Power equipment fault diagnosis method based on adaptive fuzzy reasoning

Abstract

The invention discloses a power equipment fault diagnosis method based on adaptive fuzzy reasoning. The method comprises the following steps: S1, constructing a multi-source normalized feature vector set; s2, constructing a power system diagram; s3, executing an attention graph convolution operation, and outputting a node embedding state vector set; s4, obtaining an edge synapse updating weight matrix; s5, obtaining a fuzzy rule base; s6, executing a multi-stage fuzzy reasoning process, and outputting a fault diagnosis judgment result; s7, writing a fault diagnosis judgment result into a state memory unit of a corresponding node to form a node state memory record and an edge state evolution record; and S8, when the confidence coefficient of a fault diagnosis judgment result is lower than a set threshold value, calling a node state memory record and an edge state evolution record, executing time sequence sliding window backtracking processing, updating a node embedding state vector set, and repeating the step S6 to execute re-judgment. According to the invention, the graph neural network and adaptive fuzzy reasoning are fused, and intelligent diagnosis of power faults is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of power systems, and in particular to a method for diagnosing power equipment faults based on adaptive fuzzy reasoning. Background Art

[0002] With the continuous expansion of power systems and the increasing complexity of equipment structures, real-time perception of power equipment operating status and fault diagnosis have become critical to ensuring safe and stable system operation. Traditional fault diagnosis methods rely on manual inspections, static rule bases based on expert experience, or single-point feature analysis based on signal processing. These methods often suffer from low diagnostic accuracy, slow response, and poor scalability when faced with multi-source heterogeneous data, large-scale complex electrical topologies, and multiple types of coupled faults. These methods struggle to meet the growing demand for high reliability and intelligence in modern power systems.

[0003] In recent years, with the increasing application of technologies such as artificial intelligence, graph neural networks, and fuzzy reasoning in industrial intelligent operation and maintenance, some research has begun to attempt to combine graph structure modeling with deep learning methods to perform structured modeling of power systems and, based on this, conduct intelligent diagnosis. However, existing graph neural network methods mostly focus on the deep extraction and classification of node features, ignoring the dynamic role of connection relationships between devices in the state evolution process. In particular, the mechanism of how node state changes affect edge connection weights is poorly modeled, making it difficult to accurately characterize the potential causal transmission chain in the electrical system.

[0004] In the area of ​​fuzzy reasoning, traditional methods are mostly based on preset rule bases, relying on domain experts to manually design fuzzy rules and set parameters. This static, fixed rule-building approach is difficult to cope with the dynamic changes in equipment status, connection modes, and fault mechanisms in power systems. In particular, when faced with the intersection of multi-source fuzzy information and multi-node linkage anomalies, the coverage and adaptability of the rule base are seriously insufficient, resulting in reduced reasoning accuracy and even misjudgment. In addition, existing fuzzy systems generally lack effective utilization of system structure information during rule activation and reasoning, and are unable to combine topological dependencies between devices to make causal constraint judgments, making it difficult to achieve high-confidence reasoning for actual fault propagation paths.

[0005] To address these issues, some research has begun to introduce the technical approach of integrating graph neural networks with fuzzy systems. These approaches treat power equipment as nodes in the graph, model electrical connections as edge structures, and use multi-source monitoring data as input features to achieve structure-aware state representation and enhanced reasoning. However, these methods still have several shortcomings. For one thing, the interaction between node states and edge connection weights has not yet been systematically modeled, resulting in a lack of a time series representation of the evolution of edge synaptic states. Furthermore, fuzzy rule construction lacks a joint screening mechanism for historical state coverage, state consistency, and causal relationships in the graph, resulting in low rule reliability. Furthermore, current methods generally lack effective self-feedback and backtracking mechanisms when reasoning confidence is insufficient, making it difficult to support fault re-diagnosis and model adaptive optimization.

[0006] Therefore, how to provide a power equipment fault diagnosis method based on adaptive fuzzy reasoning is a problem that those skilled in the art urgently need to solve. Summary of the Invention

[0007] One purpose of the present invention is to propose a method for diagnosing power equipment faults based on adaptive fuzzy reasoning. The present invention integrates graph neural networks and adaptive fuzzy reasoning methods, constructs a power system graph structure and dynamically models node states and edge synaptic weight changes, automatically generates high-confidence fuzzy rules, and realizes multi-level fault type identification, location positioning and risk ranking. It has the advantages of strong model self-learning ability, good interpretability of reasoning results, and high adaptability to complex faults, effectively improving the intelligence, precision and reliability of power equipment fault diagnosis.

[0008] According to an embodiment of the present invention, a method for diagnosing faults of electric power equipment based on adaptive fuzzy reasoning includes the following steps:

[0009] S1. Collect multi-source data from devices in the power system, perform time alignment and normalization processing, and construct a multi-source normalized feature vector set;

[0010] S2. constructing a power system graph structure based on the multi-source normalized feature vector set and initializing the fuzzy state vector of each node;

[0011] S3. Input the power system graph structure and the fuzzy state vector into a graph neural network, perform a structure-dependent attention graph convolution operation, and output a set of node embedding state vectors;

[0012] S4. Calculate the synaptic plasticity weight change of each edge in the power system diagram and generate an edge synaptic update weight matrix based on the fuzzy state change of the edge-connected nodes;

