Research Article  Open Access
Junbo Long, Haibin Wang, Peng Li, "Applications of Fractional Lower Order Frequency Spectrum Technologies to Bearing Fault Analysis", Mathematical Problems in Engineering, vol. 2019, Article ID 7641383, 24 pages, 2019. https://doi.org/10.1155/2019/7641383
Applications of Fractional Lower Order Frequency Spectrum Technologies to Bearing Fault Analysis
Abstract
The traditional spectral analysis method is used to study the characteristics of bearing fault signals in frequency domain, which is reasonable and effective in general cases. However, it is proved that the fault signals have heavy tails in this paper, which are stable distribution, and , and even the noises belong to stable distribution. Then the conventional spectral analysis methods degenerate and even fail under stable distribution environment. Several improved frequency spectral analysis methods are proposed employing fractional lower order covariation or fractional lower order covariance in this paper, including fractional lower order BlackmanTukey covariation spectrum (FLOBTCS), fractional lower order periodogram covariation spectrum (FLOPCS), and fractional lower order welch covariation spectrum (FLOWCS). In order to suppress side lobe and improve resolution, we present novel fractional lower order autoregression (FLOAR) and fractional lower order autoregressive moving average (FLOARMA) parameter model frequency spectrum methods, and the calculation steps are summarized. The proposed spectrum methods are compared with the existing methods based on secondorder statistics under Gaussian and distribution environments, and the results show that the new algorithms have better performance than the traditional methods. Finally, the improved methods are applied to estimate frequency spectrums of the normal and outer race fault signals, and it is demonstrated that they are effective for fault diagnosis.
1. Introduction
The mechanical fault vibration signals are a nonstationary and nonGaussian process; their spectrum characteristic change with the time and their probability density function (PDF) have heavy tails. Hence, it is necessary to study their parameter characteristics and new methods suitable for the signals. Bearing fault diagnosis is a technology that can monitor the running state of mechanical equipment in real time. The frequency spectrum analysis is an important means for fault feature extraction in engineering application. The traditional spectrum estimation methods mainly include BlackmanTukey (BT), Periodogram, Welch, multiple signal classification, wavelet spectrum, and parameter model spectrum estimation. The spectral analysis methods have been widely applied in radar signal processing [1], passive sonar signal processing [2], target recognition [3], seismic analysis [4], mechanical failure analysis, and other fields [5–7]. Recently, some improved methods were also applied to machine breakdown analysis, such that singular spectrum analysis was used for roller element bearing fault diagnosis [8, 9]. A new wavelet spectrum analysis was applied to incipient bearing fault diagnostics [10]. The frequency spectrum analysis method employing spectrum sparseness was proposed in [11], which was applied to the rolling bearing performance degradation assessment. Combining the segmentation point obtained from pretreatment and the spectrum analysis, SU WeiJun et al. proposed a new rolling bearing fault feature extraction method employing the local spectrum [12]. A new detection method of the change in amplitude was presented based on principal component analysis, which was used to detect the rolling element bearing fault [13]. Pandarakone and the others proposed a distinct motor bearing analysis method employing support vector machine for frequency spectrum determination [14], although the existing frequency spectrum methods have been well applied, which cannot work in peaked noise environment.
The parameter model spectrum estimation method is a modern technology with practical significance, which mainly includes AR (autoregression), MA (moving average), and ARMA (autoregressive moving average) model methods. Recently, the methods are developed rapidly. A new method employing empirical mode decomposition (EMD) and AR model was proposed for fault feature extraction in [15]; firstly, EMD was applied to decompose the fault signals, and then each intrinsic mode function component was analyzed for feature extraction. The spectral kurtosis method based on AR model spectrum was applied to fault diagnosis and condition monitoring for rolling bearings [16]. A novel method based on fuzzy cluster analysis and AR model was presented for the bearing fault diagnosis; firstly, the parameters of the AR model were estimated based on the higherorder statistics, and then fuzzy cluster analysis was used for classification and pattern recognition [17]. The bispectrum analysis method based on ARMA model and fuzzy cmeans was proposed in [18], which was applied to fault identification, and an algorithm employing AR neural networks and ARMA model has been applied to fault diagnosis in [19]. Although the traditional parameter model methods have been applied in many fields, which still have some defects, the methods degrade in performance under impulsive signal or noise environment and even fail. Therefore, it is of great significance to explore high performance and efficient ARMA model spectrum estimation algorithm.
