CN112613493A - Rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM - Google Patents

Abstract

The invention discloses a fault diagnosis method for a rolling bearing based on feature fusion and WOA-ELM. The steps are as follows: step 1, using the official bearing data of Casey Western Reserve University as original vibration data; step 2, performing time domain analysis and frequency spectrum on the original data Analysis and wavelet packet decomposition to extract time domain, frequency domain and time-frequency domain features; step 3, normalize the mixed domain feature set obtained in step 2; step 4, use the manifold learning LPP algorithm to normalize the normalized The dimension of the high-dimensional feature set is reduced, and the low-dimensional feature sample set is obtained; in step 5, the ELM network parameters are optimized by the WOA algorithm, and the low-dimensional feature data is input into the WOA-ELM model for fault diagnosis. The invention effectively solves the problems of insufficient feature extraction, redundant feature information in multi-feature samples, and insufficient stability caused by the random generation of network parameters by the extreme learning machine. The invention has obvious advantages for improving the bearing fault diagnosis rate.

Description

Rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM

Technical Field

The invention relates to a bearing fault diagnosis method, in particular to a rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM.

Background

The rotary machine is the power for supporting the stable development of national economy, and the serious economic loss can be brought when the machine breaks down. Rolling bearings are important components of large-scale mechanical equipment and are very important for safe operation of the mechanical equipment, so that an effective and reliable bearing fault diagnosis method needs to be researched urgently. Aiming at the characteristics of nonlinearity, instability and the like of early failure of the rolling bearing, researchers at home and abroad continuously perform exploratory research on the aspects of a characteristic extraction and diagnosis model of the rolling bearing.

At present, although the time domain analysis method can effectively retain the characteristics of the original signal, the time domain analysis method is not sensitive to non-stationary signals. The fast fourier transform is only suitable for analysis of stationary signals, and it is difficult to simultaneously embody the overall and local features of two time-frequency domains. The complexity of energy in the vibration signal cannot be well reflected by pure wavelet packet energy, and the rolling bearing is insensitive to early failure of the rolling bearing. In the case of multi-feature extraction, there is also a problem of redundant features. Therefore, there is a need for efficient feature extraction methods and diagnostic models.

Disclosure of Invention

Aiming at the problem of improving the fault diagnosis rate of the rolling bearing, the rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM is provided. And respectively carrying out time domain analysis, frequency spectrum analysis and wavelet packet decomposition on the vibration signals of the rolling bearing to form a mixed fault characteristic set. And (5) applying local preserving projection to reduce the dimension of the mixed feature set and eliminate redundant features. And optimizing network parameters of the extreme learning machine by introducing a whale algorithm, and establishing a WOA-ELM rolling bearing diagnosis model to classify and diagnose faults. The method can effectively improve the fault diagnosis rate of the rolling bearing.

The technical scheme for realizing the invention is as follows:

a rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM comprises the steps of feature extraction and classifier construction, and specifically comprises the following steps:

step one, adopting official bearing data of Kaixi storage university as original vibration data;

performing time domain analysis, frequency spectrum analysis and wavelet packet decomposition on the original data, and extracting time domain, frequency domain and time-frequency domain characteristics;

step three, carrying out normalization processing on the mixed domain feature set obtained in the step two;

fourthly, carrying out dimension reduction on the normalized high-dimensional feature set by using a manifold learning LPP algorithm to obtain a low-dimensional feature sample set;

and fifthly, optimizing ELM network parameters by using a WOA algorithm, and inputting the low-dimensional characteristic data to a WOA-ELM model for fault diagnosis.

The concrete content of the second step is as follows: the time domain features are dimensionless feature indexes including peak factors, pulse factors, form factors, margin factors and kurtosis factors. And converting the time domain signal into a frequency domain signal by using FFT (fast Fourier transform), carrying out spectrum analysis, and selecting the mean frequency and the frequency divergence as frequency domain characteristics. And (4) performing three-layer wavelet packet decomposition by using db3 mother wavelets, and extracting the energy of the wavelet packets of the sub-bands with larger difference between the bands to serve as the time-frequency domain characteristics.

