Sparsity-Oriented Nonconvex Nonseparable Regularization for Rolling Bearing Compound Fault under Noisy Environment

Li, Xiaocheng; Wang, Jingcheng; Wang, Hongyuan

doi:https://doi.org/10.1155/2020/8823102

Shock and Vibration

On this page

Abstract Introduction Conclusions Data Availability Conflicts of Interest Acknowledgments References Copyright Related Articles

Research Article | Open Access

Volume 2020 | Article ID 8823102 | https://doi.org/10.1155/2020/8823102

Sparsity-Oriented Nonconvex Nonseparable Regularization for Rolling Bearing Compound Fault under Noisy Environment

Xiaocheng Li,¹Jingcheng Wang,^1,2and Hongyuan Wang¹

Academic Editor: Rafał Burdzik

Received20 Mar 2020

Revised09 May 2020

Accepted04 Jun 2020

Published30 Jun 2020

Abstract

Rolling bearing is widely used in rotating machinery and, at the same time, it is easy to be damaged due to harsh operating environments and conditions. As a result, rolling bearing is critical to the safe operation of the machinery devices. Compound fault of rolling bearing is not a simple superimposition of multiple single faults, but the coupling of multiple fault features, making the vibration signal, becomes complicated. In our study, sparsity-oriented nonconvex nonseparable regularization (SONNR) method is proposed to rolling bearing compound fault diagnosis under noisy environment. Firstly, a theoretical model of rolling bearing compound fault is established, and the vibration characteristics of rolling bearing compound fault are analyzed. Secondly, four-layer structure of the SONNR method is proposed: input layer, nonconvex sparse regularization layer, signal reconstruction layer, and compound faults isolation layer. Finally, the validity of the method is verified by simulation data and actual data, and it is compared with the traditional time domain diagnostic methods and artificial intelligence methods.

1. Introduction

Rolling bearing is widely used in many mechanical devices. With increase in complexity of machinery, the requirements of reliability for rolling bearing are increasing. Bearing damage, even minor damage, will affect the normal function of the devices [1–3]. A rolling bearing is often composed of an inner ring, an outer ring, a number of rolling elements, and a cage. The inner ring acts to match the shaft and rotates around the shaft. The outer ring acts to cooperate with the bearing seat to provide support. The cage evenly distributes the rolling elements, prevents the rolling elements from falling off, and guides the rolling elements to rotate. Fortunately, the vibration signal generated by a single fault of the rolling bearing has been extensively studied and a series of powerful diagnostic methods have been made. However, in the course of the rolling bearing, there is not only single fault; with the change of the operating environment and running time increasing, bearing often has two or more faults that are named compound fault of the rolling bearing. Compound fault is not a simple superimposition of multiple single faults, but the coupling of multiple fault features, making the vibration signal complicated. In addition, rolling bearings are often disturbed by various background noises during operation. When there are early weak faults in the bearing, the fault characteristics are often covered by strong noise, making it difficult to obtain the fault features. The background noise is often assumed to be additive white Gaussian noise (AWGN) [4]. In the classical techniques, the assumption of Gaussian noise is taken under consideration [5–7]. The research signal in this paper already contains noise signal. In order to verify the effectiveness of the method, we artificially added Gaussian white noise. Therefore, the study of compound fault signals detection under noisy environment technologies becomes urgent to overcome.

To monitor and detect bearing faults and damage, vibration signals are often used. In the past decades, various signal processing techniques were developed for extracting fault information about vibration signals. At present, the methods of compound fault diagnosis of rolling bearing are mainly empirical mode decomposition (EMD) method, wavelet analysis method, and sparsity-assisted filtering methods [8–11]. With the rapid development of artificial intelligence and computer technology, intelligent diagnostic methods based on deep learning (DL) [16–19] were developed, including convolutional neural network (CNN), stacked auto encoder (SAE), deep belief network (DBN), deep Boltzmann machine (DBM), and recurrent neural network (RNN). However, the existing methods have the following limitations. The intrinsic mode function (IMF) component extracted by the EMD method is essentially a band-amplified signal that has a sharp cut in the impact. In addition, wavelet analysis and EMD are both signal processing methods based on band division. When the center frequency of the impact signal overlaps between the center frequencies of other signals, the impact component cannot be effectively extracted. L1 norm regularization methods have become widely used in various applications due to the convexity of the L1 norm. However, L1 norm solutions often underestimate the high-amplitude components, yet these comprise the signal of interest in most cases. Thus, L1 norm solutions are not ideal for certain applications, including bearing fault diagnosis. A vast number of deep models had been proposed and verified with various types of bearing fault signals. For artificial intelligence, all the above publications stopped at the problem of a single fault, while the more challenging case, the multiple fault bearing diagnosis, has not been verified yet.

