Rolling bearing fault diagnosis method based on autoregression data generation

Through the autoregressive data generation model CSQ-VAE based on wavelet packet transformation and cascade random quantization, the problem of unbalanced sample number in rolling bearing fault diagnosis is solved, high-quality pseudo-samples are generated, fault diagnosis performance is improved, and a small number of faults is achieved Effective diagnosis of samples and large amounts of normal samples.

CN118936887APending Publication Date: 2024-11-12JIANGNAN UNIV
View PDF 0 Cites 6 Cited by

Patent Information

Application Number
CN202411135347.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

In actual industrial production, the number of samples is unbalanced in rolling bearing fault diagnosis, normal data accounts for the majority, and the number of fault samples is limited, making it difficult to effectively carry out fault diagnosis methods based on deep learning, and existing data generation methods such as GAN during training Convergence is difficult and computationally expensive.

Method used

Using the autoregressive data generation model CSQ-VAE based on wavelet packet transformation and cascaded random quantization, pseudo-samples are generated through a hierarchical encoder, a cascaded random quantization module, a parallel attention-guided quantization module and a multi-scale quantization feature fusion module, and a pseudo-sample sample is generated by the hierarchical encoder, a cascaded random quantization module, and a multi-scale quantization feature fusion module. To expand the unbalanced rolling bearing fault diagnosis data set, improve the performance of the model.

Benefits of technology

It improves the generation quality and diversity of pseudo-samples, can better extract features from the original signal, realizes rolling bearing fault diagnosis in the case of unbalanced sample number, and improves the interpretability and diagnostic performance of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118936887A_ABST
    Figure CN118936887A_ABST
Patent Text Reader

Abstract

The invention discloses a rolling bearing fault diagnosis method based on autoregression data generation, and relates to the technical field of bearing fault diagnosis, and the method comprises the steps: obtaining a large number of labeled vibration signals in a normal working state of a bearing and a small number of labeled vibration signals in a fault working state of the bearing; an autoregression data generation model CSQ-VAE is designed and comprises a parallel attention guidance quantization module and a multi-scale quantization feature fusion module so as to better integrate latent variable features of different scales. In order to increase diversity of pseudo samples, wavelet packet transformation decomposition is performed on original input signals to obtain high and low frequency information, the high and low frequency information is used as training data of CSQ-VAE, and then the generated high and low frequency information is reconstructed to obtain the pseudo samples. And storing the trained generation model, and generating a pseudo sample of a fault signal through the model, thereby expanding the unbalanced data set, and finally realizing the fault diagnosis of the rolling bearing under the unbalanced sample of a small number of fault samples and a large number of normal samples.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of bearing fault diagnosis, and in particular to a rolling bearing fault diagnosis method based on autoregressive data generation. Background Art

[0002] Rolling bearings are important components in mechanical equipment. They ensure smooth rotation and reduce friction between mechanical parts. However, over time and with continuous use, bearings may develop various problems. If not discovered and resolved in time, it may lead to degraded equipment performance or even serious mechanical failures. Deep learning-based fault diagnosis methods have attracted much attention due to their powerful end-to-end feature learning capabilities. However, the learning ability of deep learning depends largely on whether the training samples are sufficient. Unfortunately, in actual industrial production, machines work in a healthy state most of the time, which means that it is extremely difficult to obtain enough fault samples. In addition, in actual work, it is also a challenge to obtain a sufficient number of labeled fault samples, and this work is time-consuming and labor-intensive, which hinders the progress of the work.

[0003] The above problems also mean that the collected data set is unbalanced, with normal data accounting for the majority and a limited number of fault samples, which brings great difficulties to the fault diagnosis method based on deep learning. Therefore, it is of great value to study how to use vibration signal samples to realize rolling bearing fault diagnosis when the number of samples is unbalanced. In order to meet these challenges, the mainstream data enhancement methods include data sampling and data generation. The goal of data generation technology is to capture the basic characteristics of the original signal and generate pseudo samples with probability distribution similar to the original signal. Typical methods include a series of improved methods represented by generative adversarial networks (GANs) and variational autoencoders (VAEs).