And step three, specifically, constructing a mixed feature matrix Y by using the three-domain features obtained in the step two, and then carrying out normalization processing on the mixed feature matrix Y in order to avoid the influence of a great numerical value difference between feature indexes on the diagnosis of the classifier:

where max and min are the sample data maximum and minimum values.

The specific content of the fourth step is that the LPP is used for reducing the dimension of the normalized high-dimensional feature set obtained in the third step, and the specific steps are as follows:

computationally deriving the transformation matrix A such that the data points x in the high-dimensional space_i(i-1, 2, …, m) to a low-dimensional spatial data point y_i(i ═ 1,2, …, m), i.e. y_i＝A^Tx_i

The matrix A is obtained by minimizing the objective function by calculation, i.e.

W_ijAnd assigning values for the weight matrix between the connection nodes by adopting a thermonuclear mode.

Adding constraint y^TDy is 1, and A is argminA^TXLX^TA

Then converting the general characteristic value into the problem of solving the general characteristic valueI.e. XLX^TA＝λXDX^TA

The first l eigenvectors can thus be solved, i.e. a ═ a₀,a₁,…,a_l-1]。

The concrete content of the fifth step is as follows:

and setting the whale population number and the algorithm iteration number, and initializing an input weight W and a hidden layer threshold b of the extreme learning machine to serve as an initial position vector of whales in the WOA.

And calculating individual fitness values in the population, finding the optimal whale individual, and recording the position of the current optimal individual.

And if the iteration times or the minimum fitness value is not met, updating the position between the whale and the target, and entering the next iteration.

And when the conditions are met, the current optimal individual position of the whale is reserved, and the optimal parameters of the ELM model are obtained.

And then, obtaining a diagnosis model of the WOA-ELM, and dividing the low-dimensional mixed feature set into a training set and a testing set after setting a transfer function and the number of network layers of the extreme learning machine. And inputting the training set into a diagnosis model for training, and inputting the training set into a test set for verifying the diagnosis performance of the WOA-ELM.

The invention provides a rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM, which is characterized in that a mixed fault feature set is formed by carrying out time domain analysis, frequency spectrum analysis and wavelet packet decomposition on a vibration signal of a rolling bearing. And then, applying local preserving projection to reduce the dimension of the mixed feature set and eliminate redundant features. Finally, introducing whale algorithm to optimize network parameters of extreme learning machine, establishing WOA-ELM rolling bearing diagnosis model to classify and diagnose faults

The method effectively solves the problems of insufficient feature extraction, redundant feature information in multi-feature samples, insufficient stability caused by random generation of network parameters by an extreme learning machine and the like. The method has obvious advantages for improving the bearing fault diagnosis rate.

Drawings

FIG. 1 is a diagnostic flow chart of the present invention;

FIG. 2 is a graph of the results of an unoptimized ELM diagnosis;

FIG. 3 is a graph of the diagnostic results of WOA-ELM;

FIG. 4 is an iterative graph of the WOA-ELM algorithm.

Detailed Description

The present invention will be described in detail below.

Rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM

Step one, adopting official bearing data of Kaixi storage university as original vibration data;

performing time domain analysis, frequency spectrum analysis and wavelet packet decomposition on the original data, and extracting time domain, frequency domain and time-frequency domain characteristics;

step three, carrying out normalization processing on the mixed domain feature set obtained in the step two;

fourthly, carrying out dimension reduction on the normalized high-dimensional feature set by using a manifold learning LPP algorithm to obtain a low-dimensional feature sample set;

and fifthly, optimizing ELM network parameters by using a WOA algorithm, and inputting the low-dimensional characteristic data to a WOA-ELM model for fault diagnosis.

