CN115630280A - Rolling bearing fault diagnosis method based on CEEMD multi-scale diffusion entropy and PSO-ELM - Google Patents

Abstract

The invention discloses a rolling bearing fault diagnosis method based on combination of complementary set empirical mode decomposition (CEEMD) multi-scale dispersion entropy and Particle Swarm Optimization (PSO) optimized Extreme Learning Machine (ELM). The method comprises the steps of firstly decomposing a non-stationary original acceleration vibration signal of the rolling bearing by adopting a CEEMD method to obtain a plurality of stationary Intrinsic Mode Function (IMF) components, screening the IMF components according to a correlation coefficient principle, then extracting the characteristics of the screened IMF components by utilizing multi-scale dispersion entropy, finally sending the characteristic set into an ELM optimized by PSO for classification and identification, and finally realizing the extraction and diagnosis of fault characteristics. The method has the advantages of achieving a good effect on the Kaiser university data set, being high in operation speed and high in anti-interference capability, solving the problem of inaccurate classification caused by random selection of traditional ELM parameters, and greatly improving the fault diagnosis precision.

Description

A Rolling Bearing Fault Diagnosis Based on CEEMD Multi-scale Dispersion Entropy and PSO-ELM method

technical field

The invention belongs to the field of mechanical fault diagnosis and relates to a fault diagnosis method for rolling bearings.

Background technique

Rolling bearing is one of the essential parts in many mechanical equipments and plays an important role. At the same time, it is also one of the most fragile parts. Once the bearing is damaged, it will have a great impact on the service life and performance of the machine, and even a major accident will cause major economic losses and endanger personal safety. Studies have shown that rolling bearing failures account for 45% to 55% of the total failures of mechanical equipment, which seriously affects the operating efficiency of machinery. Therefore, the study of early fault diagnosis methods for rolling bearings is of great significance for improving the safety of mechanical operation and predicting the risk of failure in advance. Due to the nonlinear and non-stationary vibration signals of rolling bearings, it is difficult to obtain a large number of typical fault samples. Therefore, effective fault diagnosis methods are the hotspots of current research.

In order to make the fault diagnosis results more accurate, YEH J proposed a new signal processing method - Complementary Ensemble Empirical Mode Decomposition (CEEMD) method. Confusion problem, and the operation time is shorter than the integrated empirical mode decomposition (EEMD), so it is widely used in mechanical fault diagnosis. The signal after CEEMD decomposition needs feature extraction. With the efforts of researchers, many feature extraction methods have emerged, such as: sample entropy, permutation entropy, fuzzy entropy, symbol entropy, basic scale entropy, and scatter entropy. The complexity of the signal can be measured, which effectively improves the accuracy of fault diagnosis. However, the above-mentioned entropy is usually measured on a single scale of the signal, and the fault information in the rolling bearing fault signal is often distributed on different scales, and only a single scale of measurement is likely to cause the loss of fault information, and thus cannot be accurately analyzed. Describe the fault information. Therefore, the feature extraction method that can overcome the loss of fault information and has strong anti-interference ability is the key to improve the accuracy of fault diagnosis.

With the advent of the digital age, machine learning methods can solve problems such as prediction, classification, clustering, and feature extraction. Among them, the effect of solving classification problems is remarkable. The extreme learning machine (ELM) is a classification method proposed by Huang et al. It does not need iterative updates, greatly reduces the running time, and is an efficient classification method. However, since the initial value and threshold of the ELM algorithm are both random, it will bring uncertainty to the model and affect the classification accuracy. Therefore, it is an urgent problem to be solved to optimize the method of selecting parameters of ELM to improve the classification accuracy.

Contents of the invention

In view of the above deficiencies, the present invention provides a rolling bearing fault diagnosis method based on the combination of CEEMD multi-scale dispersive entropy and PSO-ELM. The present invention decomposes the original signal through the complementary set empirical mode, and then performs feature Extract, and then use the extreme learning machine optimized by the particle swarm optimization algorithm to train the extracted feature values, thereby improving the efficiency of fault diagnosis.

In order to achieve the above object, the present invention adopts following technology to realize:

A rolling bearing fault diagnosis method based on the combination of CEEMD multi-scale scatter entropy and PSO-ELM. Including the following steps:

S1: Obtain the original signal of the rolling bearing.

S2: Use CEEMD to decompose the original signal to obtain multiple IMF components.

S3: Use the principle of correlation coefficient to screen the IMF components.

S4: The components screened in S3 are extracted using multi-scale distribution entropy.

S5: Using the PSO optimization algorithm to optimize the parameters of the ELM network.

S6: Use the optimized ELM to train the previously obtained rolling bearing feature set to achieve accurate classification of fault features.

In step S2, Complementary Ensemble Empirical Mode Decomposition (CEEMD) is performed on the collected original vibration signal to obtain multiple Intrinsic Mode Function (IMF) components.

