Graphical variational autoencoder feature extraction method for multi-source monitoring data of transmission system

Abstract

A graph variational autoencoder feature extraction method for multi-source monitoring data of a transmission system is disclosed, comprising: collecting signals from a fault simulation experimental system using multiple sensors, collecting normal signals and fault signals respectively; constructing multi-sensor graph signals using the obtained data according to a multi-sensor network graph data construction method, and dividing the data into training and testing sets; inputting the training set data into a graph attention variational autoencoder model for training, extracting and encoding features from the multimodal heterogeneous graph signals, and obtaining low-dimensional embedding fusion feature representations of the graph signals; further inputting the extracted low-dimensional embedding fusion features of the graph signals into a global average pooling layer for graph readout operation, and then inputting the graph readout features into a fully connected layer classifier for collaborative training and classification with the encoder part of the graph attention variational autoencoder model, and testing the model with the highest test accuracy as the optimal multi-source data fusion and fault feature extraction model.

Description

Technical Field

[0001] This disclosure pertains to the field of transmission system fault diagnosis, specifically relating to a graph variational autoencoder feature extraction method for multi-source monitoring data of a transmission system. Background Technology

[0002] As a crucial component of the transmission systems in major equipment such as artillery, helicopters, and aero engines, the health of the transmission gearbox is vital to the operational safety and reliability of these systems. With the continuous development of intelligent operation and maintenance and health management technologies, using deep learning models to construct fault feature mining and state mapping has become a mainstream approach. By inputting a large amount of collected normal and fault data into a deep network for training, a good classification and mapping model can be obtained. However, most current intelligent diagnostic models are trained using single-sensor data. Even those models that use multiple sensor data simply splice and fuse the features or decision results trained from single-sensor data, without considering the spatial structural relationships between the sensors. Transmission systems have numerous components, operate under harsh conditions, and collect abundant information from both built-in and external sensors. Therefore, effectively utilizing all sensor information for intelligent diagnosis of system components and designing strategies for fusing multi-source sensor information and extracting fault features are key to successful diagnosis.

[0003] While existing models such as convolutional neural networks can effectively extract features from signals, they struggle to uncover relationships between signals from different sensors, making it difficult to fuse and extract features from multi-source information. Graph neural networks (GNNs) are excellent at mining the spatial relationships between sensor nodes in a graph, making them a natural approach for sensor network data mining. However, commonly used GNNs suffer from oversmoothing, resulting in poor fused features and significant information loss. Furthermore, spectral-based GNNs typically employ transductive learning, requiring the entire graph as input for training. They are based on a specific graph; changing the graph signal necessitates retraining from scratch. Consequently, their generalization ability is generally poor, a crucial requirement in fault diagnosis.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of the present invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this disclosure is to provide a graph variational autoencoder feature extraction method for multi-source monitoring data of a transmission system. Graph attention networks are a type of inductive learning, a spatial domain-based method that can be generalized to other graphs. Attention mechanisms can also effectively perform message passing and sensor information fusion on graph signals. Graph variational autoencoders can better extract low-dimensional graph embedding information, minimizing information loss during feature extraction, thereby completing multi-source data fusion and fault feature extraction.

[0006] To achieve the above objectives, this disclosure provides the following technical solutions:

[0007] A graph variational autoencoder feature extraction method for multi-source monitoring data of a transmission system includes the following steps:

[0008] Step S100: Use multiple sensors to collect normal signals and fault signals from the transmission system fault simulation experimental system;

[0009] Step S200: Based on the normal signal and the fault signal, construct a multi-sensor graph signal according to the multi-sensor network graph data construction method, and divide it into a training set and a test set;

[0010] Step S300: Input the training set data into the graph attention variational autoencoder model (GAT-VAE) for training, extract and encode features of the multimodal heterogeneous graph signal, and obtain the low-dimensional embedded fusion feature representation of the multimodal heterogeneous graph signal;

[0011] Step S400: Input the low-dimensional embedding fusion features into the global average pooling layer to perform graph readout operation to obtain graph readout features, and then input the graph readout features into the fully connected layer classifier for collaborative training and classification with the encoder part of the graph attention variational autoencoder model GAT-VAE. Test the model with the highest test accuracy and save it as the optimal multi-source data fusion and fault feature extraction model.