In general, it is challenging to extract features of vibration signals to indicate bearing faults, especially for compound faults. The coupling characteristics of rolling bearing compound fault signals are mainly manifested in three aspects: (1) coupling characteristic caused by the change of structural parameters, that is, when the rolling bearing fails, the geometric parameters of the bearing structure change due to the fault, which causes a certain difference between different frequency components. The ratio relationship, the ratio, and the phase corresponding to the frequency is consistent. (2) Coupling occurring between different frequency components. Since different rolling bearing fault signals have corresponding characteristic frequencies and their harmonic components, and the phases corresponding to different signal components have a certain correlation, some signal frequencies may have a certain correlation with other harmonic frequencies. (3) Convex spare solutions often underestimate the high-amplitude components, yet these comprise the signal of interest in most cases. Thus, convex spare solutions are not ideal for certain applications, including bearing fault diagnosis.

In order to overcome these issues, sparsity-oriented nonconvex nonseparable regularization (SONNR) for rolling bearing compound fault under noisy environment is studied in this paper. The SONNR method uses a nonconvex penalty and simultaneously maintains the convexity of the least squares cost function to be minimized. Nonconvex sparse regularization (NSR) that maintains the convexity of the cost function has been recently studied, to capture the advantages of both nonconvex regularization and convex optimization [20, 21]. The SONNR method has a four-layer structure: input layer, nonconvex sparse regularization layer, signal reconstruction layer, and compound faults isolation layer. The measured time-domain signal is input layer. In NSR decomposition layer, nonconvex sparse regularization by K-singular value decomposition (K-SVD) is used to improve denoising performance of raw signal. In signal reconstruction layer, more sparse and noise-reduced signals are reconstructed. In compound faults isolation layer, the measured time-domain signal is decomposed into high and low components using overcomplete rational-dilation wavelet transforms (RADWT).

The main contributions of this paper are summarized as follows:(1)The physical model of the compound fault is established, and the fault impact signal, the envelope signal, and the envelope signal spectrum of the compound fault signal are derived. The modulation characteristics of the compound fault signal are obtained from the analysis of the envelope spectrum, which lays a theoretical foundation for the subsequent fault information extraction.(2)The nonconvex optimization method for compound faults under noisy environment has seldom been considered in previous studies. The problem of the characteristic coupling of the rolling bearing compound faults signals and the interference caused by the noisy environment makes this work a great challenge. Exploiting the high-amplitude components property, a K-SVD-based nonconvex nonseparable regularization denoising algorithm is presented to extract bearing fault features.(3)The RADWT decomposes a complex signal into the sum of an oscillatory component and a transient component. For compound faults, RADWT can well reveal the amplitude modulation frequency of the fault shock response.(4)We present a verification based on real damage (accelerated lifetime) data set to validate the effectiveness of the proposed method in the diagnosis of compound faults in rolling bearing. In particular, we have artificially added noise based on the acquired signals.

The main advantage compared to conventional filtering methods is that the proposed method can better preserve the bearing fault signal while reducing noise and other interference components; thus it can significantly improve the estimation accuracy of the bearing fault signal. In addition, the compound fault signal can be separated to extract more obvious fault features.

The rest of this paper is organized as follows: Section 2 introduces the characteristics of the rolling bearing fault signal. Section 3 details the proposed SONNR method. Section 4 presents a simulation study and a real damage study, which is followed by concluding remarks in Section 5.

2. Analysis of Modulation Characteristics of Rolling Bearing Compound Fault Signal

We detect the signals from the fault signals collected by the sensors. The detection is based on the modulation characteristics of the vibration signals. Therefore, before detecting the fault signals, we need to understand the signal modulation characteristics in the theoretical analysis.

2.1. Compound Fault Theory Model

The response of the impact sequence caused by the spalling fault at the sensor is a series of unilateral exponentially decaying oscillations. “Resonant demodulation” of this series of exponentially decaying sequences will obtain its envelope spectrum, which is the characteristic spectrum of rolling bearing fault. When a compound bearing has a compound fault, there is more than one location of the spalling fault. The position of the spalling point relative to the acceleration sensor changes continuously with the rotation. The spalling point will enter the load area and exit the load area. The magnitude and direction of the received series of vibration impact forces also change. In this case, the vibration signal of the rolling bearing compound fault collected by the acceleration sensor is more complicated, and the corresponding envelope signal and its spectrum are also more complicated. In order to make a more realistic interpretation of the envelope signal and the envelope spectrum and improve the accuracy of the diagnosis, a model for analyzing the envelope signal and its frequency spectrum of the compound bearing compound fault is proposed in the following.

In this chapter, the theoretical model is studied by taking the compound fault of the inner ring spalling and the rolling element spalling as an example. Figure 1 shows the theoretical model of compound fault of inner ring spalling and rolling element spalling.