In step S3, since the eigenmode function components of S2 have a certain correlation with the original signal, the correlation coefficient between each eigenmode function component and the original signal is calculated using the principle of correlation coefficient, so as to filter out the most relevant The eigenmode function component of the eigenmode function is recombined into a new vibration signal. Compared with the original signal, the new vibration signal removes the noise interference and the vibration signal is purer.

In step S4, the improved distribution entropy (DE), that is, multi-scale distribution entropy is used to extract fault features and form a feature set. Multi-scale distribution entropy has more advantages than distribution entropy in measuring the complexity and irregularity of time series. Prominent, very suitable for signal feature extraction.

In step S5, since the initial value of the extreme learning machine (ELM) algorithm and the threshold are both random, it will have a certain impact on the training results. Here, the particle swarm optimization algorithm (PSO) is used to search for the optimal parameters. The operation is to set the ELM input weight IW and the hidden layer neuron bias IB as the particles of the PSO algorithm, so as to avoid random training of the ELM model.

In step S6, the feature set is trained using the improved extreme learning machine, and finally the accurate classification of the fault is obtained.

Among them, in the optimization of the ELM network parameters described in S5, the input weight IW and the hidden layer neuron bias IB are set as the particles of the PSO algorithm to avoid random training of the ELM model. The implementation process is divided into the following steps:

In the first step, parameters such as the population size N ₁ and the number of population updates it max are initialized.

In the second step, the inertia parameter w is randomly generated according to the sample data, and the mean square error between the ELM test sample output and the predicted output is used as the fitness function to calculate the fitness value of each particle. After comparison and optimization, the particle position and velocity Update, when the mean square error is the smallest or the maximum number of iterations is reached, the ELM network parameters after particle swarm optimization are finally obtained.

Features and beneficial effects of the present invention:

1. The present invention adopts the CEEMD method, which not only can effectively suppress the problem of modal confusion, but also has shorter operation time.

2. Multi-scale distribution entropy is used for feature extraction, which has fast operation speed and strong anti-interference ability, overcomes the shortcoming of fault information loss easily caused by a single scale, and effectively improves the accuracy of fault diagnosis.

3. The PSO algorithm is used to optimize the ELM network parameters, avoid random training of the ELM model, and improve its classification accuracy.

4. Combining the fault diagnosis method with the optimized fault classification method, the efficiency of fault diagnosis is effectively improved.

Description of drawings

Fig. 1 is a flowchart of the present invention;

Figure 2 is a CEEMD decomposition flow chart;

Figure 3 is one of the IMF component diagrams decomposed by CEEMD;

Fig. 4 is the correlation coefficient figure of IMF component and original signal;

Figure 5 is a diagram of the number of iterations of PSO-ELM;

Figure 6 is the ELM confusion matrix matrix before and after PSO optimization;

Figure 7 shows the PSO-ELM confusion matrix under different loads.

Detailed ways

The present invention is a rolling bearing fault diagnosis method based on the combination of CEEMD multi-scale dispersive entropy and PSO-ELM. Its flow chart is shown in Figure 2, and its process includes the following steps:

The CEEMD decomposition process includes the following steps:

其中，H_m、J_m为加入白噪声之后的信号序列。n_m(t)代表第m次添加的白噪声。Step 1, introduce a set of additive and subtractive white noise signals into the original signal x(t), and get:

Among them, H _m and J _m are signal sequences after adding white noise. n _m (t) represents the white noise added for the mth time.

c_j，m(t)代表第m次添加白噪声，EMD分解后获取的第j个IMF分量，q为IMF个数。Step 2 uses EMD to decompose H _m and J _m , and then obtain the IMF components of the two, as follows:

c _{j, m} (t) represents the m-th addition of white noise, the j-th IMF component obtained after EMD decomposition, and q is the number of IMFs.

Step 3, repeat step 1, step 2, calculate the average value of the IMF components obtained after m decompositions, that is, the jth IMF component c _j

(2) Multi-scale scatter entropy feature extraction includes the following steps:

Step 1, the time series of the original signal is {u(i), i=1, 2, ..., L}, compound the coarse-grained sequence, and the k-th coarse-grained sequence under the scale factor τ is represented by x _k ^τ means that the following is the expression of the sequence:

其中c为类别，d为时间延迟，m为嵌入维数，X为初始时间序列信号。Step 2, under each scale factor τ, calculate the DE(x _k ^τ , m, c, d) of each coarse-grained sequence according to the DE principle, then the MDE is expressed as:

where c is the category, d is the time delay, m is the embedding dimension, and X is the initial time series signal.

(3) Particle swarm optimization ELM network parameter steps:

Step 1: Initialize parameters such as population size N ₁ , population update times it max and so on.