[0013] S5. Based on the node fuzzy state vector set and the edge synaptic update weight matrix, a candidate fuzzy rule set is constructed, and effective fuzzy rules are screened according to the coverage, state consistency and causal relationship in the graph of the candidate fuzzy rule set to obtain a fuzzy rule base;

[0014] S6. Based on the fuzzy rule base and the node embedded state vector set, a multi-level fuzzy reasoning process is executed to output the fault diagnosis judgment result, including the fault type judgment result, the fault location judgment result and the priority ranking result;

[0015] S7. Writing the fault diagnosis result into the state memory unit of the corresponding node, and updating the edge synapse history state in the power system graph structure according to the edge synapse update weight matrix to form a node state memory record and an edge state evolution record;

[0016] S8. When the confidence level of the fault diagnosis result is lower than the set threshold, call the node state memory record and the edge state evolution record, perform time series sliding window backtracking processing, update the node embedded state vector set, and repeat step S6 to perform re-determination.

[0017] Optionally, the multi-source data includes voltage data, current data, temperature rise data, vibration data, partial discharge data and acoustic signal data.

[0018] Optionally, the power system graph structure is constructed with power equipment as nodes and electrical connection relationships between electronic devices as edges. The fuzzy state vector of each node is initialized according to the characteristic values ​​corresponding to each electronic device in the multi-source normalized feature vector set. The fuzzy state vector is composed of the fuzzy membership of the electronic device in normal state, overheating fault state, partial discharge fault state and mechanical vibration abnormal state.

[0019] Optionally, it is characterized in that S3 specifically includes:

[0020] S31. Constructing a graph neural network based on device topology relationships, wherein the graph neural network uses the fuzzy state vector of each node in the power system diagram as an initial input feature;

[0021] S32. Introducing a structure-dependent attention mechanism into the graph neural network, dynamically calculating the attention weight coefficient based on the topological connection relationship and feature similarity between each node and its adjacent nodes;

[0022] S33, based on the attention weight coefficient, performing weighted aggregation on the fuzzy state vectors of the adjacent nodes of each node to obtain an intermediate feature representation of structure perception;

[0023] S34, introduce the node's own historical embedded state vector and fuse it with the current intermediate feature representation to enhance the node representation's ability to express time-evolving features;

[0024] S35, using a nonlinear activation function to process the fused intermediate feature representation to obtain the node embedding state vector of the current propagation layer;

[0025] S36. Repeat steps S32 to S35 to perform multiple rounds of graph convolution propagation until the set number of propagation layers is reached, and the output node embedding state vector set is output.

[0026] Optionally, the S4 specifically includes:

[0027] S41, obtaining the current fuzzy state vector of the starting node and the target node connected to each side of the power system diagram;

[0028] S42, based on the change of the fuzzy state vector of the edge-connected node, calculate the relative change rate of the fuzzy state of the edge at the current moment as a state coupling change indicator;

[0029] S43. Based on the side synaptic plasticity adjustment function, dynamically generate a side synaptic plasticity weight value according to the change amplitude and direction of the state coupling change indicator:

[0030] w ij (t) = η·tanh(γ·||[μ i (t)-μ i (t-1)]⊙[μ j (t)-μ j (t-1)]||2);

[0031] Among them, w ij (t) represents the edge synaptic plasticity weight value between the i-th node and the j-th node, η represents the global learning rate factor, tanh represents the hyperbolic tangent function, γ represents the gain coefficient of the adjustment sensitivity, ⊙ represents the Hadamard product, μ i (t) represents the fuzzy state vector of the i-th node at the current moment, μ i (t-1) represents the fuzzy state vector of the i-th node at the previous moment, μ j (t) represents the fuzzy state vector of the jth node at the current moment, μ j (t-1) represents the fuzzy state vector of the jth node at the previous moment, ||·||2 represents the L2 norm;

[0032] S44: Normalize the edge synapse plasticity weight values, fill the normalized edge synapse plasticity weight values ​​into the adjacency matrix of the corresponding power system graph, and generate and output an edge synapse update weight matrix.

[0033] Optionally, the S5 specifically includes:

[0034] S51. Based on the node fuzzy state vector set, extract the current fuzzy state information from each node, and locate the directly connected adjacent nodes in combination with the topological connection relationship in the power system diagram;

[0035] S52. Perform weighted processing on the topological connection relationship according to the edge synaptic plasticity weight value recorded in the edge synapse update weight matrix, and construct a triple set including edge connection direction, connection strength, and associated node state change;

[0036] S53. Utilize a fuzzy logic rule construction mechanism to perform premise-result modeling on each triplet, and generate a candidate fuzzy rule set based on the following pattern: when the fuzzy state of node A is in the first state category, the edge synaptic plasticity weight value between node A and node B is a set strength, and the fuzzy state of node B shows a trend of changing to the second state category, it is inferred that the state of node B will evolve to the third state category. The first state category corresponds to the normal state of the power equipment, the second state category corresponds to the overheating fault state, and the third state category corresponds to the partial discharge state or the abnormal mechanical vibration state. The edge synaptic plasticity weight value is used to reflect the influence of state conduction between nodes and participates in the construction and correction of fuzzy causal relationships.

[0037] S54. Counting the proportion of samples matched by each candidate fuzzy rule in the historical state evolution record as the coverage of the candidate fuzzy rule;

[0038] S55. Calculate a state consistency score by performing a similarity analysis on the fuzzy membership between the premise state and the result state involved in the candidate fuzzy rules. The state consistency score reflects the confidence level of the fuzzy logic match. Simultaneously, perform graph structure consistency verification on the edges and node paths involved in each rule to exclude false rules that violate the actual electrical connection logic of the power system.

