLS of the vibration signal with heavy outer-race defect. (1) when s(1) = 0, the inverse transform of the lifting scheme is shown in Fig. 8. The even samples d̂ e = (1) (1) (1) {d̂ 2i } and the odd samples d̂ o = {d̂ 2i+1 } of the reconstructed signal d̂ equations: (1) Ñ (1) (1) d̂ 2i = − uj d i−Ñ/2+j−1 are expressed by the following (13) j=1 Fig. 27. The Hilbert spectrum of detail signal r(1) in Fig. 26. 330 L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 Fig. 28. Maximum normalized values of bearing under different conditions using the proposed method. (1) (1) d̂ 2i+1 = d̂ i + N (1) pr d̂ 2i−N+2r (14) r=1 (1) (1) From the forward transform of redundant lifting scheme, the even samples ḋ e = {ḋ 2i } and the odd samples (1) (1) = {ḋ 2i+1 } of the detail signal ḋ are described as, (1) ḋ o (1) ḋ 2i = x2i − N (15) pr x2i−N+2r−1 r=1 (1) ḋ 2i+1 = x2i+1 − N pr x2i−N+2r (16) r=1 (1) (1) (1) (1) The even samples ṡe = {ṡ2i } and the odd samples ṡo = {ṡ2i+1 } of the approximation signal ṡ(1) are expressed as follows: (1) ṡ2i = x2i + Ñ (1) uj ḋ 2i−Ñ+2j−1 (17) j=1 (1) ṡ2i+1 = x2i+1 + Ñ (1) (18) uj ḋ 2i−Ñ+2j j=1 From Eqs. (9)–(11), we obtain Ñ N (1) ŝ2i = x2i + uj x2i−Ñ+2j−1 − pr x2i−Ñ+2j−2−N+2r j=1 (19) r=1 Considering Eqs. (16) and (17), we have (1) ṡ2i = x2i + Ñ j=1 uj (x2i−Ñ+2j−1 − N pr x2i−Ñ+2j−2−N+2r ) (20) r=1 (1) (1) From Eqs. (19) and (20), we can see that the even samples ṡe = {ṡ2i } of the approximation signal ṡ(1) are equal (1) (1) (1) (1) to the even samples ŝ(1) e = {ŝ2i } of reconstructed signal ŝ . Moreover, compared with the odd samples ŝo = {ŝ2i+1 } L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 (1) 331 (1) of reconstructed signal ŝ(1) , the odd samples ṡo = {ṡ2i+1 } of the approximation signal ṡ(1) is calculated by different interpolation method. The approximation signal ṡ(1) cannot bring the error propagation. Therefore, the error propagation (1) of decomposition results using redundant lifting scheme is caused by the distortion of the detail signal ḋ . 3.2. The algorithm of the IRLS The IRLS is proposed to control the error propagation by adding the normalization factors. The forward and inverse transforms of IRLS are shown in Fig. 9. The al+1 and bl+1 are the normalization factors at level l + 1. The decomposition results of the approximation signal ṡ(l) at level l with IRLS are expressed as follows: (l+1) = ṡ(l) − P (l) ṡ(l) (l+1) = ṡ ḋ ṡ r (l+1) (21) (l) (l+1) + U ḋ (l) = bl+1 ḋ (22) (l+1) (23) c(l+1) = al+1 ṡ(l+1) (24) (l+1) and ṡ(l+1) are computed by the classical redundant lifting scheme. r (l+1) and c(l+1) are the detail signal where ḋ and approximation signal at level l + 1 using IRLS. Suppose the total energy of the approximation signal ṡ(l) is described as Esl = n (l) 2 [ṡi ] . (25) i=1 The total energies of ṡ(l+1) and ḋ Esl+1 = n (l+1) are given by Esl+1 and Edl+1 , respectively (l+1) 2 [ṡi ] , (26) i=1 Edl+1 = n (l+1) 2 [ḋ i ] (27) i=1 The total energies of c(l+1) and r (l+1) are expressed by Ecl+1 and Erl+1 , respectively Ecl+1 = n (l+1) 2 [ci ] (28) i=1 Erl+1 = n (l+1) 2 [ri ] (29) i=1 If the error propagation is controlled, the total energy of original signal should be approximately equal to that of decomposition results, i.e. Esl = Ecl+1 + Erl+1 . Furthermore, the approximation signal ṡ(l+1) cannot bring the error propagation in decomposition results using redundant lifting scheme. The normalization factors are calculated by the following equations: al+1 = 1 bl+1 = (30) Esl − Esl+1 Edl+1 (31) The inspected signal shown in Fig. 6 is performed 1-level decomposition by using IRLS. The results are given in Fig. 10. Compared with Fig. 6, the total energy of detail signal r(1) and approximation signal c(1) is 98.26, which is equal to that of the original signal. The range of the detail signal varies from [−1.5, 1.5] to [−0.8, 0.8], and the distortion of the decomposition results using redundant lifting scheme is eliminated by the normalization factors. 