[0012] In the method described, in step S100, multiple vibration sensors are used to collect experimental data from the transmission system fault simulation experiment system. The transmission system fault simulation experiment system simulates two different speed conditions. Each speed condition includes one normal gear state and four different fault degrees of gear health states, resulting in 2 normal signals and 8 fault signals, for a total of 10 health states. The data sampling frequency is 10kHz.

[0013] In the method described above, step S200, the multi-sensor network graph data construction method includes the following steps:

[0014] S201: For a multivariate time series X = [X1, X2, ..., X] consisting of τ sensors, ...τ First, for a single sensor X i The original data of length l is divided into subsamples of length J by partitioning the univariate sequence. Each subsample is assigned a corresponding label, and there is no overlap between the samples. The resulting subsample set can be represented as:

[0015]

[0016] In the formula, ∏ is the obtained subsample set, x represents the subsample, y represents the label, and n represents the number of subsamples;

[0017] S202: Based on the obtained subset ∏, each subset is treated as a node in a multi-sensor graph network to construct a graph with τ nodes. Sensor neighbors are found by calculating cosine distance, denoted as the RadiusGraph method. The method for obtaining the adjacency matrix using the RadiusGraph method can be represented as follows:

[0018] ε = -radius(x) i x j )

[0019]

[0020] Wherein, -radius(·) computes node x i and node x j The cosine similarity ε, where A(i,j) is the value in the i-th row and j-th column of the adjacency matrix of the graph, γ p To determine the threshold for whether an edge exists between nodes, this method sets the threshold to 0.

[0021] In the method described, step S300, the graph attention variational autoencoder model GAT-VAE includes:

[0022] S30l: The encoder part uses a graph attention layer (GAT) and a non-linear activation function (ReLU) to generate a low-dimensional feature matrix for the graph signal.

[0023]

[0024] Where X is the node feature matrix and A is the adjacency matrix.

[0025] S302: The mean μ and log-variance log σ of the latent variables generated by the second GAT layer 2 :

[0026]

[0027] S303: Sample the latent variable z from the mean μ and variance σ of the latent variables so that the model can continue training:

[0028] z = μ + σ × ε, ε ~ N(0, 1)

[0029] Where N(0,1) represents a normal distribution, and ε is sampled from the normal distribution.

[0030] S304: Perform GATPooling graph pooling on the latent variable z:

[0031] Z = GATPooling(Z, A)

[0032] Here, GATPooling(·) represents the graph attention pooling layer.

[0033] S305: The decoder obtains the node feature vectors through two fully connected layers (FCLayers):

[0034] X′=FCLayers(FCLayers(Z)),

[0035] S306: The loss function of the graph attention variational autoencoder model GAT-VAE can be expressed as:

[0036] L = MSELoss(X′, X)

[0037] Where MSELoss(·) refers to the mean squared error loss, expressed as:

[0038]

[0039] In the method described, in step S301, the GAT layer is a graph attention network, and its steps are as follows:

[0040] S3011: Calculate the node attention coefficient based on node characteristics:

[0041]

[0042] in, For node features, The parameters representing a single-layer neural network, Let α represent the neighbors of node i. ij This represents the attention coefficient between node i and node J. || denotes the cascading operation, Θ represents the linear transformation weight matrix applied to each node, LeakReLU(·) is a non-linear activation function, and exp(·) represents the natural exponential operation;

[0043] S3012: Calculate multi-head attention based on the attention coefficient calculation method described above. The multi-head attention graph convolutional layer GAT is defined as follows:

[0044]

[0045] Where M represents the number of bullish investors. It is the normalized attention coefficient calculated by the Mth attention mechanism. Let σ(·) represent the final learned node representation, and let σ(·) represent a non-linear activation function.

[0046] In the method described, in step S304, GATPooling is a graph attention pooling layer, and the pooled graph (X′, A′) is represented as follows:

[0047]

[0048] Where GATConv is the GAT graph attention convolutional network, Z represents the attention score obtained through the graph attention network, topk(·) is the topk mechanism, which can return the k nodes with the highest attention scores, returning an exponent [kn], k∈(0,1), idx represents the indexing operation, and Z mask X represents the attention mask obtained. idx, : represents the feature matrix indexed by nodes, and ⊙ represents element-wise matrix multiplication.