When t = 0, the position of the spalling point of the inner ring is exactly at the peak of the load distribution density curve, that is, the lowest position in the axis direction, and it just collides with the rolling element. Since the fit between the inner ring and the shaft is an interference fit during assembly, the inner ring will rotate as the shaft rotates. The spalling point on the inner ring also rotates with the rotation of the bearing, so that the position of the spalling point does not always fall within the load distribution area as shown in Figure 1; most of the time it will fall outside the load distribution area.

Due to the different load, the resulting pulse force is different. When the spalling point of the inner ring falls within the load distribution area as shown in the figure, the magnitude and direction of the vibration pulse force are related to the position of the spalling point and the angle of the axis. When the spalling point is outside the load area shown in Figure 1, it can be known from the force analysis that if the rotational inertia force generated when the rolling bearing rotates is ignored at this time, no vibration pulse force is generated in this area. In summary, when the internal ring has a spalling fault, the vector value of the generated vibration pulse force is not only affected by the load distribution but also related to the location of the spalling fault point.

From this analysis, it can be seen that the magnitude and direction of the impulse force caused by the damage point of the inner ring and rolling element are affected by the load distribution and the position of the damage point. These factors are discussed in the following sections.

2.1.1. Regardless of Load Distribution and Location of Spalling Point

When the spalling point of the inner ring appears, its vibration characteristics appear as pulse vibrations in equal time intervals. If the impact of the load distribution and the position of the spallation point on the pulse force is ignored, it can be determined that each collision between the defect and the rolling element produces one vibration pulse, and the magnitude of each pulse force is equal. The expression of a series of pulses caused by spalling of the inner ring is shown inwhere d_i is intensity of pulse force caused by inner ring damage, is unit pulse function, T_i = 1/f_i is interval between pulses, and f_i is inner ring fault characteristic frequency.

2.1.2. Load Distribution Expression

Rolling bearing is mainly subject to radial load forces during operation. When a rolling bearing is subjected to radial loads, the load distribution is shown inwhere , , is radial clearance, and is maximum contact deformation , .

2.1.3. Expression of Influence Coefficient of Pulse Force Acting Direction

Assuming that the magnitude of the pulse force is proportional to the load distribution density at the position where it is applied, the magnitude of the pulse force acting on the sensor axis is

Since the phases of and are the same, let

The vibration envelope signal due to spalling damage of the inner ring is ( indicates convolution)

The envelope spectrum corresponding to is

For the rolling element, the amplitude spectrum is shown in

Then, the envelope spectrum of the compound fault of the inner ring and the rolling element can be processed by the superimposition principle, and the vector sum of the total amplitude is shown as follows:

According to (8) and (9), the envelope amplitude spectrum of the combined fault of the inner ring spalling and the rolling element spalling can be obtained. The envelope spectrum diagram is shown in Figure 2. The frequency component is the same as in the case of a single fault: the characteristic spectrum of the inner ring spalling is a series of main spectral lines with the inner circle fault characteristic frequency f_i as the distance, and at the same time, the two sides of each main spectral line are separated by the frequency f_r and its multiples. The amplitudes of the main and sidelines decrease with increasing frequency. It should be noted that there is a peak at the frequency f_r. The frequency spectrum is a series of main spectral lines with the characteristic frequency f_b of the rolling element as the interval. At the same time, each of the main spectral lines is separated from the revolution frequency f_f and its side frequency with smaller amplitude side spectral lines. In addition, there is a peak at the revolution frequency f_f.

(a)

(b)

(c)

(d)

By analyzing the modulation characteristics of the compound fault, we can know that the fault characteristics of the inner ring and the rolling element need to be execrated in order to prevent missed diagnosis and misdiagnosis.

2.2. Coupling Characteristics and Challenges of Rolling Bearing Compound Fault Signals

The coupling characteristics of rolling bearing compound fault signals are mainly manifested in four aspects: (1) Due to the difference in the location and size of the fault, the response waveforms generated by different faults will also affect each other, coupling multiple fault characteristics and interfering with each other. (2) The coupling caused by changes in structural parameters, that is, when the rolling bearing fails, the fault causes the geometric parameters of the bearing structure to change, which causes a certain degree of difference between different frequency components. The relationship between this ratio and the phase corresponding to the frequency is the same. (3) The coupling between different frequency components, that is, different rolling bearing fault signals generally appear to have corresponding characteristic frequencies and their harmonic components. Because the phases corresponding to different signal components have a certain correlation, some signal frequencies may have a certain correlation with other harmonic frequencies. (4) For rolling bearing compound fault under noisy environment, the signal modulation characteristics of the time domain signal are not obvious. It will be strengthened, and the impact component with low energy may be overwhelmed by the fault characteristics with high energy. These four aspects will lead to the incomplete diagnosis of fault characteristics when the single fault diagnosis method is used to diagnose the rolling bearing compound fault.

This section illustrates these challenges using actual signals as examples.