332 L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 4. Hilbert transform and SPM and the proposed method 4.1. Hilbert transform As is well known, for a signal x(t), the Hilbert transform is defined by 1 ∞ x(t) H[x(t)] = dτ π −∞ t − τ Demodulation is accomplished by forming a complex-valued signal A[x(t)], which is defined as √ A[x(t)] = x(t) + iH[x(t)] = a(t) eiφ(t) , i = −1 (32) (33) The complex time-domain signal can be converted from the real/imaginary format to the magnitude/phase format, a(t) = x2 (t) + H 2 [x(t)], (34) φ(t) = arctan H[x(t)] x(t) (35) The envelope represents an estimate of the modulation phenomenon in the signal. In addition, it removes carrier signals and decreases the influence of irrelevant information for bearing fault detection. 4.2. The shock pulse method (SPM) The shock pulse method (SPM) [31] has been developed by SPM Instrument AB in the early 1970s in Sweden. It relies on the fact that when a ball or roller contacts with a damaged area of raceway or with debris in the bearing, an impulse of vibration is generated. The maximum normalized shock value (dB) gives an indication of bearing condition. It is defined as, dB = 20 log 2000 SV ND0.6 (36) where N denotes rotating speed of the bearing, and D is the inner diameter. SV is the shock value, which should be calculated by demodulation of bearing vibration signals. With dB value, the bearing running state can be estimated as [31]: ⎫ (1) 0 ≤ dB < 20 healthy bearing ⎪ ⎬ (2) 20 ≤ dB < 35 weak fault (37) ⎪ ⎭ (1) 35 ≤ dB < 60 heavy fault The bounds of three condition zones were empirically established by testing bearings under variable operating condition in the early 1970s. For over 35 years, based on three condition zones, the shock pulse method (SPM) has been very successfully used to obtain a reliable diagnosis of the operating condition of rolling element bearings [1,12,13,18,26]. Additionally, the bounds have been applied for the design of shock pulse meters, which have achieved wide acceptance in industries. 4.3. The proposed method The SPM combined with demodulation method is a useful method of estimating bearing running state. The maximum normalized shock value (dB) is used as a measure of bearing condition. However, the direct demodulation may mistakenly estimate the shock value indicating the bearing condition. The IRLS is considered to preprocess the vibration signal of rolling bearing before demodulation. Fig. 11 shows the flow chart of the proposed method for bearing condition monitoring. L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 333 Table 1 The geometric parameters of the tested bearing Ball diameter d (mm) Pitch diameter E (mm) Contact angle α (◦ ) Number of rolling elements z 68 450 0 17 5. The analysis of simulation experiment To verify the feasibility of the proposed method for bearing condition monitoring, we first use a simulated signal. The simulated signal is composed of two periodic impulse signals. One is described as 0.1 e−2π×16t sin(2π × 2000t) with a period of 0.0167 s, and the other is expressed as 0.1 e−2π×25t sin(2π × 5000t) with a period of 0.0125 s. The periodic impulse signal with a period of 0.0167 s and its Hilbert spectrum are shown in Fig. 12. The peak value of 60 Hz (1/0.0167 s) is 0.0147. The other with a period of 0.0125 s and its Hilbert spectrum are shown in Fig. 13. The peak value of 80 Hz (1/0.0125 s) is 0.0105. The simulated signal is shown in Fig. 14a. It is directly demodulated by the Hilbert transform. The Hilbert spectrum is shown in Fig. 14b. Theoretically, the demodulation spectrum should have two peaks at 60 Hz and 80 Hz, and their values are equal to 0.0147 and 0.0105. But from Fig. 14b, the peak values of 60 Hz and 80 Hz are 0.0096 and 0.0064, respectively. They are all less than real values. It indicates that the direct demodulation may mistakenly estimate useful information. To extract accurately the useful information from the simulated signal, the simulated signal is