A method for fault diagnosis-oriented relational graph contrastive learning and autonomous optimization fusion

Abstract

The application provides a fault diagnosis-oriented relational graph contrast learning and autonomous optimization fusion method, and belongs to the technical field of power grid system fault diagnosis.The application constructs and optimizes a fusion topological relationship graph and a feature relationship graph, establishes a fusion relationship graph capable of reflecting the physical connection relationship of the power grid node and the fault feature correlation relationship, and thus accurately describes the fault propagation path; the feature relationship graph is autonomously generated based on the node fault feature through relational graph autonomous learning, the structural optimization and the relationship representation consistency of the two relationship graphs are strengthened through relational graph deep learning representation and relational graph contrast reinforcement, and the topological relationship graph is dynamically updated by the feature relationship graph through relational graph adaptive fusion.The fusion relationship graph constructed by the application can better represent the fault propagation path between the nodes under the complex power grid, and is more suitable for the fault diagnosis of the complex power grid.

Description

Technical Field

[0001] This invention belongs to the field of power grid system fault diagnosis technology, and relates to a relationship graph comparison learning and autonomous optimization fusion method for fault diagnosis, and particularly to a multi-dimensional information fusion and relationship graph autonomous optimization method for fault diagnosis of complex power grids. Background Technology

[0002] Existing technologies have attempted to abstract power system nodes and branches into graph structures and combine them with graph neural networks to achieve fault prediction / diagnosis. For example, Chinese invention patent CN111461392A discloses a power fault prediction method based on graph neural networks, which obtains prediction results by constructing the association relationship between power system nodes and inputting it into a graph neural network. In addition, some studies have used methods such as "statistical correlation / mutual information" to construct weighted association graphs and then use graph depth models to complete fault identification and classification [Qian Qiang, Ma Ping, Wang Nini, et al. Fault diagnosis of chemical process based on graph convolutional network under imbalanced sample [J]. Journal of Harbin Institute of Technology, 2025, 57(09):76-86].

[0003] However, the above methods generally suffer from the following drawbacks: (1) the relationship graph is often determined at once by electrical connections or statistical similarity, which is prone to inconsistencies with the actual fault propagation logic and misjudgments caused by edge weight / adjacency noise; (2) most works focus on "classification / prediction on a given relationship graph" without optimizing the relationship graph for fault diagnosis. In response to the above shortcomings, the relationship graph comparison learning and autonomous optimization fusion mechanism proposed in this invention can constrain and optimize the relationship graph to better characterize fault propagation and node relationships and improve diagnostic reliability.

[0004] To address the above issues, scholars both domestically and internationally have conducted extensive research and proposed several solutions. For example, Chinese invention patent CN111191333B provides an automatic distribution network topology identification method based on node electrical distance. By acquiring the electrical distance between nodes, it determines the node connection type based on the numerical relationship between the distances, thereby automatically constructing a global topology diagram of the distribution network. Chinese invention patent CN111932396B provides an automatic distribution network topology identification method. This method establishes an initial adjacency matrix, constructs parameter and state constraint relationships by combining DC power flow equations, and uses a SOM-BP neural network to identify branch opening and closing states, thereby dynamically updating the topology matrix and generating a real-time distribution network topology.