[0039] S56. Candidate rules with a coverage higher than a set range threshold, a state consistency score higher than a set score threshold, and satisfying causal reachability in the graph structure are screened as valid fuzzy rules, and a fuzzy rule base is constructed.

[0040] Optionally, the S6 specifically includes:

[0041] S61, using the fuzzy rule base and the node embedded state vector set as rule input and state input of multi-level fuzzy reasoning respectively;

[0042] S62, fuzzifying the node embedded state vector set, converting the current state representation of each node into membership values ​​corresponding to multiple state categories, for adapting the precondition structure of the fuzzy rule;

[0043] S63. In the first-level fuzzy reasoning stage, the activation strength is calculated based on the matching degree between the premise of each fuzzy rule and the fuzzified node state, forming a primary fuzzy reasoning result for identifying a set of fault types and outputting a confidence score for each type.

[0044] S64. In the second-level fuzzy reasoning stage, the spatial distribution positions and topological dependency paths of the nodes involved in the activation rules are further combined to calculate the state impact transmission direction, determine the node location where the fault occurs, and output the fault probability of each node;

[0045] S65. In the third-level fuzzy reasoning stage, a risk weighted assessment is performed on the identified faults based on the severity of the fault type, the importance index of the fault location, and the current load status, and a fault event priority ranking result is generated;

[0046] S66. Structuring the output results of the multi-level fuzzy reasoning process to generate fault diagnosis results, including fault type determination results, fault location determination results, and fault event priority ranking results.

[0047] Optionally, the S7 specifically includes:

[0048] S71. Match the fault diagnosis result with each node in the power system graph structure, generate structured state update data for the nodes with matching results, and bind a timestamp to record the current time;

[0049] S72, writing the state update data into the state memory unit of the node, wherein the state memory unit is organized in a time series form and records the fuzzy state vector, fault diagnosis result label and confidence output of the node at multiple consecutive time points;

[0050] S73. Extracting structural information of the edge connections between the node and other nodes based on the state change content of the newly written node in the state memory unit, and querying the corresponding edge synaptic plasticity weight value in the edge synapse update weight matrix;

[0051] S74. For each edge connected to a node whose state has changed, calculate the incremental change value of the historical state of the edge synapse based on the magnitude and direction of the state change before and after the connected node, combined with the edge synapse plasticity weight value, and form a difference vector.

[0052] S75. Write the differential vector into the edge state evolution recording unit, organize it by edge identifier index, and record the change sequence of the edge synaptic plasticity weight value at different time steps, including the change trend, historical state strength, and update trigger source node number;

[0053] S76. The updated node state memory unit and the edge state evolution recording unit are stored in association in the graph structure in the form of key-value pairs.

[0054] The beneficial effects of the present invention are:

[0055] First, the present invention models various types of equipment in the power system as nodes in a graph structure, constructs edges in the graph based on electrical connection relationships, and combines multi-source feature vectors after time alignment and normalization. This can accurately characterize the structural dependencies and state coordination relationships between various devices in the power system, providing a structured input basis for subsequent graph neural network analysis.

[0056] Secondly, this paper introduces a structure-dependent attention graph convolution mechanism into graph neural networks, effectively extracting high-order feature interactions between node fuzzy states and adjacency relationships. This allows for stronger spatiotemporal dynamic expression of node embeddings while maintaining the physical connection context. Furthermore, by calculating the plasticity weight changes of edge synapses and combining them with node state evolution, the edge synapse update weight matrix is ​​dynamically generated, enabling time-varying modeling of connection relationships. This overcomes the expressiveness and adaptability limitations of traditional static graph diagnostic models.

[0057] Furthermore, the present invention has designed a fuzzy rule construction method that integrates node state vectors and edge weight structures. This method no longer relies on manually designed rules, but instead automatically extracts candidate rules from the system state evolution process. Rules are then screened based on coverage, state consistency, and causal structure constraints, significantly improving the accuracy, practicality, and scalability of fuzzy inference rules. During the fuzzy inference stage, a multi-level progressive inference process is constructed to implement fault type identification, fault location determination, and priority sorting, enabling the system to respond to complex fault patterns in a hierarchical and high-confidence manner.

[0058] Furthermore, the present invention introduces a state memory mechanism and a structure for recording the state evolution of edge synapses. This mechanism writes diagnostic results to node state storage and updates the historical state evolution trajectory of edge synapses in real time, establishing a dynamic and traceable memory framework for the system. When diagnostic confidence is insufficient, backtracking reasoning on state and edge information can be performed based on a sliding time window, effectively enabling self-feedback and re-judgment during the diagnostic process, improving the system's stability and resilience in atypical failure scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0060] Figure 1 This is an overall flow chart of a method for diagnosing power equipment faults based on adaptive fuzzy reasoning proposed by the present invention;

[0061] Figure 2 This is a flowchart of candidate fuzzy rule construction and effective rule screening for a power equipment fault diagnosis method based on adaptive fuzzy reasoning proposed by the present invention. DETAILED DESCRIPTION

[0062] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0063] refer to Figure 1 and Figure 2 , a method for fault diagnosis of power equipment based on adaptive fuzzy reasoning, comprising the following steps:

