Application of VPMCD method based on PLS for rolling bearing fault diagnosis

1. Introduction

Rolling element bearings are significant components of rotary machines, and their working condition can directly affect the operation of the machine. The failure of a bearing can lead to the failure of the entire structure. Therefore, the early fault diagnosis of rolling element bearings can improve the safety of operating machinery [1-3].

The first step in fault diagnosis is to extract fault features from rolling bearing signals. Because the vibration signal carries large amounts of information representing the health conditions of mechanical equipment, vibration analysis has been established as the most common and reliable method of analysis in the field of condition monitoring and diagnostics of rotating machinery. When a fault occurs, the generated vibration signals are mostly nonlinear and non-stationary [3-5]; therefore, the key to bearing fault diagnosis is determining how to extract fault features from nonlinear and non-stationary signals. Conventional signal processing techniques, such as time-domain statistical analysis, Fourier transforms, and Wigner-Viller distributions, are based on the assumption that the signals are stationary and linear, which is not realistic. Wavelet transforms can be used process nonstationary signals. However, energy leakage will occur in the wavelet transformation, and the selection of the wavelet base function in the wavelet transform is difficult. The EMD method is a time-frequency analysis method suited to addressing nonlinear and non-stationary signals and can decompose signals into several stable intrinsic mode function (IMF) components [6, 7]. EMD has been widely applied in the fault diagnosis of rolling element bearings [8-10]. The collected signals are often mixed with noise, which can increase the number of layers of the EMD and enhance end effects, mixing fault feature signals and noise and increasing the difficulty of extracting fault features [10]. Therefore, selecting appropriate de-noising methods to remove the noise signals is very important and improves the accuracy of feature extraction. De-noising methods based on SVD represent effective nonlinear filtering methods with high robustness and are also used in the fields of image processing and signal filtering [11-13]. Entropy can be used to not only represent the complexity of signals but also measure the non-determinacy of a system or a piece of information [1, 14, 15]. When different bearing faults occur, the energy distributions of signals will change in different bands. Therefore, the energy entropy of vibration signals can be used as the eigenvector to extract fault features.

After extracting fault features from vibration signals, many scholars adopt neural networks, support vector machines (SVMs) or other intelligent methods to identify the fault features of bearings [1, 15, 16]. However, neural network identification methods are built based on large training samples. Their training speed is low, and such methods can easily fall into local extrema [17]. SVMs represent a method of small-sample learning that can obtain a globally optimal solution; however, the kernel function and parameters are not easy to confirm. Because the internal relations between the features of different faults obviously differ, the relations can be used to perform fault diagnosis. Thus, a new pattern recognition method based on the VPMCD method was proposed by Raghuraj and Lakshminarayanan [18, 19]. Cheng adopted the VPMCD method to study the fault diagnosis of bearings and to improve diagnosis performance [20]. LS regression is used in the VPMCD method to estimate the model parameters based on the assumption that no highly linear correlations exist between independent parameters. However, the fault feature data are limited in practical situations such that a linear correlation between independent variables is inevitable for small samples. A high degree of correlation influences the accuracy of parameter estimation, therein increasing model error and destabilizing the model. To address this problem, instead of the LS method, PLS regression is used to estimate the model parameters. The PLS method benefits from a strong processing capacity for high-dimensional data and can estimate parameters given a linear dependence between independent variables to improve the estimation accuracy [21].

In this paper, an improved VPMCD method is used for bearing fault diagnosis. First, an SVD de-noising method is adopted to facilitate the filtering of vibration signals. Then, the EMD method is used to decompose the signals into several IMF components. When a fault appears, some useful faulty information can be extracted from the high-frequency bands of vibration signals. The energy entropy of the first several orders of IMF components is selected to construct the fault eigenvector. The PLS method is selected to estimate the VPMCD model parameters, and the prediction model is applied to facilitate bearing fault identification.

