A fault diagnosis method based on feature-topology two-dimensional graph attention mechanism

Abstract

A fault diagnosis method based on a feature-topology dual-dimensional graph attention mechanism belongs to the field of power grid system fault diagnosis technology. First, a power grid system relationship graph is constructed to comprehensively represent the physical connection relationships and fault feature correlations of the power grid system. Second, using the power grid system relationship graph as input, a power grid system fault diagnosis model based on the feature-topology dual-dimensional graph attention mechanism is established. Third, during the training phase, the power grid system fault diagnosis model is trained using a complete fault feature dataset without missing data. Finally, for fault feature datasets with varying degrees of randomly missing node data, the power grid system fault diagnosis model is used to achieve robust fault diagnosis under different degrees of randomly missing node data. This invention achieves robust fault diagnosis under different degrees of randomly missing node data through joint modeling of the power grid feature-topology dual-dimensional graph attention mechanism and graph convolutional neural network without data completion.

Description

Technical Field

[0001] This invention belongs to the field of power grid system fault diagnosis technology, and relates to a fault diagnosis method based on a feature-topology dual-dimensional graph attention mechanism. Background Technology

[0002] To address the performance degradation in fault diagnosis caused by random missing node data in power grid systems, scholars both domestically and internationally have conducted extensive research and proposed several solutions. For example, Chinese invention patent CN114779015B provides a method for fault diagnosis and location in distribution networks based on super-resolution and graph neural networks. This method first utilizes a super-resolution model constructed using a graph convolutional network (GCN), reconstructs the missing features of all nodes in the network based on measurement data from some key nodes, and then uses a graph attention network (GAT) to achieve fault diagnosis and location based on the reconstructed complete data. Chinese invention patent CN117609824B provides a method, device, and equipment for active distribution network topology identification and fault diagnosis analysis. This method directly aggregates node and topology information based on a graph convolutional network and introduces a multi-head attention mechanism to enhance the model's ability to perceive fault features, thus addressing the impact of distributed power source integration.

[0003] While existing technical solutions improve diagnostic reliability through approaches such as "completing data before diagnosis" or "enhancing model attention," neither achieves synergistic optimization and robust perception of topological structure and node fault characteristics under conditions of missing data. Chinese invention patent CN114779015B relies on data reconstruction, and reconstruction errors propagate backward, interfering with the final diagnosis. While Chinese invention patent C117609824B utilizes an attention mechanism, its attention focuses on the topological relationships between nodes, lacking the ability to filter the importance of the fault characteristics themselves. Both fail to simultaneously guarantee high-precision and robust diagnosis of key fault characteristics and key topological locations in complex scenarios with randomly missing node data.

[0004] To address this, this invention targets random missing node data of varying degrees, emphasizing information aggregation and robust diagnosis through two-dimensional graph attention using both feature and topological dimensions without performing missing data completion. This solves the problems of existing technologies, such as easy error accumulation under missing conditions and insufficient dimensionality adjustment of attention mechanisms. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a fault diagnosis method based on a feature-topology dual-dimensional graph attention mechanism. This fault diagnosis method addresses the issue of varying degrees of random data loss at nodes in a power grid system, i.e., the complete data loss of different numbers of nodes. Without performing data completion, it achieves robust fault diagnosis under varying degrees of random data loss at nodes through joint modeling of the power grid feature-topology dual-dimensional graph attention mechanism and graph convolutional neural network.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A fault diagnosis method based on a feature-topology dual-dimensional graph attention mechanism includes the following steps: The first step is to construct a power grid system relationship diagram, which comprehensively represents the physical connections and fault characteristic correlations of the power grid system; specifically: A power grid system relationship graph is constructed based on existing relationship graph optimization methods, providing accurate physical topology and node characteristic relationships as input for subsequent models. The power grid system relationship graph is... , ,in It is a set of power grid monitoring nodes, representing the status monitoring points of key power grid equipment such as transformers and busbars in fault diagnosis; This is a set of connection relationships, representing the association between physical connection lines and fault characteristics between nodes; The node fault feature matrix contains the fault-sensitive electrical characteristics of the three-phase voltage and three-phase current of each node, which is used to characterize the power grid operating status. It is a topology connection weight matrix, whose weight values ​​comprehensively represent the physical topology strength and electrical characteristic correlation of the power grid system, and are used to guide the propagation and aggregation of fault information under data missing conditions.