Figure 3 is the acceleration vibration signal and the frequency spectrum of the bearing when the inner and outer raceways and rolling elements have spalling fault at the same time. The shaft speed is 30.01 Hz, the load is 5 Kg, the sampling frequency is 51.2 KHz, and the number of sampling points is 65536. The bearing’s outer ring characteristic frequency is 91.47 Hz, the inner ring characteristic frequency is 148.55 Hz, the rolling element characteristic frequency is 59.78 Hz, and the cage rotation frequency is 11.34 Hz. As can be seen from Figure 4, the vibration of time-domain signal of the rolling bearing compound fault is relatively chaotic. Although several impacts can be observed, it is difficult to determine the fault location. Envelope analysis is performed on this signal, and the envelope spectrum obtained is shown in Figure 5.

(a)

(b)

(a)

(b)

As can be seen in Figure 4(b), there is a spectral peak of 91.41 Hz and its multiples, which is consistent with the characteristic frequency of the outer ring fault. The analysis can only detect the inner ring fault, while other simultaneous outer ring and rolling element faults are omitted. This determines that the traditional single-point fault method can only be used to diagnose the compound bearing of a rolling bearing, and it is difficult to analyze and diagnose it comprehensively and accurately.

The autocorrelation processing is performed on the envelope signal of Figure 4(b) to obtain the self-power spectrum shown in Figure 5. Figure 5 eliminates noise glitches and makes the fault frequency in the spectrum more prominent. We can see that there is a spectral peak of 91.41 Hz and its frequency doubling, which can more significantly determine that the bearing has an outer ring spalling fault. The energy caused by the impact of the fault is distributed in the entire frequency domain. The overall analysis can only display the characteristics of the compound fault in the same result. The fault features with lower energy may be overwhelmed by the fault features with higher energy. Therefore, compound faults should be expressed in different channels as much as possible.

3. The Proposed Algorithm SONNR

Figure 6 gives a flowchart of the proposed methodology. The SONNR method has a four-layer structure: input layer, nonconvex sparse regularization layer, signal reconstruction layer, and compound faults isolation layer. The measured time-domain signal is input layer. In NSR decomposition layer, nonconvex sparse regularization by K-SVD is used to improve the denoising performance of raw signal. In signal reconstruction layer, more sparse and noise-reduced signals are reconstructed. In compound faults isolation layer, the measured time-domain signal is decomposed into high and low components using RADWT. The high-Q component contains a sustained oscillation cycle signal; the low-Q component contains the rolling fault impact transient signal.

3.1. Nonconvex Sparse Regularization by K-SVD

In order to enhance sparsity, many NSR methods have been proposed, and a general way to model the problem is the nonconvex regularization least squares problem:where is a nonconvex sparsity-inducing penalty. In this section, we use the nonconvex generalized minimax-concave (GMC) penalty in [22].

In our SONNR method, we use the GMC penalty , which is defined as

Using the GMC penalty in (11), the cost function in (10) can be expressed as

In our SONNR method, we use the GMC penalty , which is defined aswhere , . is generalized Huber function. The GMC penalty is shown in Figure 7 and the generalized Huber function is shown in Figure 8.

The minimization of the GMC-regularized least squares problem can be written as a saddle-point problem:wherewith , which ensures that the iterative algorithm converges.

The linear transformation A in (11) is important for sparse representation. An effective transformation can promote the sparsity of coefficients. For bearing fault diagnosis, we use K-SVD as the linear transformation A in this paper. The detailed description of K-SVD can be found in [23]. In K-SVD, the linear transformation A can be solved by

K-SVD can be regarded as a generalized form of K-means. In the K-means algorithm, each semaphore can only be approximated by one atom, while each signal in K-SVD is a linear combination of multiple atoms. Solve the sparse representation coefficients based on the given signal y and the initial dictionary D. This problem can be solved by K-SVD, and the closest solution A can be found by the basis pursuit (BP) algorithm.

Iterative algorithm for nonconvex sparse regularization by K-SVD problem is shown in Algorithm 1.

(1)	initialize
(2)
(3)	calculate the coefficients using a simple gradient descent procedure
(4)	update the dictionary
(5)
(6)
(7)
(8)
(9)	end for
(10)	return

3.2. Principle of Overcomplete Rational-Dilation Wavelet Transforms

The authors in [24] proposed an overcomplete rational expansion from the perspective of the frequency domain, wavelet transform (RADWT), overcomplete rational expansion wavelet transform by changing the upsampling of the signal and downsampling the value of q to adjust the quality factor of the wavelet, greatly expanding the range of the quality factor and increasing. The flexibility of quality factor selection: the quality factor Q of this method is determined by the following equation:

The quality factor is a measure of the degree of frequency aggregation of a signal. The higher the quality factor, the more concentrated the frequency of the signal, and the higher the number of oscillations in the time domain.