[0004] The current mainstream data generation methods are basically based on GAN or VAE combined with GAN. However, GAN has difficulty converging during training, has high computational cost and is difficult to evaluate. VAE models the uncertainty of data by introducing the probability distribution of latent variables, thereby capturing the underlying structure of the data. This enables it to have better performance when processing complex data. This means that the inherent structure and laws of the data can be clarified by studying the latent variable space, thereby improving the interpretability of the model. Therefore, this application draws on the VAE series methods in the field of image generation to study a more effective autoregressive data generation method that can be directly used for one-dimensional time domain signals for rolling bearing fault diagnosis under unbalanced samples. Summary of the invention

[0005] In view of the above problems and technical requirements, the inventors have proposed a rolling bearing fault diagnosis method based on autoregressive data generation, thereby solving the technical problem of unbalanced sample quantity in the bearing diagnosis process in the prior art. The technical solution of the present invention is as follows:

[0006] A rolling bearing fault diagnosis method based on autoregressive data generation comprises the following steps:

[0007] Step 1: Collect a certain amount of labeled vibration signals under normal working conditions of the rolling bearing, which are normal samples; collect a small amount of labeled vibration signals under faulty working conditions of the rolling bearing, which are fault samples; the vibration signal is one-dimensional time series data, and its label is set according to the fault category of the vibration signal source;

[0008] Step 2: A data set is prepared by randomly windowing a certain length of data samples from the labeled vibration signal. Each sample is preprocessed by maximum and minimum normalization. The data set is divided into a training set and a test set. The training set is an unbalanced data set consisting of a small number of fault samples and a sufficient number of normal samples, and the test set is a balanced data set consisting of a certain number of fault samples and normal samples.

[0009] Step 3: Build a CSQ-VAE model for rolling bearing fault diagnosis under unbalanced samples. The CSQ-VAE model is an autoregressive data generation model based on wavelet packet transform and cascaded random quantization, including a hierarchical encoder for compressing the dimension of the original input signal, a cascaded random quantization module, a parallel attention-guided quantization module, a multi-scale quantization feature fusion module, and a decoder for reconstructing the original input signal.

[0010] The input of the hierarchical encoder is the high and low frequency signals obtained after the original input signal is decomposed by wavelet packets, and the output is feature vectors of different sizes; the cascade random quantization module receives feature vectors of different sizes as input, adopts a random quantization strategy, and outputs latent variable quantization features that correspond to the input scale one by one; the parallel attention-guided quantization module adopts a parallel attention mechanism in the time domain and channel domain to weight the quantization features of latent variables of different scales respectively; the multi-scale quantization feature fusion module fuses the weighted quantization features of latent variables of different scales through upsampling; the decoder reconstructs the fused latent variable quantization features to obtain high and low frequency information, and then uses wavelet packets to reconstruct a one-dimensional time series signal of the same size as the hierarchical encoder input, which is the reconstructed pseudo sample;

[0011] Step 4: Based on the autoregressive data generation model CSQ-VAE in step 3, the fault samples are classified according to the fault category, and after wavelet packet decomposition, they are used as input signals to train CSQ-VAE. The mean square error is used as the loss function to evaluate the distribution difference between the pseudo samples and the real fault samples, and the CSQ-VAE model weight files obtained by training different fault categories are saved respectively;

[0012] Step 5: Use the CSQ-VAE model weight files of different fault categories trained in step 4 to generate autoregressive data for the fault samples respectively, and then use the wavelet packet to reconstruct the generated high and low frequency information to obtain pseudo samples. The generated pseudo samples and real fault samples are used as fault samples together, thereby expanding the unbalanced data set to improve the performance of the fault diagnosis model;

[0013] Step 6: Use the expanded data set obtained in step 5 as the training set of the fault diagnosis model, and verify the performance of the CSQ-VAE model through the test set to obtain the fault diagnosis results.

