Bearing fault diagnosis method and system, storage medium and electronic equipment

Abstract

The invention provides a bearing fault diagnosis method and system, a storage medium and electronic equipment. The method comprises the following steps: acquiring source domain data and target domain data; denoising the source domain data and the target domain data; extracting features of the source domain data and the target domain data to obtain an initial feature vector; performing initialization processing on the initial feature vector; iteratively updating nodes of the source domain data and the target domain data to obtain a source domain graph and a target domain graph, and calculating a difference value; calculating the similarity loss of the two types, calculating the real distribution probability and the predicted distribution probability of the bearing type, obtaining a cross entropy loss function, obtaining a maximum mean value difference function, and calculating a total loss function; carrying out gradient calculation on parameters of the metric function, the source domain graph, the target domain graph, the nodes and the graph kernel model; and supervising the result, and performing joint optimization on the graph kernel model in combination with a graph kernel loss function. According to the method, the graph morphological characterization difference between the source domain graph and the target domain graph is reduced, and the generalization ability of the recognition model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fault diagnosis, and particularly relates to a bearing fault diagnosis method, system, storage medium and electronic device. Background Art

[0002] The axle box bearing is one of the most important rotating components in a rail train, and its health status directly affects the performance and service life of the equipment. Early diagnosis of bearing faults can effectively improve the accuracy and timeliness of fault detection, reduce operation and maintenance costs, and improve production efficiency.

[0003] In the prior art, the geometric features between the vibration signals of the axle box bearing are not effectively utilized. Traditional domain adaptation methods mostly rely on distribution metrics in the statistical sense (such as MMD, LMMD, etc.). Such explicit metric functions have limited ability to characterize the non-linear and non-stationary characteristics of vibration signals under complex working conditions. Especially when there are significant modal differences between the source domain and the target domain, it is difficult to effectively capture deep-level domain-invariant features. Secondly, the existing feature alignment strategies mostly focus on shallow statistical matching and lack the ability to model the inherent physical correlations of vibration signals, resulting in being easily interfered by local pseudo-correlated features during cross-device and cross-condition migration. Summary of the Invention Aiming at the deficiencies of the prior art, the purpose of the present invention is to provide a bearing fault diagnosis method, aiming to solve the technical problems mentioned in the background art.

[0004] To achieve the above purpose, the present invention is implemented by the following technical solutions: A bearing fault diagnosis method includes the following steps: Obtain the signal data of the bearing in a known environment, and use the signal data in the known environment as source domain data. Obtain the signal data of the bearing in other environments, and use the signal data in other environments as target domain data; Perform denoising processing on the source domain data and the target domain data; Extract features from the denoised source domain data and target domain data through a deep residual neural network combined with a self-attention mechanism to obtain an initial feature vector; Regard the initial feature vector as a node, and perform weighted processing on the node in the time channel and the frequency channel to initialize the label embedding of the node; Introduce a graph kernel model to iteratively update the nodes after label embedding initialization in the source domain data and the target domain data respectively to obtain a source domain graph and a target domain graph, and calculate the difference value between the source domain graph and the target domain graph through a graph morphological difference metric function; Calculate the similarity loss between the source domain graph and the target domain graph according to the difference value. Based on the cross-entropy loss function, obtain the true distribution probability of the bearing category according to the source domain data and the target domain data, and through the fully connected layer in the deep residual neural network, obtain the predicted distribution probability of the bearing category during the iterative update of the nodes in the source domain data and the target domain data, so as to calculate the distribution difference loss between the true distribution probability of the bearing category and the predicted distribution probability of the bearing category. Introduce a kernel function, and calculate the distance of the mean embeddings of the source domain data and the target domain data in space through the maximum mean discrepancy function, so as to obtain the mean discrepancy loss. Combine the similarity loss, the distribution difference loss and the mean discrepancy loss to calculate the total loss function; Based on the total loss function, calculate the gradients of the graph morphology difference metric function, the source domain graph and the target domain graph, the nodes, and the parameters of the graph kernel model in sequence; Use the cross-entropy loss function to perform supervised learning on the classification results during the iterative update of the nodes in the source domain data and the target domain data, and combine the graph morphology difference metric function and the maximum mean discrepancy function to jointly optimize the graph kernel model.

[0005] According to one aspect of the above technical solution, the feature extraction of the denoised source domain data and target domain data through the deep residual neural network combined with the self-attention mechanism to obtain the initial feature vector specifically includes: Introduce a deep residual neural network combined with the self-attention mechanism, where the deep residual neural network includes a multi-scale residual feature extraction module, a channel-temporal dual-path attention mechanism module, a normalization layer, a global average pooling layer, and a global maximum pooling layer; Use the multi-scale residual feature extraction module to extract the time-domain and frequency-domain features in the denoised source domain data and target domain data; Normalize the denoised source domain data and target domain data through the normalization layer; Use the global average pooling layer and the global maximum pooling layer to reflect continuous fault features and capture transient shock features; Calculate the global average value of the inner channels of the input features in the deep residual neural network through the global average pooling layer to obtain the initial feature vector.

[0006] According to one aspect of the above technical solution, regarding the initial feature vector as a node, perform weighted processing on the node in the time channel and the frequency channel to initialize the label embedding of the node, specifically including: Regarding the initial feature vector as a node, calculate the temporal-channel attention feature weight value of the node based on the channel-temporal dual-path attention mechanism module; ; ; where; represents the node \(v\) of the source domain data in the deep residual neural network, represents the node \(i\) of the target domain data in the deep residual neural network, \(W\) 1 and \(W\) 2 are the fully connected layer parameters in the deep residual neural network, represents the activation function, GAP represents the global average pooling operation, and ReLU represents the non-linear activation function; Calculate the frequency-channel attention feature weight value of the node based on the channel-temporal dual-path attention mechanism module; ; ; where, \(W\) 3 is the fully connected layer parameter, and GRU is the gated recurrent unit in the deep residual neural network; Fuse the temporal-channel attention feature weight value and the frequency-channel attention feature weight value to obtain the fused feature weight value; ; ; Calculate the internal learning parameter \(\gamma\) in the deep residual neural network; ; ; where, GlobalPool(·) is the global average pooling operation, N ( v ) represents the node neighborhood feature; \(W_r\) represents the weight matrix, \(W_r\in R\) (d×2d) , and \(d\) represents the dimension of the output feature; Initialize the label embedding for the nodes in the source domain data and the target domain data; ; .

