A bearing life prediction method based on a multi-stage canonical correlation analysis network

Abstract

The application discloses a bearing life prediction method based on a multi-stage canonical correlation analysis network, first acquires double-channel vibration signal data collected by bearing vibration x-direction and y-direction sensors, and carries out data preprocessing; obtains double-channel multi-band information by using a multi-band feature attention module, and carries out first-stage constraint on the double-channel multi-band information to obtain double-channel collaborative multi-band information, which is then respectively input into a hollow convolution spatial feature extraction module to extract spatial features; second-stage constraint is carried out on the double-channel spatial feature sequences obtained after extraction to obtain double-channel collaborative spatial feature sequences; the double-channel collaborative spatial feature sequences are respectively input into a convolution time sequence feature extraction module to extract time sequence features; third-stage constraint is carried out on the double-channel space-time feature sequences obtained after extraction to obtain double-channel collaborative space-time feature sequences; finally, a life prediction regression layer is constructed, the double-channel collaborative space-time feature sequences are input into the regression layer, and the residual life of the bearing is predicted.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of bearing remaining life prediction, in particular to a bearing life prediction method based on a multistage canonical correlation analysis network. BACKGROUND

[0002] As the most basic and core component of a rotating mechanical equipment, the health state of a bearing often determines whether the whole mechanical equipment can safely operate. Investigations show that more than 2 / 3 of rotating equipment failures are related to bearing failures. Therefore, the research on bearing health monitoring and life prediction methods and the accurate prediction of the remaining useful life (RUL) of bearings and other key components can greatly ensure the safe and reliable production of rotating equipment, reduce maintenance and downtime losses, and bring huge economic benefits to enterprises.

[0003] At present, relevant research institutions and scholars at home and abroad have carried out a large amount of research work on the RUL prediction of bearings. The RUL prediction methods studied can be mainly divided into three categories: a prediction method based on a physical model, a data-driven prediction method, and a hybrid method of the two. The prediction method based on the physical model needs a large amount of expert prior knowledge and is difficult to establish an accurate degradation model under complex working conditions. Thanks to the continuous progress of machine learning methods, the data-driven RUL prediction method has become a research hotspot in the current academic and industrial circles. Among them, the method based on deep learning is more and more concerned by researchers due to its strong feature extraction capability and prediction accuracy.

[0004] Most of the existing bearing RUL prediction methods based on deep learning directly input the double-channel original data into the deep learning model for training, ignoring the correlation between the two channels, and the original vibration signal contains a large amount of noise, so direct input will reduce the prediction accuracy and limit the generalization ability of the model. SUMMARY

[0005] The technical problem to be solved by the application is to provide a bearing life prediction method based on a multistage canonical correlation analysis network, which can automatically learn the deep degradation features of bearings, constrain the double-channel vibration signals of bearings at multiple levels, effectively improve the accuracy of bearing remaining life prediction, timely process and maintain the components, avoid deep damage to the mechanical equipment, prolong the service life of the bearings, and reduce the economic losses caused by bearing damage.

[0006] To solve the above technical problems, the technical scheme adopted by the application is as follows: a bearing life prediction method based on a multistage canonical correlation analysis network, comprising the following steps:

[0007] Step S1, acquire the double-channel vibration signal data collected by the bearing vibration x-direction and y-direction sensors, and perform data preprocessing; design and utilize a multi-band feature attention module to automatically screen bearing degradation sensitive information, and obtain double-channel multi-band information; design and utilize a canonical correlation analysis constraint module to perform first-level constraint on the screened double-channel multi-band information, and obtain double-channel collaborative multi-band information;

[0008] Step S2, input the double-channel collaborative multi-band information screened in step S1 into a hollow convolution spatial feature extraction module respectively to extract spatial features; utilize the canonical correlation analysis constraint module to perform second-level constraint on the double-channel spatial feature sequence obtained after extraction, and obtain a double-channel collaborative spatial feature sequence;

[0009] Step S3, input the double-channel collaborative spatial feature sequence obtained in step S2 into a convolution time sequence feature extraction module respectively to extract time sequence features; utilize the canonical correlation analysis constraint module to perform third-level constraint on the double-channel space-time feature sequence obtained after extraction, and obtain a double-channel collaborative space-time feature sequence;

[0010] Step S4, construct a life prediction regression layer, input the double-channel collaborative space-time feature sequence obtained in step S3 into the regression layer, and predict the bearing residual life.