The concrete content of the second step is as follows: the time domain features are dimensionless feature indexes including peak factors, pulse factors, form factors, margin factors and kurtosis factors. And converting the time domain signal into a frequency domain signal by using FFT (fast Fourier transform), carrying out spectrum analysis, and selecting the mean frequency and the frequency divergence as frequency domain characteristics. And (4) performing three-layer wavelet packet decomposition by using db3 mother wavelets, and extracting the energy of the wavelet packets of the sub-bands with larger difference between the bands to serve as the time-frequency domain characteristics.

And step three, specifically, constructing a mixed feature matrix Y by using the three-domain features obtained in the step two, and then carrying out normalization processing on the mixed feature matrix Y in order to avoid the influence of a great numerical value difference between feature indexes on the diagnosis of the classifier:

where max and min are the sample data maximum and minimum values.

The specific content of the fourth step is that the LPP is used for reducing the dimension of the normalized high-dimensional feature set obtained in the third step, and the specific steps are as follows:

computationally deriving the transformation matrix A such that the data points x in the high-dimensional space_i(i-1, 2, …, m) to a low-dimensional spatial data point y_i(i ═ 1,2, …, m), i.e. y_i＝A^Tx_i

The matrix A is obtained by minimizing the objective function by calculation, i.e.

W_ijAnd assigning values for the weight matrix between the connection nodes by adopting a thermonuclear mode.

Adding constraint y^TDy is 1, and A is argminA^TXLX^TA

Then converted into a generalized eigenvalue problem, XLX^TA＝λXDX^TA

The first l eigenvectors can thus be solved, i.e. a ═ a₀,a₁,...,a_l-1]。

The concrete content of the fifth step is as follows:

and setting the whale population number and the algorithm iteration number, and initializing an input weight W and a hidden layer threshold b of the extreme learning machine to serve as an initial position vector of whales in the WOA.

And calculating individual fitness values in the population, finding the optimal whale individual, and recording the position of the current optimal individual.

And if the iteration times or the minimum fitness value is not met, updating the position between the whale and the target, and entering the next iteration.

And when the conditions are met, the current optimal individual position of the whale is reserved, and the optimal parameters of the ELM model are obtained.

And then, obtaining a diagnosis model of the WOA-ELM, and dividing the low-dimensional mixed feature set into a training set and a testing set after setting a transfer function and the number of network layers of the extreme learning machine. And inputting the training set into a diagnosis model for training, and inputting the training set into a test set for verifying the diagnosis performance of the WOA-ELM.

As can be seen from FIG. 2, the fault diagnosis rate of the rolling bearing based on the multi-feature extraction and WOA-ELM reaches 99.67%, the accuracy rate is improved by 4.42% compared with the ELM method of the multi-feature extraction, and as can be seen from FIG. 4, the fitness curve of the WOA-ELM algorithm tends to be stable when the WOA-ELM algorithm is iterated to the 4 th time, and the rapid convergence performance of the WOA-ELM algorithm is verified. Therefore, the rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM is effective.

The invention provides a rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM, which is characterized in that a mixed fault feature set is formed by carrying out time domain analysis, frequency spectrum analysis and wavelet packet decomposition on a vibration signal of a rolling bearing. And then, applying local preserving projection to reduce the dimension of the mixed feature set and eliminate redundant features. And finally, introducing a whale algorithm to optimize network parameters of the extreme learning machine, and establishing a WOA-ELM rolling bearing diagnosis model to classify and diagnose faults.

The method effectively solves the problems of insufficient feature extraction, redundant feature information in multi-feature samples, insufficient stability caused by random generation of network parameters by an extreme learning machine and the like. The method has obvious advantages for improving the bearing fault diagnosis rate.

The invention realizes the diagnosis of the rolling bearing fault. Aiming at the difficulty in extracting the characteristics of the vibration signals, a multi-characteristic fusion method is adopted to extract characteristic information, and the dimension reduction is carried out on the characteristic set of the mixed domain, so that the original characteristics of the signals are extracted more comprehensively. The whale optimization algorithm is utilized to overcome the defects that the extreme learning machine is poor in stability and easy to fall into local optimum, and the accuracy and the stability of the algorithm are improved, so that the fault diagnosis rate is improved. The invention includes and is not limited to effect diagrams in the simulation experiment.