[0014] Its further technical solution is that in step three:

[0015] Wavelet packet decomposition is used to perform wavelet transform on the original input signal to obtain high-frequency information containing local features of the original input signal and low-frequency information containing global features. The wavelet transform is implemented using Haar wavelet, which is defined as shown in the following formula:

[0016]

[0017] Where ψ(t) represents the wavelet scaling function, and h(t) represents the wavelet function;

[0018] The hierarchical encoder consists of three encoders connected in sequence, and both the encoder and the decoder include a convolution layer, a pooling layer, a batch normalization layer, and an attention mechanism that uses parallelism in the time domain and channel domain, and introduces a residual structure; each encoder is used to perform feature encoding on the input data after wavelet packet decomposition, and obtains feature vectors of 1×256, 1×64, and 1×32 respectively;

[0019] The parallel attention-guided quantization module calculates attention on the quantized features of latent variables of different scales after random quantization in the time domain and channel domain respectively; in the channel dimension, the weights of latent variable spaces of different scales are calculated using softmax respectively, so that the features of different latent variable spaces can be better focused on; in the time domain dimension, due to the different scales of latent variable space features, the latent space features are first downsampled to the same dimension, and then softmax is used to calculate the importance of latent space features of different scales in the time dimension, thereby assigning different weights; finally, the latent variable space features weighted by the channel dimension and the time domain dimension are added together to obtain the final attention-guided latent variable space features of different scales;

[0020] Assume Q t , Q m and Q b Represent the spatial feature representation of latent variables of different scales, Q t * , Q m * and Q b * represent the spatial feature representations of latent variables of different scales guided by attention, then Q t * , Q m * and Q b * Calculate as follows:

[0021] Q t * =Q t ×CW t +Q t ×TW t ,

[0022] Q m * =Q m ×CW m +Ups(Downs(Q m )×TW m ),

[0023] Q b * =Q b ×CW b +Ups(Downs(Q b )×TW b );

[0024] CW t , CW b and CW m Respectively represent the weights of the channel dimensions; TW t、TW b and TW m They represent the weights of the time domain dimensions respectively; Downs represents the method using maximum pooling and average pooling, and Ups is a transposed convolution with a step size of 2 or 4;

[0025] In the multi-scale quantization feature fusion module, assuming that Q * is the fused latent variable quantized feature representation as the decoder input, then Q * The calculation formula is as follows:

[0026] Q mu * =Conv([Ups(Q t * ),Q m * ]),

[0027] Q * =Conv([Ups(Q mu * ),Q b * ]);

[0028] Conv stands for convolution, and its convolution kernel is 1, which is used for dimensionality reduction.

[0029] The further technical solution is that in step 4, the model parameters are initialized before training, including setting the training rounds, model iteration times, batch size and learning rate of the autoregressive data generation model; the metric loss function of the network is expressed as follows:

[0030]

[0031] where y i , They represent real fault samples and pseudo samples respectively, and n represents the length of sample data.

[0032] Its further technical solution is that in step 2:

[0033] The size of the sliding window is 2048, and the interception method is random interception; the category labels of the labeled vibration signals in the data set are set to 1,...,T, where T is the number of rolling bearing fault types. The maximum and minimum normalization performs the following operations on each data:

[0034]

[0035] Where x is the original vibration signal data, x′ is the sample after preprocessing, min represents the minimum value of each data, max represents the maximum value of each data, and each data refers to sample data of a certain length intercepted by a random sliding window.

[0036] Its further technical solution is that the method further comprises:

[0037] Step 4 uses mean square error as the loss function of the generative model CSQ-VAE, while step 6 classifies the expanded data set and uses cross entropy as the loss function. The network training of both minimizes the corresponding loss function and selects the Adam optimizer for gradient back propagation to update the network parameters. The iterative process of the Adam optimizer is as follows:

[0038]

[0039] In the formula, m k and n k is the objective function gradient g k The first-order matrix and second-order matrix of , k represents the current iteration batch, k-1 represents the previous iteration batch; β1 and β2 represent the matrix exponential decay rate, and It is for m k and n k Correction; θ represents the network parameters, η represents the learning rate, and ε is a constant;

[0040] The learning rate is dynamically clipped, and the upper and lower bounds of the learning rate are given to stabilize the late model convergence.

[0041] Its further technical solution is that in step six, the fault diagnosis model is implemented based on a one-dimensional signal fault diagnosis classifier in the form of WDCNN, MCNN_LSTM, RDD_WDCNN or MSDARN.

[0042] The beneficial technical effects of the present invention are:

[0043] 1) In the rolling bearing fault diagnosis method based on autoregressive data generation proposed in this application, a cascaded autoregressive generation model CSQ-VAE for high-dimensional one-dimensional signals is given. The hierarchical structure can better learn the multi-scale deep features of the input signal, and random quantization is introduced to accelerate the convergence of the model.