2. Feature extraction of bearing faults based on SVD and EMD

2.1. Effect of noise on EMD

The EMD method proposed by Huang et al. in 1998 was found to be remarkably effective in analysing nonlinear and non-stationary signals [6]. The method can decompose any nonlinear, non-stationary signal into several IMF components and a remainder [8, 9]:

1

$x\left(t\right)=\sum _{i=1}^n{c}_i\left(t\right)+r\left(t\right),$

where $c_i\left(t\right)$ represents the $i$th IMF and $r\left(t\right)$ represents the remainder.

During EMD, the upper and lower envelopes are obtained via the cubic spline interpolation of extreme points [10]. Because of the uncertainty of whether the endpoints are the extremes, the spline has a fitting error at the endpoints in each IMF. The error continues to spread to the data internally during the decomposition process, which can lead to the IMF losing its original physical meaning and to false IMF components. Because the vibration signals are polluted by various types of noise to some extent, the number of spline interpolations and layers in the EMD increases, gradually accumulating the error caused by the end effect and seriously affecting the quality of the EMD. Moreover, signals that contain no interference by EMD cannot be well separated into IMF components because of the influence of noise; thus, it is extremely difficult to effectively extract fault features from interfered IMF components.

The effect of interference on EMD is presented in the following example. The simulated signal is:

2

$x\left(t\right)=\left(3+\mathrm{c}\mathrm{o}\mathrm{s}\left(15\pi t\right)\right)\mathrm{c}\mathrm{o}\mathrm{s}\left(380\pi t-3\mathrm{s}\mathrm{i}\mathrm{n}\left(5\pi t\right)\right)+2t\mathrm{c}\mathrm{o}\mathrm{s}\left(70\pi t+0.5\mathrm{c}\mathrm{o}\mathrm{s}2\pi t\right)+\tilde{n}\left(t\right),$

where $x\left(t\right)$ is the time-domain signal, as shown in Fig. 1, and $\tilde{n}\left(t\right)$ is a Gaussian white noise signal with an SNR of 14.5 dB. The sampling frequency is 5 kHz, and the sampling duration is 1 s. The IMF components obtained from the EMD are shown in Fig. 2.

Fig. 1Time-domain signal

Fig. 2IMF components

Fig. 2 shows that the 1st-3rd IMF components are mainly noise signals and that the 4th-8th IMF components contain the active ingredient. Due to the influence of noise, frequency aliasing appears in the IMF components. Therefore, the signals containing noise pollution must be preprocessed before applying EMD to improve the accuracy.

2.2. De-noising method based on SVD

The de-noising method based on SVD provides good stability and can reduce noise and improve the SNR [11-13]. Assuming that the vibration signal of a rolling bearing is $\mathbf{x}=\left[x_1,x_2,\cdots ,x_M\right]$, the reconstructed track matrix of the attractor ${\mathbf{D}}_m$ is as follows:

3

${\mathbf{D}}_m=\left[\begin{array}{cccc}{x}_1& x_2& \cdots & x_n\\ x_2& x_3& \cdots & x_{n+1}\\ ⋮& ⋮& ⋮& ⋮\\ x_m& x_{m+1}& \cdots & x_{m+n-1}\end{array}\right].$

According to SVD theory, the matrix ${\mathbf{D}}_m$ is decomposed as ${\mathbf{D}}_m=\mathbf{U}\mathbf{S}{\mathbf{V}}^T$, where $\mathbf{U}\in {\mathbf{R}}^{m\times m}$, ${\mathbf{V}}^T\in {\mathbf{R}}^{n\times n}$, $\mathbf{U}{\mathbf{U}}^T\in \mathbf{I}$, and $\mathbf{V}{\mathbf{V}}^T\in \mathbf{I}$. $\mathbf{S}$ is an $m\times n$ diagonal matrix with diagonal elements of ${\lambda }_1$, ${\lambda }_2$,…, ${\lambda }_k$, which are called the singular values of the matrix ${\mathbf{D}}_m$, where ${\lambda }_1>{\lambda }_2>\cdots >{\lambda }_k$ and $k=\mathrm{m}\mathrm{i}\mathrm{n}\left(m,n\right)$. For the signal containing noise, its reconstructed track matrix of the attractor ${\mathbf{D}}_m$ must be a column full-rank matrix, namely, rank $k=m$. Based on SVD theory and the matrix optimal approximation theorem based on the Frobenius norm, if the first $k\text{'}$ singular values are retained and if the other singular values are set to zero, then the matrix ${\mathbf{D}}_m^{\text{'}}$ can be obtained via the inverse process of SVD, which is the best approximation matrix of ${\mathbf{D}}_m$. Thus, the de-noised signals can be obtained from the matrix ${\mathbf{D}}_m^{\text{'}}$.