[0007] According to one aspect of the above technical solution, the introduced graph kernel model iteratively updates the nodes after initializing the label embeddings in the source domain data and the target domain data respectively to obtain a source domain graph and a target domain graph, and calculates the difference value between the source domain graph and the target domain graph through a graph morphological difference metric function, specifically including: Taking the nodes after initializing the label embeddings as the initial embedding values, and respectively performing neighbor aggregation calculations layer by layer on the initial embedding values in the source domain data and the target domain data through the graph kernel model; ; ; where k represents the number of neighbor points in the source domain data, N represents the total number of nodes in the source domain data, represents the embedding value corresponding to node u in the (k - 1)-th layer in the source domain data, represents the neighbor aggregation feature of node v in the k-th layer in the source domain data, j represents the number of neighbor points in the target domain data, C represents the total number of nodes in the target domain data, represents the embedding value corresponding to node f in the (j - 1)-th layer in the target domain data, represents the neighbor aggregation feature of node i in the j-th layer in the target domain data; Based on the results of the neighbor aggregation calculations, respectively update the nodes with the initially embedded nodes in the source domain data and the target domain data; ; ; where, represents an activation function, W represents a weight, and b represents a bias, represents the updated node v in the k-th layer in the source domain data, represents the updated node i in the j-th layer in the target domain data; By respectively performing pooling on the updated nodes in the source domain data and the target domain data, construct the embedding representation of the graph to obtain the source domain graph and the target domain graph respectively; ; ; where, represents the updated node v in the last layer K in the source domain data, V G represents the total number of updated nodes in the source domain data, represents the source domain graph, represents the updated node i in the last layer J in the target domain data, I H represents the total number of updated nodes in the target domain data, represents the target domain graph; Introduce a graph morphological difference metric function to calculate the difference value between the source domain graph and the target domain graph; ; Among them, δ represents the bandwidth of the graph morphological difference metric function.

[0008] According to one aspect of the above technical solution, calculating the similarity loss between the source domain graph and the target domain graph according to the difference value, based on the cross-entropy loss function, obtaining the true distribution probability of the bearing category according to the source domain data and the target domain data, and through the fully connected layer in the deep residual neural network, obtaining the predicted distribution probability of the bearing category in the process of iteratively updating the nodes in the source domain data and the target domain data, so as to calculate the distribution difference loss between the true distribution probability of the bearing category and the predicted distribution probability of the bearing category, introducing a kernel function, and calculating the distance of the mean embeddings of the source domain data and the target domain data in space through the maximum mean discrepancy function to obtain the mean discrepancy loss, specifically including: Calculating the similarity loss between the source domain graph and the target domain graph according to the difference value; ; Customize the fault category items of the bearing; Based on the definition of the fault category items, obtaining the true distribution probability of the bearing category according to the source domain data and the target domain data; Based on the definition of the fault category items, using the fully connected layer in the deep residual neural network to obtain the predicted distribution probability of the bearing category in the process of iteratively updating the nodes in the source domain data and the target domain data; Based on the cross-entropy loss function, calculating the distribution difference loss between the true distribution probability of the bearing category and the predicted distribution probability of the bearing category according to the true distribution probability of the bearing category and the predicted distribution probability of the bearing category; ; Among them, represents the probability that the bearing category in the true distribution is , represents the probability that the bearing category in the predicted distribution is , and F(P, Q) represents the cross-entropy loss function; Introducing a kernel function, and calculating the mean discrepancy loss according to the feature vectors of the vibration signals in the source domain data and the feature vectors of the vibration signals in the target domain data; ; Among them, x c represents the feature vector of the c-th vibration signal in the source domain data, y βIt represents the eigenvector of the β-th vibration signal in the target domain data, φ represents the kernel function, g represents the number of vibration signals in the source domain data, and n represents the number of vibration signals in the target domain data.

[0009] According to one aspect of the above technical solution, the total loss function is calculated by combining the similarity loss, the distribution difference loss, and the mean difference loss, and the definition of the total loss function is as follows: 。

[0010] According to one aspect of the above technical solution, the gradient is calculated for the graph morphology difference metric function, the source domain graph, the target domain graph, the nodes, and the parameters of the graph kernel model in sequence based on the total loss function, which specifically includes: Calculating the gradient of the graph morphology difference metric function based on the total loss function; ; where, ‌ represents the symbol of partial derivative; Based on the gradient of the graph morphology difference metric function and through the chain rule, the gradients of the source domain graph and the target domain graph are calculated respectively; ; ; where, sign represents the sign function, which is used to extract the direction of the node embedding, and the output of the sign function is ±1 or 0; Calculating the gradient of the nodes based on the gradients of the source domain graph and the target domain graph; ; ; Calculating the gradients of the weight parameter and the bias parameter of the graph kernel model; ; ; ; ; where, represents the node v after the last update of neighbor aggregation, represents the node i after the last update of neighbor aggregation, and R represents the transpose.