Step 2: Randomly generate the inertial parameter w according to the sample data, use the mean square error between the ELM test sample output and the predicted output as the fitness function, and calculate the fitness value of each particle. Update, when the mean square error is the smallest or the maximum number of iterations is reached, the ELM network parameters after particle swarm optimization are finally obtained.

In order to verify the fault diagnosis rate of this combined method of fault diagnosis, this patent adopts the public rolling bearing data set of Case Western Reserve University, and conducts tests through Matlab. The experimental bearing is the driving end bearing, the fault diameter is 0.007 inches, and the sampling frequency is 12kHz. The motor speed is 1797rpm. The experiment is divided into four groups of 0, 1, 2, and 3 according to the motor load. Each group has 10 categories of experiments, and each category includes 40 samples. The vibration signal data description is shown in Table 1.

Table 1 Vibration signal data

Firstly, after the CEEMD decomposition with short operation time and effective suppression of modal confusion, a set of component diagrams is shown in Figure 3; each component has a certain correlation with the original signal, and the correlation is measured by the correlation coefficient. The correlation coefficient between each IMF component and the original signal is shown in Figure 4, and then the multi-scale distribution entropy is used to extract the eigenvalues to form a feature matrix; the iterative times diagram of the PSO optimized extreme learning machine is drawn as shown in Figure 5, it can be seen that , when the PSO algorithm iterates to the 73rd time, the convergence is reached, and the optimal value of the fitness function is 0.001, and the optimal parameters obtained at this time are substituted into the ELM. Then use PSO-ELM to identify the state of the feature set, and compare it with the unoptimized extreme learning machine. As shown in Figure 6, after optimization, the number of wrongly identified samples is greatly reduced, and the accuracy rate is significantly increased. Finally, in Experimental tests were carried out under four loads of the Case Western Reserve University dataset, and the results are shown in Figure 7. The accuracy rates of the algorithms in the four cases are shown in Table 2. It can be clearly seen that all of them have achieved very high accuracy rates. , so this patent adopts a rolling bearing fault diagnosis method based on the combination of CEEMD multi-scale dispersive entropy and PSO-ELM, which can improve the efficiency of fault diagnosis.

Table 2 PSO-ELM accuracy rate under different loads

Claims (6)

1. A rolling bearing fault diagnosis method based on CEEMD multi-scale diffusion entropy and PSO-ELM is characterized by comprising the following steps:

s1: acquiring an original signal of a rolling bearing;

s2: decomposing an original signal by adopting CEEMD to obtain a plurality of IMF components;

s3: screening IMF components by adopting a correlation coefficient principle;

s4: carrying out feature extraction on the components screened in the S3 by adopting a multi-scale diffusion entropy;

s5: optimizing ELM network parameters by adopting a PSO optimization algorithm;

s6: and (4) training the rolling bearing feature set obtained in the front by using the optimized ELM to realize accurate classification of fault features.

2. The CEEMD multi-scale entropy and PSO-ELM-based rolling bearing fault diagnosis method of claim 1, wherein in step S2, a complementary set empirical mode decomposition (CEEMD) is performed on the acquired original vibration signal to obtain a plurality of Intrinsic Mode Function (IMF) components.

3. The method as claimed in claim 1, wherein in step S3, since the eigenmode function component of S2 has a certain correlation with the original signal, the correlation coefficient between each eigenmode function component and the original signal is calculated by using the correlation coefficient principle, so as to screen out the eigenmode function component with the largest correlation, and then the eigenmode function component is recombined into a new vibration signal, and the new vibration signal is compared with the original signal, so as to remove noise interference and make the vibration signal more pure.

4. The CEEMD multi-scale entropy diffusion and PSO-ELM based rolling bearing fault diagnosis method as claimed in claim 1, wherein in step S4, improved entropy Diffusion (DE) is adopted, namely multi-scale entropy diffusion is adopted to perform fault feature extraction, a feature set is formed, and the multi-scale entropy diffusion is more prominent than the entropy diffusion in measuring time series complexity and irregularity, and is very suitable for signal feature extraction.

5. The CEEMD multi-scale entropy and PSO-ELM based rolling bearing fault diagnosis method of claim 1, wherein in step S5, since both the initial value and the threshold of the Extreme Learning Machine (ELM) algorithm are random, which has a certain effect on the training result, the Particle Swarm Optimization (PSO) algorithm is used to search out the optimal parameters, and the specific operation is to set the ELM input weight IW and the hidden layer neuron bias IB as the particles of the PSO algorithm, so as to avoid the random training of the ELM model.

6. The CEEMD multi-scale entropy and PSO-ELM based rolling bearing fault diagnosis method of claim 1, wherein in step S6, the feature set is trained by using an improved extreme learning machine, and finally, an accurate classification of the fault is obtained.

Images