[0007] The second step is to use the power grid system relationship diagram constructed in the first step. Using the input as input, a power grid system fault diagnosis model based on a feature-topology dual-dimensional graph attention mechanism is established. The power grid system fault diagnosis model includes a graph feature extraction module, a dual-dimensional graph attention mechanism module, and a fault classification module. The power grid system fault diagnosis model adopts a multi-layer cascaded graph convolutional-attention layer structure. Each graph convolutional-attention layer includes a graph feature extraction module and a two-dimensional graph attention mechanism module. The graph feature extraction module of each graph convolutional-attention layer uses a Chebyshev graph convolutional layer to aggregate fault features of neighboring nodes on the power grid topology, transforming the input node features from their original dimensions into high-dimensional graph convolutional features with richer fault feature information. The two-dimensional graph attention mechanism module filters key fault features and key node locations based on these high-dimensional graph convolutional features. Specifically: Step 2.1: The graph feature extraction module uses a Chebyshev graph convolutional layer to aggregate the fault features of neighboring nodes to the current node in the power grid topology, thereby extracting the graph fault features of each node in the power grid system. Step 2.2: The dual-dimensional graph attention mechanism module sequentially utilizes the power grid sensitive fault feature filtering function based on graph feature attention and the power grid key monitoring node filtering function based on graph topology attention to process the graph fault features extracted in Step 2.1. The graph feature extraction module and the dual-dimensional graph attention mechanism module work together to filter key information from the fault feature dimension and the topology structure dimension, respectively, to address varying degrees of random missing node data. The graph fault features processed by the dual-dimensional graph attention mechanism module... and reshape it into new graph features ,in This represents the batch sample size. This represents the number of nodes in the power grid system. Output the fault feature dimensions for the current graph feature extraction module. Specifically: Step 2.2.1, the power grid sensitive fault feature screening function is achieved through dimensionality reduction compression and dimensionality increase reconstruction. It learns the distribution and response patterns among fault features, filters out the fault feature information most important for fault diagnosis, and generates graph feature weights. .

[0008] Step 2.2.1.1: Perform global average pooling on all nodes of the entire topology for each fault feature of the power grid to obtain the global response level of each fault feature, characterizing the overall health status of each fault feature in the power grid, and outputting the compressed fault feature representation. As shown below: (1) in, This indicates a global average pooling operation; Step 2.2.1.2: Compressed characterization of fault features Feature dimensionality reduction is performed using convolutional layers, reducing the number of features from... Compress to Forced fault feature compression representation retains the most important fault feature information for fault diagnosis during the compression process, and outputs the compressed important fault features. , The proportion of dimensionality reduction for channels: (2) in This is a convolution operation.

[0009] Step 2.2.1.3, for important fault characteristics By increasing the dimensionality through convolutional layers, and based on the key feature information learned in the compression stage, the complete fault feature dimension weight distribution is reconstructed. : (3) Step 2.2.1.4: Weight distribution of fault feature dimensions By performing sigmoid normalization, graphical feature weights are generated to quantify the importance of each fault feature. .

[0010] Step 2.2.2: The key power grid monitoring node screening function learns the fault propagation patterns between adjacent nodes through feature splicing and topology dimensionality reduction, filters out the monitoring node information most important for fault diagnosis, and generates graph topology weights. .

[0011] Step 2.2.2.1: Perform mean pooling and max pooling on each feature dimension of each node of the fault characteristics to capture the overall health status of the power grid and the significant fault features, as shown below: (4) (5) in, Used to represent the average response of all fault characteristics at each node, indicating the overall health of the power grid; i represents the index of the fault characteristic; This represents the fault characteristic value of the i-th node; Used to represent the most significant fault characteristics of each node, i.e., to represent the significant fault characteristics.

[0012] Step 2.2.2.2, will and By concatenating the features along the feature dimension, we obtain the comprehensive fault features of the nodes. .

[0013] Step 2.2.2.3: Through topological relationship learning convolutional layers, the topological dependencies and fault propagation relationships between nodes are learned, and the node comprehensive fault characteristics are integrated. Compression yields the topological dimension weight distribution : (6) Step 2.2.2.4 involves learning the fault propagation patterns between adjacent nodes during the compression process. This includes calculating the weight distribution for the topology dimensions. Sigma-Oblique normalization is used to generate graph topology weights for quantifying the importance of each monitoring node. .