performed 4-level decomposition by using IRLS. The results are shown in Fig. 15. The two periodic impulse signals are at the different levels. The Hilbert spectrums of detail signals r(1) and r(2) are shown in Fig. 16. The peak value of 60 Hz (1/0.0167 s) is 0.0145 in Fig. 16a, and the peak value of 80 Hz (1/0.0125 s) is 0.0102 in Fig. 16b. The relative errors of the peak values at 60 Hz and 80 Hz with their real values are 1.361% and 2.857%, respectively. They are approximately equal to the real values. Therefore, the IRLS is very effective as a preprocessor for feature extraction. To demonstrate the effectiveness of the IRLS, the same signal is analyzed by using redundant lifting scheme for comparison. Fig. 17 shows decomposition results using redundant lifting scheme. Since the error propagation is not (1) (2) controlled, it brings the distortion of detail signals. The Hilbert spectrums of detail signals ḋ and ḋ are shown in Fig. 18. The peak value of 60 Hz (1/0.0167 s) is 0.0289 in Fig. 18a, and the peak value of 80 Hz (1/0.0125 s) is 0.0191 in Fig. 18b. They far exceed the real values. The wavelet transform based on lifting scheme, which was introduced in [7], is also used to preprocess the simulated signal. The decomposition results are shown in Fig. 19. The peak value of 60 Hz is 0.0132 in Fig. 20a, and the peak value of 80 Hz is 0.0094 in Fig. 20b. Compared with direct demodulation, the Jiang’s method is effective. However, the relative errors of the peak values at 60 Hz and 80 Hz with their real values are 10.20% and 10.38%, respectively. They are more than the relative errors of decomposition results using IRLS. 6. The analysis of practical bearing vibration signals In the present work, the propose method is used to evaluate the bearings under different conditions. The geometric parameters of the roller bearing are listed in Table 1. The vibration signals are picked up at a constant speed of 390 rpm. Based on the geometric parameters and the rotating speed of the bearing, the characteristic frequency of the outer-race defects can be calculated by the following formula, fo = 1 f 2 1− d cos α z E (38) Using the above formula, the characteristic frequency of outer-race defect is calculated to be at 46.88 Hz. Fig. 21 shows the bearing vibration signals under the condition of health, weak outer-race defect and heavy outerrace defect. Although the evident impulses occur in the vibration signal with heavy outer-race defect, it is hardly possible to evaluate the bearing condition only through such signals. 334 L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 Fig. 29. Maximum normalized values of bearing under different conditions using redundant lifting scheme decomposition. The vibration signals are performed 4-level decomposition by using IRLS. The decomposition results of the vibration signal of healthy bearing are shown in Fig. 22. The amplitude of characteristic frequency is very small in the Hilbert spectrum shown in Fig. 23. The decomposition results of the vibration signal with weak outer-race defect are shown in Fig. 24. The detail signal r(1) contains the predominant impulses, and its Hilbert spectrum is shown in Fig. 25. The characteristic frequency of 46.88 Hz is exacted clearly. Fig. 26 illustrates the decomposition results of the vibration signal with heavy outer-race defect. The Hilbert spectrum of detail signal at the first level is shown in Fig. 27. From Fig. 27, the amplitude of the characteristic frequency (46.88 Hz) is far more than that in Fig. 25. Based on the IRLS decomposition and Hilbert spectrum, the maximum normalized shock values (dB) of bearings under different conditions are given in Fig. 28. The dB value of healthy bearing is 16.5, indicating good bearing condition. The dB value of bearing with weak outer-race defect is 31.6, indicating caution zone. The maximum normalized value of bearing with heavy outer-race defect is 49.8, indicating the damaged bearing condition. For