When constructing the best approximation matrix ${\mathbf{D}}_m^{\text{'}}$, the de-noising effect differs when the selected order $k\text{'}$ changes. If the order is too low, the information of the filtered signal is not complete, and if the order is too high, the filtered signal continues to contain an excessive noise signal. The singular entropy of the signal can reflect the degree of information contained in the singular value. The singular entropy $S_i$ is defined as follows:

4

$S_i=-p_i^s\mathrm{l}\mathrm{o}\mathrm{g}{p}_i^s,$

where $p_i^s={\lambda }_i/\sum _{j=1}^k{\lambda }_j$ ($i=$1, 2,…, $k$). Because the first $k\text{'}$ singular value corresponds to the effective components of the signal, the singular entropy is relatively large. After reaching a certain order, the singular value that corresponds to the components of the noise and the singular entropy is relatively small; therefore, the distribution diagram of the singular entropy can be used to determine the order of the effective components. Then, the signal can be reconstructed to effectively and simultaneously retain the signal information and remove the noise. The distribution diagram of the singular entropy is shown in Fig. 3 for the time-domain signal $x\left(t\right)$ in Fig. 1. The figure shows that after the singular entropy curve decreases to the asymptotic value, the small singular value can be considered as that caused by noise signals. Fig. 3 shows that the order of denoising used to reconstruct the signal is 28.

The SNR after de-noising is 55 dB. The first 3 orders of IMF components obtained by EMD from the de-noising signals are shown in Fig. 4.

The figure shows that the de-noising method based on SVD provides good stability and can both reduce noise and improve the SNR. The IMF components obtained from EMD are the effective parts of the signal.

Fig. 3Distribution diagram of singular entropy

Fig. 4IMF components of filtered signal

3. Improved VPMCD method

4. A fault diagnosis approach for rolling bearings based on the improved VPMCD method

In this paper, SVD, EMD and the PLS-VPMCD are combined to provide fault diagnosis for rolling bearings. Using the SVD method, noise is reduced in the original vibration signals. Then, the EMD method is used to decompose the vibration signals of the rolling bearings into several IMF components. The energy entropy of the first several orders of IMF components is extracted to construct the fault eigenvector, which is combined with the PLS-VPMCD method for pattern recognition. The flow chart of the proposed fault diagnosis approach is shown in Fig. 5.

The fault diagnosis process based on PLS-VPMCD is as follows:

(1) The vibration signals at a certain sample frequency $f_s$ are collected under four types of conditions: the rolling bearing is normal, the bearing has outer race faults, the bearing has inner race faults and the bearing has rolling ball faults. The number of samples is $n$ under each condition.

(2) The de-noising method based on SVD is used to pre-process the collected signals, and the reconstruction order is determined by the singular entropy and used to reconstruct the vibration signals.

(3) The EMD method is used to decompose the reconstructed signals to obtain several IMF components. The first $q$ orders are selected, and the energy entropy is calculated to construct the fault eigenvector.

(4) The $n$ fault eigenvectors under each fault condition are used as the training sample. The corresponding variable predictive model $VP{M}_i^{\left(k\right)}$ can be obtained using the PLS method as shown in Eqs. (8)-(22), where $k=$ 1, 2, 3, 4 represent the normal state, the outer race fault state, the inner race fault state and the rolling ball fault state, respectively, and $i$ denotes different eigenvalues.