[0049] In the method described, in step S305, the fully connected layers (FCLayers) are represented as follows:

[0050] FCLayers(X) = W T X+b,

[0051] Among them, W T Let b be the weight matrix to be learned, and let b be the bias vector.

[0052] In the method described, in step S400, the readout layer uses global average pooling, which converts the node representation of the subgraph into a graph representation and returns a graph embedding.

[0053]

[0054] Where x is the node feature matrix, r is the graph embedding after global average pooling, and N i This represents the number of neighbors of the i-th node.

[0055] In the method described, in step S400, the fully connected classification layer is represented as follows:

[0056] FCLayers(r) = W T ×BacthNrom(r)+b,

[0057] Among them, W TLet b be the weight matrix to be learned, BatchNorm(·) represents batch normalization of the read graph embedding features, and b is the bias vector.

[0058] In the method described above, in step S400, the fully connected layer and the encoder part of the graph attention variational autoencoder model GAT-VAE are trained using the cross-entropy loss function.

[0059]

[0060] Where M is the number of categories, y ic p is a sign function, which is either 0 or 1. It takes the value 1 if the true class of sample i is equal to c, and 0 otherwise. ic Let be the predicted probability that observed sample i belongs to category c.

[0061] Compared with the prior art, the beneficial effects of this disclosure are: this disclosure can effectively realize the ability to fuse multi-source data of multi-sensor signals and extract fault features, improve the fault diagnosis accuracy of transmission gearboxes of major equipment, reduce the manpower and material resources for feature engineering design of fault modes, and thus obtain an end-to-end fault diagnosis model. Attached Figure Description

[0062] Figure 1 This disclosure provides a flowchart of a method for multi-source data fusion and fault feature extraction based on a graph attention variational autoencoder, according to one embodiment.

[0063] Figure 2 This is a schematic diagram of a gearbox test bench and gear health status provided in one embodiment of this disclosure;

[0064] Figure 3 This is a schematic diagram illustrating the diagnostic accuracy of various models provided in one embodiment of this disclosure;

[0065] Figure 4 This is a schematic diagram of the confusion matrix of a multi-sensor diagnostic model provided in one embodiment of this disclosure;

[0066] Figure 5 This is a schematic diagram of t-SNE dimensionality reduction visualization of a multi-sensor diagnostic model provided in one embodiment of this disclosure;

[0067] Figure 6 This is a schematic diagram illustrating the diagnostic accuracy of a single-sensor diagnostic model and a multi-sensor diagnostic model provided in another embodiment of this disclosure. Detailed Implementation

[0068] The following will refer to the appendix. Figures 1 to 6Specific embodiments of this disclosure are described in detail. While specific embodiments of this disclosure are shown in the accompanying drawings, it should be understood that this disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0069] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this disclosure is determined by the appended claims.

[0070] To facilitate understanding of the embodiments of this disclosure, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of this disclosure.

[0071] In one embodiment, such as Figure 1 As shown, a graph variational autoencoder feature extraction method for multi-source monitoring data of a transmission system includes the following steps:

[0072] S100: Use multiple sensors to collect signals from the fault simulation experimental system, collecting normal signals and fault signals separately. Multiple sensors should be two or more, and are not limited to the same type of sensor.

[0073] In this step, data collected from 12 vibration sensors was used to simulate two different rotational speed conditions. Each speed condition included one normal gear state and four different gear health states with varying degrees of fault, such as... Figure 2 As shown, there are a total of 10 health states, making this a 10-class classification task, with a data sampling frequency of 10kHz.