The RADWT decomposes a complex signal into the sum of an oscillatory component and a transient component [25]. The RADWT nonlinearly separates the components in the signal according to the oscillation characteristics and establishes the optimal sparse representation of each of the high resonance components and the low resonance components.

It is assumed that the observed signal x can be expressed as the sum of two signals x₁ and x₂:

The purpose of the RADWT analysis is to estimate the source signals x₁ and x₂ from the observed signal x, respectively.

3.3. Solving SONNR Problem

In our study, sparsity-oriented nonconvex nonseparable regularization (SONNR) method is proposed to rolling bearing compound fault diagnosis under noisy environment. Four-layer structure of SONNR method is proposed: input layer, nonconvex sparse regularization layer, signal reconstruction layer, and compound faults isolation layer.

Iterative algorithm for our SONNR problem is shown in Algorithm 2.

(1)	initialize
(2)	do
(3)	calculate the coefficients using a simple gradient descent procedure
(4)	update the dictionary
(5)
(6)
(7)
(8)
(9)
(10)	end for
(11)	return

In view of the coupling characteristics of rolling bearing compound faults and the problem that traditional fault diagnosis methods cannot comprehensively diagnose compound faults, the coupling signals need to be decoupled in order to more effectively diagnose the fault characteristics one by one. As shown in Algorithm 2, The SONNR method has a four-layer structure: input layer, nonconvex sparse regularization layer, signal reconstruction layer, and compound faults isolation layer. The SONNR method can better preserve the bearing fault signal while reducing noise and other interference components; thus it can significantly improve the estimation accuracy of the bearing fault signal. In addition, the compound fault signal can be separated to extract more obvious fault features.

4. Applications in Bearing Compound Fault Diagnosis

4.1. Simulated Compound Fault under Noisy Environment

In order to verify the effectiveness of the SONNR extraction fault response, this method is first used to analyze a simulated compound fault signal.

The rolling bearing fault model is used to simulate the compound fault of the inner ring and the outer ring under the strong noise background. The mathematical model of the inner ring spalling fault of the rolling bearing is shown in

The mathematical model of the rolling bearing outer ring spalling fault is shown in

The mathematical model of the compound fault of the inner ring and the outer ring is shown inwhere is a small fluctuation of the i-th impact with respect to the average period T. Assume that the bearing fault is a steel peeled off from the external raceway; the rotating speed of the bearing is 3660 rpm. It is also assumed that the fault characteristic frequency of inner ring is and fault characteristic frequency of outer ring is . The natural frequency f_d of the rolling bearing system is 9225 Hz, the random slip is assumed to be positively distributed over a standard deviation from 0.5% of the rotational speed, and the sampling frequency f_s is 60 KHz. The simulated signal without noise is shown in Figure 9.

In Figure 10, the random white noise distributed over the positive is added (the signal-to-noise ratio is −10 dB), and the time domain waveform after adding noise is shown in Figure 10. As can be seen from Figure 10, the fault feature is almost overwhelmed due to the background noise variable.

Determine the decomposition parameters of the SONNR: , Q₁ = 4.5, Q₂ = 1, J₁ = 60, J₂ = 15, r₁ = r₂ = 3.5. SONNR sparse decomposition of the fault signal is shown in Figures 11 and 12. It can be seen from Figure 11(b) that there is a spike in the outer ring fault frequency of 500 Hz and its multiplier (1000 Hz, 1500 Hz, 2000 Hz, etc.), and it can be diagnosed that there is an outer ring fault. The inner ring fault frequency is 600 Hz and its multiple frequency 1200 Hz. It can be seen from Figure 12(b) that there is a main peak at 1800 Hz, etc., and a side band appears at the frequency shift (200 Hz) on both sides of the main peak, indicating that there is an inner ring fault in the rolling bearing. The results show that it is effective to extract the compound fault information of rolling bearing by the SONNR method.

(a)

(b)

(a)

(b)

4.2. Results Based on Accelerated Lifetime (Real Damage) Data Set

The proposed method is validated by benchmark data set of Paderborn University [26], which involves compound damage to various operating conditions. This benchmark generates real bearing damage samples by accelerating lifetime tests. The option to compound fault with measurement data onto real bearing damage: different damage symptoms occur in the bearing or identical damage symptoms occur on different bearing components. Our test was carried out by including data sets from bearing with multiple damage at both raceways, at inner and outer rings (KB23). The compound fault form of KB23 is shown in Figure 13. The computational complexity is O (n·d), the length of the sequence n is 3840, and the computational dimension d is 1.

The manufacturer specific information about bearing KB23 is shown in Table 1.

4.2.1. Step 1: Noise Signal Preparation

In order to verify the effectiveness of the proposed method of fault classification in noisy environments, we add Gaussian white noise to each data set. The noise n is white Gaussian noise with standard deviation . As shown in Figure 14, the bearing KB23 fault signal, the additive white Gaussian noise signal, and the frequency domain signal of the KB23 fault signal are shown from top to bottom. It can be seen that the signal is seriously polluted by noise, and the time domain signal and the frequency domain signal cannot distinguish the vibration characteristics of the original model. Therefore, it is very difficult to extract effective fault information about a signal with noise.