[0011] Further improvement of the technical scheme of the application is that step S1 comprises the following specific steps:

[0012] Step S11, acquire double-channel vibration signal data through bearing vibration x-direction and y-direction sensors, and remove abnormal points to obtain x-direction signal input V x and y-direction signal input V y .

[0013] Step S12, the multi-band feature attention module first processes the original double-channel data respectively through an m-layer "db1" wavelet packet decomposition function, and the given input original data sample size is f l ∈R d×c (l∈1,2,…,N), after wavelet packet decomposition, 2 m frequency bands are obtained, and the length of each frequency band is L=f i / D; each input sample is decomposed into an L×D coefficient matrix as a model input, and finally x-direction and y-direction double-channel multi-band inputs are obtained respectively;

[0014] Step S13, utilize a global average pooling layer to aggregate information of the input, and the calculation formula is as follows:

[0015]

[0016] wherein, The output obtained by passing the i-th frequency band of the first sample through the global average pooling layer is represented as W i,j The j-th wavelet packet coefficient of the i-th frequency band is represented as W

[0017] Then, a l is input into two convolutional layers, the size of the convolution kernel is 1, and the number of feature maps is 2 m-1 and 2 m respectively; and the frequency band weight W hs is obtained through a sigmoid activation function σ The calculation formula is as follows:

[0018]

[0019]

[0020] Finally, the obtained weight is multiplied by the unprocessed multi-band time-frequency information to obtain double-channel weighted time-frequency information and

[0021] Step S14, a canonical correlation analysis constraint module is designed, the canonical correlation analysis constraint module includes a flattening layer and an aggregated fully connected layer, the number of neurons of the fully connected layer is set to r, and a canonical correlation loss is added to the total loss of model training during model training, wherein the optimization target of the canonical correlation loss is

[0022] min(-corr(X1,X2))

[0023]

[0024] Wherein, X1=W1 T H1+b1, is the input of the canonical correlation analysis, W1 and W2 are the weight matrices of the input signals corresponding to the channel x and the channel y respectively, b1 and b2 are the bias matrices, r1 and r2 are the regularization parameters, I is the unit matrix, and n is the sample number.

[0025] Step S15, using the canonical correlation analysis constraint module, the correlation of the double-channel weighted time-frequency information is constrained to obtain double-channel collaborative multi-band information and

[0026] Further improvement of the technical scheme of the application is that the step S2 includes the following specific steps:

[0027] Step S21, the dual-channel cooperative multi-band information obtained in step S15 is respectively input into a mixed hollow convolution network with three different hollow rates dr1, dr2 and dr3 cyclically stacked to automatically learn the correlation between different frequency bands, and dual-channel spatial feature sequences F1 and F2 are obtained;

[0028] Step S22, the dual-channel spatial feature sequences are subjected to correlation constraint by using a canonical correlation analysis constraint module, and dual-channel cooperative spatial feature sequences g cca (F1) and g cca (F2) are obtained.

[0029] The further improvement of the technical scheme of the application is that the step S3 comprises the following specific steps:

[0030] S31, the dual-channel cooperative spatial feature sequences obtained in step S22 still have a time sequence relationship, and convolution time sequence feature extraction modules are respectively used to learn time sequence features and perform space-time fusion, wherein the convolution time sequence feature extraction modules comprise three convolution layers, the convolution kernel sizes used are the same, and a maximum value pooling layer with the same pooling size is added after each convolution, and dual-channel space-time feature sequences TF1 and TF2 of bearing degradation are obtained.

[0031] S32, the dual-channel space-time feature sequences are subjected to correlation constraint by using a canonical correlation analysis constraint module, and dual-channel cooperative space-time feature sequences g cca (TF1) and

[0032] The further improvement of the technical scheme of the application is that the step S4 comprises the following specific steps:

[0033] S41, bearing residual life prediction is defined as a regression prediction problem;

[0034] S42, the dual-channel cooperative space-time feature sequences obtained in step S32 are spliced in the channel dimension, and then converted into a two-dimensional matrix and input into a regression layer with a mean absolute error loss function, and the trained model is used to evaluate the residual useful life of the test bearing; wherein the calculation formula of the mean absolute error is as follows: Wherein y i represents a true value, represents a predicted value, and the optimization target of the trained model is:

[0035]

[0036] Due to the adoption of the above technical scheme, the application has the following technical progress:

[0037] The application provides a bearing life prediction method based on a multi-level canonical correlation analysis network, which is different from existing bearing life prediction methods based on deep learning. First, bearing vibration x-direction and y-direction sensor signals are collected, and data preprocessing is performed on the signals respectively. The double-channel vibration signal samples obtained after preprocessing are filtered through an attention mechanism to obtain weighted double-channel signals, and the correlation between the two channels after weighting is constrained by a canonical correlation analysis module. Then, the weighted double-channel signals are input into a cavity convolution network to extract spatial features, and the extracted spatial feature sequence is constrained by a canonical correlation analysis. The double-channel spatial feature sequence is then input into a convolution time sequence feature extraction module to obtain a double-channel bearing space-time feature sequence, which is again constrained by a canonical correlation analysis. Finally, the bearing remaining life is predicted by a regression layer. The application can constrain the multi-level correlation of the bearing double-channel signals on the basis of extracting deep degradation features of the bearing, improve the accuracy and reliability of bearing life prediction, and thus maintain or replace the bearing in time to avoid deep damage to large components and reduce losses. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a flowchart of the application;

[0039] Figure 2 is a schematic diagram of a multi-band feature attention module of the application;

[0040] Figure 3 is a schematic diagram of a spatial and time sequence feature extraction module of the application;

[0041] Figure 4 is a schematic diagram of a canonical correlation analysis constraint module of the application. DETAILED DESCRIPTION

[0042] The application will be further described in detail below in combination with embodiments:

[0043] As shown in Figure 1 , a bearing life prediction method based on a multi-level canonical correlation analysis network comprises the following steps:

[0044] Step S1: Obtain double-channel vibration signal data collected by bearing vibration x-direction and y-direction sensors, and perform data preprocessing. Design and use a multi-band feature attention module to automatically filter bearing degradation sensitive information and obtain double-channel multi-band information. Design and use a canonical correlation analysis constraint module to constrain the first level of the filtered double-channel multi-band information and obtain double-channel collaborative multi-band information.

[0045] Step S1 specifically comprises the following steps:

[0046] Step S11, obtain double-channel vibration signal data through bearing vibration x-direction and y-direction sensors, and remove abnormal points to obtain x-direction signal input V x and y-direction signal input V y ;

[0047] Step S12, as shown in Figure 2 , the multi-band feature attention module first processes the original double-channel data respectively through an m-layer "db1" wavelet packet decomposition function, and the given input original data sample size is f l ∈R d×c (l∈1,2,…,N). After wavelet packet decomposition, 2 m bands are obtained, and the length of each band is L=f i / D; each input sample is decomposed into an L×D coefficient matrix as a model input, and finally x-direction and y-direction double-channel multi-band inputs are obtained respectively;

[0048] Step S13, use a global average pooling layer to aggregate information of the input, and the calculation formula is as follows:

[0049]

[0050] Among them, represents the output of the lth sample after the global average pooling layer, and W i,j represents the jth wavelet packet coefficient of the ith band;

[0051] Then input a l to two convolution layers, and the size of the convolution kernel is 1, and the number of feature maps is 2 m-1 and 2 m respectively; then obtain the band weight through the sigmoid activation function σ hs The calculation formula is as follows:

[0052]

[0053]

[0054] Finally, multiply the obtained weight with the unprocessed multi-band time-frequency information to obtain double-channel weighted time-frequency information and

[0055] Step S14, as shown in Figure 4 ​As shown, the canonical correlation analysis constraint module is designed, which includes a flattening layer and an aggregated fully connected layer, the number of neurons of the fully connected layer is set to r, and a canonical correlation analysis loss is added to the total loss of model training during model training, wherein the optimization target of the canonical correlation loss is

[0056] min(-corr(X1,X2))

[0057]

[0058] Wherein, X1=W1 T H1+b1, is the input of the canonical correlation analysis, W1 and W2 are the weight matrices of the input signals corresponding to the channel x and the channel y respectively, b1 and b2 are the bias matrices, r1 and r2 are the regularization parameters, I is the unit matrix, and n is the sample number;

[0059] Step S15, using the canonical correlation analysis constraint module to constrain the correlation of the double-channel weighted time-frequency information, to obtain double-channel collaborative multi-band information and

[0060] Step S2, as Figure 3 shown, the double-channel collaborative multi-band information screened in step S1 is respectively input into the hollow convolution spatial feature extraction module to extract spatial features; the double-channel spatial feature sequence obtained after extraction is subjected to a second level constraint by using the canonical correlation analysis constraint module, to obtain a double-channel collaborative spatial feature sequence;

[0061] Step S2 specifically includes the following steps:

[0062] Step S21, the double-channel collaborative multi-band information obtained in step S15 is respectively input into a mixed hollow convolution network with three different sizes of hole rates dr1, dr2 and dr3 stacked in a loop to automatically learn the correlation between different frequency bands, to obtain double-channel spatial feature sequences F1 and F2;

[0063] Step S22, using the canonical correlation analysis constraint module to constrain the correlation of the double-channel spatial feature sequence, to obtain a double-channel collaborative spatial feature sequence g cca (F1) and g cca (F2).