[0064] S1. Collect multi-source data from devices in the power system, perform time alignment and normalization processing, and construct a multi-source normalized feature vector set;

[0065] S2. constructing a power system graph structure based on the multi-source normalized feature vector set and initializing the fuzzy state vector of each node;

[0066] S3. Input the power system graph structure and the fuzzy state vector into a graph neural network, perform a structure-dependent attention graph convolution operation, and output a set of node embedding state vectors;

[0067] S4. Calculate the synaptic plasticity weight change of each edge in the power system diagram and generate an edge synaptic update weight matrix based on the fuzzy state change of the edge-connected nodes;

[0068] S5. Based on the node fuzzy state vector set and the edge synaptic update weight matrix, a candidate fuzzy rule set is constructed, and effective fuzzy rules are screened according to the coverage, state consistency and causal relationship in the graph of the candidate fuzzy rule set to obtain a fuzzy rule base;

[0069] S6. Based on the fuzzy rule base and the node embedded state vector set, a multi-level fuzzy reasoning process is executed to output the fault diagnosis judgment result, including the fault type judgment result, the fault location judgment result and the priority ranking result;

[0070] S7. Writing the fault diagnosis result into the state memory unit of the corresponding node, and updating the edge synapse history state in the power system graph structure according to the edge synapse update weight matrix to form a node state memory record and an edge state evolution record;

[0071] S8. When the confidence level of the fault diagnosis result is lower than the set threshold, call the node state memory record and the edge state evolution record, perform time series sliding window backtracking processing, update the node embedded state vector set, and repeat step S6 to perform re-determination.

[0072] This technical solution forms a fault diagnosis framework that integrates data-driven and structural expression by constructing a power system graph structure driven by multi-source normalized feature vectors and combining node fuzzy state vectors with graph neural networks for deep graph reasoning. This method can effectively integrate multi-source heterogeneous data such as voltage, current, temperature rise, vibration, partial discharge, and acoustic signals, and improve the consistency of feature scales and model generalization capabilities through normalization processing; the introduction of graph structure modeling can depict the actual electrical connection relationship between devices, improving the authenticity and interpretability of fault propagation chain modeling; the integration of fuzzy logic and neural network reasoning mechanisms makes the model have good adaptability and fault tolerance when dealing with multi-state nonlinear systems, thereby significantly improving the robustness, accuracy, and response speed of the diagnostic system under complex working conditions.

[0073] In this embodiment, the multi-source data includes voltage data, current data, temperature rise data, vibration data, partial discharge data and acoustic signal data.

[0074] In this embodiment, the power system graph structure is constructed with power equipment as nodes and the electrical connection relationship between electronic equipment as edges. The fuzzy state vector of each node is initialized according to the characteristic value corresponding to each electronic equipment in the multi-source normalized feature vector set. The fuzzy state vector is composed of the fuzzy membership of the electronic equipment in normal state, overheating fault state, partial discharge fault state and mechanical vibration abnormal state.

[0075] This step uses electrical equipment as nodes in the graph structure and the electrical connection relationships between electronic devices as edges to perform graph modeling, comprehensively reflecting the structural topology and physical dependency relationships between devices. On this basis, the state values ​​of each device in the multi-source feature vector are combined to initialize the fuzzy state vector of the node, effectively representing the fuzzy membership information of the device under different operating states. This state expression method not only retains the physical structural characteristics of the electrical system, but also introduces a state distribution representation of fuzzy semantics, improving the model's expressiveness and robustness in the face of various non-deterministic inputs, equipment aging, or signal fluctuations. Through the fusion of structural modeling and fuzzy states, the solution has stronger diagnostic adaptability and interpretability, which helps to achieve accurate extraction and propagation analysis of fault modes.

[0076] In this embodiment, it is characterized in that the S3 specifically includes:

[0077] S31. Constructing a graph neural network based on device topology relationships, wherein the graph neural network uses the fuzzy state vector of each node in the power system diagram as an initial input feature;

[0078] S32. Introducing a structure-dependent attention mechanism into the graph neural network, dynamically calculating the attention weight coefficient based on the topological connection relationship and feature similarity between each node and its adjacent nodes;

[0079] S33, based on the attention weight coefficient, performing weighted aggregation on the fuzzy state vectors of the adjacent nodes of each node to obtain an intermediate feature representation of structure perception;

[0080] S34, introduce the node's own historical embedded state vector and fuse it with the current intermediate feature representation to enhance the node representation's ability to express time-evolving features;

[0081] S35, using a nonlinear activation function to process the fused intermediate feature representation to obtain the node embedding state vector of the current propagation layer;

[0082] S36. Repeat steps S32 to S35 to perform multiple rounds of graph convolution propagation until the set number of propagation layers is reached, and the output node embedding state vector set is output.

[0083] This step introduces a structurally dependent attention mechanism into the graph neural network, and uses the node's historical state to participate in the embedded feature construction, thus achieving dynamic, multi-level feature extraction of the node's fuzzy state. Under the weighting of the attention mechanism, the degree of influence of different adjacent nodes on the target node can be adaptively adjusted according to their connection strength and state similarity, avoiding the information loss caused by feature averaging. Through multiple rounds of graph convolution propagation and nonlinear activation, the final embedded state vector of the node not only contains the current state information, but also integrates the neighborhood context and time evolution trend, which is conducive to the subsequent context perception and rule matching accuracy of fuzzy reasoning. The overall method takes into account both feature depth expression and structural dependency modeling, enhancing the modeling ability of the state evolution path of equipment in complex power networks.