(5) The testing signals are collected, and the eigenvector $\mathbf{X}\text{'}$ is constructed according to steps (1)-(3) as the input of the $VP{M}_i$ classifier. Then, the working condition and fault classes can be identified by the output of the $VP{M}_i$ classifier.

Fig. 5Flow chart of the fault diagnosis model based on PLS-VPMCD

5. Experiment

In this paper, PLS-VPMCD method is used for the fault diagnosis of rolling bearings. Rolling bearing experimental data from the Bearing Data Center of Case Western Reserve University are adopted to verify the validity and superiority of this method. A type 6205-2RS JEM SKF bearing is used. The sampling frequency is 12 kHz, the motor load is 0.746 kW, and the rotational speed is 1797 rpm. The fault types are the normal state, the outer race fault state, the inner race fault state and the rolling ball fault state. The diameter of the fault point is 0.1778 mm, and the depth of the fault is 0.2794 mm. The sampling time of each group is 0.1 s. The acceleration signals of the rolling bearing vibration in the normal state, the outer race fault state, the inner race fault state and the rolling ball fault state are shown in Figs. 6(a)-(d), respectively.

Fig. 6Vibration signal of the rolling bearing

The collected acceleration signal of the rolling bearing vibration is inevitably influenced by noise. In this paper, the SVD method is used to reduce the noise components in the rolling bearing vibration signal. We reconstruct the vibration signals, which correspond to the normal state, the outer race fault state, the inner race fault state and the rolling ball fault state, and calculate the singular entropy of the vibration signal under each working condition, as shown as Fig. 7.

There are different laws for the singular entropy value distribution in each component of a signal. The singular entropy values of the smooth signals and fault signals are larger and mainly appear during the initial period of the singular entropy diagram; thus, the singular entropy value that decreases to the flat region is caused by noise. We select the order for which the singular entropy value decreases to the flat region of the singular entropy curve as the de-noising order. This ensures the validity of the noise filtering process. Fig. 7 shows that the distribution of the singular entropy value exhibits different characteristics during each working state. However, after the 15th order, the singular entropy curve becomes flat; thus, the de-noising order is selected as 15 in this paper. The rolling bearing vibration signals after de-noising in the normal state, the outer race fault state, the inner race fault state and the rolling ball fault states are shown in Figs. 8(a)-(d), respectively.

Comparing Fig. 6 with Fig. 8 shows that the stochastic noise of the signal is largely reduced. The signal after filtering is decomposed using EMD, and the energy entropy of the first 4 orders of the IMF components is used to construct the bearing eigenvector. The first 4 orders of the IMF components decomposed using EMD in the normal state, the outer race fault state, the inner race fault state and the rolling ball fault state are shown in Figs. 9(a)-(d), respectively.

Fig. 7Singular entropy of the vibration signal

Fig. 8Vibration signal after de-noising

Fig. 9 shows that the SVD de-noising process decreases the influence of noise on the EMD and that each order of the IMF components is the main component of the vibration signal. The characteristics of the energy distributions of the IMF components differ under different fault conditions. The fault eigenvector constructed using the energy entropy of the IMF components can effectively provide information about the bearing work state.

The key point of the VPMCD method used in the fault diagnosis of bearings is obtaining an effective prediction model VPM via training sample regression. With fewer training samples, the number of sample points could be nearly equal to and sometimes less than the number of variables. When the number of samples is smaller, a linear correlation between independent variables is inevitable. If the linear correlation is strong and if LS regression is used to construct the regression model, the accuracy and reliability of the prediction model cannot be easily guaranteed.

The variance inflation factor is the most commonly used diagnostic method when addressing multiple correlations. When the variance inflation factor is larger than 10, multiple correlations strongly influence the estimated value of the LS method. The average variance inflation factors between independent variables are shown in Table 5.

Table 5 shows that the average variance inflation factors between independent variables are much larger than 10. When the model parameters are estimated via the LS method, large deviations appear and subsequently influence the identification precision of the VPM model. The number of samples is $n=$ 7. VPM models are constructed using the LS method and the PLS method. VPM models are then used to identify 20 new groups of samples, and the VPM model identification errors are shown in Fig. 10.