[0011] The present invention also provides a bearing fault diagnosis system, including: An acquisition module: used to acquire the signal data of the bearing in a known environment, and use the signal data in the known environment as the source domain data, acquire the signal data of the bearing in other environments, and use the signal data in other environments as the target domain data; Denoising module; used for denoising the source domain data and the target domain data; Extraction module: used for extracting features of the denoised source domain data and target domain data through a deep residual neural network combined with a self-attention mechanism to obtain initial feature vectors; Specifically, the extraction module is used to introduce a deep residual neural network combined with a self-attention mechanism, where the deep residual neural network includes a multi-scale residual feature extraction module, a channel-temporal dual-path attention mechanism module, a normalization layer, a global average pooling layer, and a global maximum pooling layer; Use the multi-scale residual feature extraction module to extract time-domain and frequency-domain features in the denoised source domain data and target domain data; Normalize the denoised source domain data and target domain data through the normalization layer; Use the global average pooling layer and the global maximum pooling layer to reflect continuous fault features and capture transient shock features; Calculate the global average value of the input features in the channels of the deep residual neural network through the global average pooling layer to obtain initial feature vectors; Initialization module: used to regard the initial feature vectors as nodes, and perform weighted processing on the nodes in the time channel and frequency channel to initialize the label embedding of the nodes; Specifically, the initialization module is used to regard the initial feature vectors as nodes, and calculate the time channel attention feature weight values of the nodes based on the channel-temporal dual-path attention mechanism module; ; ; where; represents the node v of the source domain data in the deep residual neural network, represents the node i of the target domain data in the deep residual neural network, W 1 and W 2 are the full connection layer parameters in the deep residual neural network, represents the activation function, GAP represents the global average pooling operation, and ReLU represents the non-linear activation function; Calculate the frequency channel attention feature weight values of the nodes based on the channel-temporal dual-path attention mechanism module; ; ; where, W 3 is the full connection layer parameter, and GRU is the gated recurrent unit in the deep residual neural network; Fuse the time-channel attention feature weight value and the frequency-channel attention feature weight value to obtain a fused feature weight value; ; ; Calculate the internal learning parameter γ in the deep residual neural network; ; ; where GlobalPool(·) is the global average pooling operation, N ( v ) represents the node neighborhood feature; Wr represents the weight matrix, Wr ∈ R (d×2d) , and d represents the dimension of the output feature; Initialize the label embedding for the nodes in the source domain data and the target domain data; ; ; Graph calculation module: used to introduce a graph kernel model to iteratively update the nodes with initialized label embeddings in the source domain data and the target domain data respectively to obtain a source domain graph and a target domain graph, and calculate the difference value between the source domain graph and the target domain graph through a graph morphology difference metric function; The graph calculation module is specifically used for: taking the nodes with initialized label embeddings as the initial embedding values, and performing neighbor aggregation calculations layer by layer on the initial embedding values in the source domain data and the target domain data respectively through the graph kernel model; ; ; where k represents the number of neighbor points in the source domain data, N represents the total number of nodes in the source domain data, represents the embedding value corresponding to node u in the (k - 1)-th layer in the source domain data, represents the neighbor aggregation feature of node v in the k-th layer in the source domain data, j represents the number of neighbor points in the target domain data, C represents the total number of nodes in the target domain data, represents the embedding value corresponding to node f in the (j - 1)-th layer in the target domain data, represents the neighbor aggregation feature of node i in the j-th layer in the target domain data; Based on the results of the neighbor aggregation calculations, update the nodes with initially embedded labels in the source domain data and the target domain data respectively; ; ; where, represents the activation function, W represents the weight, and b represents the bias. represents the updated node v in the k-th layer of the source domain data. represents the updated node i in the j-th layer of the target domain data. By aggregating the updated nodes in the source domain data and the target domain data respectively, an embedded representation of the graph is constructed to obtain the source domain graph and the target domain graph respectively. ; ; Among them, represents the updated node v in the last layer K of the source domain data, V G represents the total number of updated nodes in the source domain data. represents the source domain graph. represents the updated node i in the last layer J of the target domain data, I H represents the total number of updated nodes in the target domain data. represents the target domain graph. Introduce a graph morphology difference metric function to calculate the difference value between the source domain graph and the target domain graph. ; Among them, δ represents the bandwidth of the graph morphology difference metric function. Loss calculation module: used to calculate the similarity loss between the source domain graph and the target domain graph according to the difference value, based on the cross-entropy loss function, obtain the true distribution probability of the bearing category according to the source domain data and the target domain data, and through the fully connected layer in the deep residual neural network, obtain the predicted distribution probability of the bearing category during the iterative update of the nodes in the source domain data and the target domain data, so as to calculate the distribution difference loss between the true distribution probability of the bearing category and the predicted distribution probability of the bearing category, introduce a kernel function, calculate the distance of the mean embedding in space between the source domain data and the target domain data through the maximum mean discrepancy function, so as to obtain the mean difference loss, and combine the similarity loss, the distribution difference loss and the mean difference loss to calculate the total loss function. The loss calculation module is specifically used to: calculate the similarity loss between the source domain graph and the target domain graph according to the difference value. ; Customize the fault category items of the bearing. Based on the definition of the fault category items, obtain the true distribution probability of the bearing category according to the source domain data and the target domain data. Based on the definition of the fault category items, by using the fully connected layer in the deep residual neural network, the bearing category prediction distribution probability is obtained during the iterative update of the nodes in the source domain data and the target domain data; Based on the cross-entropy loss function, according to the true distribution probability of the bearing category and the predicted distribution probability of the bearing category, the distribution difference loss between the true distribution probability of the bearing category and the predicted distribution probability of the bearing category is calculated; ; Among them, represents the probability that the bearing category in the true distribution is , represents the probability that the bearing category in the predicted distribution is , and F(P,Q) represents the cross-entropy loss function; Introduce a kernel function, and calculate the mean difference loss according to the eigenvectors of the vibration signals in the source domain data and the eigenvectors of the vibration signals in the target domain data; ; Among them, x c represents the eigenvector of the c-th vibration signal in the source domain data, y β represents the eigenvector of the β-th vibration signal in the target domain data, φ represents the kernel function, g represents the number of vibration signals in the source domain data, and n represents the number of vibration signals in the target domain data; The definition of the total loss function is as follows: ; Gradient calculation module: used to calculate the gradients of the graph morphological difference metric function, the source domain graph and the target domain graph, the nodes, and the parameters of the graph kernel model based on the total loss function in sequence; The gradient calculation module is specifically used for: calculating the gradient of the graph morphological difference metric function based on the total loss function; ; Among them, ⁣ represents the partial derivative symbol; Based on the gradient of the graph morphological difference metric function and through the chain rule, calculate the gradients of the source domain graph and the target domain graph respectively; ; ; Among them, sign represents the sign function, which is used to extract the direction of the node embedding, and the output of the sign function is ±1 or 0; Based on the gradients of the source domain graph and the target domain graph, calculate the gradient of the nodes; ; ; Perform gradient calculation on the weight parameters and bias parameters of the graph kernel model; ; ; ; ; wherein, represents the node v after the last update of neighbor aggregation, represents the node i after the last update of neighbor aggregation, and R represents transpose; Optimization module: used to use the cross-entropy loss function to perform supervised learning on the classification results during the iterative update of the nodes in the source domain data and the target domain data, and combine the graph morphological difference metric function and the maximum mean difference function to jointly optimize the graph kernel model.