The conventional frequency spectrum estimation methods mentioned above are suitable for rotating machine condition monitoring and performance evaluation in general. However, probability density functions (PDFs) of the bearing fault signals have heavy trails in some special cases, even the same with the noise in the signals, which belong to stable distribution [20–25]. Because stable distribution has no finite second moment, the existing methods based on Gaussian hypothesis and secondorder statistics degenerate under stable distribution environment. We cannot reuse the frequency domain analysis and its extended analysis employing the subsistent power spectrum methods. Therefore, new spectral estimation methods applicable to stable distribution are needed to seek. Recently, it is verified that the bearing fault signals belong to stable distribution [26–28]. Furthermore, the method combining stable distribution and support vector machine was proposed in [29], which was used in fault diagnosis of the gearbox. A new empirical mode decomposition method based on the parameters of stable distribution was presented for fault diagnosis of the lowspeed rolling bearings, which has better diagnosis accuracy and operational efficiency [30]. A short time Fourier transform local maxima (STFTLM) spectrum based on FLOC was proposed for local damage detection [31]; the method combines the local maxima enhancement with dependency analysis in the frequency bands, which do not require second moment to be finite and can work in case when the sample distribution processes heavy tails.
Aiming at the degradation of the traditional frequency spectrum methods under stable distribution environment, the improved fractional lower order BlackmanTukey covariation, fractional lower order Periodogram covariation, and fractional lower order Welch covariation methods are proposed based on fractional lower order covariation or fractional lower order covariance in this paper. The results show that the proposed frequency spectrum methods can better work in Gaussian noise and stable distribution noise environment, which are robust, and their performance is better than the existing frequency spectrum methods based on secondorder statistics. In order to suppress the side lobe and improve the resolution, fractional lower order autoregression (FLOAR) and fractional lower order autoregressive moving average (FLOARMA) frequency spectrum methods are presented, which can effectively estimate out frequency spectrum of the signal from stable distribution noise. Their mixed mean square error is significantly lower than that of the existing AR and ARMA methods under different characteristics index and generalized signal noise ratio (); especially, when or , the performance advantage of the FLOAR and FLOARMA methods is more obvious. Finally, the improved methods are applied to analyze the outer race fault signal; the results show that the methods can better estimate out transient harmonic vibration components of the bearing fault signals under stable noise environment, but the conventional methods degenerate.
In this paper, several improved frequency spectrum estimation methods employing fractional lower order covariation or covariance are proposed for the bearing fault diagnosis under Gaussian or stable distribution environment. This paper is structured in the following manner. stable distribution and the bearing fault signals are introduced in Section 2. The improved fractional lower order frequency spectrum estimation methods are demonstrated, and simulation comparisons with the conventional spectrum analysis algorithms based on secondorder statistics are performed to demonstrate justifiability of the improved methods in Section 3. The conjoint analysis of the outer race fault signals employing the proposed frequency spectrum and timefrequency representation methods in [28] are demonstrated in Section 4. Finally, conclusions and future research are given in Section 5.
2. Stable Distribution and Bearing Fault Signals
2.1. Stable Distribution
Probability density function (PDF) of stable distribution has heavy tail, and its characteristic function is defined aswhere . stable distribution is a generalized Gaussian process, and is its characteristic index, when , which is Gaussian distribution, and when , which is lower order stable distribution. is location parameter, and and are symmetry and dispersion coefficient, respectively. When , and , which is called standard symmetric stable distribution (). The waveforms of stable distribution are shown in Figure 1 under = 0.5, 1.0, 1.5, and 2.0.
Figure 2 is PDFs of stable distribution under , 1.0, 1.5, and 2.0, respectively. From Figure 2(a), we see that when , the value of under (Gaussian distribution) is the smallest, and the value of under is the largest, which have pulse characteristics. Figure 2(b) shows that the value of under is equal to 0 when is approximately 4.75, but the value of is not equal to 0 under , which have tails. Hence, we can know that the smaller the characteristic index of stable distribution is, the thicker its tail is, and the more significant its pulse characteristic is. On the contrary, when the value of increases, the tail becomes thinner and the pulse characteristic decreases.
(a)
(b)
2.2. Bearing Fault Signals
The actual bearing fault signals data are obtained from the Case Western Reserve University (CWRU) bearing data center [32]. The test stand consists of a 2 hp motor (left), a torque transducer/encoder (center), a dynamometer (right), and control electronics (not shown). The test bearings support the motor shaft. Experiments are conducted using a 2 hp Reliance Electric motor, and acceleration data is measured at locations near to and remote from the motor bearings. The experimental equipment adopts 62052RS JEM SKF type bearing, the motor bearings are seeded with faults using electrodischarge machining, and the vibration data is collected using accelerometers, which are attached to the housing with magnetic bases. The outer race diameter is 20.472 inches, and the inner race and the ball diameter are 0.9843 inches and 0.3126 inches, respectively. The bearing outer race thickness is 0.5906 inches, motor load is 0 HP, and motor speed is 1797 rpm. The bearing faults of inner race, ball, and out race are set, and the fault diameters are all 0.021 inches. The fault data are collected at 12,000 samples per second, and the outer race position relative to load zone centered at 6:00. The vibration signals are collected using a 16 channel DAT recorder, and the speed and horsepower data are collected using the torque transducer/encoder and were recorded by hand. The normal signals are given in Figure 3(a), and the fault signals of inner race, ball, and out race are shown in Figures 3(b), 3(c), and 3(d), respectively. We can know that the waveform of the fault signals has a certain impulse.