[0074] S200: Using the obtained data, construct multi-sensor graph signals according to the multi-sensor network graph data construction method, and divide the training set and test set;

[0075] In this embodiment, signals from 12 sensors were used to construct a multi-sensor graph signal. The RadiusGraph method was used to construct a multi-sensor spatiotemporal graph signal from the multi-source sensor signals. Each sensor was assigned 1024 points as graph node features, resulting in 12 nodes for the 12 sensors. The RadiusGraph method was used to generate the edge set and edge weights of the graph signal. The final sample consisted of 10,000 samples, with 1,000 samples per class. The test and training sets were split at a ratio of 0.5, resulting in a training set of 5,000 samples (500 samples per class) and a test set of 5,000 samples (500 samples per class). Furthermore, to compare the advantages of multi-source sensor signal fusion diagnosis, a single vibration sensor signal was used for single-sensor fault diagnosis. A univariate time series graphing method was used, and the RadiusGraph method was used to generate the edge set and edge weights of the single-sensor graph signal. There are a total of 1000 samples, with 100 samples in each class. The ratio of the test set to the training set is 1:4. Therefore, there are 800 training samples and 200 test samples.

[0076] S300: Input the training set data into the graph attention variational autoencoder model (GAT-VAE) for training, extract and encode features of multimodal heterogeneous graph signals, and obtain low-dimensional embedded fusion feature representations of multimodal heterogeneous graph signals;

[0077] In this step, the network architecture parameters of the graph attention variational autoencoder model GAT-VAE are shown in Table 1.

[0078] Table 1.GFA - VAE network architecture

[0079]

[0080] S400: Continue to input the low-dimensional embedding fusion features of the extracted graph signals into the global average pooling layer for graph readout operation, and then input the graph readout features into the fully connected layer classifier for collaborative training and classification with the encoder part of the graph attention variational autoencoder model GAT-VAE. The model with the highest test accuracy is saved as the optimal multi-source data fusion and fault feature extraction model.

[0081] In this step, test accuracy is defined as: the number of correctly classified samples on the test set divided by the total number of samples, expressed by the following formula:

[0082] Where 1[·] is the indicator function, and the expression within the parentheses is 1 if true and 0 if false; For the predicted class of the i-th sample, y i Let M be the true i-th sample category, and M be the number of samples in the diagnostic dataset.

[0083] This disclosure implements a novel graph variational autoencoder feature extraction method for multi-source monitoring data of a transmission system. By representing multi-source sensor data as graph signals, it combines the powerful feature extraction capabilities of graph neural networks to capture low-dimensional graph embeddings of multimodal sensor signals. Furthermore, an attention mechanism is employed in the entire model to enhance the network's representational ability. In addition, the architecture of the graph variational autoencoder can better extract and encode features from multimodal heterogeneous graph signals, and its reconfigurability ensures the minimization of information loss during feature extraction.

[0084] In another embodiment, step S200, the specific steps of the multi-sensor network graph data construction method are as follows:

[0085] S201: For a multivariate time series X = [X1, X2, ..., X] consisting of τ sensors, ... τ ], for a single sensor X i The original data of length l is divided into subsamples of length d by partitioning the univariate sequence, and each subsample is assigned a corresponding label. There is no overlap between the samples. The resulting subsample set can be represented as:

[0086]

[0087] Where ∏ is the obtained sample set, x represents a subsample, y represents a label, and n represents the number of subsamples.

[0088] S202: Based on the obtained subset ∏, each subset is treated as a node in a multi-sensor graph network to construct a graph with τ nodes. Sensor neighbors are found by calculating cosine distance, denoted as the RadiusGraph method. The method for obtaining the adjacency matrix using the RadiusGraph method can be represented as follows:

[0089] ε = -radius(x) i x j )

[0090]

[0091] Wherein, -radius(·) computes node x i and node x j The cosine similarity ε, where A(i,j) is the value in the i-th row and j-th column of the adjacency matrix of the graph, γ p To determine the threshold for whether an edge exists between nodes, this method sets the threshold to 0.

[0092] In another embodiment, step S300, the graph attention variational autoencoder model GAT-VAE includes the following steps:

[0093] S301: The encoder part uses a graph attention layer (GAT) and a non-linear activation function (ReLU) to generate a low-dimensional feature matrix for the graph signal.

[0094]

[0095] Where X is the node feature matrix of the graph, and A is the adjacency matrix of the graph.