[0012] The present invention also provides a storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the bearing fault diagnosis method described above is implemented.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, the bearing fault diagnosis method described above is implemented.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: By combining the distribution characteristics of the source domain and target domain graph structures, and integrating traditional domain difference measurement methods and kernel function concepts, the differences between graph structures are comprehensively captured. During the calculation process, the global distribution difference and local structure difference are dynamically balanced through an adaptive weight adjustment mechanism, and at the same time, the feature mapping in the high-dimensional Hilbert space is optimized to minimize the distribution difference between the source domain and the target domain, thereby improving the adaptation ability and diagnostic accuracy of the model in the target domain; The present invention constructs a topological representation of the time-domain signal graph of the axle box bearing vibration signal, deeply excavates the inherent morphological features in the process of equipment state evolution, and establishes a hierarchical domain-invariant feature extraction mechanism based on neural networks. This method breaks through the limitations of traditional statistical distribution matching, realizes cross-domain knowledge transfer from the essence structure of data, and provides a new theoretical tool and technical path for equipment state migration diagnosis in complex industrial environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flowchart of the bearing fault diagnosis method in the first embodiment of the present invention; Figure 2It is the structural block diagram of the bearing fault diagnosis system in the second embodiment of the present invention; Figure 3 It is the structural block diagram of the electronic device in the third embodiment of the present invention; The following specific embodiments will further illustrate the present invention in conjunction with the above-mentioned drawings. Specific Embodiments

[0016] To facilitate the understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. Several embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present invention more thorough and comprehensive.

[0017] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly on the other element or there can also be a middle element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be a middle element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are only for the purpose of illustration.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs. The terms used herein in the description of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0019] Please refer to Figure 1 , which shows a bearing fault diagnosis method in the first embodiment of the present invention, including the following steps: S10, Obtain the signal data of the bearing in a known environment, and use the signal data in the known environment as the source domain data. Obtain the signal data of the bearing in other environments, and use the signal data in other environments as the target domain data; S20, Denoise the source domain data and the target domain data; S30, Extract features from the denoised source domain data and target domain data through a deep residual neural network combined with a self-attention mechanism to obtain an initial feature vector; S40, Regard the initial feature vector as a node, and perform weighted processing on the node in the time channel and the frequency channel to initialize the label embedding of the node; S50. Introduce the graph kernel model to iteratively update the nodes after initializing the label embeddings in the source domain data and the target domain data respectively, so as to obtain the source domain graph and the target domain graph, and calculate the difference value between the source domain graph and the target domain graph through the graph morphology difference metric function; S60. Calculate the similarity loss between the source domain graph and the target domain graph according to the difference value. Based on the cross - entropy loss function, obtain the true distribution probability of the bearing category according to the source domain data and the target domain data, and through the fully - connected layer in the deep residual neural network, obtain the predicted distribution probability of the bearing category during the process of iteratively updating the nodes in the source domain data and the target domain data, so as to calculate the distribution difference loss between the true distribution probability of the bearing category and the predicted distribution probability of the bearing category. Introduce the kernel function, calculate the distance of the mean embeddings of the source domain data and the target domain data in space through the maximum mean discrepancy function, so as to obtain the mean discrepancy loss, and combine the similarity loss, the distribution difference loss and the mean discrepancy loss to calculate the total loss function; S70. Based on the total loss function, calculate the gradients of the graph morphology difference metric function, the source domain graph and the target domain graph, the nodes, and the parameters of the graph kernel model in sequence; S80. Use the cross - entropy loss function to perform supervised learning on the classification results during the process of iteratively updating the nodes in the source domain data and the target domain data, and combine the graph morphology difference metric function and the maximum mean discrepancy function to jointly optimize the graph kernel model.

[0020] It can be understood that according to the distribution characteristics of the source domain and target domain graph structures, the present invention combines the traditional domain difference measurement method and the kernel function concept to comprehensively capture the differences between graph structures. During the calculation process, through the adaptive weight adjustment mechanism, the global distribution difference and the local structure difference are dynamically balanced, and at the same time, the feature mapping in the high - dimensional Hilbert space is optimized to minimize the distribution difference between the source domain and the target domain, thereby improving the adaptability and diagnostic accuracy of the model in the target domain; The present invention constructs a topological representation of the time - domain signal graph of the axle - box bearing vibration signal, deeply excavates the inherent morphological features in the process of equipment state evolution, and establishes a hierarchical domain - invariant feature extraction mechanism based on neural networks. This method breaks through the limitations of traditional statistical distribution matching, realizes cross - domain knowledge transfer from the essence structure of data, and provides a new theoretical tool and technical path for equipment state migration diagnosis in complex industrial environments.

[0021] Further, the specific steps of step S20 include: Divide the source domain data into a source domain training set and a source domain test set, and divide the target domain data into a target domain training set and a target domain test set; it should be noted that both the source domain data and the target domain data are vibration signal data; Divide the source domain training set, the source domain test set, the target domain training set, and the target domain test set into data samples according to a preset data length, and perform standardization processing on the data samples; this step is to enhance the robustness of the image recognition model; Use the abnormal amplitude replacement method to correct the outliers in the standardized data samples.

[0022] It can be understood that the purpose of this step is mainly to reduce the interference of noise on the signal data, thereby improving the ability to diagnose bearings in the later stage.

[0023] Further, the step S30 specifically includes: Introduce a deep residual neural network combined with a self-attention mechanism, where the deep residual neural network includes a multi-scale residual feature extraction module, a channel-temporal dual-path attention mechanism module, a normalization layer, a global average pooling layer, and a global maximum pooling layer; Use the multi-scale residual feature extraction module to extract the time-domain and frequency-domain features in the denoised source domain data and target domain data; Normalize the denoised source domain data and target domain data through the normalization layer; this step can make the data distribution more standard, improve the convergence speed of the model and the generalization ability of the model; Use the global average pooling layer and the global maximum pooling layer to reflect continuous fault features and capture transient impact features; this is achieved by the dual-path operation of the global average pooling layer and the global maximum pooling layer; Calculate the global average value of the input features in the deep residual neural network through the global average pooling layer for each channel to obtain an initial feature vector; this initial feature vector is a one-dimensional feature vector.