(a)
(b)
(c)
(d)
In order to further verify the pulse characteristics of the bearing fault signals, we have applied stable distribution statistical model to estimate the parameters of the normal signal and the fault signals of inner race, ball, and outer race, respectively, the results are given in Table 1. As it can be seen, the characteristic index of the normal signals is equal to 2, which are Gaussian distribution. However, the characteristic index of the bearing fault signals is greater than 1 but smaller than 2, which belongs to nonGaussian stable distribution ().

PDFs of the signals of inner race fault, ball fault, and outer race fault are shown in Figures 4(a), 4(b), and 4(c), respectively. From the PDFs of normal and fault signals, we can see that PDFs of fault signals have heavy tails. Most of the parameters are approximately equal to zero in Table 1, and Figure 4 shows PDFs of the fault signals are near symmetric. Hence, distribution is a more concise and accurate statistical model for the bearing fault signals.
(a)
(b)
(c)
3. Fractional Lower Order Frequency Spectrum Estimation Methods
3.1. Fractional Lower Order Covariation Spectrum Estimation Method
3.1.1. Covariation Spectrum Principle
The autocorrelation function and power spectrum of a secondorder random process can be expressed as
From (2)(4), we can see that power spectrum of is Fourier transform of its correlation function, which is called WienerKhinchin theorem. demonstrates secondorder statistical properties of in time domain, and is the statistical properties of in frequency domain. The corresponding discrete correlation function and power spectrum are given by
When is stable distribution process, the autocorrelation function and power spectrum based on secondorder statistics are no longer applicable. Hence, fractional lower order covariation was proposed in [21], which was obtained by calculating fractional low order moments.where is a given real constant, [19–22], and denotes order moment of . When is a real signal, , , and when is a complex signal, , denotes conjugate operation. is dispersion coefficient of . According to (7), the autocovariation function of can be expressed as When is Gaussian process, then and , and (8) degenerates into autocorrelation functionWe take Fourier transform on both sides of (8) and getEquation (10) is known as fractional low order covariation spectrum under stable distribution process, which is corresponding to power spectrum under Gaussian distribution. When and , (10) changes as power spectrum; hence, fractional low order covariation spectrum is a generalized power spectrum.
3.1.2. Fractional Lower Order BlackmanTukey Covariation Spectrum Estimation
We assume that is samples of a discrete stable distribution random process, and the sample mean is 0, thenEquation (12) can be applied to estimate the covariation function of . By taking discrete Fourier transform of (12), BlackmanTukey covariation spectrum estimation can be obtained, as shown in (13).
3.1.3. Fractional Lower Order Periodogram Covariation Spectrum Estimation
Letting and , substituting them into (13),where is discrete Fourier transform of . The discrete Fourier transform of is given by . From (14), we firstly compute discrete Fourier transform of and and take their product; then fractional lower order Periodogram covariation spectrum estimation can be obtained.
3.1.4. Fractional Lower Order Welch Covariation Spectrum Estimation
Covariation spectrums of the fractional lower order BlackmanTukey and fractional lower order Welch methods can be gotten by Fourier transform, the principle is simple, and the calculation quantity is also small. However, FLOBTCS and FLOPCS have higher side lobe, and the variance is larger, which results in poor performance and affects their practical application. Therefore, the traditional Welch method needs to be improved.
By dividing discrete signal into partially overlapped segments, the length of each segment is , the overlap length between two adjacent segments is , and then the nd segment can be expressed asBy taking the windowed Fourier transform of and , the following equations are obtained:Substituting (17)(18) into (14), the nd covariation spectrum can be written asCalculating the average of the covariation spectrum of all segments, then
Equation (20) is called fractional low order Welch covariation spectrum method. In order to reduce variance of the covariation spectrum estimation, the number of segments can be increased appropriately, which is bound to reduce length of the segment , so that the main lobe of the covariation spectrum is widened to reduce the resolution. In real application, and can be set according to the specific needs.
3.1.5. Application Review
We will now apply the FLOBTCS and FLOPCS and FLOWelch covariation spectrum methods to estimate frequency spectrum of the test signal. The test signal is defined as three sinusoidal signals with Gaussian or stable distribution noise.where , , , , and . When is Gaussian noise, generalized signal to noise ratio () is used to instead of .where is the dispersion coefficient of stable distribution noise. According to the given , the amplitude of the signal can be written as
Letting , the fractional lower order covariation spectrum methods and traditional power spectrum methods are used to estimate frequency spectrum of the signal under Gaussian distribution noise and stable distribution noise; the simulation results of ten independent experiments are shown in Figures 5–8.