[0096] S302: The mean μ and log-variance log σ of the latent variables generated by the second GAT layer 2 :

[0097]

[0098] S303: Sample the latent variable z from the mean μ and variance σ of the latent variables so that the model can continue training:

[0099] z = μ + σ × ε, ε ~ N(0, 1)

[0100] Where N(0,1) represents a normal distribution, and ε is sampled from the normal distribution.

[0101] S304: Perform GATPooling graph pooling on the latent variable z:

[0102] Z = GATPooling(Z, A)

[0103] Here, GATPooling(·) represents the graph attention pooling layer.

[0104] S305: The decoder obtains the node feature vectors through two fully connected layers (FCLayers):

[0105] X′=FCLayers(FCLayers(Z)),

[0106] S306: The loss function of the graph attention variational autoencoder model GAT-VAE can be expressed as:

[0107] L = MSELoss(X′, X)

[0108] Where MSELoss(·) refers to the mean squared error loss, expressed as:

[0109]

[0110] In another embodiment, the GAT layer in step S301 refers to the graph attention network, and its construction steps are as follows:

[0111] S3011: Calculate the node attention coefficient based on node characteristics:

[0112]

[0113] in, For node features, The parameters representing a single-layer neural network, Let α represent the neighbors of node i. ij This represents the attention coefficient between node i and node j. || denotes the cascading operation, Θ represents the linear transformation weight matrix applied to each node, LeakReLU(·) is a non-linear activation function, and exp(·) represents the natural exponential operation.

[0114] S3012: Based on this attention coefficient calculation method, multi-head attention can be calculated. The GAT definition for the multi-head attention graph convolutional layer is:

[0115]

[0116] Where M represents the number of bullish investors. It is the normalized attention coefficient calculated by the Mth attention mechanism. Let represent the node representation of the final learned node, and let □(·) represent a non-linear activation function.

[0117] In another embodiment, in step S304, GATPooling refers to the graph attention pooling layer, and the pooled graph (X′, A′) can be represented as:

[0118]

[0119] Where GATConv is the GAT graph attention convolutional network, Z represents the attention score obtained through the graph attention network, topk(·) is the topk mechanism, which can return the k nodes with the highest attention scores, returning an exponent [kn], k∈(0,1), idx represents the indexing operation, and Z mask X represents the attention mask obtained. idx， : represents the feature matrix indexed by nodes, and ⊙ represents element-wise matrix multiplication.

[0120] In another embodiment, in step S305, the fully connected layers FCLayers can be represented as follows:

[0121] FCLayers(X) = W T X+b

[0122] Among them, W T Let b be the weight matrix to be learned, and let b be the bias vector.

[0123] In another embodiment, in step S400, the readout layer is a global average pooling layer, which converts the node representation of the subgraph into a graph representation and returns a graph embedding:

[0124]

[0125] Where x is the node feature matrix, r is the graph embedding after global average pooling, and N i This represents the number of neighbors of the i-th node.

[0126] In another embodiment, in step S400, the fully connected classification layer can be represented as:

[0127] FCLayers(r) = W T ×BacthNrom(r)+b

[0128] Among them, W T Let b be the weight matrix to be learned, BatchNorm(·) represents batch normalization of the read graph embedding features, and b is the bias vector.

[0129] In another embodiment, in step S400, training the fully connected layer and the encoder part of the graph attention variational autoencoder model GAT-VAE means training it using the cross-entropy loss function:

[0130]

[0131] Where M is the number of categories, y ic p is a sign function, which is either 0 or 1. It takes the value 1 if the true class of sample i is equal to c, and 0 otherwise. ic Let be the predicted probability that observed sample i belongs to category c.

[0132] It should also be noted that in the above embodiments, the test accuracy is defined as: the number of samples correctly classified on the test set divided by the total number of samples.

[0133] The following is combined Figures 2 to 6 The technical solution disclosed herein will be further described.

[0134] In a specific embodiment, such as Figure 2 As shown, gear fault diagnosis was performed on a gearbox fault testing test bench. The specific diagnostic process is as follows:

[0135] Four sets of gear experiments under different conditions were set up on a gearbox fault testing test bench. The test bench and fault modes are as follows: Figure 2As shown, each group operates at two different speeds. Data under the same state share the same classification label. For each state, a sample graph is constructed using signals from 12 sensors to obtain a sensor network graph signal. The dataset is shown in Table 2. Furthermore, to demonstrate the advantages of using multi-sensor signals for fault diagnosis, one vibration sensor is used as a single-sensor fault diagnosis method for comparison. The graph construction method is similar, and the resulting dataset is shown in Table 3.