[0044] 2) In order to improve the quality and diversity of pseudo sample generation and facilitate feature extraction from the original signal, this application uses wavelet packet transform (WPT) to decompose the original input signal to obtain high-frequency and low-frequency signals as training data for the generation model. These data are then used to reconstruct the generated high-frequency and low-frequency information to obtain pseudo samples to expand the unbalanced data set, and finally realize rolling bearing fault diagnosis under unbalanced samples with a small number of fault samples and a large number of normal samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1This is a structural diagram of an autoregressive data generation model based on wavelet packet transform and cascade random quantization provided by the present application.

[0046] Figure 2 This is a structural diagram of the parallel attention-guided quantization module provided in this application.

[0047] Figure 3 This is a flow chart of autoregressive data generation based on wavelet packet transform provided by this application.

[0048] Figure 4 This is a fault diagnosis flowchart for data set expansion using the CSQ-VAE model provided in this application. DETAILED DESCRIPTION

[0049] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings.

[0050] The present application provides a rolling bearing fault diagnosis method based on autoregressive data generation, and the specific implementation of the method includes the following steps:

[0051] Step 1: Collect a certain amount of labeled vibration signals under the normal working state of the rolling bearing, which are recorded as normal samples. Collect a small amount of labeled vibration signals under the faulty working state of the rolling bearing, which are recorded as fault samples. The vibration signal is a one-dimensional time series data, and its label is set according to the fault category of the vibration signal source. In this embodiment, the labeled fault categories are four working states under three different speeds (600r, 800r and 1000r), and the fault types are inner ring fault IR, outer ring fault OR and rolling element fault RB under three speeds, respectively. A total of 12 vibration signals under normal working state NC at three speeds are collected, that is, T=12. The 12 categories correspond to labels 0-11, and unique hot encoding is used during training. For example, if it is a label "1", the corresponding unique hot encoding is [010000000000].

[0052] Step 2: Randomly window the collected vibration signal data and repeatedly intercept data samples with a length of 2048. Perform maximum and minimum normalization preprocessing on each sample to make a training set and a test set. The training set is an unbalanced data set consisting of a small number of fault samples and a sufficient number of normal samples, and the test set is a balanced data set consisting of a certain number of fault samples and normal samples. The divided data sets are shown in Table 1.

[0053] Table 1 Dataset

[0054]

[0055] Step 3: Build a CSQ-VAE model for rolling bearing fault diagnosis under unbalanced samples.

[0056] The CSQ-VAE (Variational Autoencoder with Cascaded StochasticQuantization) model is an autoregressive data generation model based on wavelet packet transform and cascaded stochastic quantization. This embodiment uses the Pytorch deep learning framework to build the model. Figure 1 As shown, the CSQ-VAE model includes a hierarchical encoder 1 for compressing the dimension of the original input signal, a cascaded random quantization module 2, a parallel attention-guided quantization module 3, a multi-scale quantization feature fusion module 4, and a decoder 5 for reconstructing the original input signal.

[0057] The hierarchical encoder 1 includes three encoders connected in sequence, and the decoder 5 includes a convolution layer, a pooling layer, a batch normalization layer, and an attention mechanism that uses parallelism in the time domain and channel domain, and introduces a residual structure. The encoder input is the high and low frequency signals obtained after the original input signal is decomposed by wavelet packets, and the feature vectors of 1×256, 1×64 and 1×32 are output respectively after feature encoding. The cascade random quantization module 2 receives feature vectors of different sizes as input, adopts a random quantization strategy, and outputs latent variable quantization features that correspond to the input scale one by one. The random quantization strategy is implemented based on the existing technical solution and will not be introduced in detail here.

[0058] like Figure 2 As shown, the parallel attention-guided quantization module 3 adopts a parallel attention mechanism in the time domain and channel domain to weight the quantization features of latent variables of different scales respectively. In the channel dimension Channel Attention, the weights of latent variable spaces of different scales are calculated using softmax respectively. In the time domain dimension Time Domain Attention, due to the different scales of the latent variable space features, the latent space features are first downsampled to the same dimension, and then softmax is used to calculate the importance of latent space features of different scales in the time dimension, thereby assigning different weights. Finally, the latent variable space features weighted by the channel dimension and the time domain dimension are added to obtain the final latent variable space features of different scales guided by attention. In the figure, L t , L m , L b is the output of the cascaded random quantization module. The AMMP module represents the weighted average of the maximum pooling and average pooling means in the channel dimension. t , Q m and Q b Represent the spatial feature representation of latent variables of different scales, Q t * , Q m * and Q b* represent the spatial feature representations of latent variables of different scales guided by attention, then Q t * , Q m * and Q b * Calculate as follows:

[0059] Q t * =Q t ×CW t +Q t ×TW t ,

[0060] Q m * =Q m ×CW m +Ups(Downs(Q m )×TW m ),

[0061] Q b * =Q b ×CW b +Ups(Downs(Q b )×TW b );

[0062] CW t , CW b and CW m Respectively represent the weights of the channel dimensions; TW t 、TW b and TW m They represent the weights of the time domain dimensions respectively; Downs represents the method of using maximum pooling and average pooling, and Ups is a transposed convolution with a step size of 2 or 4.

[0063] The multi-scale quantitative feature fusion module 4 fuses the weighted quantitative features of latent variables of different scales through upsampling. Assuming Q * is the fused latent variable quantized feature representation as the decoder input, then Q * The calculation formula is as follows:

[0064] Q mu * =Conv([Ups(Q t * ),Q m * ]),

[0065] Q * =Conv([Ups(Q mu* ),Q b * ]);

[0066] Conv stands for convolution, and its convolution kernel is 1, which is used for dimensionality reduction.

[0067] The decoder 5 reconstructs the quantized features of the fused latent variables to obtain high and low frequency information, and then uses wavelet packets to reconstruct a one-dimensional time series signal of the same size as the encoder input, which is the reconstructed pseudo sample. The original input signal is sequentially decomposed by wavelet packets 6, CSQ-VAE model and wavelet packet reconstruction 7. Figure 3 As shown in the figure, L represents the sample length and B represents the batch size. After wavelet packet decomposition, high-frequency information of local features of the original input signal (as shown in the first row in the figure) and low-frequency information of global features (as shown in the second row in the figure) are obtained. After decomposition, the sample length becomes L / 2.

[0068] Step 4: Classify the fault samples according to the fault category, decompose them into wavelet packets and use them as input signals to train the CSQ-VAE model constructed in step 3. The mean square error is used as the loss metric function to evaluate the distribution difference between the generated pseudo samples and the real fault samples, and save the CSQ-VAE model weight files obtained by training different fault categories. The loss function is expressed as follows:

[0069]

[0070] where y i , They represent real fault samples and pseudo samples respectively, and n represents the length of sample data.

[0071] Before training, the CSQ-VAE model parameters were initialized, the learning rate was set to 0.001, the Adam optimizer was used, the batch size was 64, and in order to prevent overfitting of the source domain, the number of training rounds in the training phase was finally determined to be 500 epochs through experimental adjustment.

[0072] Step 5: Use the CSQ-VAE model weight files of different fault categories trained in step 4, such as Figure 3As shown, autoregressive data is generated for the fault samples respectively, and then the high and low frequency information generated by wavelet packet reconstruction is used to obtain pseudo samples, and the generated pseudo samples and real fault samples are used as fault samples together, so as to expand the data of the unbalanced data set to improve the performance of the fault diagnosis model. In this embodiment, since the wavelet packet transform can better extract the local and global features of the original input signal, and reduce the dimension of the input data (the dimension is reduced from 2048 to 1024), compared with directly using the original input signal to generate pseudo samples, the use of wavelet packet transform can increase the diversity and generation quality of the generated pseudo samples, and at the same time improve the training efficiency of the data generation model.

[0073] Step 6: Use the expanded data set obtained in step 5 as the training set of the fault diagnosis model. The expansion process is as follows: Figure 4 As shown, a one-dimensional signal fault diagnosis classifier such as WDCNN, MCNN_LSTM, RDD_WDCNN or MSDARN can be used for training, and the performance of the CSQ-VAE model can be verified through the test set to obtain the fault diagnosis results after the data set is expanded.