[0014] Step 2.2.3: In the two-dimensional graph attention mechanism module, the original node fault features are weighted by the graph features. Graph topological weights By performing element-wise multiplication sequentially, an enhanced feature is obtained that simultaneously highlights important fault characteristics and important monitoring node information. : (7) in, Represents graph characteristics; This indicates element-wise multiplication. Step 2.3: The fault classification module flattens the node fault features after processing by multiple cascaded graph convolutional-attention layers using a fully connected layer. Further fault information fusion and feature refinement are then performed, among which the first... The fully connected layer performs dimensionality reduction on the input features and finally outputs precise fault location information. The third step, during the training phase, involves using a complete fault feature dataset without missing data to train the power grid system fault diagnosis model constructed in the second step, which consists of multiple graph convolutional-attention layers and a fault classification decision module.

[0015] Specifically: The training of the power grid system fault diagnosis model based on the feature-topology dual-dimensional graph attention mechanism uses the cross-entropy loss function to measure the gap between the predicted and actual fault diagnosis results. The formula for the cross-entropy loss function is as follows: (8) in, This represents the fault mode predicted by the fault diagnosis model. Represents the actual failure mode; It is a loss function used to measure the difference between the predicted result and the actual result; That is, batch represents the size of each batch participating in training, b represents the index of the sample in the batch; C represents the dimension of the fault mode, i.e. the number, and c represents the index of the fault category. This represents the true label vector of the b-th sample; Let represent the predicted probability vector for the b-th sample. A smaller loss value indicates that the predicted failure mode is closer to the actual failure mode.

[0016] The fourth step involves using the fault diagnosis model trained and optimized in the third step to achieve robust fault diagnosis under varying degrees of random missing node data in fault feature datasets.

[0017] The beneficial effects of this invention are as follows: (1) This invention proposes a fault diagnosis method based on a feature-topology dual-dimensional graph attention mechanism. Addressing the issue of varying degrees of random missing node data in power grid systems, this method, through joint modeling of the power grid feature-topology dual-dimensional attention mechanism and graph convolutional neural network, enables robust fault diagnosis without data completion, avoiding the error accumulation problem introduced by first completing the data and then diagnosing. (2) This invention proposes a dual-dimensional graph attention mechanism module. By sequentially introducing a graph feature attention mechanism and a graph topology attention mechanism in the same graph convolution-attention layer, it achieves multi-level filtering and enhancement of power grid fault information. The former can learn the distribution and response patterns among fault features through dimensionality reduction compression and dimensionality increase reconstruction, and filter out the fault feature information most important for fault diagnosis; the latter can learn the fault propagation patterns between adjacent nodes through feature concatenation and topology dimensionality reduction, and filter out the monitoring node information most important for fault diagnosis. The two functions work together to achieve multi-level information filtering and compensation. The two attention mechanisms work together through element-wise weighting, enabling the model to effectively utilize the remaining node features and topology information even when node data is randomly missing to varying degrees, thus achieving structural compensation for missing information.

[0018] (3) This invention proposes a multi-layer graph convolution-attention layer cascaded power grid system fault diagnosis model structure, which can continuously extract stable graph fault features under different degrees of random missing node data, and maintain effective attention to key fault features and key topological locations. Simulation verification shows that the method of this invention can maintain high and stable diagnostic performance in power grid system fault diagnosis tasks under different degrees of random missing node-level data, demonstrating good robustness and engineering applicability. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is the structure of the core module of this invention - graph convolution - attention layer. Detailed Implementation

[0020] The present invention will be further described below with reference to specific implementation examples.

[0021] A fault diagnosis method based on a feature-topology dual-dimensional graph attention mechanism includes the following steps: The first step is to construct a power grid system relationship diagram, which comprehensively represents the physical connections and fault characteristic correlations of the power grid system; specifically: A power grid system relationship graph is constructed based on existing relationship graph optimization methods, providing accurate physical topology and node characteristic relationships as input for subsequent models. The constructed power grid system relationship graph is as follows: , .in It is a set of power grid monitoring nodes, representing the status monitoring points of key power grid equipment such as transformers and busbars in fault diagnosis; This is a set of connection relationships, representing the association between physical connection lines and fault characteristics between nodes; The node fault feature matrix contains the fault-sensitive electrical characteristics of the three-phase voltage and three-phase current of each node, which is used to characterize the power grid operating status. It is a topology connection weight matrix, whose weight values ​​comprehensively represent the physical topology strength and electrical characteristic correlation of the power grid system, and are used to guide the propagation and aggregation of fault information under data missing conditions.