[0005] While existing technologies can construct or identify power grid topology based on physical distance or electrical parameters, the method CN111191333B, based on node electrical distance, relies entirely on static, prior electrical distance data to derive connection relationships, failing to reflect the dynamic and implicit fault correlations between node features in actual operation. The method CN111932396B, a dynamic update method for the topology matrix based on DC power flow constraints and branch status identification, focuses on updating the topology by determining the on / off state of physical branches through data; its edge weights are determined by fixed reactance parameters, failing to autonomously learn deeper, quantifiable correlation strengths from node feature data. Neither method integrates prior physical topology connections with data-driven node feature correlations within a unified framework for joint learning and optimization. This results in limitations in representing propagation paths under complex fault scenarios, affecting the accuracy of subsequent fault diagnosis. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a method for fusing relational graph comparative learning and autonomous optimization for fault diagnosis. It constructs a topological relational graph and a feature relational graph, respectively, and constrains the consistency between the topological relational graph and the feature relational graph through comparative learning. Finally, it fuses to obtain a relational graph that simultaneously represents the topological connection relationship and the feature connection relationship between nodes.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A graph comparison learning and autonomous optimization fusion method for fault diagnosis is proposed. This method, tailored to the physical structure of complex power grids, constructs an original topology graph and builds a graph optimization fusion model based on graph comparison learning. It then generates a fused graph by combining node fault characteristics. The specific steps include: The first step is to construct the original topology graph: Based on the physical structure of the complex power grid and the characteristics of node faults, the original topology graph is constructed. It includes node fault characteristics and a topological adjacency matrix, where, It is a set of nodes in the power grid system, used to characterize monitoring nodes in the power grid system such as transformers; It is a set of edges in the power grid system, representing the fault propagation paths between power grid monitoring nodes; Set of fault characteristics of power grid system nodes; This is the topological adjacency matrix, representing the fault propagation strength between nodes in the topological graph. Details are as follows: Step 1.1: Construct a set of power grid system nodes, using each monitoring node of the power grid system, such as a substation, distribution point, or load, as a node. , where subscript This indicates the number of nodes in the power grid system. This represents the Nth node in the power grid system.

[0008] Step 1.2: Using the transmission lines between power grid system monitoring nodes as edges, represent the fault propagation paths of the power grid system monitoring nodes, and construct the power grid system edge set. , where subscript This represents the relationship between any two nodes; This represents the Mth edge of the power grid system.

[0009] Step 1.3: Based on the node fault characteristics consisting of physical quantities such as voltage and current at each monitoring node of the power grid system, construct a set of node fault characteristics for the power grid system. The number of nodes included in the power grid system is The feature dimension collected by each node is 1. A total of Group samples, each group of samples collected is numbered as follows: Then for the first The nth sample, the th The node of the first Each fault characteristic is denoted as The node fault characteristics of all samples are represented as follows: In the actual construction process, the original node fault features collected from each node are first denoised and averaged using a sliding window to obtain stable node fault features. Then, the node fault features of all nodes are concatenated according to the sample number, node number, and feature number to form a complete set of node fault features. Finally, regarding Normalization is performed to distribute the feature values ​​within a uniform range, which facilitates subsequent model training and feature extraction.

[0010] use Represents the normalized i-th The nth sample, the th The node of the first One fault characteristic; and Let J and V represent the minimum and maximum values ​​obtained statistically for the same feature dimension j of all samples B and all nodes N, respectively.

[0011] Step 1.4: Measure the fault propagation intensity between power grid system nodes using the physical distance between each monitoring node, and construct an adjacency matrix between power grid system nodes. The initial adjacency matrix is ​​an empty set. The physical distance matrix between monitoring nodes in the power grid system. ,in Indicates the first The monitoring node and the first The physical distance between each monitoring node, for each node Choose the nearest one There are nodes, and the physical distance between them is... , This represents the number of nearest neighbor nodes.

[0012] For each node Nearest neighbor nodes Normalized distance weights The calculation formula is as follows: .in The value ranges from 0 to 1, with values ​​closer to 1 indicating a stronger connection between the two nodes. This assigns weights to all nodes. Fill the adjacency matrix with sparse matrix form In this method, by setting the diagonal elements to zero and ensuring that the elements within the matrix are symmetric, an adjacency matrix of an undirected graph is obtained. Used to describe the degree of interconnection between nodes in a power grid system.

[0013] Step 1.5: Construct the completed artificially constructed original topological relationship graph. .in It is a set of nodes used to represent monitoring nodes in power grid systems such as transformers; It is a set of edges, representing the fault propagation paths between power grid monitoring nodes; Indicates the node fault characteristics on the node; This is an adjacency matrix, representing the fault propagation intensity between power grid monitoring nodes.

[0014] The second step involves autonomous learning of the relationship graph based on the original topological relationship graph constructed in the first step: fault feature transformation and correlation measurement are performed on the node fault features to construct a feature adjacency matrix, thereby autonomously generating the feature relationship graph. , ,in, This represents the set of edges in the power grid system within the characteristic relation graph. Specifically: Step 2.1: Using a multilayer perceptron, fault feature transformation is carried out based on the node fault features in the first step, and fault feature transformation vectors are generated autonomously through graph structure learning.