[0074] By adopting the above method provided by the present application, a small number of labeled fault vibration signals and sufficient labeled normal vibration signal samples under all working conditions of rolling bearings are obtained. First, the vibration signal is normalized to the maximum value to reduce the impact of signal amplitude fluctuations caused by different speeds on the fault diagnosis model. Secondly, an autoregressive data generation model based on wavelet packet transform and cascade random quantization is constructed, and the high and low frequency information after wavelet packet decomposition is used as input. The mean square error is taken as the loss metric function to evaluate the distribution difference between the generated pseudo-samples and the real samples, and the trained model weight file is saved. Then, autoregressive data generation is performed on the fault samples respectively, and the high and low frequency information generated is reconstructed using wavelet packets to obtain pseudo-samples. The generated pseudo-samples and real fault samples are used together as fault samples, thereby expanding the data set of unbalanced rolling bearings, and realizing rolling bearing fault diagnosis under an unbalanced data set using a small number of fault samples and a large number of normal samples.

[0075] The above is only a preferred embodiment of the present application, and the present invention is not limited to the above embodiments. It is understood that other improvements and changes directly derived or associated by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included in the protection scope of the present invention.

Claims

1. A rolling bearing fault diagnosis method based on autoregressive data generation, characterized in that: The method comprises: Step 1: Collect a certain amount of labeled vibration signals under normal working conditions of the rolling bearing, which are normal samples; collect a small amount of labeled vibration signals under faulty working conditions of the rolling bearing, which are fault samples; the vibration signal is one-dimensional time series data, and its label is set according to the fault category of the vibration signal source; Step 2: A data set is prepared by randomly windowing a data sample of a certain length from the labeled vibration signal, and each sample is preprocessed by maximum and minimum normalization; the data set is divided into a training set and a test set, wherein the training set is an unbalanced data set consisting of a small number of fault samples and a sufficient number of normal samples, and the test set is a balanced data set consisting of a certain number of fault samples and normal samples; Step 3: Build a CSQ-VAE model for rolling bearing fault diagnosis under unbalanced samples. The CSQ-VAE model is an autoregressive data generation model based on wavelet packet transform and cascaded random quantization, including a hierarchical encoder for compressing the dimension of the original input signal, a cascaded random quantization module, a parallel attention-guided quantization module, a multi-scale quantization feature fusion module, and a decoder for reconstructing the original input signal. The input of the hierarchical encoder is the high- and low-frequency signals obtained by decomposing the original input signal through wavelet packets, and the output is feature vectors of different sizes; the cascade random quantization module receives the feature vectors of different sizes as input, adopts a random quantization strategy, and outputs latent variable quantization features corresponding to the input scale one by one; the parallel attention-guided quantization module adopts a parallel attention mechanism in the time domain and channel domain to weight the quantization features of latent variables of different scales respectively; the multi-scale quantization feature fusion module fuses the weighted quantization features of latent variables of different scales through upsampling; the decoder reconstructs the fused latent variable quantization features to obtain high- and low-frequency information, and then uses wavelet packets to reconstruct a one-dimensional time series signal of the same size as the hierarchical encoder input, which is the reconstructed pseudo sample; Step 4: Based on the autoregressive data generation model CSQ-VAE described in step 3, the fault samples are classified according to the fault categories, and are used as input signals to train CSQ-VAE after wavelet packet decomposition. The mean square error is used as the loss function to evaluate the distribution difference between the pseudo samples and the real fault samples, and the CSQ-VAE model weight files obtained by training different fault categories are saved respectively; Step 5: Using the CSQ-VAE model weight files of different fault categories trained in step 4, autoregressive data generation is performed on the fault samples respectively, and then the high and low frequency information generated by wavelet packet reconstruction is used to obtain pseudo samples. The generated pseudo samples and real fault samples are used together as fault samples, thereby expanding the unbalanced data set to improve the performance of the fault diagnosis model; Step 6: Use the expanded data set obtained in step 5 as the training set of the fault diagnosis model for training, and verify the performance of the CSQ-VAE model through the test set to obtain the fault diagnosis result.