[0022] In this embodiment, a power grid system relationship graph is constructed for the topology of the IEEE 123 power grid system: for each node, the four nearest neighboring nodes are selected, and the normalized weights are calculated based on the physical distance between the nodes to complete the initial construction of the relationship graph, providing topological priors for subsequent neighborhood aggregation of Chebyshev graph convolution and information propagation under missing conditions.

[0023] The second step is to use the power grid system relationship diagram constructed in the first step. Using the input as input, a power grid system fault diagnosis model based on a feature-topology dual-dimensional graph attention mechanism is established. The power grid system fault diagnosis model includes a graph feature extraction module, a dual-dimensional graph attention mechanism module, and a fault classification module. The power grid system fault diagnosis model adopts a multi-layer cascaded graph convolutional-attention layer structure. Each graph convolutional-attention layer includes a graph feature extraction module and a two-dimensional graph attention mechanism module. The graph feature extraction module of each graph convolutional-attention layer uses a Chebyshev graph convolutional layer to aggregate fault features of neighboring nodes on the power grid topology, transforming the input node features from their original dimensions into high-dimensional graph convolutional features with richer fault feature information. The two-dimensional graph attention mechanism module filters key fault features and key node locations based on these high-dimensional graph convolutional features. Specifically: Step 2.1: The graph feature extraction module uses a Chebyshev graph convolutional layer to aggregate the fault features of neighboring nodes onto the current node in the power grid topology, thereby extracting the graph fault features of each node in the power grid system. The graph convolution operation can be represented as: (9) in, For the first Fault features of nodes in the input of a layer graph convolutional layer; Indicates the first +1 layer graph convolutional layer output node fault characteristics; It is a normalized Laplace matrix; It is a Chebyshev polynomial; These are the parameters of the convolution kernel; is a non-linear activation function; K represents the maximum order of the Chebyshev polynomial, which determines the size of the convolution kernel, and k represents the order index of the Chebyshev polynomial.

[0024] Step 2.2: The dual-dimensional graph attention mechanism module sequentially utilizes the power grid sensitive fault feature filtering function based on graph feature attention and the power grid key monitoring node filtering function based on graph topology attention to process the graph fault features extracted in Step 2.1. The two modules work together to filter key information from both the fault feature dimension and the topology dimension, respectively, to address varying degrees of random missing node data. The graph fault features processed by the dual-dimensional graph attention mechanism module... and reshape it into new graph features ,in This represents the batch sample size. This represents the number of nodes in the power grid system. Output the fault feature dimensions for the current graph feature extraction module. Specifically: Step 2.2.1: The power grid sensitive fault feature screening function learns the distribution and response patterns among fault features through dimensionality reduction compression and dimensionality increase reconstruction, filters out the fault feature information most important for fault diagnosis, and generates graph feature weights. .

[0025] Step 2.2.1.1: Perform global average pooling on all nodes of the entire topology for each fault feature of the power grid to obtain the global response level of each fault feature, which characterizes the overall health status of each fault feature in the power grid. Output fault feature compression representation The overall health status As shown below: (1) in, This indicates a global average pooling operation; Step 2.2.1.2: Compressed characterization of fault features Feature dimensionality reduction is performed using convolutional layers, reducing the number of features from... Compress to Forced fault feature compression representation retains the most important fault feature information for fault diagnosis during the compression process, and outputs the compressed important fault features. , The proportion of dimensionality reduction for channels: (2) in This is a convolution operation.

[0026] Step 2.2.1.3, for important fault characteristics By increasing the dimensionality through convolutional layers, and based on the key feature information learned in the compression stage, the complete fault feature dimension weight distribution is reconstructed. : (3) Step 2.2.1.4: Weight distribution of fault feature dimensions By performing sigmoid normalization, graphical feature weights are generated to quantify the importance of each fault feature. .