(a)

(b)

(c)

4.2.2. Step 2: Nonconvex Sparse Regularization by K-SVD

We use GMC with a fixed regularization parameter λ to analyze the simulation signal of inner fault with noisy signal. In order to explore the effect of GMC on the performance of the sparse regularized method, we compare this method with the L1 regularization method.

The L1-regularized least squared problem is

In this case, the parameters of GMC nonconvexity and L1 regularization are set to . A is an oversampled inverse discrete Fourier transform and the result of Fourier coefficients is shown in Figure 15(c) by K-SVD. The performance is evaluated by the average root-mean-squared error (RMSE). Sparse result by GMC and L1 is shown in Figures 15(a) and 15(b). The RMSE of the GMC and L1-norm methods is 0.129 and 0.149. The comparison results show that the GMC is more accurate and uses fewer coefficients and yields more accurate denoising result.

(a)

(b)

(c)

4.2.3. Step 3: Compound Fault Isolation

The fault characteristic frequency of the outer ring is

The fault characteristic frequency of the inner ring is

Determine the decomposition parameters of the RADWT: Q₁ = 4, Q₂ = 1, J₁ = 30, J₂ = 15, r₁ = r₂ = 3.5. Taking the parameters of the bearing into (16) and (17), we can get f_i = 123.3, f_o = 76.25.

From Figure 16, we can know that the RADWT decomposes a complex signal into the sum of two different components. The RADWT nonlinearly separates the components in the signal according to the oscillation characteristics and establishes the optimal sparse representation of each of the high resonance components (Figure 16(b)) and the low resonance components (Figure 16(c)). We perform envelope analysis on Figures 16(b) and 16(c), respectively, and we can get Figures 17 and 18.

(a)

(b)

(c)

It can be seen from Figure 17 that there is a spike in the outer ring fault frequency of 76.25 Hz and its multiplier, and it can be diagnosed that there is an outer ring fault. Figure 18 shows the fundamental rotation frequency of the shaft and its harmonics (25 Hz), the ball pass frequency of the inner race (123.3 Hz), and its sidebands as well as its corresponding harmonics. Figure 18 indicates that there is an inner ring fault.

The results show that it is effective to extract the compound fault information of rolling bearing by the SONNR method.

4.3. Comparison with Time-Domain Diagnostic Methods

Envelope analysis, EMD, EEMD, and EEMD-PCA are commonly used to extract the rolling bearing fault information. The traditional analysis techniques are applied to the vibration signal of KB23.

Envelope analysis result of KB23 is shown in Figure 19. From Figure 19, it is difficult to find the characteristic frequency of the fault, so the traditional envelope analysis is difficult to reveal the characteristic frequency of the compound fault shock response.

EMD and EEMD analysis results of KB23 are, respectively, shown in Figures 20 and 21. Through EMD decomposition, we can get a series of IMF with different frequencies, which can be used to extract the characteristics of the original signal. From Figures 20(b) and 21(b), it is difficult to find the characteristic frequency of the fault, so the EMD and EEMD analysis is difficult to reveal the characteristic frequency of the compound fault shock response.

(a)

(b)

(a)

(b)

EEMD-PCA analysis results of KB23 are, respectively, shown in Figures 22 and 23. In this section, we reconstruct the original compound fault signal KB23 by EEMD-PCA method. The reconstructed picture is shown in Figure 22. From Figure 22, we can find that the EEMD-PCA method plays a role in noise reduction. Next, spectrum analysis is performed on the reconstructed signal, and the result is shown in Figure 23. From Figure 23, it is difficult to find the characteristic frequency of the fault, so the EEMD-PCA analysis is difficult to reveal the characteristic frequency of the compound fault shock response.

The comparison experiments with traditional envelope analysis and EMD, EEMD, and EEMD-PCA analysis show that the proposed technique SONNR of fault information extraction has fewer spectral glitches and a higher signal-to-noise ratio. And this method can reveal the amplitude modulation frequency of fault shock response more than traditional methods.

4.4. Comparison with Artificial Intelligence Methods

Further tests were compared with artificial intelligence methods. The test was carried out by including data sets from bearings with compound damage at both raceways, at inner and outer ring (KB23, KB24, and KB27) in [22]. The benchmark data set included 26 damaged bearing states (12 outer faults, 3 compound faults, and 11 inner faults) and 6 undamaged (healthy) states for reference. Each bearing state included 80 measurements. Therefore, our training types are divided into 4 categories (class 1: healthy; class 2: outer fault; class 3: inner fault; class 4: compound fault), 70% of which are used to train the network, and the remaining 30% are used for testing. The experiments are tested in Python 3.7 software platform using a laptop equipped with Intel-i5 3.2 GHz CPU, 8 GB memory. At the end of this section, experimental results are verified by comparing with commonly used diagnostic methods including stacked auto encoder (SAE), another version of stacked auto encoder DBN, and convolutional neural network (CNN). For the neural networks, parameters such as the number of hidden layers and neurons were not tuned and this explains the poor performance.