[0024] Further, the step S40 specifically includes: Regard the initial feature vector as a node, and calculate the time-channel attention feature weight value of the node based on the channel-temporal dual-path attention mechanism module; ; ; where; represents the node v of the source domain data in the deep residual neural network, represents the node i of the target domain data in the deep residual neural network, W 1and W 2 are the parameters of the fully connected layer in the deep residual neural network, represents the activation function, GAP represents the global average pooling operation, and ReLU represents the non-linear activation function; Based on the channel-temporal dual-path attention mechanism module, calculate the frequency channel attention feature weight value of the node; ; ; where, W 3 is the parameter of the fully connected layer, and GRU is the gated recurrent unit in the deep residual neural network; Fuse the time channel attention feature weight value and the frequency channel attention feature weight value to obtain the fused feature weight value; ; ; Calculate the internal learning parameter γ in the deep residual neural network; ; ; where, GlobalPool(·) is the global average pooling operation, N ( v ) represents the node neighborhood feature; Wr represents the weight matrix, Wr ∈ R (d×2d) , and d represents the dimension of the output feature; Perform label embedding initialization on the nodes in the source domain data and the target domain data; ; .

[0025] It can be understood that the label embedding initialization of nodes is a technique that maps the nodes in the network to a low-dimensional vector space. It can maintain the similarity of nodes in the network, so that the node relationships in the embedding space can approximately reflect the structure and properties of the original network. The basic idea of node embedding is to represent each node as a low-dimensional vector, and these vectors maintain a certain similarity in the network; graph representation is a data structure that constructs the relationship between nodes and edges. The graph structure can achieve fast retrieval of specific information in the graph, facilitating the extraction and analysis of graph morphological features. In traditional graph neural networks, node initialization usually directly uses the original features or generates the initial embedding through simple linear transformation. And the attention-weighted initialization dynamically allocates feature weights and fuses cross-node information through the attention mechanism, significantly improving the representation ability of the initial features.

[0026] Further, the step S50 specifically includes: Using the label-embedded initialized node as the initial embedding value, and respectively performing neighbor aggregation calculations layer by layer on the initial embedding values in the source domain data and the target domain data through the graph kernel model; ; ; where k represents the number of neighbor nodes in the source domain data, N represents the total number of nodes in the source domain data, represents the embedding value corresponding to node u in the (k - 1)-th layer in the source domain data, represents the neighbor aggregation feature of node v in the k-th layer in the source domain data, j represents the number of neighbor nodes in the target domain data, C represents the total number of nodes in the target domain data, represents the embedding value corresponding to node f in the (j - 1)-th layer in the target domain data, represents the neighbor aggregation feature of node i in the j-th layer in the target domain data; Based on the results of the neighbor aggregation calculations, respectively perform node updates on the nodes with initial embeddings in the source domain data and the target domain data; ; ; where, represents the activation function, W represents the weight, and b represents the bias, represents the updated node v in the k-th layer in the source domain data, represents the updated node i in the j-th layer in the target domain data; By respectively performing pooling on the updated nodes in the source domain data and the target domain data, construct the embedding representation of the graph to respectively obtain the source domain graph and the target domain graph; ; ; where, represents the updated node v in the last layer K in the source domain data, V G represents the total number of updated nodes in the source domain data, represents the source domain graph, represents the updated node i in the last layer J in the target domain data, I H represents the total number of updated nodes in the target domain data, represents the target domain graph; Introduce a graph morphological difference measurement function to calculate the difference value between the source domain graph and the target domain graph; ; where δ represents the bandwidth of the graph morphological difference measurement function.

[0027] It is understandable that the similarity between the source domain graph and the target domain graph embedding is measured by calculating the difference value, and the parameterized graph kernel function can better characterize the difference in graph morphological features between the source domain and the target domain, ensuring that the domain adaptation process is more accurate; This step is carried out in two paths. One path starts from the source domain data and gradually calculates the source domain graph, and the other path starts from the target domain data and gradually calculates the target domain graph.

[0028] Furthermore, in step S60, according to the difference value, calculate the similarity loss between the source domain graph and the target domain graph. Based on the cross-entropy loss function, obtain the true distribution probability of the bearing category according to the source domain data and the target domain data, and through the fully connected layer in the deep residual neural network, obtain the predicted distribution probability of the bearing category during the iterative update process of the nodes in the source domain data and the target domain data, so as to calculate the distribution difference loss between the true distribution probability of the bearing category and the predicted distribution probability of the bearing category. Introduce a kernel function, and calculate the distance of the mean embedding in space between the source domain data and the target domain data through the maximum mean discrepancy function to obtain the mean discrepancy loss, which specifically includes: Calculate the similarity loss between the source domain graph and the target domain graph according to the difference value; ; Customize the fault category items of the bearing; Based on the definition of the fault category items, obtain the true distribution probability of the bearing category according to the source domain data and the target domain data; Based on the definition of the fault category items, use the fully connected layer in the deep residual neural network to obtain the predicted distribution probability of the bearing category during the iterative update process of the nodes in the source domain data and the target domain data; Based on the cross-entropy loss function, calculate the distribution difference loss between the true distribution probability of the bearing category and the predicted distribution probability of the bearing category according to the true distribution probability of the bearing category and the predicted distribution probability of the bearing category; ; Among them, represents the probability that the bearing category in the true distribution is , represents the probability that the bearing category in the predicted distribution is , and F(P, Q) represents the cross-entropy loss function; Introduce a kernel function, and calculate the mean discrepancy loss according to the feature vectors of the vibration signals in the source domain data and the feature vectors of the vibration signals in the target domain data; ; Among them, xc represents the eigenvector of the c-th vibration signal in the source domain data, y β represents the eigenvector of the β-th vibration signal in the target domain data, φ represents the kernel function, g represents the number of vibration signals in the source domain data, and n represents the number of vibration signals in the target domain data.