(a)
(b)
(a)
(b)
(a)
(b)
(a)
(b)
3.1.6. Remarks
Figure 5(a) shows that the traditional BlackmanTukey spectrum estimation method fails under noise environment, but the improved FLOBTCS method in Figure 5(b) has good toughness. Figures 6(a) and 6(b) are frequency estimations of the signal employing the Periodogram and FLOPCS methods, respectively; the result reveals the advantage of the FLOPCS method. The Welch spectrum and FLOWelch covariation spectrum of the signal are given in Figure 7; we can know that the Welch power spectrum method cannot work under noise environment, but the proposed FLOWelch covariation method demonstrates good performance. Figure 8 is frequency spectrum of the signal employing FLOBTCS, FLOPCS, and FLOWelch covariation spectrum methods under ( and ) noise environment. The results show that the conventional power spectrum methods degenerate under noise environment, but the improved fractional low order covariation spectrum methods can better estimate out the frequency spectrum of the signal .
The FLOBTCS and FLOPCS methods have a certain side lobe and little amount of calculation, which cause the weak signals to be submerged and lead to the distortion of the signals. FLOWelch algorithm can suppress the side lobe effects by spectrum segmentation. However, the resolution and variance of the FLOWelch method are controlled by the number of segments and length like the conventional Welch method. In applications, the shorter number of segments should be used when we would like to get higher frequency resolution, but variance gets larger, and if we wish to reduce estimation variance, a longer number of segments are preferred. Hence, we can compromise the number and length of the segments by the actual situation. As a result, the power spectrum methods are only suitable to analyze the signals under Gaussian environment, but the fractional low order covariation spectrum methods can be applied under Gaussian and noise environment.
3.2. FLOAR Model Spectrum Estimation Method
3.2.1. Principle
An independent identically distributed (i.i.d) process can be expressed by order AR modelwhere is the parameters of AR model. is an i.i.d process, and its characteristic index and dispersion coefficient are and , respectively. When passes a finite impulse response (FIR) filter, thenTaking z transformation for (24)(25),where is transfer function of the filter. spectrum of the signal was defined as [21]whereWhen , substituting and (26) into (27), AR model spectrum of an stable distribution process on the unit circle can be written asThen AR model spectrum of a () process on the unit circle is given by
According to the definition in [21], we can obtain fractional low order autocovariance of the signal .The corresponding low order autocovariance estimation can be written aswhere , , and is number of the observation signal .
We define , , andwhere is call fractional low order autocovariance matrix (FLOACM). We apply FLOACM to fractional low order moment (FLOM) matrix in [21] and then getAndWe call (35) fractional low order Euler Walker equation. The parameters can be gotten by solving (35). When are substituted into (26) or (29), fractional low order AR model spectrum can be obtained. In this paper, we apply the final prediction error (FPE) criteria to determine the order of fractional low order AR model. When increases gradually from 1, FPE will be the minimum at a certain , which just is the most appropriate order. The calculation formula can be written aswhere is the variance of residuals.
We summarize the steps of fractional low order AR model spectrum method, as follows.
Step 1. Solve fractional low order AR model parameters of employing (35).
Step 2. Determine the order with the FPE criterion.
Step 3. Compute fractional low order AR model spectrum of by substituting into (29) or (30).
3.2.2. Application Review
In this simulation, the test signal is in (21). We compare the performance of the traditional AR model method and the improved fractional low order AR model method under Gaussian distribution noise () and stable distribution noise (, ); the results of the ten independent experiments are shown in Figures 910. In order to further verify the effectiveness of the fractional low order AR method, we conduct comparative experiment on two methods under different when , as show in Figure 11. When and changes from to , Figure 12 shows the change of Errors Power of the exiting AR model method and the improved FLOAR method under stable distribution noise environment.
(a)
(b)
(a)
(b)
3.2.3. Remarks
Figures 9(a) and 9(b) show the traditional AR model spectrum and the improved FLOAR model spectrum of the signal , respectively. Both methods can estimate out the frequency spectrum of the signal well under Gaussian noise environment. Figure 10(a) reveals that the AR model spectrum method fails under noise environment (), but the FLOAR model spectrum method in Figure 10(b) has good toughness. As a result, the AR model spectrum method is only suitable to analyze the signals under Gaussian environment, but the FLOAR model spectrum method can work under Gaussian and noise environment, which is robust.
Figure 11 is Errors Power of the AR and FLOAR model spectrum algorithms under different . When the characteristic index of distribution noise changes from 0 to 2, Errors Power of the improved FLOAR method is lower than the AR algorithm, which is very stable.