[0064] Step S3, the double-channel collaborative spatial feature sequence obtained in step S2 is respectively input into the convolution time sequence feature extraction module to extract time sequence features; the double-channel spatio-temporal feature sequence obtained after extraction is subjected to a third level constraint by using the canonical correlation analysis constraint module, to obtain a double-channel collaborative spatio-temporal feature sequence;

[0065] Step S3 specifically comprises the following steps:

[0066] S31, the double-channel cooperative spatial feature sequence obtained in step S22 still has a time sequence relationship, and a convolutional time sequence feature extraction module is used to learn time sequence features and perform space-time fusion, wherein the convolutional time sequence feature extraction module contains three convolutional layers, the convolutional kernel sizes used are the same, and a maximum pooling layer with the same size is added after each convolutional layer to reduce the model training parameter amount, aggregate information, and obtain double-channel space-time feature sequences TF1 and TF2 of bearing degradation;

[0067] S32, a canonical correlation analysis constraint module is used to constrain the correlation of the double-channel space-time feature sequences, and double-channel cooperative space-time feature sequences g cca (TF1) and

[0068] Step S4, a life prediction regression layer is constructed, the double-channel cooperative space-time feature sequences obtained in step S3 are input into the regression layer, and the remaining life of the bearing is predicted.

[0069] Step S4 specifically comprises the following steps:

[0070] S41, the prediction of the remaining life of the bearing is defined as a regression prediction problem;

[0071] S42, the double-channel cooperative space-time feature sequences obtained in step S32 are spliced in the channel dimension, and then converted into a two-dimensional matrix and input into a regression layer with a mean absolute error loss function, and the trained model is used to evaluate the remaining available life of the test bearing; wherein the calculation formula of the mean absolute error is as follows: Where y i represents the true value, represents the predicted value, and the optimization target of the trained model is:

[0072]

[0073] Example 1

[0074] The embodiment adopts LDK UER204 type rolling bearing, the data sampling frequency is 25.6 kHz, data is sampled once every 1 min, and the sampling time lasts 1.28 s each time. The embodiment uses three different working conditions for verification, wherein the first three bearings in each working condition are used as a training set, and the last two bearings are used as a test set. In order to effectively predict the remaining life of the bearing, the average result of ten repeated runs is used as the final prediction result in this experiment. Table 1 shows the life prediction results of the present application and other methods, and from the table, it can be seen that from the other variants of the convolutional neural network to the present application, the average absolute error MAE of the prediction is significantly reduced, the R2 value measuring the fitting degree of the predicted value and the true value is significantly improved, and the present application obtains enhanced life prediction performance. This is mainly due to the method proposed in the embodiment, which first uses wavelet packet decomposition and attention mechanism to separate from the original signal, and designs and uses a space-time feature extraction module to learn the deep degradation features in the dual-channel multi-band information, and at the same time, on the basis of extracting the deep degradation features of the bearing, the multi-level correlation constraint is performed on the dual-channel signal of the bearing, thereby improving the accuracy and reliability of the bearing life prediction.

[0075] Table 1 Comparison of bearing life prediction results of the present application and related methods

[0076]