[0027] Step 2.2.2: The key power grid monitoring node screening function learns the fault propagation patterns between adjacent nodes through feature splicing and topology dimensionality reduction, filters out the monitoring node information most important for fault diagnosis, and generates graph topology weights. .

[0028] Step 2.2.2.1, Fault characteristics For each feature dimension of each node, mean pooling and max pooling are performed respectively to capture the overall health status and significant fault features of the power grid, as shown in formulas (4) and (5): (4) (5) in, Used to represent the average response of all fault characteristics at each node, indicating the overall health of the power grid; i represents the index of the fault characteristic; This represents the fault characteristic value of the i-th node; Used to represent the most significant fault characteristics of each node, i.e., to represent the significant fault characteristics.

[0029] Step 2.2.2.2, will and By concatenating the features along the feature dimension, we obtain the comprehensive fault features of the nodes. .

[0030] Step 2.2.2.3: Through topological relationship learning convolutional layers, the topological dependencies and fault propagation relationships between nodes are learned, and the node comprehensive fault characteristics are integrated. Compression yields the topological dimension weight distribution : (6) Step 2.2.2.4 involves learning the fault propagation patterns between adjacent nodes during the compression process. This includes calculating the weight distribution for the topology dimensions. Sigma-Oblique normalization is used to generate graph topology weights for quantifying the importance of each monitoring node. .

[0031] Step 2.2.3: In the two-dimensional graph attention mechanism module, the original node fault features are weighted by the graph features. Graph topological weights By performing element-wise multiplication sequentially, an enhanced feature is obtained that simultaneously highlights important fault characteristics and important monitoring node information. : (7) in, Represents graph characteristics; This indicates element-wise multiplication. Step 2.3: The fault classification module flattens the node fault features after processing by multiple cascaded graph convolutional-attention layers using a fully connected layer. Further fault information fusion and feature refinement are then performed, among which the first... The fully connected layer performs dimensionality reduction on the input features, ultimately outputting precise fault location information; the formula for the dimensionality reduction process is as follows: (10) in, This represents the feature matrix output by the k-th fully connected layer; This represents the feature matrix input to the (k-1)th fully connected layer; This represents the learnable weight matrix of the k-th fully connected layer; This represents the learnable bias vector of the k-th fully connected layer; This indicates batch normalization operation.

[0032] In this embodiment, a power grid system fault diagnosis model is constructed using a power grid system relationship diagram as input. The model includes three cascaded graph convolution-attention layers. Each layer first extracts features through Chebyshev graph convolution, and all three graph convolution layers use 3rd-order Chebyshev graph convolution. Then, a dual-dimensional graph attention mechanism modulates the weights of the feature dimension and the topology dimension. After the three cascaded layers, four fully connected layers are connected to complete the fault classification. The structure of the fully connected layers is 1024-512-256-124, and the first three layers use BN normalization and ReLU nonlinear activation.

[0033] The third step, during the training phase, involves using a complete fault feature dataset without any missing data to train the power grid system fault diagnosis model, which was constructed in the second step and consists of multiple graph convolutional-attention layers and a fault classification decision module.

[0034] Specifically: The training of the power grid system fault diagnosis model based on the feature-topology dual-dimensional graph attention mechanism uses the cross-entropy loss function to measure the gap between the predicted and actual fault diagnosis results. The formula for the cross-entropy loss function is as follows: (8) in, This represents the fault mode predicted by the fault diagnosis model. Represents the actual failure mode; It is a loss function used to measure the difference between the predicted result and the actual result; That is, batch represents the size of each batch participating in training, b represents the index of the sample in the batch; C represents the dimension of the fault mode, i.e. the number, and c represents the index of the fault category. This represents the true label vector of the b-th sample; Let represent the predicted probability vector for the b-th sample. A smaller loss value indicates that the predicted failure mode is closer to the actual failure mode.

[0035] In this embodiment, a simulation is conducted for a "single-phase ground fault at a single node". The fault mode injection is achieved by setting parameters based on the IEEE 123 power grid system simulation model.

[0036] In this embodiment, the fault modes in this experiment include: one normal state and 123 states where single-phase grounding faults occur at each node, totaling 124 classes; 1000 samples were obtained from each class simulation. Considering that the samples in the later stages of simulation are more stable, 200 stable samples from the 801-1000 range of each class were selected, and Gaussian noise with a specified signal-to-noise ratio was added to simulate real monitoring noise; then, feature compression was performed using a sliding window with a window size of 10 and a step size of 10. Finally, 20 samples were obtained from each class, which were concatenated to form a dataset with dimensions (2480, 123, 6), and a 2480-length label set was constructed based on the fault node index; the training set and test set were divided at a ratio of 0.5, and standardized normalization was applied.