When changes from to , the Errors Power of the AR model and FLOAR model methods are given in Figure 12. The results show that Errors Power of the improved FLOAR method is lower than that of the AR method, and they are very stable with the changing of . Especially, when , the performance superiority of the FLOAR method is more obvious.
3.3. FLOARMA Model Spectrum Estimation Method
3.3.1. Principle
An i.i.d random process can be expressed by an ARMA .where and are parameters of order AR model and order MA model, respectively. Taking Z transform on (37), obtainingwhere is impulse response function, is system transfer function, and are the FIR and IIR filters, respectively. From (41), FLOARMA model spectrum of stable distribution process on the unit circle () is given byThen FLOARMA model spectrum of () process on the unit circle can be written as
Multiplying both sides of (38) and computing fractional low order covariance, thenwhereThen we can get [20]When , (45) changes asEquation (48) can be written aswhere (49) is also a generalized YuleWalker equation. Then we can obtain the parameters of the FLOAR model by solving (49).
When , (45) changes asLetting , thenBy solving the nonlinear equation (51), the parameters of the FLOMA model can be obtained.
An approximate infinite order FLOAR model can be used to be equivalent to a finite order MA model (), thenwhere is error. Letting , thenBy taking autocorrelation of both sides of (53) with one haswhere and . Equation (38) can be written asEquation (55) is a Toeplitz matrix. By solving the YuleWalker equation (55), the parameters of the FLOMA model are gotten.
We summarize the steps of the FLOARMA model spectrum method, as follows.
Step 1. Compute the parameters of the FLOAR model of the signal employing (49).
Step 2. By solving the nonlinear equation (51) or the YuleWalker equation (55), the parameters of the FLOMA model can be gotten.
Step 3. Determine the order and employing FPE criterion.
Step 4. Substitute and into (29) or (30); compute FLOARMA model spectrum of the signal .
3.3.2. Application Review
In this simulation, in (21) is used as the experimental test signal. The traditional ARMA method and modified FLOARMA method are applied to estimate the frequency spectrum of the signal under Gaussian noise () and noise environment (). The results of ten independent experiments are shown in Figures 1314.
(a)
(b)
(a)
(b)
3.3.3. Remarks
The ARMA frequency spectrum of the signal is shown in Figure 13(a), and Figure 13(b) is FLOARMA frequency spectrum of the signal under Gaussian() noise environment. The results show that both methods can well estimate out the frequency spectrum of the signal . Figure 14(a) shows that the traditional ARMA model spectrum method completely fails under () noise environment (), but the improved FLOARMA model spectrum method in Figure 14(b) has very good toughness. Hence, the ARMA spectrum method is only suitable to estimate the frequency spectrum of the signals under Gaussian environment, but the FLOARMA model method can be applied in Gaussian and noise environment, which can work in lower .
The traditional parameter model spectrum estimation methods are equivalent to solving YuleWalker equation and their extension equations based on the autocorrelation function. However, the fractional low order parameter model spectrum estimation methods construct the generalized YuleWalker equation and their extension equations employing fractional low order covariance matrix, because fractional low order autocovariance is a generalized covariance, which changes into autocorrelation when . As a result, the fractional low order parameter model spectrum estimation methods can be applied to both Gaussian and stable distribution noise environment.
3.4. Spectral Resolution of the FLO Spectrum Estimation Methods
In order to further verify the spectral resolution capability of the improved fractional low order spectrum methods, in this section, we apply the FLOBTCS, FLOPCS, FLOWelch, FLOAR, and FLOARMA model spectrum methods to estimate the spectrum of the signal with closely spaced frequencies. The signal in (21) is used as the experimental test signal, where , , , , and . is stable distribution noise, , and . The simulations of ten independent experiments are shown in Figure 15.
(a)
(b)
(c)
(d)
(e)
Figures 15(a) and 15(b) and Figure 15(c) are the FLOBTCS, FLOPCS, and FLOWelch spectrums of the signal , respectively. The FLOAR(90) model spectrum is shown in Figure 15(d), and Figure 15(e) is FLOAR(65, 3) model spectrum of the signal . The results show that the improved methods can distinguish the closely spaced frequencies well. Hence, the proposed methods have better performance of spectral estimation and higher spectral resolution, and the FLOARMA model spectrum method performs best.
4. Application Simulations
In this simulation, the experiment signal adopts from the real normal signal and bearing outer race fault signal (DE) in Section 2. 0.2 seconds data is selected as the test signal, which is collected at 12,000 samples per second, then . We apply the existing frequency spectrum methods and improved frequency spectrum methods to analyze the signals; the simulation results are shown in Figures 16–19.