[0015] Step 2.2: For all power grid system nodes, calculate the correlation metric of fault characteristic transformation vectors between nodes and autonomously generate a correlation metric matrix. Sort the correlation metrics between all nodes, determine the TOP-K method parameter K, retain the top K correlation metrics, set the remaining correlation metrics to zero, eliminate weak correlations, normalize, generate a feature adjacency matrix S, and construct a feature relationship graph. ,in This is the set of edges in the power grid system after removing edges with an affinity metric of zero.

[0016] The third step involves performing deep learning representation of the topological relationship graph obtained in the first step and the feature relationship graph obtained in the second step. This deep learning representation process includes feature extraction and feature mapping, generating node fault representation vectors for both the topological and feature relationship graphs. Specifically: Step 3.1: The deep learning representation of the relationship graph first undergoes robustness enhancement, with the feature relationship graph being the target of the robustness enhancement. and topological relationship diagram This includes node fault feature masking operations and fault propagation path blocking operations.

[0017] Step 3.2, Feature extraction of the relational graph: Through multi-layer graph convolutional layers, the node fault features of adjacent nodes are aggregated onto the current node on the topology and feature relational graph to extract features.

[0018] Step 3.3, feature mapping of the relational graph, involves mapping the features of the topological and feature relational graph extracted in Step 3.1 onto the contrastive learning space using a projection network composed of multiple fully connected layers and nonlinear activations, to obtain... and ,in, A node fault representation vector representing a node in a topological graph. The node fault representation vector represents the feature relationship graph.

[0019] The fourth step is relationship graph comparison and enhancement: The consistency difference between the two node fault representation vectors obtained in the third step is calculated, and the structure of the fusion model is autonomously optimized based on the consistency difference loss. Specifically: Step 4.1: Calculate the node fault representation vector of the topology graph. Node fault representation vectors in the feature relationship graph The consistency difference between them, and the formula for calculating the consistency difference loss function is: (1) (2) (3) in, This represents the consistency difference loss; n represents the number of nodes, and i represents the node index; The consistency difference term representing the contrastive learning form, where Indicates Match the same node for the anchor point , Indicates Match the same node for the anchor point j represents the node index used for summing the denominator. This represents the node fault representation vector of the j-th node in the topological graph; This is a cosine similarity function used to measure the correlation between two fault representation vectors; The focusing coefficient is used to adjust the sensitivity of the relational graph optimization fusion model to consistency differences between fault representation vectors. Decreasing the value amplifies consistency differences, making the graph optimization fusion model more focused on learning highly similar fault representation vectors; Increasing the value mitigates consistency discrepancies and enhances the robustness of the relationship graph optimization fusion model to noise and outliers.

[0020] Step 5, Adaptive Fusion of Relationship Graphs: After optimizing the relationship graph fusion model structure in Step 4, the feature relationship graph constructed in Step 2 is used... Dynamically update the topology graph constructed in the first step Output an autonomously optimized fusion relationship diagram As shown below: (4) Among them, fusion parameters It can control the relative weights of prior physical topology and fault-sensitive features in fault propagation path modeling.

[0021] The second and fifth steps are used to construct a graph optimization and fusion model based on graph contrast learning. Specifically, the second and fifth steps are used to generate a fused graph by combining the original topological graph and node fault features: graph autonomous learning, graph deep learning representation, graph contrast reinforcement, and graph adaptive fusion.

[0022] The sixth step involves training the constructed graph-based contrastive learning-based relationship graph optimization and fusion model. This model takes the original topological relationship graph and node fault features as input. In each iteration, it first performs autonomous graph learning, constructing a feature relationship graph based on fault feature transformations and correlation metrics between node fault features. Then, through a graph deep learning representation process, it extracts and maps graph features from the topological and feature relationship graphs, generating node fault representation vectors for graph contrast enhancement. It calculates the consistency difference between the two node fault representation vectors and dynamically updates the topological relationship graph using the feature relationship graph during adaptive fusion, outputting an autonomously optimized fused relationship graph. In each iteration, the fused relationship graph output from the previous iteration is used as the updated topological relationship graph as input to the model. During backpropagation optimization, the model parameters for fault feature transformations and graph deep learning representations are optimized through backpropagation based on consistency difference loss, ensuring that the fault representation vectors of the same node in the two relationship graphs tend to be consistent.