comparison, the redundant lifting scheme is used to preprocess the same vibration signals. The maximum normalized values of bearings are shown in Fig. 29. From Fig. 29, the dB value of bearing with weak outer-race Fig. 30. The EMD decomposed results of the vibration signal of healthy bearing. L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 335 Fig. 31. The Hilbert spectrum of IMF1 in Fig. 30. Fig. 32. The EMD decomposed results of the vibration signal with weak outer-race defect. defect is 37.3 greater than 35. It indicates that the bearing condition with weak defect is mistakenly estimated as heavy defect. The Hilbert–Huang transform is also applied to analyze the bearing vibration signals under different conditions. The vibration signals are decomposed into some intrinsic mode functions (IMFs) by empirical mode decomposition (EMD) [6]. The decomposition results using EMD are given in Figs. 30–34, respectively. Fig. 31 shows the Hilbert spectrum of IMF1 in Fig. 30. Fig. 33 shows the Hilbert spectrum of IMF1 in the EMD decomposed results of the vibration signal with weak outer-race defect. Fig. 33. The Hilbert spectrum of IMF1 in Fig. 32. 336 L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 Fig. 34. The EMD decomposed results of the vibration signal with heavy outer-race defect. Fig. 35. The Hilbert spectrum of IMF1 in Fig. 34. Fig. 36. The decomposed results using STFT of the vibration signal with weak outer-race defect. L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 337 Fig. 37. The Hilbert spectrum at the frequency range (3200 Hz and 4000 Hz) in Fig. 36. For the heavy outer-race defect, the IMF1 includes the most dominant fault information. The Hilbert spectrum of IMF1 shown in Fig. 35 can reveal clearly the characteristic frequency of outer-race defect. Compared with the results shown in Fig. 25, the fault components caused by weak outer-race defect are difficult to identify using visual inspection in Figs. 32 and 33. The short time Fourier transform (STFT) [28] and envelope detector are used to analyze the signal with weak outer-race defect. The results are given in Figs. 36 and 37. The fault features are not extracted effectively from Figs. 36 and 37. Therefore, the proposed method is more effective to evaluate the bearing condition. 7. Conclusion The method of combining IRLS and SPM has been proposed for bearing condition monitoring. Based on the analysis of the cause of decomposition distortion using the redundant lifting scheme, the classical redundant lifting scheme has been improved by adding the normalization factors. With the IRLS and SPM, the bearing condition can be accurately estimated. The proposed method is verified by simulated signals and bearing vibration signals. The results demonstrate that the proposed method is more effective for bearing condition monitoring. Acknowledgements This work was supported by the key project of National Natural Science Foundation of China (No. 50335030), National High-tech R&D Program of China (863 Program) (2006AA04Z430), National Basic Research Program of China (No. 2005CB724100). References [1] D.E. Butler, The shock pulse method for the detection of damaged rolling bearing, NDT Int. 6 (1973) 92–95. [2] J. Cheng, D. Yu, Y. Yang, Application of an impulse response wavelet to fault diagnosis of rolling bearing, Mech. Syst. Signal Process. 21 (2007) 920–929. [3] J. Cheng, D. Yu, Y. Yang, A fault diagnosis approach for roller bearings based on EMD method and AR model, Mech. Syst. Signal Process. 20 (2006) 350–365. [4] R.L. Claypoole, Adaptive wavelet transform via lifting, Ph.D. Thesis, Deportment of electrical and computer engineering, Rice University, Houston, Texas, 1999. [5] C. Duan, Z. He, H. Jiang, A sliding window feature extraction method for rotating machinery based on the lifting scheme, J. Sound Vib. 299 (2007) 774–785. [6] N.E. Huang, Z. Shen, S.R. Long, et al., The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis, Proc. Roy. Soc. Lond. 454 (1998) 903–995. [7] H. Jiang, Research on second generation wavelet construction theory and its applications in fault features extraction, Ph.D. Thesis, Xi’an Jiaotong university, China, 2006. [8] H. Jiang, Z. He, C. Duan, P. Chen, Gearbox fault diagnosis using adaptive redundant lifting scheme, Mech. Syst. Signal Process. 