[0029] It can be understood that after calculating the source domain graph and the target domain graph, and calculating the difference value, the inter-domain difference can be minimized through the kernel loss term, thereby improving the generalization performance of the model in the target domain. The graph morphology difference metric function can better characterize the graph morphology feature difference between the source domain and the target domain, ensuring that the domain adaptation process is more accurate; then customize the fault category items of the bearing. The true distribution probability of the bearing category can be obtained from the source domain data, and the graph embedding is classified using a fully connected network to obtain the fault category of the bearing. The fully connected network contains three layers, and each layer uses the ReLU activation function to enhance the non-linear expression ability. Finally, the Softmax activation function is connected to output the classification result as a probability distribution. This step can obtain the predicted distribution probability of the bearing category during the iterative update process of the nodes in the source domain data and the target domain data, and then the cross-entropy loss function can be constructed. This function can measure the distribution difference loss between the predicted distribution probability of the bearing category by the model and the true distribution probability of the bearing category; finally, introduce the kernel function, and calculate the mean difference loss according to the eigenvector of the vibration signal in the source domain data and the eigenvector of the vibration signal in the target domain data. This step maps the data to a high-dimensional space through the kernel function and calculates the distance between the mean embeddings of the source domain data and the target domain data in the space.

[0030] Furthermore, in step S60, the definition of the total loss function is as follows: .

[0031] It can be understood that by combining the similarity loss, the distribution difference loss, and the mean difference loss, the total loss function is obtained. Through the total loss function, the graph kernel function can be optimized to achieve the graph morphology difference between the source domain and the target domain, and this difference is minimized.

[0032] Furthermore, step S70 specifically includes: Calculating the gradient of the graph morphology difference metric function based on the total loss function; ; where represents the partial derivative symbol; Based on the gradient of the graph morphology difference metric function and through the chain rule, calculate the gradients of the source domain graph and the target domain graph respectively; ; ; Among them, sign represents the sign function, which is used to extract the direction of the node embedding. The output of the sign function is ±1 or 0; Calculate the gradient of the node based on the gradients of the source domain graph and the target domain graph; ; ; Calculate the gradients of the weight parameters and bias parameters of the graph kernel model; ; ; ; ; Among them, represents the node v after the last update of neighbor aggregation, represents the node i after the last update of neighbor aggregation, and R represents the transpose.

[0033] It can be understood that after minimizing the graph difference through the total loss function, the gradients of the network parameters (such as graph kernel weights, biases, kernel function bandwidths, etc.) can be calculated, and the parameters can be updated to finally achieve cross-device diagnosis of bearing faults.

[0034] The present invention can use the Python programming language and mainly relies on the PyTorch deep learning framework for model construction and training. The implementation of the CNN model includes five convolutional layers, and each convolutional layer is followed by a batch normalization layer and a ReLU activation layer. Finally, a global average pooling layer is used to generate features. The graph embedding is mapped through a custom graph embedding layer, and the node embedding update of the parameterized WL graph kernel is realized. Domain adaptation is trained through a defined adaptive loss function, combined with the kernel similarity between the source domain and the target domain, using the backpropagation process.

[0035] In summary, for the bearing fault diagnosis method in the above embodiments of the present invention, 1. By combining a deep convolutional neural network and a parameterized graph kernel model, local and global features in the bearing vibration signal can be effectively extracted, significantly improving the accuracy of fault diagnosis; 2. By introducing a domain-adaptive graph kernel loss function and a cross-entropy loss function, the distribution difference between the source domain and the target domain is reduced, making the migration effect of the model better on different data sets, improving the generalization ability of the model. The parameterized graph kernel function not only enhances the mathematical interpretability of the model by introducing trainable weights, biases, and graph kernel function bandwidths, but also can effectively quantify the graph morphology difference; 3. It has a high degree of automation. The diagnosis process does not rely on manual feature engineering, reduces the dependence on expert knowledge, and improves the diagnosis efficiency and accuracy.

[0036] Please refer toFigure 2 , the second embodiment of the present invention provides a bearing fault diagnosis system, including: An acquisition module 11: configured to acquire signal data of a bearing in a known environment, and use the signal data in the known environment as source domain data, acquire signal data of the bearing in other environments, and use the signal data in other environments as target domain data; A denoising module 12; configured to perform denoising processing on the source domain data and the target domain data; An extraction module 13: configured to perform feature extraction on the denoised source domain data and target domain data through a deep residual neural network combined with a self-attention mechanism to obtain an initial feature vector; An initialization module 14: configured to regard the initial feature vector as a node, perform weighted processing on the node in the time channel and the frequency channel to perform label embedding initialization on the node; A graph calculation module 15: configured to introduce a graph kernel model to iteratively update the nodes after label embedding initialization in the source domain data and the target domain data respectively to obtain a source domain graph and a target domain graph, and calculate the difference value between the source domain graph and the target domain graph through a graph morphology difference metric function; A loss calculation module 16: configured to calculate the similarity loss between the source domain graph and the target domain graph according to the difference value, based on the cross-entropy loss function, obtain the true distribution probability of the bearing category according to the source domain data and the target domain data, and through the fully connected layer in the deep residual neural network, obtain the predicted distribution probability of the bearing category during the iterative update of the nodes in the source domain data and the target domain data, so as to calculate the distribution difference loss between the true distribution probability of the bearing category and the predicted distribution probability of the bearing category, introduce a kernel function, calculate the distance of the mean embedding in space of the source domain data and the target domain data through the maximum mean difference function to obtain the mean difference loss, and combine the similarity loss, the distribution difference loss and the mean difference loss to calculate the total loss function; A gradient calculation module 17: configured to perform gradient calculation on the graph morphology difference metric function, the source domain graph and the target domain graph, the node, and the parameters of the graph kernel model in sequence based on the total loss function; An optimization module 18: configured to use the cross-entropy loss function to perform supervised learning on the classification result during the iterative update of the nodes in the source domain data and the target domain data, and combine the graph morphology difference metric function and the maximum mean difference function to jointly optimize the graph kernel model.

[0037] The present invention also proposes an electronic device, please refer to Figure 3, shown is an electronic device in the third embodiment of the present invention, including a memory 10, a processor 20, and a computer program 30 stored on the memory 10 and executable on the processor 20. When the processor 20 executes the computer program 30, the above-mentioned bearing fault diagnosis method is implemented.