[0023] The beneficial effects of this invention are as follows: (1) This invention proposes a fusion method of relational graph comparison learning and autonomous optimization for fault diagnosis. This method addresses the problem that traditional power grid relational graphs based on physical distance are difficult to characterize the fault propagation path between nodes in complex power grids. By constructing and optimizing the fusion topology relational graph and feature relational graph, a fusion relational graph that can simultaneously reflect the physical connection relationship of power grid nodes and the correlation relationship of fault features is established, thereby accurately describing the fault propagation path; (2) This invention proposes a graph optimization and fusion model based on graph comparison learning. The model generates a feature graph based on node fault features through graph self-learning. The model optimizes the structure and strengthens the consistency of the relationship representation by graph deep learning representation and graph comparison enhancement. The model also dynamically updates the topology graph by using the feature graph through graph adaptive fusion.

[0024] Verification has shown that the fusion relationship diagram constructed by the method of the present invention for complex power grids can better characterize the fault propagation path between nodes in complex power grids and is more suitable for fault diagnosis of complex power grids. Attached Figure Description

[0025] Figure 1 This is a complete process for a relationship graph comparison learning and autonomous optimization fusion method for fault diagnosis.

[0026] Figure 2 To optimize the fusion model structure of the relational graph based on graph contrastive learning.

[0027] Figure 3 An initial topology diagram constructed for an example.

[0028] Figure 4 This is the final fusion relationship diagram. Detailed Implementation

[0029] The present invention will be further illustrated below with specific examples.

[0030] A method for fusing relational graph comparison learning and autonomous optimization for fault diagnosis includes the following steps: The first step is to construct the original topology graph: Based on the physical structure of the complex power grid and the characteristics of node faults, the original topology graph is constructed. It includes node fault characteristics and a topological adjacency matrix, where It is a set of nodes in the power grid system, used to characterize monitoring nodes in the power grid system such as transformers; It is a set of edges in the power grid system, representing the fault propagation paths between power grid monitoring nodes; Set of fault characteristics of power grid system nodes; This is the topological adjacency matrix, representing the fault propagation strength between nodes in the topological graph. Details are as follows: Step 1.1: Construct a set of power grid system nodes, using each monitoring node of the power grid system, such as a substation, distribution point, or load, as a node. , where subscript This indicates the number of nodes in the power grid system. This represents the Nth node in the power grid system. In this embodiment, the IEEE 123 power grid system is selected, and the number of nodes in the power grid system is N=123.

[0031] Step 1.2: Using the transmission lines between power grid system monitoring nodes as edges, represent the fault propagation paths of the power grid system monitoring nodes, and construct the power grid system edge set. , where subscript This represents the relationship between any two nodes; This represents the Mth edge of the power grid system.

[0032] Step 1.3: Based on the node fault characteristics consisting of physical quantities such as voltage and current at each monitoring node of the power grid system, construct a set of node fault characteristics for the power grid system. The number of nodes included in the power grid system is The feature dimension collected by each node is 1. A total of Group samples, each group of samples collected is numbered as follows: Then for the first The nth sample, the th The node of the first Each fault characteristic is denoted as The node fault characteristics of all samples can be represented as follows: In the actual construction process, the original node fault features collected from each node are first denoised and averaged using a sliding window to obtain stable node fault features. Then, the node fault features of all nodes are concatenated according to the sample number, node number, and feature number to form a complete set of node fault features. Finally, regarding Normalization is performed to distribute the feature values ​​within a uniform range, facilitating subsequent model training and feature extraction. The normalization formula is as follows: (5) in Represents the normalized i-th The nth sample, the th The node of the first One fault characteristic; and Let J and V represent the minimum and maximum values ​​obtained statistically for the same feature dimension j of all samples B and all nodes N, respectively.