It can be seen from Table 2 that the artificial intelligence methods get the poor performance. It can be seen that the artificial intelligence algorithm is not very effective for the classification of compound faults. This also makes the application of artificial intelligence algorithms in compound fault diagnosis a challenging topic in the future.

5. Conclusions

In this paper, we present the SONNR method for rolling bearing compound fault diagnosis. The main advantage of the SONNR over conventional fault diagnosis methods is that SONNR can better preserve the bearing fault signal while reducing the interference of noise and other components; thus it can reveal the periodic fault impact mode in the bearing signal. The physical model of the compound fault is established, and the modulation characteristics of the compound fault signal are obtained from the analysis of the envelope spectrum, which lays a theoretical foundation for the subsequent fault information extraction. The nonconvex optimization method for compound faults under noisy environment has been considered in our studies. Exploiting the high-amplitude components property, a K-SVD-based nonconvex nonseparable regularization denoising algorithm is presented to extract bearing fault features. For compound faults, RADWT can well reveal the amplitude modulation frequency of the fault shock response. Finally, we present a verification based on real damage (accelerated lifetime) data set to validate the effectiveness of the proposed method in the diagnosis of compound faults in rolling bearing. In particular, we have artificially added noise based on the acquired signals. The comparison experiments with traditional envelope analysis and EMD, EEMD, and EEMD-PCA analysis show that the proposed SONNR method has more advantages in complex fault diagnosis than traditional time domain method. Further tests were compared with artificial intelligence methods. The result shows that the artificial intelligence methods are not very effective for the classification of compound faults. This also brings challenges to future research.

Data Availability

The data (a benchmark data set for condition monitoring of rolling bearings) used to support the findings of this study have been deposited in the KAtDataCenter website of the Chair of Design and Drive Technology, Paderborn University, Germany, repository (http://mb.uni-paderborn.de/kat/datacenter).

Conflicts of Interest

The authors declare no potential conflicts of interest.

Acknowledgments

This work was supported by National Natural Science Foundation of China (nos. 61533013 and 61633019), Shaanxi Provincial Key Project (2018ZDXMGY-168), and Shanghai Project (17DZ1202704).