[0038] Among them, the memory 10 includes at least one type of storage medium, and the storage medium includes flash memory, hard disk, multimedia card, card-type memory (such as SD or DX memory, etc.), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 10 can be an internal storage unit of the electronic device, such as the hard disk of the electronic device. In other embodiments, the memory 10 can also be an external storage device, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. Further, the memory 10 can also include both the internal storage unit of the electronic device and the external storage device. The memory 10 can be used not only to store application software and various types of data installed in the electronic device, but also to temporarily store data that has been output or will be output.

[0039] Among them, in some embodiments, the processor 20 can be an Electronic Control Unit (ECU, also known as a vehicle computer), a Central Processing Unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips, and is used to run the program code stored in the memory 10 or process data, such as executing an access restriction program, etc.

[0040] It should be noted that Figure 3 the shown structure does not constitute a limitation on the electronic device. In other embodiments, the electronic device can include fewer or more components than shown, or combine certain components, or have a different component arrangement.

[0041] The embodiment of the present invention also proposes a readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the above-mentioned bearing fault diagnosis method is implemented.

[0042] Those skilled in the art will understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or used in combination with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

[0043] More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection part (electronic device) having one or more wirings, a portable computer disk cartridge (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or other suitable processing as necessary, and then stored in a computer memory.

[0044] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0045] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0046] The above-described embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent for the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention shall be subject to the appended claims.

Claims

1. A bearing fault diagnosis method, characterized in that: The steps include: Acquire signal data of the bearing in a known environment, and use the signal data in the known environment as source domain data; acquire signal data of the bearing in other environments, and use the signal data in other environments as target domain data; Performing denoising processing on the source domain data and the target domain data; Performing feature extraction on the denoised source domain data and the target domain data by using a deep residual neural network combined with a self-attention mechanism to obtain an initial feature vector; The initial feature vector is regarded as a node, and weighted processing is performed on the node in a time channel and a frequency channel to initialize label embedding of the node; A graph kernel model is introduced to iteratively update the nodes after label embedding initialization in the source domain data and the target domain data, respectively, to obtain a source domain graph and a target domain graph, and a difference value between the source domain graph and the target domain graph is calculated by a graph morphology difference measurement function; The similarity loss between the source domain graph and the target domain graph is calculated according to the difference value, and based on the cross entropy loss function, the true distribution probability of the bearing category is obtained according to the source domain data and the target domain data, and the predicted distribution probability of the bearing category is obtained in the process of iteratively updating the nodes in the source domain data and the target domain data through the fully connected layer in the deep residual neural network, so as to calculate the distribution difference loss between the true distribution probability of the bearing category and the predicted distribution probability of the bearing category, introduce a kernel function, and calculate the distance between the mean embedding of the source domain data and the target domain data in space through the maximum mean difference function to obtain the mean difference loss, and calculate the total loss function by combining the similarity loss, the distribution difference loss and the mean difference loss; Based on the total loss function, sequentially performing gradient calculations on the graph morphology difference metric function, the source domain graph and the target domain graph, the nodes, and the parameters of the graph kernel model; Using the cross entropy loss function, supervised learning is performed on the classification results during the iterative update of the nodes in the source domain data and the target domain data, and the graph kernel model is jointly optimized in combination with the graph morphology difference metric function and the maximum mean difference function.

2. The bearing fault diagnosis method according to claim 1, characterized in that: The deep residual neural network combined with the self-attention mechanism is used to extract features from the denoised source domain data and the target domain data to obtain an initial feature vector, specifically including: A deep residual neural network combined with a self-attention mechanism is introduced, wherein the deep residual neural network includes a multi-scale residual feature extraction module, a channel-temporal dual-path attention mechanism module, a normalization layer, a global average pooling layer, and a global maximum pooling layer; Extracting time domain and frequency domain features from the denoised source domain data and the target domain data using the multi-scale residual feature extraction module; Normalizing the denoised source domain data and the target domain data through the normalization layer; Using the global average pooling layer and the global maximum pooling layer to reflect persistent fault characteristics and capture transient impact characteristics; The global average pooling layer is used to calculate the global average of the internal channels of the input features in the deep residual neural network to obtain an initial feature vector.

3. The bearing fault diagnosis method according to claim 2, characterized in that: The initial feature vector is regarded as a node, and weighted processing is performed on the node in the time channel and the frequency channel to initialize the label embedding of the node, specifically including: The initial feature vector is regarded as a node, and based on the channel-temporal dual-path attention mechanism module, a time channel attention feature weight value of the node is calculated; ； ； in; represents the node v of the source domain data in the deep residual neural network, represents the node i of the target domain data in the deep residual neural network, W1 and W2 are the fully connected layer parameters in the deep residual neural network, represents the activation function, GAP represents the global average pooling operation, and ReLU represents the nonlinear activation function; Based on the channel-time dual-path attention mechanism module, calculate the frequency channel attention feature weight value of the node; ； ； Wherein, W3 is a fully connected layer parameter, and GRU is a gated recurrent unit in the deep residual neural network; Fusing the time channel attention feature weight value and the frequency channel attention feature weight value to obtain a fused feature weight value; ； ； Calculating an internal learning parameter γ in the deep residual neural network; ； ； Among them, GlobalPool (·) is the global average pooling operation, N ( v ) represents the node neighborhood features; Wr represents the weight matrix, Wr∈R (d×2d) , d represents the dimension of the output feature; Initializing label embedding for nodes in the source domain data and the target domain data; ； 。 4. The bearing fault diagnosis method according to claim 3, characterized in that: The graph kernel model is introduced to iteratively update the nodes after label embedding initialization in the source domain data and the target domain data respectively to obtain a source domain graph and a target domain graph, and the difference value between the source domain graph and the target domain graph is calculated by a graph morphology difference measurement function, specifically including: The nodes after label embedding initialization are used as initial embedding values, and the initial embedding values ​​in the source domain data and the target domain data are respectively subjected to neighbor aggregation calculation layer by layer through the graph kernel model; ； ； Among them, k represents the number of neighbor points in the source domain data, N represents the total number of nodes in the source domain data, represents the embedding value corresponding to the node u in the k-1th layer in the source domain data, represents the neighbor aggregation feature of node v in the kth layer in the source domain data, j represents the number of neighbor points in the target domain data, C represents the total number of nodes in the target domain data, represents the embedding value corresponding to the node f at the j-1th layer in the target domain data, Represents the neighbor aggregation features of the target domain data at the jth layer node i; Based on the result of neighbor aggregation calculation, updating nodes initially embedded in the source domain data and the target domain data respectively; ； ； in, represents the activation function, W represents the weight, b represents the bias, represents the updated node v in the kth layer of the source domain data, represents the updated node i in the jth layer in the target domain data; By respectively aggregating updated nodes in the source domain data and the target domain data, an embedding representation of a graph is constructed to obtain a source domain graph and a target domain graph respectively; ； ； in, represents the updated node v, V in the last layer K of the source domain data G Indicates the total number of updated nodes in the source domain data, represents the source domain graph, represents the updated node i in the last layer J in the target domain data, I H Indicates the total number of updated nodes in the target domain data, represents the target domain graph; Introducing a graph morphology difference measurement function to calculate the difference value between the source domain graph and the target domain graph; ； Among them, δ represents the bandwidth of the graph morphology difference metric function.