[0033] In this embodiment, one normal state and 123 single-phase ground fault states are constructed, resulting in 124 fault modes. Simulation yields 1000 samples for each fault mode, from which 200 stable samples from slice 801-1000 are selected. Gaussian noise with a specified signal-to-noise ratio is added to these 200 samples. Then, the samples are averaged and compressed using a sliding window of 10 and a step size of 10, resulting in 20 fault feature samples for each class. The final dataset H has dimensions (2480, 123, 6), and 2480 labels are constructed based on the fault node index. The training and test sets are divided by a ratio of 0.5 and normalized.

[0034] Step 1.4: Measure the fault propagation intensity between power grid system nodes using the physical distance between each monitoring node, and construct an adjacency matrix between power grid system nodes. The initial adjacency matrix is ​​an empty set. The physical distance matrix between monitoring nodes in the power grid system. ,in Indicates the first The monitoring node and the first The physical distance between each monitoring node, for each node Choose the nearest one The physical distance between the nodes is , The nearest neighbor number. In this embodiment, the four nearest neighbors are selected for each node, i.e. , as nearest neighbors to construct a sparse topological adjacency matrix.

[0035] For each node Nearest neighbor nodes Normalized distance weights The calculation formula is as follows: ;in The value ranges from 0 to 1, with values ​​closer to 1 indicating a stronger connection between the two nodes. This assigns weights to all nodes. Fill the adjacency matrix with sparse matrix form In this method, by setting the diagonal elements to zero and ensuring that the elements within the matrix are symmetric, we obtain the adjacency matrix of an undirected graph. Used to describe the degree of interconnection between nodes in a power grid system.

[0036] Step 1.5: Construct the completed artificially constructed original topological relationship graph. .in It is a set of nodes used to represent monitoring nodes in power grid systems such as transformers; It is a set of edges, representing the fault propagation paths between power grid monitoring nodes; Indicates the node fault characteristics on the node; The adjacency matrix represents the fault propagation strength between power grid monitoring nodes. In this embodiment, the original topology graph contains 123 nodes and 654 valid edges; after being partitioned by a ratio of 0.5, the training set has 1240 samples.

[0037] The second step involves autonomous learning of the relationship graph based on the original topological relationship graph constructed in the first step: fault feature transformation and correlation measurement are performed on the node fault features to construct a feature adjacency matrix, thereby autonomously generating the feature relationship graph. , ,in, This represents the set of edges in the power grid system within the characteristic relation graph. Specifically: Step 2.1: Using a multilayer perceptron, based on the node fault features from Step 1, perform fault feature transformation based on graph structure learning, and autonomously generate fault feature transformation vectors. The fault feature transformation process is as follows: (6) (7) (8) in, Here is the weight matrix for each layer of the perceptron. For the bias vector of each layer, It is a non-linear activation function, typically... The function. Each layer of a multilayer perceptron includes a linear transformation and a nonlinear activation, ultimately resulting in a fault feature transformation vector. Feature dimension and While remaining unchanged, these fault feature transformation vectors are more discriminative and expressive than the original data, and their correlation measures are better able to characterize the fault propagation paths between power grid nodes.

[0038] Step 2.2: For all power grid system nodes, calculate the correlation metric of the fault characteristic transformation vectors between nodes. The association measurement matrix is ​​generated autonomously. The formula for calculating the association measurement using cosine similarity is as follows: (9) Relationship measurement between all nodes The associations are sorted, the TOP-K method parameter K is determined, the top K association metrics are retained, the remaining association metrics are set to zero, weaker associations are eliminated, and after normalization, a feature adjacency matrix S is generated, and a feature relationship graph is constructed. ,in This is the set of edges in the power grid system after removing edges with an association metric of zero. In this embodiment, K=4.

[0039] The third step involves performing deep learning representation of the topological relationship graph obtained in the first step and the feature relationship graph obtained in the second step. This deep learning representation process includes feature extraction and feature mapping, generating node fault representation vectors for both the topological and feature relationship graphs. Specifically: Step 3.1: The deep learning representation of the relationship graph first undergoes robustness enhancement, with the feature relationship graph being the target of the robustness enhancement. and topological relationship diagram This includes node fault feature masking operations and fault propagation path blocking operations. The specific operations for robustness enhancement are as follows: For the input node fault characteristics With mask ratio Randomly generate data masking matrix The fault characteristic data after data masking is as follows: (10) (11) in, This indicates element-wise multiplication.