[0084] In this embodiment, the S4 specifically includes:

[0085] S41, obtaining the current fuzzy state vector of the starting node and the target node connected to each side of the power system diagram;

[0086] S42, based on the change of the fuzzy state vector of the edge-connected node, calculate the relative change rate of the fuzzy state of the edge at the current moment as a state coupling change indicator;

[0087] S43. Based on the side synaptic plasticity adjustment function, dynamically generate a side synaptic plasticity weight value according to the change amplitude and direction of the state coupling change indicator:

[0088] w ij (t) = η·tanh(γ·||[μ i (t)-μ i (t-1)]⊙[μ j (t)-μ j (t-1)]||2);

[0089] Among them, w ij (t) represents the edge synaptic plasticity weight value between the i-th node and the j-th node, η represents the global learning rate factor, tanh represents the hyperbolic tangent function, γ represents the gain coefficient of the adjustment sensitivity, ⊙ represents the Hadamard product, μ i (t) represents the fuzzy state vector of the i-th node at the current moment, μ i (t-1) represents the fuzzy state vector of the i-th node at the previous moment, μ j (t) represents the fuzzy state vector of the jth node at the current moment, μ j (t-1) represents the fuzzy state vector of the jth node at the previous moment, ||·||2 represents the L2 norm;

[0090] S44: Normalize the edge synapse plasticity weight values, fill the normalized edge synapse plasticity weight values ​​into the adjacency matrix of the corresponding power system graph, and generate and output an edge synapse update weight matrix.

[0091] This step dynamically generates an edge synaptic update weight matrix by calculating the changes in the synaptic plasticity weights of the connecting edges and combining it with the trend of changes in the states of the connected nodes, thereby achieving time-varying modeling of the edge states in the power system diagram. The introduced synaptic plasticity regulation function refers to a neurophysiological mechanism and can adaptively adjust the influence weights of the connecting edges according to the direction and amplitude of the changes in the fuzzy states of the nodes, reflecting the enhanced or weakened coupling effects between devices due to state changes. This mechanism effectively characterizes the non-static state conduction relationship between devices, enabling the system to accurately reflect the potential fault diffusion paths between devices. By constructing and updating the adjacency matrix through normalization, it provides both structural and semantic support for subsequent fuzzy rule modeling, thereby improving the accuracy and flexibility of diagnostic modeling.

[0092] In this embodiment, the S5 specifically includes:

[0093] S51. Based on the node fuzzy state vector set, extract the current fuzzy state information from each node, and locate the directly connected adjacent nodes in combination with the topological connection relationship in the power system diagram;

[0094] S52. Perform weighted processing on the topological connection relationship according to the edge synaptic plasticity weight value recorded in the edge synapse update weight matrix, and construct a triple set including edge connection direction, connection strength, and associated node state change;

[0095] S53. Utilize a fuzzy logic rule construction mechanism to perform premise-result modeling on each triplet, and generate a candidate fuzzy rule set based on the following pattern: when the fuzzy state of node A is in the first state category, the edge synaptic plasticity weight value between node A and node B is a set strength, and the fuzzy state of node B shows a trend of changing to the second state category, it is inferred that the state of node B will evolve to the third state category. The first state category corresponds to the normal state of the power equipment, the second state category corresponds to the overheating fault state, and the third state category corresponds to the partial discharge state or the abnormal mechanical vibration state. The edge synaptic plasticity weight value is used to reflect the influence of state conduction between nodes and participates in the construction and correction of fuzzy causal relationships.

[0096] S54. Counting the proportion of samples matched by each candidate fuzzy rule in the historical state evolution record as the coverage of the candidate fuzzy rule;

[0097] S55. Calculate a state consistency score by performing a similarity analysis on the fuzzy membership between the premise state and the result state involved in the candidate fuzzy rules. The state consistency score reflects the confidence level of the fuzzy logic match. Simultaneously, perform graph structure consistency verification on the edges and node paths involved in each rule to exclude false rules that violate the actual electrical connection logic of the power system.

[0098] S56. Candidate rules with a coverage higher than a set range threshold, a state consistency score higher than a set score threshold, and satisfying causal reachability in the graph structure are screened as valid fuzzy rules, and a fuzzy rule base is constructed.

[0099] This step constructs a set of candidate fuzzy rules based on node states and edge synaptic weights, and introduces a triple screening mechanism of coverage, state consistency, and causal reachability to automatically extract high-quality fuzzy rules, significantly improving the expressiveness and application effectiveness of the rule base. By extracting rule fragments from historical state evolution data and calculating the proportion of matching samples, the rules are ensured to have broad adaptability; fuzzy membership similarity is used to quantify state consistency and eliminate rules with large semantic deviations; the logical rationality of the rule causal chain is verified in combination with the topological constraints of the graph structure to avoid misjudgments due to non-physical paths. This solution effectively solves the problem of rule design relying on expert experience and poor versatility in traditional fuzzy systems, and significantly improves the system's automation level and reasoning reliability in complex power fault diagnosis.