[0136] Table 2

[0137]

[0138] Table 3

[0139]

[0140] Figure 3 This diagram illustrates the diagnostic accuracy of various models trained using multi-sensor graph signals in deep network training. GAT-VAE is the diagnostic model proposed in this disclosure; GCN and ChebyNet are commonly used spectral domain graph neural networks; GAT and GraphSAGE are commonly used spatial domain graph neural networks; and GAE-M is a graph autoencoder. All models are designed with identical parameters such as the number of layers, learning rate, and hidden layer feature dimensions. Each model undergoes ten trials, and the average accuracy is used as the final model's accuracy to eliminate the influence of randomness. Figure 3 It can be seen that the diagnostic accuracy of the method proposed in this disclosure is the highest, thus demonstrating the superiority of the proposed solution.

[0141] Figure 4 This is the confusion matrix of the diagnostic results of the embodiment. According to the results of the confusion matrix, it can be seen that the model used in this invention has a good ability to fuse multi-source data and extract fault features, and has almost no diagnostic errors.

[0142] Figure 5 The features of the last layer of the diagnostic results in the embodiment are visualized using t-SNE. As can be seen from the t-SNE feature visualization, the model used in this invention can effectively distinguish different fault types, further demonstrating that this invention has good multi-source data fusion and fault feature extraction capabilities.

[0143] Figure 6The same graph attention variational autoencoder (GAT-VAE) model was used, with input datasets including multi-sensor graph signal datasets and single-sensor graph signal datasets. The diagnostic results of the multi-sensor diagnostic model and the single-sensor diagnostic model were compared. Each model was run ten times, and the average accuracy was used as the final model's accuracy to eliminate the influence of randomness. As can be seen from the figure, the diagnostic accuracy of the multi-sensor diagnostic model is significantly higher than that of the single-sensor diagnostic model. This further illustrates the advantages of using multi-sensor signals for fault diagnosis; the redundancy and complementarity of multi-sensor information can help reflect system state information.

[0144] The above are merely preferred embodiments of this disclosure and are not intended to limit the implementation methods and protection scope of this disclosure. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made using the content of this disclosure should be included within the protection scope of this disclosure.

Claims

1. A method for extracting graph variational autoencoder features from multi-source monitoring data of a transmission system, characterized in that, It includes the following steps: Step S100: Use multiple sensors to collect normal signals and fault signals from the transmission system fault simulation experimental system; Step S200: Based on the normal signal and the fault signal, construct a multi-sensor graph signal according to the multi-sensor network graph data construction method, and divide it into a training set and a test set; Step S300: Input the training set data into the graph attention variational autoencoder model GAT-VAE for training, extract and encode features of the multimodal heterogeneous graph signal, and obtain the low-dimensional embedded fusion feature representation of the multimodal heterogeneous graph signal; Step S400: Input the low-dimensional embedding fusion features into the global average pooling layer to perform graph readout operation to obtain graph readout features, and then input the graph readout features into the fully connected layer classifier to train and classify in conjunction with the encoder part of the graph attention variational autoencoder model GAT-VAE. The model with the highest test accuracy is saved as the optimal multi-source data fusion and fault feature extraction model. In step S300, the graph attention variational autoencoder model GAT-VAE includes: S301: The encoder part uses the graph attention layer GAT and the non-linear activation function ReLU to generate a low-dimensional feature matrix for the graph signal. ， Where X is the node feature matrix of the graph, and A is the adjacency matrix of the graph; S302: The second graph attention layer (GAT layer) generates the mean of latent variables. With log variance : ， S303: From the mean of latent variables With variance The latent variable z is sampled from the model so that it can continue training: ， Where N(0,1) represents a normal distribution. Sampling is performed from a normal distribution; S304: Perform GATPooling graph pooling on the latent variable z: Z = GATPooling(Z,A) Where GATPooling(·) represents the graph attention pooling layer; S305: The decoder obtains the node feature vectors through two fully connected layers (FCLayers). ， S306: The loss function of the graph attention variational autoencoder model GAT-VAE can be expressed as: ， Where MSELoss(·) refers to the mean squared error loss, expressed as: 。 2. The method according to claim 1, characterized in that, In step S100, multiple vibration sensors are used to collect experimental data from the transmission system fault simulation experiment system. The transmission system fault simulation experiment system simulates two different speed conditions. Each speed condition includes one normal gear state and four different fault degrees of gear health states, resulting in 2 normal signals and 8 fault signals, for a total of 10 health states. The data sampling frequency is 10kHz.