5. The bearing fault diagnosis method according to claim 4, characterized in that: The similarity loss between the source domain graph and the target domain graph is calculated according to the difference value, and based on the cross entropy loss function, the true distribution probability of the bearing category is obtained according to the source domain data and the target domain data, and the predicted distribution probability of the bearing category is obtained in the process of iteratively updating the nodes in the source domain data and the target domain data through the fully connected layer in the deep residual neural network, so as to calculate the distribution difference loss between the true distribution probability of the bearing category and the predicted distribution probability of the bearing category, introduce a kernel function, and calculate the distance between the mean embedding of the source domain data and the target domain data in space through the maximum mean difference function to obtain the mean difference loss, specifically including: Calculating a similarity loss between the source domain graph and the target domain graph according to the difference value; ； Customize the fault category items of bearings; Based on the definition of the fault category item, obtaining the true distribution probability of the bearing category according to the source domain data and the target domain data; Based on the definition of the fault category item, using the fully connected layer in the deep residual neural network, obtaining the predicted distribution probability of the bearing category in the process of iteratively updating the nodes in the source domain data and the target domain data; Based on the cross entropy loss function, according to the true distribution probability of the bearing category and the predicted distribution probability of the bearing category, to calculate the distribution difference loss between the true distribution probability of the bearing category and the predicted distribution probability of the bearing category; ； in, Indicates that the bearing category in the real distribution is The probability of Indicates that the bearing category in the predicted distribution is The probability of , F(P,Q) represents the cross entropy loss function; A kernel function is introduced to calculate a mean difference loss according to a feature vector of the vibration signal in the source domain data and a feature vector of the vibration signal in the target domain data; ； Among them, x c Represents the feature vector of the cth vibration signal in the source domain data, y β represents the feature vector of the βth vibration signal in the target domain data, φ represents the kernel function, g represents the number of vibration signals in the source domain data, and n represents the number of vibration signals in the target domain data.

6. The bearing fault diagnosis method according to claim 5, characterized in that: The total loss function is calculated by combining the similarity loss, the distribution difference loss and the mean difference loss. The total loss function is defined as follows: 。 7. The bearing fault diagnosis method according to claim 6, characterized in that: The step of sequentially performing gradient calculation on the graph morphology difference metric function, the source domain graph, the target domain graph, the nodes, and the parameters of the graph kernel model based on the total loss function specifically includes: Calculating the gradient of the graph morphology difference metric function based on the total loss function; ； in, Indicates the symbol for partial derivative; Based on the gradient of the graph morphology difference metric function and by using the chain rule, respectively calculating the gradients of the source domain graph and the target domain graph; ； ； Wherein, sign represents a sign function, which is used to extract the direction of embedding of the node, and the output of the sign function is ±1 or 0; Calculating the gradient of the node based on the gradients of the source domain graph and the target domain graph; ； ； Performing gradient calculation on weight parameters and bias parameters of the graph kernel model; ； ； ； ； in, Represents the node v after the last update of neighbor aggregation, represents the node i after the last update of neighbor aggregation, and R represents transpose.

8. A bearing fault diagnosis system, characterized in that: include: Acquisition module: used to acquire signal data of the bearing in a known environment, and use the signal data in the known environment as source domain data, acquire signal data of the bearing in other environments, and use the signal data in other environments as target domain data; A denoising module; used for performing denoising processing on the source domain data and the target domain data; Extraction module: used for extracting features from the denoised source domain data and the target domain data by using a deep residual neural network combined with a self-attention mechanism to obtain an initial feature vector; Initialization module: used for treating the initial feature vector as a node, performing weighted processing on the node in the time channel and the frequency channel, so as to initialize the label embedding of the node; Graph calculation module: used to introduce a graph kernel model to iteratively update the nodes after label embedding initialization in the source domain data and the target domain data, respectively, to obtain a source domain graph and a target domain graph, and calculate the difference value between the source domain graph and the target domain graph through a graph morphology difference measurement function; Loss calculation module: used to calculate the similarity loss between the source domain graph and the target domain graph according to the difference value, obtain the true distribution probability of the bearing category according to the source domain data and the target domain data based on the cross entropy loss function, and obtain the predicted distribution probability of the bearing category in the process of iteratively updating the nodes in the source domain data and the target domain data through the fully connected layer in the deep residual neural network, so as to calculate the distribution difference loss between the true distribution probability of the bearing category and the predicted distribution probability of the bearing category, introduce a kernel function, calculate the distance between the mean embedding of the source domain data and the target domain data in space through the maximum mean difference function, so as to obtain the mean difference loss, and calculate the total loss function by combining the similarity loss, the distribution difference loss and the mean difference loss; Gradient calculation module: used to perform gradient calculation on the graph morphology difference metric function, the source domain graph and the target domain graph, the nodes, and the parameters of the graph kernel model in sequence based on the total loss function; Optimization module: used to use the cross entropy loss function to supervise the classification results in the process of iteratively updating the nodes in the source domain data and the target domain data, and to jointly optimize the graph kernel model in combination with the graph morphology difference metric function and the maximum mean difference function.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the bearing fault diagnosis method according to any one of claims 1 to 7 is implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the bearing fault diagnosis method according to any one of claims 1 to 7 is implemented.

Images