References

J. Liu and Y. Shao, “Overview of dynamic modelling and analysis of rolling element bearing with localized and distributed faults,” Nonlinear Dynamics, vol. 93, no. 4, pp. 1765–1798, 2018.
View at: Publisher Site | Google Scholar
F. Cong, J. Chen, G. Dong, and M. Pecht, “Vibration model of rolling element bearings in a rotor-bearing system for fault diagnosis,” Journal of Sound and Vibration, vol. 332, no. 8, pp. 2081–2097, 2013.
View at: Publisher Site | Google Scholar
P. K. Kankar, S. C. Sharma, and S. P. Harsha, “Vibrations based performance prediction of ball bearings caused by localized defects,” Nonlinear Dynamics, vol. 69, no. 3, pp. 847–875, 2012.
View at: Publisher Site | Google Scholar
I. Gondal, M. Amar, and C. Wilson, “Vibration spectrum imaging: a novel bearing fault classification approach,” IEEE Transactions on Industrial Electronics, vol. 62, no. 1, pp. 494–502, 2014.
View at: Publisher Site | Google Scholar
J. Lundén, S. A. Kassam, and V. Koivunen, “Robust nonparametric cyclic correlation-based spectrum sensing for cognitive radio,” IEEE Transactions on Signal Processing, vol. 58, no. 1, pp. 38–52, 2009.
View at: Publisher Site | Google Scholar
V. Girondin, K. M. Pekpe, H. Morel, and J.-P. Cassar, “Bearings fault detection in helicopters using frequency readjustment and cyclostationary analysis,” Mechanical Systems and Signal Processing, vol. 38, no. 2, pp. 499–514, 2013.
View at: Publisher Site | Google Scholar
P. Borghesani and M. R. Shahriar, “Cyclostationary analysis with logarithmic variance stabilisation,” Mechanical Systems and Signal Processing, vol. 70-71, pp. 51–72, 2016.
View at: Publisher Site | Google Scholar
B. Pang, G. Tang, T. Tian et al., “Rolling bearing fault diagnosis based on an improved HTT transform,” Sensors, vol. 18, no. 4, pp. 1203–1208, 2018.
View at: Publisher Site | Google Scholar
Y. Lv, Q. Zhu, and R. Yuan, “fault diagnosis of rolling bearing based on fast nonlocal means and envelop spectrum,” Sensors, vol. 15, no. 1, pp. 1182–1198, 2015.
View at: Publisher Site | Google Scholar
X. Chen, F. Feng, and B. Zhang, “weak fault feature extraction of rolling bearings based on an improved kurtogram,” Sensors, vol. 16, no. 9, p. 1482, 2016.
View at: Publisher Site | Google Scholar
L. Yong, H. Zhang, and C. Yi, “Trivariate empirical mode decomposition via convex optimization for rolling bearing condition identification,” Sensors, vol. 18, no. 7, pp. 2325–2330, 2018.
View at: Publisher Site | Google Scholar
X. Chen, D. Liu, G. Xu, K. Jiang, and L. Liang, “Application of wavelet packet entropy flow manifold learning in bearing factory inspection using the ultrasonic technique,” Sensors, vol. 15, no. 1, pp. 341–351, 2014.
View at: Publisher Site | Google Scholar
H. Liu, C. Liu, and Y. Huang, “Adaptive feature extraction using sparse coding for machinery fault diagnosis,” Mechanical Systems and Signal Processing, vol. 25, no. 2, pp. 558–574, 2011.
View at: Publisher Site | Google Scholar
W. He, Y. Ding, Y. Zi, and I. W. Selesnick, “Sparsity-based algorithm for detecting faults in rotating machines,” Mechanical Systems and Signal Processing, vol. 72-73, pp. 46–64, 2016.
View at: Publisher Site | Google Scholar
Y. Qin, “A new family of model-based impulsive wavelets and their sparse representation for rolling bearing fault diagnosis,” IEEE Transactions on Industrial Electronics, vol. 65, no. 3, pp. 2716–2726, 2018.
View at: Publisher Site | Google Scholar
R. A. Kanai, R. G. Desavale, and S. P. Chavan, “Experimental-based fault diagnosis of rolling bearing using artificial neural network,” Journal of Tribology, vol. 138, no. 3, 2016.
View at: Publisher Site | Google Scholar
Q. Fu, B. Jing, P. He, S. Si, and Y. Wang, “Fault feature selection and diagnosis of rolling bearing based on EEMD and optimized Elman_AdaBoost algorithm,” IEEE Sensors Journal, vol. 18, no. 12, 2018.
View at: Publisher Site | Google Scholar
H. Cao, L. Niu, S. Xi, and X. Chen, “Mechanical model development of rolling bearing-rotor systems: a review,” Mechanical Systems and Signal Processing, vol. 102, pp. 37–58, 2018.
View at: Publisher Site | Google Scholar
X. Li, W. Zhang, and Q. Ding, “A robust intelligent fault diagnosis method for rolling element bearing based on deep distance metric learning,” Neurocomputing, vol. 310, pp. 77–95, 2018.
View at: Publisher Site | Google Scholar
I. Selesnick and M. Farshchian, “Sparse signal approximation via nonseparable regularization,” IEEE Transactions on Signal Processing, vol. 65, no. 10, pp. 2561–2575, 2017.
View at: Publisher Site | Google Scholar
I. Selesnick, “Sparse regularization via convex analysis,” IEEE Transactions on Signal Processing, vol. 65, no. 17, pp. 4481–4494, 2017.
View at: Publisher Site | Google Scholar
S. Wang, I. W. Selesnick, G. Cai et al., “Nonconvex sparse regularization and convex optimization for bearing fault diagnosis,” IEEE Transactions on Industrial Electronics, vol. 65, no. 9, p. 7332, 2018.
View at: Publisher Site | Google Scholar
M. Aharon, M. Elad, and A. Bruckstein, “$rm K$-SVD: an algorithm for designing overcomplete dictionaries for sparse representation,” IEEE Transactions on Signal Processing, vol. 54, no. 11, pp. 4311–4322, 2006.
View at: Publisher Site | Google Scholar
I. Bayram and I. W. Selesnick, “Frequency-domain design of overcomplete rational-dilation wavelet transforms,” IEEE Transactions on Signal Processing, vol. 57, no. 8, pp. 2957–2972, 2009.
View at: Publisher Site | Google Scholar
I. W. Selesnick and I. Bayram, “Oscillatory + transient signal decomposition using overcomplete rational-dilation wavelet transforms,” in Proceedings of SPIE the International Society for Optical Engineering, p. 7446, San Diego, CA, USA, August 2009.
View at: Publisher Site | Google Scholar
C. Lessmeier, J. K. Kimotho, D. Zimmer, and W. Sextro, “Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: a benchmark data set for data-driven classification,” in Proceedings of the 3rd European Conference of the Prognostics and Health Management Society, Bilbao, Spain, July 2016.
View at: Google Scholar

Copyright

Copyright © 2020 Xiaocheng Li 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.

PDF Download Citation

Download other formats

Order printed copies