3. The method according to claim 1, characterized in that, In step S200, the multi-sensor network graph data construction method includes the following steps: S201: For a multivariate time series X=[X1, X2, ..., X] consisting of τ sensors, τ First, for a single sensor X i The original data of length l is divided into subsamples of length d by partitioning the univariate sequence, and each subsample is assigned a corresponding label. There is no overlap between the samples. The resulting subsample set can be represented as: ， In the formula, ∏ is the obtained subsample set, x represents the subsample, y represents the label, and n represents the number of subsamples; S202: Based on the obtained subset ∏, each subset is treated as a node in a multi-sensor graph network to construct a graph with τ nodes. Sensor neighbors are found by calculating cosine distance, hereinafter referred to as the RadiusGraph method. The method for obtaining the adjacency matrix using the RadiusGraph method can be represented as follows: ， Wherein, -radius(·) computes node x i and node x j cosine similarity Let A(i,j) be the value in the i-th row and j-th column of the adjacency matrix of the graph. To determine the threshold for whether an edge exists between nodes, this method sets the threshold to 0.

4. The method according to claim 1, characterized in that, In step S301, the graph attention layer GAT is a graph attention network, and its steps are as follows: S3011: Calculate node attention coefficients based on node characteristics : ， in, For node features, The parameters representing a single-layer neural network, Indicates the neighbors of node i. The || represents the attention coefficient between node i and node j, and || represents the cascading operation. Let represent the linear transformation weight matrix applied to each node, LeakReLU(·) is a non-linear activation function, and exp(·) represents the natural exponential operation; S3012: Calculate multi-head attention based on the attention coefficient calculation method described above. The multi-head attention graph convolutional layer GAT is defined as follows: ， Where M represents the number of bullish investors. It is the normalized attention coefficient calculated by the Mth attention mechanism. This represents the node representation of the final learning process. (·) represents a non-linear activation function.

5. The method according to claim 1, characterized in that, In step S304, GATPooling is a graph attention pooling layer, and the pooled graph (X', A') can be represented as: ， Where GATConv is the GAT graph attention convolutional network, Z represents the attention score obtained through the graph attention network, topk(·) is the topk mechanism, which can return the k nodes with the highest attention scores, returning the exponent [kn], k∈(0,1), idx represents the indexing operation, and Z mask This represents the attention mask obtained. This represents the feature matrix indexed by nodes. This represents element-wise matrix multiplication.

6. The method according to claim 1, characterized in that, In step S305, the fully connected layers (FCLayers) are represented as follows: ， Among them, W T Let b be the weight matrix to be learned, and let b be the bias vector.

7. The method according to claim 1, characterized in that, In step S400, the readout layer uses global average pooling, which converts the node representation of the subgraph into a graph representation and returns a graph embedding. ， Where x is the node feature matrix, r is the graph embedding after global average pooling, and N i This represents the number of neighbors of the i-th node.

8. The method according to claim 1, characterized in that In step S400, the fully connected classification layer is represented as follows: ， Among them, W T Let b be the weight matrix to be learned, BatchNorm(·) represents batch normalization of the read graph embedding features, and b is the bias vector.

9. The method according to claim 1, characterized in that, In step S400, the fully connected layer and the encoder part of the graph attention variational autoencoder model GAT-VAE are trained using the cross-entropy loss function: ， Where M is the number of categories, y ic p is a sign function, which is either 0 or 1. It takes the value 1 if the true class of sample i is equal to c, and 0 otherwise. ic Let be the predicted probability that observed sample i belongs to category c.

Images