[0040] For the adjacency matrix of the original topological graph Adjacency matrix of the feature relation graph According to the blocking ratio Randomly block some edges, the specific formula is as follows: (12) (13) in, The probability of matching is Bernoulli distribution.

[0041] After the above two steps, the enhanced node fault characteristics are obtained. , Adjacency Matrix , .

[0042] Step 3.2, Feature extraction of the relational graph: Through multi-layer graph convolutional layers, the node fault features of adjacent nodes are aggregated onto the current node on the topology and feature relational graph to extract features; Relationship graph feature extraction uses multi-layer graph convolutional layers to aggregate the node fault features of adjacent nodes in two relationship graphs and extract features from the current node. and : (14) (15) The basic operations of a graph convolutional layer are represented as follows: (16) in, Indicates the first Layer node failure characteristics, Indicates the first The learnable weight matrix of the layer, This represents a non-linear activation function.

[0043] Step 3.2, feature mapping of the relational graph, involves mapping the features of the topological and feature relational graph extracted in Step 3.1 onto the contrastive learning space using a projection network composed of multiple fully connected layers and nonlinear activations, to obtain... and ,in, A node fault representation vector representing a node in a topological graph. The node fault representation vector represents the feature relationship graph.

[0044] The mapping process of the projection network is represented as follows: (17) (18) in, This represents a projection network.

[0045] Step 4, Relationship Graph Comparison and Enhancement: Calculate the consistency difference between the fault representation vectors of the two nodes obtained in Step 3, and autonomously optimize the relationship graph fusion model structure based on the consistency difference loss. Specifically: Step 4.1: Calculate the node fault representation vector of the topology graph. Node fault representation vectors in the feature relationship graph The consistency difference between them, and the formula for calculating the consistency difference loss function is: (1) (2) (3) in, This represents the consistency difference loss; n represents the number of nodes, and i represents the node index; The consistency difference term representing the contrastive learning form, where Indicates Match the same node for the anchor point , Indicates Match the same node for the anchor point j represents the node index used for summing the denominator. This represents the node fault representation vector of the j-th node in the topological graph; This is a cosine similarity function used to measure the correlation between two fault representation vectors; The focusing coefficient is used to adjust the model's sensitivity to consistency differences between fault representation vectors. Decreasing the value amplifies consistency differences, making the model more focused on learning highly similar fault representation vectors; Increasing the value mitigates consistency discrepancies and enhances the model's robustness to noise and outliers.

[0046] Step 5, Adaptive Fusion of Relationship Graphs: After optimizing the model structure in step 4, the feature relationship graphs constructed in step 2 are used. Dynamically update the topology graph constructed in the first step Output an autonomously optimized fusion relationship diagram As shown below: (4) Among them, fusion parameters It can control the relative weights of prior physical topology and fault-sensitive features in fault propagation path modeling.

[0047] The process from step two to step five is used to construct a graph optimization and fusion model based on graph contrastive learning. The construction process combines the original topological graph and node fault features to generate a fused graph: graph autonomous learning, graph deep learning representation, graph contrastive reinforcement, and graph adaptive fusion.

[0048] The sixth step involves training the constructed graph-based contrastive learning-based relationship graph optimization and fusion model. This model takes the original topological relationship graph and node fault features as input. In each iteration, it first performs autonomous graph learning, constructing a feature relationship graph based on fault feature transformations and correlation metrics between node fault features. Then, through a graph deep learning representation process, it extracts and maps graph features from the topological and feature relationship graphs, generating node fault representation vectors for graph contrast enhancement. It calculates the consistency difference between the two node fault representation vectors and dynamically updates the topological relationship graph using the feature relationship graph during adaptive fusion, outputting an autonomously optimized fused relationship graph. In each iteration, the fused relationship graph output from the previous round is used as the updated topological relationship graph as input to the model. During backpropagation optimization, the model parameters for fault feature transformations and graph deep learning representations are optimized through backpropagation based on consistency difference loss, ensuring that the fault representation vectors of the same node in the two relationship graphs tend to be consistent.