[0100] In this embodiment, S6 specifically includes:

[0101] S61, using the fuzzy rule base and the node embedded state vector set as rule input and state input of multi-level fuzzy reasoning respectively;

[0102] S62, fuzzifying the node embedded state vector set, converting the current state representation of each node into membership values ​​corresponding to multiple state categories, for adapting the precondition structure of the fuzzy rule;

[0103] S63. In the first-level fuzzy reasoning stage, the activation strength is calculated based on the matching degree between the premise of each fuzzy rule and the fuzzified node state, forming a primary fuzzy reasoning result for identifying a set of fault types and outputting a confidence score for each type.

[0104] S64. In the second-level fuzzy reasoning stage, the spatial distribution positions and topological dependency paths of the nodes involved in the activation rules are further combined to calculate the state impact transmission direction, determine the node location where the fault occurs, and output the fault probability of each node;

[0105] S65. In the third-level fuzzy reasoning stage, a risk weighted assessment is performed on the identified faults based on the severity of the fault type, the importance index of the fault location, and the current load status, and a fault event priority ranking result is generated;

[0106] S66. Structuring the output results of the multi-level fuzzy reasoning process to generate fault diagnosis results, including fault type determination results, fault location determination results, and fault event priority ranking results.

[0107] This step designs a multi-level fuzzy reasoning process, which realizes full-link fault reasoning from state identification, location positioning to risk ranking through layer-by-layer rule activation and causal aggregation. The first layer activates the preliminary fault type by matching fuzzy rules with node states. The second layer infers the fault location based on the node topology path in the graph. The third layer calculates the comprehensive risk level of the fault event based on the fault type, location and load status. The multi-level reasoning method not only decomposes complex reasoning tasks, but also enhances the model's adaptability to complex working conditions and output interpretation capabilities. This method outputs structured diagnostic results, which can be used in automatic alarm systems and provide a reliable decision-making basis for manual intervention, improving the integrity, accuracy and actual engineering deployability of the diagnosis.

[0108] In this embodiment, the S7 specifically includes:

[0109] S71. Match the fault diagnosis result with each node in the power system graph structure, generate structured state update data for the nodes with matching results, and bind a timestamp to record the current time;

[0110] S72, writing the state update data into the state memory unit of the node, wherein the state memory unit is organized in a time series form and records the fuzzy state vector, fault diagnosis result label and confidence output of the node at multiple consecutive time points;

[0111] S73. Extracting structural information of the edge connections between the node and other nodes based on the state change content of the newly written node in the state memory unit, and querying the corresponding edge synaptic plasticity weight value in the edge synapse update weight matrix;

[0112] S74. For each edge connected to a node whose state has changed, calculate the incremental change value of the historical state of the edge synapse based on the magnitude and direction of the state change before and after the connected node, combined with the edge synapse plasticity weight value, and form a difference vector.

[0113] S75. Write the differential vector into the edge state evolution recording unit, organize it by edge identifier index, and record the change sequence of the edge synaptic plasticity weight value at different time steps, including the change trend, historical state strength, and update trigger source node number;

[0114] S76. The updated node state memory unit and the edge state evolution recording unit are stored in association in the graph structure in the form of key-value pairs.

[0115] This step establishes a graph-oriented temporal memory mechanism by writing the fault diagnosis results into the node state memory unit and synchronously updating the historical state of the edge synapses. On the node side, the state memory unit records the embedding vector, fuzzy state, and diagnosis results at each moment, making the historical state evolution of the device traceable. On the edge side, by recording the change sequence of synaptic plasticity weights, a dynamic evolution trajectory of the edge state is formed, capturing the evolution of the coupling strength between state changes. This mechanism not only supports subsequent model correction and retrospective reasoning, but also provides the original basis for retraining, model updates, and abnormal behavior tracking, significantly enhancing the system's adaptability and memory retention capabilities during long-term operation.

[0116] Example 1:

[0117] To verify the feasibility of the present invention, it was applied to the intelligent operation and maintenance system for the main transformer of a 220 kV substation. This system covers key equipment such as the main transformer, GIS switches, busbars, and lightning arresters. The system operates on a large scale and has a complex structure, making it prone to complex faults under high load and strong interference scenarios. Traditional diagnostic methods have difficulty meeting the requirements of real-time performance and accuracy. Existing manual inspection systems or systems based on static expert rules have significant deficiencies in multi-source data fusion, state evolution modeling, and causal path reasoning. These systems are prone to misjudgments, missed judgments, and delayed responses, seriously affecting the safety of equipment operation.

[0118] In this embodiment, multiple types of sensors are deployed to collect multi-source state data, including voltage, current, temperature rise, vibration, partial discharge, and acoustic signals. The acquisition cycle is controlled within seconds to ensure the full acquisition of early fault characteristics. The system uniformly aligns and normalizes the original multi-source data in time, constructs a multi-source feature vector set, and serves as the basis for modeling the power system graph structure. In the graph structure, nodes represent equipment units, and edges represent their electrical connection relationships. The system initializes the node fuzzy state vector through feature-driven initialization, and combines the graph neural network to perform structure-aware feature extraction to generate a node embedded state vector set.

[0119] Simultaneously, the system dynamically calculates the synaptic plasticity weights of connected edges and constructs an edge synaptic update weight matrix based on node state changes and edge connection dependencies, accurately expressing the propagation channels and impact of fault conditions. State triples are constructed by combining node states with the edge synaptic matrix. The system automatically generates candidate fuzzy rules and screens them based on coverage, state consistency, and causal path consistency in historical data. This creates a fuzzy rule base tailored to the current operating state, completing the self-learning process of the rule base.