Claims

1. A bearing life prediction method based on a multi-level canonical correlation analysis network, characterized in that: Includes the following steps: Step S1: Acquire dual-channel vibration signal data collected by sensors in the x and y directions of the bearing vibration and perform data preprocessing; design and utilize a multi-band feature attention module to automatically filter bearing degradation sensitive information to obtain dual-channel multi-band information; design and utilize a canonical correlation analysis constraint module to apply first-level constraints to the filtered dual-channel multi-band information to obtain dual-channel collaborative multi-band information. Step S1 includes the following specific steps: Step S11: Acquire dual-channel vibration signal data using bearing vibration sensors in the x and y directions, and remove outliers to obtain the x-direction signal input V. x and y-direction signal input V y ; Step S12: The multi-band feature attention module first processes the original dual-channel data separately using the m-layer "db1" wavelet packet decomposition function, given that the original input data sample size is f. l ∈R d×c (l∈1,2,…,N), after wavelet packet decomposition, a total of 2 m There are 1 frequency band, and the length of each frequency band is L = f. i / D; Each input sample is decomposed into an L×D coefficient matrix as the model input, and finally the dual-channel multi-band inputs in the x and y directions are obtained respectively; Step S13: Aggregate the input information using a global average pooling layer. The calculation formula is as follows: in, W represents the output obtained by passing the i-th frequency band of the l-th sample through the global average pooling layer. i,j This represents the coefficient of the j-th wavelet packet in the i-th frequency band; Then a l The input is fed into two convolutional layers, both with a kernel size of 1 and two feature maps. m-1 and 2 m Then activate the sigmoid function σ. hs Obtaining frequency band weights The calculation formula is as follows: Finally, the obtained weights are multiplied by the unprocessed multi-band time-frequency information to obtain the dual-channel weighted time-frequency information. and Step S14: Design a canonical correlation analysis constraint module. This module includes a flattened layer and an aggregated fully connected layer. The number of neurons in the fully connected layer is set to r. During model training, the canonical correlation analysis loss is added to the total training loss. The optimization objective of the canonical correlation loss is... min(-corr(X1,X2)) Where X1 = W1 T H1+b1, X2=W2 T H2+b2 is the input for canonical correlation analysis, W1 and W2 are the weight matrices of the input signals corresponding to channel x and channel y, respectively, b1 and b2 are the bias matrices, r1 and r2 are the regularization parameters, I is the identity matrix, and n is the number of samples. Step S15: Using the canonical correlation analysis constraint module, apply correlation constraints to the dual-channel weighted time-frequency information to obtain dual-channel cooperative multi-band information. and Step S2: Input the dual-channel collaborative multi-band information filtered in step S1 into the dilated convolution spatial feature extraction module to extract spatial features; use the canonical correlation analysis constraint module to apply a second-level constraint to the extracted dual-channel spatial feature sequence to obtain the dual-channel collaborative spatial feature sequence. Step S3: Input the dual-channel collaborative spatial feature sequence obtained in step S2 into the convolutional temporal feature extraction module to extract temporal features; use the canonical correlation analysis constraint module to apply the third-level constraint to the extracted dual-channel spatiotemporal feature sequence to obtain the dual-channel collaborative spatiotemporal feature sequence. Step S4: Construct a life prediction regression layer. Input the dual-channel collaborative spatiotemporal feature sequence obtained in step S3 into the regression layer to predict the remaining life of the bearing.

2. The bearing life prediction method based on a multi-level canonical correlation analysis network according to claim 1, characterized in that: Step S2 includes the following specific steps: Step S21: Input the dual-channel collaborative multi-band information obtained in step S15 into a hybrid dilated convolutional network composed of three stacked loops with different dilation rates dr1, dr2 and dr3 to automatically learn the correlation between different frequency bands and obtain dual-channel spatial feature sequences F1 and F2. Step S22: Using the canonical correlation analysis constraint module, apply correlation constraints to the dual-channel spatial feature sequence to obtain the dual-channel cooperative spatial feature sequence g. cca (F1) and g cca (F2) 3. The bearing life prediction method based on a multi-level canonical correlation analysis network according to claim 2, characterized in that: Step S3 includes the following specific steps: The dual-channel collaborative spatial feature sequences obtained in steps S31 and S22 still have temporal relationships. Temporal features are learned using the convolutional temporal feature extraction module and spatiotemporal fusion is performed. The convolutional temporal feature extraction module contains three convolutional layers with the same kernel size. A maximum pooling layer with the same pooling size is added after each convolutional layer to obtain the dual-channel spatiotemporal feature sequences TF1 and TF2 of bearing degradation. S32. Using the canonical correlation analysis constraint module, correlation constraints are applied to the dual-channel spatiotemporal feature sequence to obtain the dual-channel cooperative spatiotemporal feature sequence g. cca (TF1) and 4. The bearing life prediction method based on a multi-level canonical correlation analysis network according to claim 3, characterized in that: Step S4 includes the following specific steps: S41. Define bearing remaining life prediction as a regression prediction problem; S42. The dual-channel collaborative spatiotemporal feature sequence obtained in step S32 is concatenated along the channel dimension and then transformed into a two-dimensional matrix. This matrix is ​​then input into a regression layer with a mean absolute error loss function. The trained model is used to evaluate the remaining usable life of the test bearing. The formula for calculating the mean absolute error is shown below: Where y i Represents the actual value. The predicted value is the target value. The optimization objective of the training model is:

Images