20 (2006) 1992–2006. [9] C.J. Li, J. Ma, Wavelet decomposition of vibrations for detection of bearing-localized defect, NDT E Int. 30 (1997) 143–149. [10] L. Li, L. Qu, Cyclic statistics in rolling bearing diagnosis, J. Sound Vib. 267 (2003) 253–265. [11] S.G. Mallat, A Wavelet Tour of Signal Processing, Academic, San Diego, 1998. [12] J. Mathew, R.J. Alfredson, The condition monitoring of rolling element bearing using vibration signal, Trans. ASME, J. Vibr. Acoust. Stress Reliab. Des. 106 (1984) 145–154. [13] L.E. Morando, Measuring shock pulses is ideal for bearing condition monitoring, Pulp Paper 62 (1988) 96–98. 338 L. Zhen et al. / Mathematics and Computers in Simulation 79 (2008) 318–338 [14] D.E. Newland, Wavelet analysis of vibration, Part 1: Theory, ASME Trans., J. Vib. Acoustics 116 (1994) 409–416. [15] N.G. Nikolaou, I.A. Antoniadis, Rolling element bearing fault diagnosis using wavelet packet, NDT E Int. 34 (2002) 197–205. [16] Z.K. Peng, F.L. Chu, Application of the wavelet transform in machine condition monitoring and fault diagnosis: a review with bibliography, Mech. Syst. Signal Process. 18 (2004) 199–221. [17] J.C. Pesquet, H. Krim, H. Carfantan, Time-invariant orthonormal wavelet representations, IEEE Trans. Signal Process. 44 (1996) 1964–1970. [18] R. Prabhu, Rolling element diagnostics, in: Proceeding of the Indo-US Symposium on Emerging Trends in Vibration and Noise Engineering, New Delhi, 1996, pp. 311–320. [19] Y. Shao, K. Nezu, Design of mixture de-noising for detecting faulty bearing signals, J. Sound Vib. 282 (2005) 899–917. [20] T. Sliwa, Y. Voisin, A. Diou, Adaptivity with near orthogonality constraint for high compression rates in lifting scheme framework, in: Proceeding of SPIE: Mathematics of Data/Image Coding, Compression, and Encryption VI, with Applications, Vol. 5208, 2004, pp. 107–115. [21] Q. Sun, T. Yang, Singularity analysis using continuous wavelet transform for bearing fault diagnosis, Mech. Syst. Signal Process. 16 (2002) 1025–1041. [22] W. Sweldens, The lifting scheme: a construction of second generation wavelets, SIAM J. Math. Anal. 29 (1997) 511–546. [23] W. Sweldens, The lifting scheme: a custom-design construction of biorthogonal wavelets, Appl. Comp. Harmonic Anal. 3 (1996) 186–200. [24] W. Sweldens, P. SchrÖder, Building your own wavelets at home, Wavelets in Computer Graphics, ACM SIGGRAPH Course Notes, (1996) 15–87. [25] N. Tandon, A. Choudhury, A review of the vibration and acoustic measurement methods for detection of defects in rolling element bearing, Tribol. Int. 32 (1999) 469–480. [26] N. Tandon, G.S. Yadava, K.M. Ramakrishna, A comparison of some condition monitoring techniques for the detection of defect in induction motor ball bearings, Mech. Syst. Signal Process. 21 (2007) 244–256. [27] P. Tse, Y.H. Peng, R. Yam, Wavelet analysis and envelop detection for rolling element bearing fault diagnosis—their effectiveness and flexibility, ASME Trans., J. Vib. Acoustics 123 (2001) 303–310. [28] X.D. Zhang, Modern Signal Processing, Tsinghua University Press, Beijing, 2002. [29] L. Zhen, Z. He, Y. Zi, Y. Wang, Customized wavelet denoising using intra- and inter-scale dependency for bearing fault detection, J. Sound Vib. 313 (2008) 342–359. [30] G.T. Zheng, W.J. Wang, A new cesptral analysis procedure of recovering excitations for transient components of vibration signals and applications to rotating machinery condition monitoring, J. Vib. Acoustics 123 (2001) 222–229. [31] The shock pulse method for determining condition of anti-friction bearings, SPM Technical Information, Sweden, SPM Instruments AB.