[0037] In this embodiment, the cross-entropy loss function is used during the training phase; the optimizer is Adam, the learning rate is 0.005, the epoch is 250, and the batch is 620, completing model training and parameter optimization under the condition of complete samples.

[0038] The fourth step involves using the fault diagnosis model trained and optimized in the third step to achieve robust fault diagnosis under varying degrees of random missing node data in fault feature datasets.

[0039] In this embodiment, during the testing phase, based on the trained fault diagnosis model, scenarios with 1-5 nodes "completely missing" are constructed for the test set samples, and the diagnostic performance under different degrees of missingness is evaluated. Each test involves completely losing the feature data of the missing nodes, and this process is repeated 200 times. The model outputs whether a fault exists and the fault node's location number: 0 indicates no fault, and 1-123 indicate a single-phase grounding fault at the corresponding node. The diagnostic effect is quantified using four metrics: accuracy, precision, recall, and F1 score.

[0040] Table 1: Average index results under the condition of 1-5 missing nodes

[0041] The above results demonstrate that the fault diagnosis method based on the feature-topology dual-dimensional graph attention mechanism proposed in this invention exhibits good diagnostic performance under varying degrees of node data loss conditions. The above specific embodiments further illustrate the purpose, technical solution and beneficial effects of this application. It should be understood that the above are only specific embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of this application should be included within the scope of protection of this application.

Claims

1. A fault diagnosis method based on a feature-topology dual-dimensional graph attention mechanism, characterized in that, The fault diagnosis method includes the following steps: The first step is to construct a power grid system relationship diagram. The physical connection relationships and fault characteristic correlations of the power grid system are comprehensively represented by the power grid system relationship diagram; The second step is to use the power grid system relationship diagram constructed in the first step. Using the input as input, a power grid system fault diagnosis model based on a feature-topology dual-dimensional graph attention mechanism is established. The power grid system fault diagnosis model includes a graph feature extraction module, a dual-dimensional graph attention mechanism module, and a fault classification module. The power grid system fault diagnosis model adopts a multi-layer cascaded graph convolutional-attention layer structure. Each graph convolutional-attention layer includes a graph feature extraction module and a two-dimensional graph attention mechanism module. The graph feature extraction module of each graph convolutional-attention layer uses a Chebyshev graph convolutional layer to aggregate the fault features of neighboring nodes on the power grid topology and converts the input node features from the original dimension into high-dimensional graph convolutional features. The two-dimensional graph attention mechanism module filters key fault features and key node locations based on the high-dimensional graph convolutional features. The third step, during the training phase, is to train the power grid system fault diagnosis model constructed in the second step using a complete fault feature dataset without any missing data. The fourth step involves using the power grid system fault diagnosis model trained and optimized in the third step to achieve robust fault diagnosis under varying degrees of random missing node data, targeting fault feature datasets with different levels of node data loss.

2. The fault diagnosis method based on a feature-topology dual-dimensional graph attention mechanism according to claim 1, characterized in that, The first step is specifically as follows: Constructing a power grid system relationship graph based on existing relationship graph optimization methods. , ,in A set of power grid monitoring nodes; It is a set of connection relationships; The node fault feature matrix contains the fault-sensitive electrical characteristics of the three-phase voltage and three-phase current of each node, which is used to characterize the power grid operating status. This is the topology connection weight matrix.

3. The fault diagnosis method based on a feature-topology dual-dimensional graph attention mechanism according to claim 2, characterized in that, The second step is specifically as follows: Step 2.1: The graph feature extraction module uses a Chebyshev graph convolutional layer to aggregate the fault features of neighboring nodes to the current node in the power grid topology, thereby extracting the graph fault features of each node in the power grid system. Step 2.2: The two-dimensional graph attention mechanism module sequentially uses the power grid sensitive fault feature screening function based on graph feature attention and the power grid key monitoring node screening function based on graph topology attention to process the graph fault features extracted in step 2.