2. The rolling bearing fault diagnosis method based on autoregressive data generation according to claim 1 is characterized in that: In step three: The wavelet packet decomposition is used to perform wavelet transform on the original input signal to obtain high-frequency information containing local features and low-frequency information containing global features of the original input signal. The wavelet transform is implemented using Haar wavelet, which is defined as shown in the following formula: Where ψ(t) represents the wavelet scaling function, and h(t) represents the wavelet function; The hierarchical encoder includes three encoders connected in sequence, and the decoder includes a convolution layer, a pooling layer, a batch normalization layer, and an attention mechanism that uses parallelism in the time domain and channel domain, and introduces a residual structure; each of the encoders is used to perform feature encoding on the input data after wavelet packet decomposition, and obtains feature vectors of 1×256, 1×64, and 1×32 respectively; The parallel attention-guided quantization module performs attention calculations on the quantized features of latent variables of different scales after random quantization in the time domain and channel domain respectively; in the channel dimension, the weights of latent variable spaces of different scales are calculated using softmax respectively; in the time domain dimension, due to the different scales of latent variable space features, the latent space features are first downsampled to the same dimension, and then softmax is used to calculate the importance of latent space features of different scales in the time dimension, thereby assigning different weights; finally, the latent variable space features weighted by the channel dimension and the time domain dimension are added to obtain the final attention-guided latent variable space features of different scales; Assume Q t , Q m and Q b Represent the spatial feature representation of latent variables of different scales, Q t * , Q m * and Q b * represent the spatial feature representations of latent variables of different scales guided by attention, then Q t * , Q m * and Q b * Calculate as follows: Q t * =Q t ×CW t +Q t ×TW t , Q m * =Q m ×CW m +Ups(Downs(Q m )×TW m ), Q b * =Q b ×CW b +Ups(Downs(Q b )×TW b ); CW t , CW b and CW m Respectively represent the weights of the channel dimensions; TW t 、TW b and TW m They represent the weights of the time domain dimensions respectively; Downs represents the method using maximum pooling and average pooling, and Ups is a transposed convolution with a step size of 2 or 4; In the multi-scale quantization feature fusion module, it is assumed that Q * is the fused latent variable quantized feature representation as the decoder input, then Q * The calculation formula is as follows: Q mu * =Conv([Ups(Q t * ),Q m * ]), Q * =Conv([Ups(Q mu * ),Q b * ]); Conv stands for convolution, and its convolution kernel is 1, which is used for dimensionality reduction.

3. The rolling bearing fault diagnosis method based on autoregressive data generation according to claim 1 is characterized in that: In the step 4, the model parameters are initialized before training, including setting the training rounds, model iteration times, batch size and learning rate of the autoregressive data generation model; the metric loss function of the network is expressed as follows: where y i , represent the real fault sample and the pseudo sample respectively, and n represents the sample data length.

4. The rolling bearing fault diagnosis method based on autoregressive data generation according to claim 1 is characterized in that: In the step 2: The size of the sliding window is 2048, and the interception method is random interception; the category labels of the labeled vibration signals in the data set are set to 1,...,T, where T is the number of rolling bearing fault types, and the maximum and minimum normalization performs the following operations on each data: Where x is the original vibration signal data, x′ is the sample after preprocessing, min represents the minimum value of each data, max represents the maximum value of each data, and each data refers to sample data of a certain length intercepted by a random sliding window.

5. The rolling bearing fault diagnosis method based on autoregressive data generation according to claim 1 is characterized in that: The method further comprises: In step 4 and step 6, the network training is performed by minimizing the corresponding loss function and selecting the Adam optimizer for gradient back propagation to update the network parameters. The iterative process of the Adam optimizer is as follows: In the formula, m k and n k is the objective function gradient g k The first-order matrix and second-order matrix of , k represents the current iteration batch, k-1 represents the previous iteration batch; β1 and β2 represent the matrix exponential decay rate, and It is for m k and n k Correction; θ represents the network parameters, η represents the learning rate, and ε is a constant; The learning rate is dynamically clipped, and the upper and lower bounds of the learning rate are given to stabilize the late model convergence.

6. The rolling bearing fault diagnosis method based on autoregressive data generation according to claim 1 is characterized in that: In step six, the fault diagnosis model is implemented based on a one-dimensional signal fault diagnosis classifier in the form of WDCNN, MCNN_LSTM, RDD_WDCNN or MSDARN.

Citation Information

Cited By

  • Bearing fault diagnosis method based on wavelet transform and mixed attention mechanism

    CN120277463A

  • Rolling bearing fault diagnosis method and system

    CN120873874A

  • A method and system for diagnosing faults in rolling element bearings

    CN120873874B

  • Cable multi-mode anomaly detection and positioning method based on layered attention VAE

    CN121276242A

  • Cable multi-modal anomaly detection and localization method based on hierarchical attention VAE

    CN121276242B