(a)
(b)
(c)
(d)
(e)
(a)
(b)
(c)
(d)
(e)
(a)
(b)
(c)
(d)
(e)
(a)
(b)
(c)
(d)
(e)
(f)
Figures 16 and 17 are frequency spectrum estimations of the normal bearing signal employing the conventional frequency spectrum methods and fractional lower order frequency spectrum methods, respectively. The figures show that the methods all can estimate out the frequency components of the normal signal; the transient harmonic vibration components is about 1060Hz. We can see that the BTCS, PCS, FLOBTCS, and FLOPCS methods have a certain side lobe in Figures 16(a), 16(b), 17(a), and 17(b), respectively, which lead to the distortion of the signal. The performance of the proposed methods is comparable to that of the traditional methods. The Welch and FLOWelch methods in Figures 16(c) and 17(c) suppress the side lobe effects because of the spectrum segmentation. The FLOAR and FLOARMA model methods in Figures 17(d) and 17(e) can better demonstrate the normal signal than the AR and ARMA method in Figures 16(d) and 16(e). Hence, the proposed methods have an important advantage in the normal bearing signal analysis.
The frequency spectrums of the bearing outer race fault signal (DE) employing the traditional methods and the improved methods are shown in Figures 18(a)–18(e) and Figures 19(a)–19(e), respectively, and Figure 19(f) is fractional lower order ChoiWilliams timefrequency representation of the bearing outer race fault signal (DE). We can see that the transient harmonic vibration components about 600 Hz, 2800 Hz, and 3500 Hz dominate frequency domain from Figures 18(a)–18(e) and Figures 19(a)–19(e), and Figure 19(f) clearly reveals the gap between the impact. The interval between the pulse points A, B, C, D, E, and F is approximately 30ms, then the characteristic frequency of the bearing outer race fault is about 33.333Hz. However, the improved methods for the pulsive signals have better performance than the conventional methods in terms of variance and resolution; especially, the FLOAR and FLOARMA model methods have higher resolution than the existing AR and ARMA methods, which can better demonstrate the frequency spectrums of the impulsive signals. Furthermore, the improved methods have their respective advantages and disadvantages, as follows. The FLOBTCS and FLOPCS methods in Figures 19(a) and 19(b) have some side lobes, the FLOBTCS method has lager variance, and their resolution is lower. The FLOWelch method in Figure 19(c) has some improvement in suppressing side lobes and variance, but its resolution is not ideal. The parameter model frequency spectrum methods in Figures 19(d)–19(e) have better performance than the previous methods, which have higher resolution, less computation, and no side lobe interference.
In order to further verify the advantages of the proposed methods, distribution noise (, ) is added in the stable distribution outer race fault signal as the actual working environment background noise. The improved methods and existing methods are applied to analyze frequency spectrum feature of the out race fault signal; ten independent experiments are shown in Figures 20 and 21, respectively. The simulation results show that the conventional methods in Figures 20(a)–20(e) fail, but the proposed methods in Figures 21(a)–21(e) have good performance, and we can know that the transient harmonic vibration components are about 600 Hz, 2800 Hz, and 3500 Hz.
(a)
(b)
(c)
(d)
(e)
(a)
(b)
(c)
(d)
(e)
5. Conclusions
stable distribution is a more appropriate statistical model for the bearing fault signals, and . We have proposed the improved FLOBTCS, FLOPCS, FLOWCS, FLOAR, and FLOARMA parameter model frequency spectrum analysis methods employing fractional low order statistics; the methods are suitable for Gaussian and stable distribution noise environment and have good toughness; their performance is superior to the existing spectral estimation methods, which can effectively suppress distribution noise and work in low . The methods are very stable in frequency estimation, which have better performance of spectral estimation and higher spectral resolution. The advantages of the FLOARMA model spectrum method are more obvious, which can effectively suppress side lobe and improve resolution. The methods have their respective advantages and disadvantages, the FLOBTCS and FLOPCS methods have a certain side lobe, and their resolutions are lower. The FLOBTCS method has lager variance, but it is very easy to compute it directly. The FLOWCS can effectively reduce the side lobe, which decreases the resolution of spectrum estimation. The FLOAR and FLOARMA parameter model frequency spectrum methods have advantages of specific sense of mathematics, higher resolution, and inexistent problem of side lobe interference, and the calculation is not very complicated. Hence, the parameter model frequency spectrum methods have high performance. In view of the stable distribution bearing fault signal, even stable distribution noise environment, the proposed methods can be applied to analyze their spectrum features. Combining the improved frequency spectrum methods with the fractional low order timefrequency methods, we can get the transient harmonic vibration components and fault characteristic frequency of the bearing fault signals and achieve their joint fault diagnosis. Therefore, the joint fault diagnosis method may melt the bright side of each, which is significance for the bearing fault diagnosis. In the future, we can also further study the methods, such as developing them into fractional low order multispectral, inverse spectrum, spectrum autocorrelation, and combined analysis based on EMD.