Claims (5)

1. A rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM is characterized by comprising the following steps:

step one, adopting official bearing data of Kaixi storage university as original vibration data;

performing time domain analysis, frequency spectrum analysis and wavelet packet decomposition on the original data, and extracting time domain, frequency domain and time-frequency domain characteristics;

step three, carrying out normalization processing on the mixed domain feature set obtained in the step two;

fourthly, carrying out dimension reduction on the normalized high-dimensional feature set by using a manifold learning LPP algorithm to obtain a low-dimensional feature sample set;

and fifthly, optimizing ELM network parameters by using a WOA algorithm, and inputting the low-dimensional characteristic data to a WOA-ELM model for fault diagnosis.

2. The rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM as claimed in claim 1, wherein the time domain features of the second step are dimensionless feature indicators including a peak factor, a pulse factor, a form factor, a margin factor and a kurtosis factor. And converting the time domain signal into a frequency domain signal by using FFT (fast Fourier transform), carrying out spectrum analysis, and selecting the mean frequency and the frequency divergence as frequency domain characteristics. And (4) performing three-layer wavelet packet decomposition by using db3 mother wavelets, and extracting the energy of the wavelet packets of the sub-bands with larger difference between the bands to serve as the time-frequency domain characteristics.

3. The rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM as claimed in claim 1, wherein the three-domain features obtained in step two are used to construct a mixed feature matrix Y, and then in order to avoid the large numerical value difference between the feature indexes from affecting the diagnosis of the classifier, the mixed feature matrix Y is normalized:

where max and min are the sample data maximum and minimum values.

4. The rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM as claimed in claim 1, wherein the high dimensional feature set obtained after normalization in step three is subjected to dimension reduction by LPP, and the specific steps are as follows: computationally deriving the transformation matrix A such that the data points x in the high-dimensional space_i(i1,2, …, m) to a low-dimensional spatial data point y_i(i ═ 1,2, …, m), and is expressed by the formula:

y_i＝A^Tx_i (1)

wherein the matrix A is obtained by minimizing the objective function by calculation, i.e.

Wherein W_ijAnd assigning values for the weight matrix between the connection nodes by adopting a thermonuclear mode.

Substituting equation (1) into (2) and adding constraint y^TDy 1, obtained

A＝arg min A^TXLX^TA (3)

Then, the formula (3) is converted into a generalized eigenvalue solving problem, namely

XLX^TA＝λXDX^TA (4)

The first i eigenvectors of equation (4) can thus be solved, i.e., a ═ a₀,a₁,…,a_l-1]。

The matrix A is solved and then is input into a formula (1) to obtain the mapping after dimension reduction, namely a low-dimension mixed feature set.

5. The rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM as claimed in claim 1, wherein the process of optimizing ELM by whale algorithm in step five is:

(1) and setting the whale population number and the algorithm iteration number, and initializing an input weight W and a hidden layer threshold b of the extreme learning machine to serve as an initial position vector of whales in the WOA.

(2) And calculating individual fitness values in the population, finding the optimal whale individual, and recording the position of the current optimal individual.

(3) And if the iteration times or the minimum fitness value is not met, updating the position between the whale and the target, and entering the next iteration.

(4) And when the conditions are met, the current optimal individual position of the whale is reserved, and the optimal parameters of the ELM model are obtained.

And then, obtaining a diagnosis model of the WOA-ELM, and dividing the low-dimensional mixed feature set into a training set and a testing set after setting a transfer function and the number of network layers of the extreme learning machine. And inputting the training set into a diagnosis model for training, and inputting the training set into a test set for verifying the diagnosis performance of the WOA-ELM.

Images