1. Graph Fault Features Processed by the Two-Dimensional Graph Attention Mechanism Module and reshape it into new graph features. ,in For the batch sample size, This represents the number of nodes in the power grid system. Output the fault feature dimensions to the current graph feature extraction module; Step 2.3: The fault classification module flattens the node fault features after processing by multiple cascaded graph convolutional-attention layers using a fully connected layer. And perform fault information fusion and feature refinement, among which the first The fully connected layer performs dimensionality reduction on the input features and finally outputs accurate fault location information.

4. The fault diagnosis method based on a feature-topology dual-dimensional graph attention mechanism according to claim 3, characterized in that, Step 2.2 specifically refers to: Step 2.2.1, the power grid sensitive fault feature screening function is achieved through dimensionality reduction compression and dimensionality increase reconstruction. It learns the distribution and response patterns among fault features, filters out the fault feature information most important for fault diagnosis, and generates graph feature weights. ; Step 2.2.2: The key power grid monitoring node screening function learns the fault propagation patterns between adjacent nodes through feature splicing and topology dimensionality reduction, filters out the monitoring node information most important for fault diagnosis, and generates graph topology weights. ; Step 2.2.3: In the two-dimensional graph attention mechanism module, the original node fault features are weighted by the graph features. Graph topological weights By performing element-wise multiplication sequentially, an enhanced feature is obtained that simultaneously highlights important fault characteristics and important monitoring node information. : （7） in, Represents graph characteristics; This indicates element-wise multiplication.

5. The fault diagnosis method based on a feature-topology dual-dimensional graph attention mechanism according to claim 4, characterized in that, Step 2.2.1 specifically refers to: Step 2.2.1.1: Perform global average pooling on all nodes of the entire topology for each fault feature of the power grid to obtain the global response level of each fault feature, characterizing the overall health status of each fault feature in the power grid, and outputting the compressed fault feature representation. As shown below: （1） in, This indicates a global average pooling operation; Step 2.2.1.2: Compressed characterization of fault features Feature dimensionality reduction is performed using convolutional layers, reducing the number of features from... Compress to Forced fault feature compression representation retains the most important fault feature information for fault diagnosis during the compression process, and outputs the compressed important fault features. , The proportion of dimensionality reduction for channels: （2） in This is a convolution operation; Step 2.2.1.3, for important fault characteristics By increasing the dimensionality through convolutional layers, and based on the key feature information learned in the compression stage, the complete fault feature dimension weight distribution is reconstructed. : （3） Step 2.2.1.4: Weight distribution of fault feature dimensions By performing sigmoid normalization, graphical feature weights are generated to quantify the importance of each fault feature. .

6. The fault diagnosis method based on a feature-topology dual-dimensional graph attention mechanism according to claim 5, characterized in that, Step 2.2.2 specifically refers to: Step 2.2.2.1: Perform mean pooling and max pooling on each feature dimension of each node of the fault feature; （4） （5） in, This represents the average response of all fault characteristics at each node; i represents the index of the fault characteristic. This represents the fault characteristic value of the i-th node; Used to represent the most significant fault characteristics of each node; Step 2.2.2.2, will and By concatenating the features along the feature dimension, we obtain the comprehensive fault features of the nodes. ; Step 2.2.2.3: Through topological relationship learning convolutional layers, the topological dependencies and fault propagation relationships between nodes are learned, and the node comprehensive fault characteristics are integrated. Compression yields the topological dimension weight distribution : （6） Step 2.2.2.4: Learn the fault propagation pattern between adjacent nodes during the compression process; calculate the weight distribution of the topology dimension. Sigma-Oblique normalization is used to generate graph topology weights for quantifying the importance of each monitoring node. .

7. The fault diagnosis method based on a feature-topology dual-dimensional graph attention mechanism according to claim 6, characterized in that, The third step is specifically as follows: The cross-entropy loss function is used to measure the gap between the predicted results and the actual results of power grid system fault diagnosis, and to train the power grid system fault diagnosis model. The formula for the cross-entropy loss function is as follows: （8） in, This represents the fault mode predicted by the fault diagnosis model. Represents the actual failure mode; It is a loss function; That is, batch represents the size of each batch participating in training, b represents the index of the sample in the batch; C represents the dimension of the fault mode, i.e. the number, and c represents the index of the fault category. This represents the true label vector of the b-th sample; Let represent the predicted probability vector of the b-th sample.

Images