Data Availability
The data used to support the findings of this study are available from the corresponding author upon request.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
This work is financially supported by Natural Science Foundation of China (61261046 and 61362038), Natural Science Foundation of Jiangxi Province China (20192BAB207002, 20151BAB207013), the Research Foundation of Health Department of Jiangxi Province China (20175561), Science and Technology Project of Provincial Education Department of Jiangxi (GJJ170954), and Science and Technology Project of Jiujiang University China (2016KJ001 and 2016KJ002).
References
 G. Xiong, C. Xi, D. Li, and W. Yu, “Cross correlation singularity power spectrum theory and application in radar target detection within sea clutters,” IEEE Transactions on Geoscience and Remote Sensing, pp. 1–14, 2019. View at: Publisher Site  Google Scholar
 J. J. Hall, M. R. AzimiSadjadi, S. G. Kargl, Y. Zhao, and K. L. Williams, “Underwater unexploded ordnance (UXO) classification using a matched subspace classifier with adaptive dictionaries,” IEEE Journal of Oceanic Engineering, pp. 1–14, 2018. View at: Google Scholar
 X. Zhang, S. Lu, J. Sun, and W. Shangguan, “LowAltitude and slowspeed small target detection based on spectrum zoom processing,” Mathematical Problems in Engineering, vol. 2018, 10 pages, 2018. View at: Google Scholar
 K. Ji, N. Bouaanani, R. Wen, and Y. Ren, “Introduction of conditional mean spectrum and conditional spectrum in the practice of seismic safety evaluation in China,” Journal of Seismology, vol. 22, no. 4, pp. 1005–1024, 2018. View at: Publisher Site  Google Scholar
 W. Yuehai, Y. Yongzheng, and W. Qinyong, “WaveletBased feature extraction in fault diagnosis for biquad highpass filter circuit,” Mathematical Problems in Engineering, vol. 2016, 13 pages, 2016. View at: Google Scholar
 L. Collins, A. Belianinov, S. Somnath et al., “Multifrequency spectrum analysis using fully digital G ModeKelvin probe force microscopy,” Nanotechnology, vol. 27, no. 10, article no. 105706, 2016. View at: Google Scholar
 J. Qingxian, C. Wen, Z. Yingchun, and J. Yu, “Robust fault reconstruction in discretetime lipschitz nonlinear systems via eulerapproximate proportional integral observers,” Mathematical Problems in Engineering, vol. 2015, Article ID 741702, 14 pages, 2015. View at: Publisher Site  Google Scholar
 H. Pan, Y. Yang, X. Li, J. Zheng, and J. Cheng, “Symplectic geometry mode decomposition and its application to rotating machinery compound fault diagnosis,” Mechanical Systems and Signal Processing, vol. 114, pp. 189–211, 2019. View at: Publisher Site  Google Scholar
 X. Yan and M. Jia, “Improved singular spectrum decompositionbased 1.5dimensional energy spectrum for rotating machinery fault diagnosis,” Journal of the Brazilian Society of Mechanical Sciences and Engineering, vol. 41, no. 1, pp. 1–20, 2019. View at: Google Scholar
 J. Liu, W. Wang, F. Golnaraghi, and K. Liu, “Wavelet spectrum analysis for bearing fault diagnostics,” Measurement Science and Technology, vol. 19, no. 1, pp. 105–118, 2007. View at: Publisher Site  Google Scholar
 C. Liu, X. Wu, T. Yang et al., “Spectrum sparseness: a new method for frequency spectrum analysis applied to equipment condition,” in Proceedings of the International Conference on Information Science & Control Engineering, pp. 1417–1419, Beijin, China, 2016. View at: Google Scholar
 S. U. WeiJun, F. Yang, Y. U. ChongChong et al., “Rolling bearing fault feature extraction method based on local spectrum,” Tien Tzu Hsueh Pao/Acta Electronica Sinica, vol. 46, no. 1, pp. 160–166, 2018. View at: Google Scholar
 M. Yusof, C. Nizwan, S. Ong et al., “lustering of Frequency Spectrum from Different Bearing Fault using Principle Component Analysis,” in Proceedings of the 2nd International Conference on Automotive Innovation and Green Vehicle (AiGEV 2016), vol. 90, p. 01006, MATEC Web of Conferences, Cyberjaya, Malaysia, 2017. View at: Publisher Site  Google Scholar
 S. E. Pandarakone, Y. Mizuno, and H. Nakamura, “Distinct fault analysis of induction motor bearing using frequency spectrum determination and support vector machine,” IEEE Transactions on Industry Applications, vol. 53, no. 3, pp. 3049–3056, 2017. View at: Publisher Site  Google Scholar