[0120] During the fault diagnosis phase, the system performs multi-level fuzzy reasoning: the first level identifies the fault type, the second level infers the fault's impact location based on the graph structure, and the third level outputs a priority ranking based on device importance and current system load. The reasoning results are presented as structured alarm information to assist on-duty personnel in making response decisions. The system also incorporates a state memory mechanism that archives each reasoning result and node status, while also recording the historical evolution of edge synapses, providing data support for subsequent confidence backtracking and re-determination.

[0121] During the operational cycle, the system identified 35 fault-related events, including 26 correct diagnoses, 3 false alarms, and 1 missed alarm. The system's average response time was kept under 30 seconds, achieving a diagnostic accuracy rate exceeding 92%. Compared to the existing static rule-based system, the false alarm rate was reduced by approximately two-thirds, the missed alarm rate was halved, and the average response time was shortened by over 70%. The system successfully predicted the fault path in a partial discharge anomaly event, prioritizing it as a Level 1 warning. This enabled operations and maintenance personnel to intervene quickly, preventing the fault from escalating, demonstrating the system's adaptive diagnostic capabilities driven by multi-source data.

[0122] Through this embodiment, the present invention outperforms traditional technical solutions in key performance areas such as fault diagnosis accuracy, inference response time, and system stability. It is particularly suitable for intelligent power operation and maintenance scenarios involving complex connections between large-scale devices, multi-source data fusion, and uncertain fault evolution paths. Its deep integration of fuzzy reasoning mechanisms and graph structure learning provides a new solution for intelligent, structured, and high-confidence identification of power faults.

[0123] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of power equipment based on adaptive fuzzy reasoning, characterized in that: The steps include: S1. Collect multi-source data from devices in the power system, perform time alignment and normalization processing, and construct a multi-source normalized feature vector set; S2. constructing a power system graph structure based on the multi-source normalized feature vector set and initializing the fuzzy state vector of each node; S3. Input the power system graph structure and the fuzzy state vector into a graph neural network, perform a structure-dependent attention graph convolution operation, and output a set of node embedding state vectors; S4. Calculate the synaptic plasticity weight change of each edge in the power system diagram and generate an edge synaptic update weight matrix based on the fuzzy state change of the edge-connected nodes; S5. Based on the node fuzzy state vector set and the edge synaptic update weight matrix, a candidate fuzzy rule set is constructed, and effective fuzzy rules are screened according to the coverage, state consistency and causal relationship in the graph of the candidate fuzzy rule set to obtain a fuzzy rule base; S6. Based on the fuzzy rule base and the node embedded state vector set, a multi-level fuzzy reasoning process is executed to output the fault diagnosis judgment result, including the fault type judgment result, the fault location judgment result and the priority ranking result; S7. Writing the fault diagnosis result into the state memory unit of the corresponding node, and updating the edge synapse history state in the power system graph structure according to the edge synapse update weight matrix to form a node state memory record and an edge state evolution record; S8. When the confidence level of the fault diagnosis result is lower than the set threshold, call the node state memory record and the edge state evolution record, perform time series sliding window backtracking processing, update the node embedded state vector set, and repeat step S6 to perform re-determination.

2. The method for diagnosing faults of power equipment based on adaptive fuzzy reasoning according to claim 1, characterized in that: The multi-source data includes voltage data, current data, temperature rise data, vibration data, partial discharge data and acoustic signal data.

3. The method for diagnosing faults of power equipment based on adaptive fuzzy reasoning according to claim 1, characterized in that: The power system graph structure is constructed with power equipment as nodes and the electrical connection relationships between electronic devices as edges. The fuzzy state vector of each node is initialized according to the characteristic values ​​corresponding to each electronic device in the multi-source normalized feature vector set. The fuzzy state vector is composed of the fuzzy membership of the electronic device in a normal state, an overheating fault state, a partial discharge fault state, and an abnormal mechanical vibration state.

4. The method for diagnosing faults of power equipment based on adaptive fuzzy reasoning according to claim 1, characterized in that: The S3 specifically includes: S31. Constructing a graph neural network based on device topology relationships, wherein the graph neural network uses the fuzzy state vector of each node in the power system diagram as an initial input feature; S32. Introducing a structure-dependent attention mechanism into the graph neural network, dynamically calculating the attention weight coefficient based on the topological connection relationship and feature similarity between each node and its adjacent nodes; S33, based on the attention weight coefficient, performing weighted aggregation on the fuzzy state vectors of the adjacent nodes of each node to obtain an intermediate feature representation of structure perception; S34, introduce the node's own historical embedded state vector and fuse it with the current intermediate feature representation to enhance the node representation's ability to express time-evolving features; S35, using a nonlinear activation function to process the fused intermediate feature representation to obtain the node embedding state vector of the current propagation layer; S36. Repeat steps S32 to S35 to perform multiple rounds of graph convolution propagation until the set number of propagation layers is reached, and the output node embedding state vector set is output.