Fig. 9IMF components of the bearing vibration signal

Fig. 10VPM model identification errors. (VPM1: VPM model constructed using samples in the normal state; VPM2: VPM model constructed using samples in the outer race fault state; VPM3: VPM model constructed using samples in the inner race fault state; VPM4: VPM model constructed using samples in the rolling ball fault state)

Fig. 11The recognition rates of different methods. A is the recognition rate when the number of samples and when the signal is unfiltered, B is the recognition rate when the number of samples and when the signal is filtered, C is the recognition rate when the number of samples and when the signal is unfiltered, D is the recognition rate when the number of samples and when the signal is filtered, E is the recognition rate when the number of samples and when the signal is unfiltered, and F is the recognition rate when the number of samples and when the signal is filtered

Fig. 10 shows that, because of the multiple correlations amongst independent variables, larger errors appear in the estimated values of the LS-VPM model. Especially for the rolling ball fault state, large errors appear such that it becomes difficult to identify the fault state. The PLS method can effectively solve the problem of multiple correlations amongst independent variables. The error of the estimated value from the PLS-VPM model is stable; thus, the model can properly identify the faulty state of a bearing.

The BP-NN and LS-SVM methods are effective identification methods that possess strong identification abilities and high robustness. Using different numbers of samples when the vibration signals are pre-processed using SVD de-noising and identified using the BP-NN, LS-SVM, LS-VPMCD and PLS-VPMCD methods, the identification error values of 20 new groups of samples are determined, as listed in Table 6. The total recognition rates of 80 groups of samples for different methods are shown in Fig. 11.

Table 7 shows the number of model identification errors for various orders (10, 15 and 20) of de-noising when the sample number $n=\text{7}$.

The following is demonstrated by the data in Fig. 11, Table 6 and Table 7:

(1) When the number of training samples $n=$ 5 or $n=$ 7, independent variables strongly affect each other through the multiple correlations, making the LS-VPM model less stable, especially as shown by the low identification rate of rolling ball bearing faults. The PLS-VPM model effectively avoids the multi-correlation influence between independent variables. Under the different fault state types, the method achieves high identification accuracy.

(2) After de-noising the original vibration signal using SVD, the fault eigenvector extracted from the original vibration signal can provide information about the bearing’s state. Compared with the unfiltered state, the identification accuracy of each method is strongly increased. Increasing the number of samples without performing signal de-noising using the SVD method cannot improve the identification accuracy. Table 7 shows that selecting the 15th order of the de-noising can improve the identification accuracy.

(3) When the number of samples $n=$ 10, the identification accuracy of the PLS-VPMCD method is 100 % after the signal is filtered. Therefore, the method can accurately identify the fault state of a bearing. The identification accuracy of the PLS-VPMCD method is higher than that of the BP-NN, LS-SVM and LS-VPMCD methods, which are more suitable for identification when using small samples.

6. Conclusions

A bearing fault diagnosis method based on the SVD de-noising method and the PLS-VPMCD method is proposed in this paper to address the vulnerabilities faced by rolling bearings, the non-stationary and nonlinear characteristics of vibration signals and the presence of noise signals found in the collected signals. The results prove that, when performing fault diagnosis using small samples, the PLS-VPMCD can effectively avoid unstable model parameter identification caused by multiple correlations amongst independent variables. The PLS-VPMCD method achieves a higher diagnostic accuracy than the BP-NN and LS-SVM methods, and the PLS-VPMCD method does not require complicated parameter adjustments, thus avoiding the parameter optimization problem. As a result of the de-noising process using SVD, the fault eigenvector extracted from the original vibration signals can effectively provide fault feature information, reducing the influence of noise on bearing fault diagnosis and improving identification accuracy. Selecting the appropriate order of de-noising can improve the identification accuracy. In this paper, we provide a new analysis method based on PLS-VPMCD for diagnosing ball bearing faults.