[0049] In this embodiment, the number of edges in the fused relationship graph is reduced from 654 to 314 (a reduction of 51.99%), and the graph density is reduced from 0.0436 to 0.0209; the mean edge weight is increased from 0.5609 to 0.8154, and the standard deviation is reduced from 0.2756 to 0.1387; the original edge retention rate is 48.0%, and the average weight of the retained edges is enhanced by 0.0153, with a maximum enhancement of 0.2575; the average degree is reduced from 5.32 to 2.55, and the standard deviation of the degree distribution is reduced from 1.77 to 1.60.

[0050] 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 method for fusing relational graph comparison learning and autonomous optimization for fault diagnosis, characterized in that, The relationship graph comparison learning and autonomous optimization fusion method is designed for complex power grid physical structures and includes the following steps: The first step is to construct the original topology graph: Based on the physical structure of the complex power grid and the characteristics of node faults, the original topology graph is constructed. It includes node fault characteristics and a topological adjacency matrix, where, It is a set of nodes in the power grid system, used to characterize monitoring nodes in the power grid system such as transformers; It is a set of edges in the power grid system, representing the fault propagation paths between power grid monitoring nodes; Set of fault characteristics of power grid system nodes; This is the topological adjacency matrix, representing the fault propagation strength between nodes in the topological graph; The second step involves autonomous learning of the relationship graph based on the original topological relationship graph constructed in the first step: fault feature transformation and correlation measurement are performed on the node fault characteristics to construct a feature adjacency matrix, thereby autonomously generating the feature relationship graph. , ,in, This represents the set of edges in the power grid system within the feature graph. The third step is to perform deep learning representation of the topological relationship graph obtained in the first step and the feature relationship graph obtained in the second step. The deep learning representation process includes feature extraction and feature mapping of the relationship graph, generating node fault representation vectors for the topological relationship graph and the feature relationship graph. The fourth step is to strengthen the relationship graph by comparing and contrasting the two node fault representation vectors obtained in the third step. The consistency difference between the two nodes is calculated, and the structure of the fusion model is optimized autonomously based on the consistency difference loss. Step 5, Adaptive Fusion of Relationship Graphs: After optimizing the relationship graph fusion model structure in Step 4, the feature relationship graph constructed in Step 2 is used... Dynamically update the topology graph constructed in the first step Output an autonomously optimized fusion relationship diagram As shown below: （4） in, Indicates the fusion parameters; The sixth step is to train the constructed graph-based comparison learning-based relational graph optimization and fusion model to obtain the trained relational graph optimization and fusion model.

2. The method for fusing relational graph comparison learning and autonomous optimization for fault diagnosis as described in claim 1, characterized in that, The first step is as follows: Step 1.1: Construct a set of power grid system nodes, using each monitoring node of the power grid system, such as a substation, distribution point, or load, as a node. , where subscript This indicates the number of nodes in the power grid system. This represents the Nth node in the power grid system; Step 1.2: Using the transmission lines between power grid system monitoring nodes as edges, represent the fault propagation paths of the power grid system monitoring nodes, and construct the power grid system edge set. , where subscript This represents the relationship between any two nodes; Indicates the Mth edge of the power grid system; Step 1.3: Construct a set of node fault characteristics for the power grid system based on the node fault characteristics at each monitoring node of the power grid system. The number of nodes included in the power grid system is The feature dimension collected by each node is 1. A total of Group samples, each group of samples collected is numbered as follows: Then for the first The nth sample, the th The node of the first Each fault characteristic is denoted as The node fault characteristics of all samples are represented as follows: In the actual construction process, the original node fault features collected from each node are first denoised and averaged using a sliding window to obtain stable node fault features. Then, the node fault features of all nodes are concatenated according to the sample number, node number, and feature number to form a complete set of node fault features. Finally, regarding Perform normalization processing; use Represents the normalized i-th The nth sample, the th The node of the first One fault characteristic; and Let J represent the minimum and maximum values ​​obtained by statistically analyzing the same feature dimension j for all samples B and all nodes N, respectively. Step 1.4: Measure the fault propagation intensity between power grid system nodes using the physical distance between each monitoring node, and construct an adjacency matrix between power grid system nodes. The initial adjacency matrix is ​​an empty set; the physical distance matrix between monitoring nodes in the power grid system. ,in Indicates the first The monitoring node and the first The physical distance between each monitoring node, for each node Choose the nearest one There are nodes, and the physical distance between them is... , This represents the number of nearest neighbor nodes. For each node Nearest neighbor nodes Normalized distance weights The calculation formula is as follows: ;in The value is between 0 and 1; the weight between all nodes. Fill the adjacency matrix with sparse matrix form In this method, by setting the diagonal elements to zero and ensuring that the elements within the matrix are symmetric, we obtain the adjacency matrix of an undirected graph. ; Step 1.5: Construct the completed artificially constructed original topological relationship graph. .