 C. Junsheng, Y. Dejie, and Y. Yu, “A fault diagnosis approach for roller bearings based on EMD method and AR model,” Mechanical Systems and Signal Processing, vol. 20, no. 2, pp. 350–362, 2006. View at: Publisher Site  Google Scholar
 F. Cong, J. Chen, and G. Dong, “Spectral kurtosis based on AR model for fault diagnosis and condition monitoring of rolling bearing,” Journal of Mechanical Science and Technology, vol. 26, no. 2, pp. 301–306, 2012. View at: Publisher Site  Google Scholar
 J. Ladou, “Degradation assessment and fault diagnosis for roller bearing based on AR model and fuzzy cluster analysis,” Shock & Vibration, vol. 18, no. 12, pp. 127–137, 2015. View at: Google Scholar
 H.B. Xu, G.H. Chen, and X.H. Wang, “Fault identification of bearings based on bispectrum distribution of ARMA model and FCM method,” Journal of South China University of Technology, vol. 40, no. 7, pp. 78–89, 2012. View at: Google Scholar
 H. Xiao, D. Huang, Y. Pan, Y. Liu, and K. Song, “Fault diagnosis and prognosis of wastewater processes with incomplete data by the autoassociative neural networks and ARMA model,” Chemometrics and Intelligent Laboratory Systems, vol. 161, pp. 96–107, 2017. View at: Publisher Site  Google Scholar
 X. Wang, J. Ren, and S. Liu, “Distribution Adaptation and Manifold Alignment for complex processes fault diagnosis,” KnowledgeBased Systems, vol. 156, pp. 100–112, 2018. View at: Publisher Site  Google Scholar
 X. Ma and C. L. Nikias, “Joint estimation of time delay and frequency delay in impulsive noise using fractional lower order statistics,” IEEE Transactions on Signal Processing, vol. 44, no. 11, pp. 2669–2687, 1996. View at: Publisher Site  Google Scholar
 T.H. Liu and J. M. Mendel, “A subspacebased direction finding algorithm using fractional lower order statistics,” IEEE Transactions on Signal Processing, vol. 49, no. 8, pp. 1605–1613, 2001. View at: Publisher Site  Google Scholar  MathSciNet
 G. Zhang, T. Yi, T. Zhang, and L. Cao, “A multiscale noise tuning stochastic resonance for fault diagnosis in rolling element bearings,” Chinese Journal of Physics, vol. 56, no. 1, pp. 145–157, 2018. View at: Publisher Site  Google Scholar
 J. Long, H. Wang, P. Li, and H. Xie, “A WHT signal detectionbased FLOTFUBSS algorithm under impulsive noise environment,” Circuits, Systems and Signal Processing, vol. 37, no. 7, pp. 2997–3022, 2018. View at: Publisher Site  Google Scholar
 J. Long, H. Wang, D. Zha, H. Fan, Z. Lao, and H. Wu, “Applications of an improved timefrequency filtering algorithm to signal reconstruction,” Mathematical Problems in Engineering, vol. 2017, Article ID 1805091, 14 pages, 2017. View at: Publisher Site  Google Scholar
 C. Li and G. Yu, “A new statistical model for rolling element bearing fault signals based on alphastable distribution,” in Proceedings of the Proceeding of the International Conference on Computer Modeling and Simulation (ICCMS '10), pp. 386–390, China, January 2010. View at: Publisher Site  Google Scholar
 G. Yu, C. N. Li, and J. F. Zhang, “A new statistical modeling and detection method for rolling element bearing faults based on alphastable distribution,” Mechanical Systems and Signal Processing, vol. 41, no. 12, pp. 155–175, 2013. View at: Publisher Site  Google Scholar
 J. Long, H. Wang, P. Li, and H. Fan, “Applications of fractional lower order timefrequency representation to machine bearing fault diagnosis,” IEEE/CAA Journal of Automatica Sinica, vol. 4, no. 4, pp. 734–750, 2017. View at: Publisher Site  Google Scholar
 Y. XiangMei and S. Tong, “Fault diagnosis method for gearbox based on αstable distribution parameters and support vector machines,” Measurement and Control Technology, vol. 31, no. 8, pp. 21–30, 2012. View at: Google Scholar
 Q. Xiong, Y. Xu, Y. Peng, W. Zhang, Y. Li, and L. Tang, “Lowspeed rolling bearing fault diagnosis based on EMD denoising and parameter estimate with alpha stable distribution,” Journal of Mechanical Science and Technology, vol. 31, no. 4, pp. 1587–1601, 2017. View at: Publisher Site  Google Scholar
 G. Zak, A. Wylomanska, and R. Zimroz, “Alphastable distribution based methods in the analysis of the crusher vibration signals for fault detection,” IFACPapersOnLine, vol. 50, no. 1, pp. 4696–4701, 2017. View at: Publisher Site  Google Scholar
 “CWRU bearing data center,” http://csegroups.case.edu/bearingdatacenter/pages/downloaddatafile. View at: Google Scholar
Copyright
Copyright © 2019 Junbo Long et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.