5. The method for diagnosing faults of power equipment based on adaptive fuzzy reasoning according to claim 1, characterized in that: The S4 specifically includes: S41, obtaining the current fuzzy state vector of the starting node and the target node connected to each side of the power system diagram; S42, based on the change of the fuzzy state vector of the edge-connected node, calculate the relative change rate of the fuzzy state of the edge at the current moment as a state coupling change indicator; S43. Based on the side synaptic plasticity adjustment function, dynamically generate a side synaptic plasticity weight value according to the change amplitude and direction of the state coupling change indicator: w ij (t)=η·tanh(γ·||[μ i (t)-m i (t-1)]⊙[μ j (t)-m j (t-1)]||2); Among them, w ij (t) represents the edge synaptic plasticity weight value between the i-th node and the j-th node, η represents the global learning rate factor, tanh represents the hyperbolic tangent function, γ represents the gain coefficient of the adjustment sensitivity, ⊙ represents the Hadamard product, μ i (t) represents the fuzzy state vector of the i-th node at the current moment, μ i (t-1) represents the fuzzy state vector of the i-th node at the previous moment, μ j (t) represents the fuzzy state vector of the jth node at the current moment, μ j (t-1) represents the fuzzy state vector of the jth node at the previous moment, ||·||2 represents the L2 norm; S44: Normalize the edge synapse plasticity weight values, fill the normalized edge synapse plasticity weight values ​​into the adjacency matrix of the corresponding power system graph, and generate and output an edge synapse update weight matrix.

6. The method for diagnosing faults of power equipment based on adaptive fuzzy reasoning according to claim 1, characterized in that: The S5 specifically includes: S51. Based on the node fuzzy state vector set, extract the current fuzzy state information from each node, and locate the directly connected adjacent nodes in combination with the topological connection relationship in the power system diagram; S52. Perform weighted processing on the topological connection relationship according to the edge synaptic plasticity weight value recorded in the edge synapse update weight matrix, and construct a triple set including edge connection direction, connection strength, and associated node state change; S53. Utilize a fuzzy logic rule construction mechanism to perform premise-result modeling on each triplet, and generate a candidate fuzzy rule set based on the following pattern: when the fuzzy state of node A is in the first state category, the edge synaptic plasticity weight value between node A and node B is a set strength, and the fuzzy state of node B shows a trend of changing to the second state category, it is inferred that the state of node B will evolve to the third state category. The first state category corresponds to the normal state of the power equipment, the second state category corresponds to the overheating fault state, and the third state category corresponds to the partial discharge state or the abnormal mechanical vibration state. The edge synaptic plasticity weight value is used to reflect the influence of state conduction between nodes and participates in the construction and correction of fuzzy causal relationships. S54. Counting the proportion of samples matched by each candidate fuzzy rule in the historical state evolution record as the coverage of the candidate fuzzy rule; S55. Calculate a state consistency score by performing a similarity analysis on the fuzzy membership between the premise state and the result state involved in the candidate fuzzy rules. The state consistency score reflects the confidence level of the fuzzy logic match. Simultaneously, perform graph structure consistency verification on the edges and node paths involved in each rule to exclude false rules that violate the actual electrical connection logic of the power system. S56. Candidate rules with a coverage higher than a set range threshold, a state consistency score higher than a set score threshold, and satisfying causal reachability in the graph structure are screened as valid fuzzy rules, and a fuzzy rule base is constructed.

7. The method for diagnosing faults of power equipment based on adaptive fuzzy reasoning according to claim 1, characterized in that: The S6 specifically includes: S61, using the fuzzy rule base and the node embedded state vector set as rule input and state input of multi-level fuzzy reasoning respectively; S62, fuzzifying the node embedded state vector set, converting the current state representation of each node into membership values ​​corresponding to multiple state categories, for adapting the precondition structure of the fuzzy rule; S63. In the first-level fuzzy reasoning stage, the activation strength is calculated based on the matching degree between the premise of each fuzzy rule and the fuzzified node state, forming a primary fuzzy reasoning result for identifying a set of fault types and outputting a confidence score for each type. S64. In the second-level fuzzy reasoning stage, the spatial distribution positions and topological dependency paths of the nodes involved in the activation rules are further combined to calculate the state impact transmission direction, determine the node location where the fault occurs, and output the fault probability of each node; S65. In the third-level fuzzy reasoning stage, a risk weighted assessment is performed on the identified faults based on the severity of the fault type, the importance index of the fault location, and the current load status, and a fault event priority ranking result is generated; S66. Structuring the output results of the multi-level fuzzy reasoning process to generate fault diagnosis results, including fault type determination results, fault location determination results, and fault event priority ranking results.

8. The method for diagnosing faults of power equipment based on adaptive fuzzy reasoning according to claim 1, characterized in that: The S7 specifically includes: S71. Match the fault diagnosis result with each node in the power system graph structure, generate structured state update data for the nodes with matching results, and bind a timestamp to record the current time; S72, writing the state update data into the state memory unit of the node, wherein the state memory unit is organized in a time series form and records the fuzzy state vector, fault diagnosis result label and confidence output of the node at multiple consecutive time points; S73. Extracting structural information of the edge connections between the node and other nodes based on the state change content of the newly written node in the state memory unit, and querying the corresponding edge synaptic plasticity weight value in the edge synapse update weight matrix; S74. For each edge connected to a node whose state has changed, calculate the incremental change value of the historical state of the edge synapse based on the magnitude and direction of the state change before and after the connected node, combined with the edge synapse plasticity weight value, and form a difference vector. S75. Write the differential vector into the edge state evolution recording unit, organize it by edge identifier index, and record the change sequence of the edge synaptic plasticity weight value at different time steps, including the change trend, historical state strength, and update trigger source node number; S76. The updated node state memory unit and the edge state evolution recording unit are stored in association in the graph structure in the form of key-value pairs.

Images