3. The method for fusing relational graph comparison learning and autonomous optimization for fault diagnosis according to claim 2, characterized in that, The second step is as follows: Step 2.1: Using a multilayer perceptron, based on the node fault features in the first step, perform fault feature transformation based on graph structure learning, and autonomously generate fault feature transformation vectors. Step 2.2: For all power grid system nodes, calculate the correlation metric of fault characteristic transformation vectors between nodes, and autonomously generate a correlation metric matrix; sort the correlation metric values ​​between all nodes, determine the TOP-K method parameter K, retain the top K correlation metric values, set the remaining correlation metric values ​​to zero, normalize them to generate a feature adjacency matrix S, and construct a feature relationship graph. ,in This is the set of edges in the power grid system after removing edges with an affinity metric of zero.

4. The method for fusing relational graph comparison learning and autonomous optimization for fault diagnosis according to claim 3, characterized in that, The third step is as follows: Step 3.1: The deep learning representation of the relationship graph first undergoes robustness enhancement, with the feature relationship graph being the target of the robustness enhancement. and topological relationship diagram This includes node fault feature masking operations and fault propagation path blocking operations; Step 3.2, Feature extraction of the relational graph: Through multi-layer graph convolutional layers, the node fault features of adjacent nodes are aggregated onto the current node on the topology and feature relational graph to extract features; Step 3.3, feature mapping of the relational graph, involves mapping the features of the topological and feature relational graph extracted in Step 3.1 onto the contrastive learning space using a projection network composed of multiple fully connected layers and nonlinear activations, to obtain... and ,in, A node fault representation vector representing a node in a topological graph. The node fault representation vector represents the feature relationship graph.

5. The method for fusing relational graph comparison learning and autonomous optimization for fault diagnosis according to claim 4, characterized in that, The fourth step is as follows: Step 4.1: Calculate the node fault representation vector of the topology graph. Node fault representation vectors in the feature relationship graph The consistency difference between them, and the formula for calculating the consistency difference loss function is: （1） （2） （3） in, This represents the consistency difference loss; n represents the number of nodes, and i represents the node index; The consistency difference term representing the contrastive learning form, where Indicates Match the same node for the anchor point , Indicates Match the same node for the anchor point j represents the node index used for summing the denominators. This represents the node fault representation vector of the j-th node in the topological graph; The cosine similarity function; This is the focusing coefficient.

6. The method for fusing relational graph comparison learning and autonomous optimization for fault diagnosis according to claim 5, characterized in that, The sixth step is as follows: The aforementioned graph optimization and fusion model takes the constructed original topological graph and node fault features as input. In each round of iterative optimization, it first performs autonomous graph learning, constructing a feature graph based on fault feature transformation and correlation measurement between node fault features. Then, through a graph deep learning representation process, it extracts graph features and maps graph features to the topological graph and feature graph, generating node fault representation vectors for graph comparison and enhancement. It calculates the consistency difference between the two node fault representation vectors. In the graph adaptive fusion process, it dynamically updates the topological graph using the feature graph, outputting an autonomously optimized fusion graph. In each iteration, the fusion relationship graph output from the previous round is used as the updated topology relationship graph as the input to the model. In the backpropagation optimization process of the relationship graph optimization fusion model, the model parameters of the fault feature transformation and the deep learning representation of the relationship graph are optimized by backpropagation based on the consistency difference loss, so that the fault representation vector of the same node in the two relationship graphs tends to be consistent.

Images