A Deep Learning-Based Fault Diagnosis Method for Power Transmission and Distribution Equipment

Abstract

This invention discloses a deep learning-based fault diagnosis method for power transmission and distribution equipment, comprising: collecting and preprocessing multi-source operating data to generate a standardized equipment status dataset; constructing a heterogeneous fault graph of the equipment, associating topology, protected sections, and fault events; constructing an improved meta-path aggregation graph neural network to generate fault embedding and propagation boundary matrices; calculating fault category probabilities, equipment risk scores, and diagnostic confidence; using a multi-objective porcupine optimization algorithm to perform time difference defense switching and fault source optimization; jointly correcting diagnostic data to generate candidate fault sources and propagation chains; performing maintenance priority ranking, generating diagnostic results, and providing feedback updates. This invention, by introducing an improved meta-path aggregation graph neural network and a multi-objective porcupine optimization algorithm, achieves multi-source operating data fusion, fault propagation boundary discrimination, fault source tracking, and maintenance priority generation for power transmission and distribution equipment.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power transmission and distribution equipment, and in particular to a fault diagnosis method for power transmission and distribution equipment based on deep learning. Background Technology

[0002] With the increasing intelligence, digitalization, and automation of power systems, monitoring the operational status and fault diagnosis of transmission and distribution equipment has become a crucial aspect of power grid safety management. Transmission and distribution equipment includes transformers, circuit breakers, switchgear, cables, distribution lines, ring main units, protection devices, and monitoring terminals. During operation, this equipment generates electrical operation data, equipment status monitoring data, protection action data, alarm data, historical fault data, and maintenance records. Existing fault diagnosis methods for transmission and distribution equipment typically collect data on voltage, current, temperature, partial discharge, infrared thermography, protection actions, and alarm records. They then employ threshold judgment, expert rules, statistical analysis, machine learning classification models, or deep learning models to identify equipment status and output abnormal equipment states, fault types, or risk levels. Compared to manual inspections and single alarm judgment methods, diagnostic methods based on data analysis and deep learning can process large amounts of operational data, improving the efficiency of equipment anomaly detection and the accuracy of fault identification.

[0003] Existing fault diagnosis methods for power transmission and distribution equipment still have shortcomings. The operational data of power transmission and distribution equipment comes from multiple sources, is complex in type, and has inconsistent time granularity. There is a lack of a unified modeling method among electrical quantities, status monitoring quantities, protection actions, alarm events, historical faults, and maintenance records. Existing methods often rely on data splicing or single-item judgments, making it difficult to express the correlation between equipment status, protected sections, alarm events, and maintenance results. Power transmission and distribution equipment has topological connections, upstream and downstream power supply relationships, and protected section attribution relationships. After a fault occurs, it may propagate along the line section or be interrupted by circuit breakers, sectionalizing switches, protection actions, or switch open / closed states. Existing deep learning diagnostic models often focus on single-device feature extraction or fixed graph structure aggregation, making it difficult to dynamically generate diagnostic paths that fit the fault mechanism and to determine the fault propagation direction by combining alarm time differences. Existing optimization algorithms are mostly used for ordinary parameter adjustments, making it difficult to collaboratively identify fault source equipment, affected equipment, and fault propagation chains. This leads to unclear fault source location and inaccurate assessment of the affected area when multiple devices alarm simultaneously, resulting in a deviation between maintenance priorities and actual operation and maintenance needs.

[0004] Therefore, how to provide a deep learning-based fault diagnosis method for power transmission and distribution equipment is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based fault diagnosis method for power transmission and distribution equipment. This invention utilizes multi-source operational data preprocessing of power transmission and distribution equipment, construction of heterogeneous fault graphs, an improved meta-path aggregation graph neural network, and a multi-objective crowned porcupine optimization algorithm. It details the implementation process of unified modeling, fault feature extraction, propagation boundary discrimination, and fault source tracing optimization for the operating status, protection actions, alarm events, historical faults, and maintenance records of power transmission and distribution equipment. Through dynamic meta-path generation of fault mechanisms, alarm time difference direction aggregation, and protection zone propagation boundary discrimination, it generates equipment fault embedding representations, alarm time difference direction embeddings, and fault propagation boundary matrices. Combined with alarm time difference defense switching and fault source tracing collaborative optimization, it generates an optimized diagnostic parameter set, enabling fault category identification, risk level judgment, fault propagation chain analysis, and maintenance priority ranking for power transmission and distribution equipment. Compared to traditional single-equipment diagnosis or fixed graph structure diagnosis methods, this invention has advantages such as strong multi-source heterogeneous data fusion capability, accurate fault mechanism expression, clear alarm timing analysis, clear fault propagation boundary identification, and strong support for maintenance decision-making.

[0006] A method for fault diagnosis of power transmission and distribution equipment based on deep learning according to an embodiment of the present invention includes:

[0007] Collect multi-source operation data of power transmission and distribution equipment, and perform preprocessing on the multi-source operation data to generate a standardized equipment status dataset;

[0008] Based on a standardized equipment status dataset, a heterogeneous fault diagram of equipment is constructed according to the topological connection relationship of power transmission and distribution equipment, the ownership relationship of protected sections, and the correlation relationship of fault events.

[0009] An improved meta-path aggregation graph neural network is constructed to perform dynamic meta-path generation of fault mechanisms, alarm time difference direction aggregation, and protection zone propagation boundary discrimination on heterogeneous fault graphs of equipment, generating equipment fault embedding representation, alarm time difference direction embedding, and fault propagation boundary matrix;

[0010] Based on the equipment fault embedding representation and alarm time difference direction embedding, the fault category probability, equipment risk score and diagnostic confidence are calculated to generate equipment diagnostic evaluation data;

[0011] The equipment diagnostic evaluation data, alarm time difference direction embedding, and fault propagation boundary matrix are used as candidate individual evaluation data for the multi-objective crown porcupine optimization algorithm. Alarm time difference defense switching and fault source tracing collaborative optimization are performed to generate an optimized diagnostic parameter set.

[0012] Based on the optimized set of diagnostic parameters, joint correction is performed on the fault category probability, equipment risk score and fault propagation boundary matrix to screen out blocking propagation relationships and generate candidate fault source equipment, candidate affected equipment and candidate fault propagation chain;

[0013] Based on candidate fault source equipment, candidate affected equipment, and candidate fault propagation chain, the power transmission and distribution equipment is prioritized for maintenance, generating fault diagnosis results for the power transmission and distribution equipment. Maintenance feedback data is received to update the standardized equipment status dataset and the heterogeneous fault diagram of the equipment.

[0014] Optionally, generating the standardized device status dataset includes:

[0015] Collect multi-source operating data of power transmission and distribution equipment, including electrical operating data, equipment status monitoring data, protection action data, alarm data, historical fault data, and maintenance record data;

[0016] Multi-source operational data are correlated according to equipment number and collection time to generate equipment operational data sequences;

[0017] The equipment operation data sequence is processed to remove outliers, fill in missing values, and delete duplicate records, generating a cleaned equipment operation data sequence.

[0018] The cleaned equipment operation data sequence is processed by timestamp alignment, unit unification and numerical standardization to generate a standardized equipment operation data sequence.

[0019] Standardized equipment operation data sequences are matched with corresponding fault labels, alarm labels, and maintenance labels to generate standardized equipment status datasets.

[0020] Optionally, the construction of the heterogeneous fault diagram of the device includes:

[0021] Read the device number, device type, collection time and status characteristics from the standardized device status dataset, establish a set of device nodes, and write the status characteristics into the node attributes of the corresponding device nodes;

[0022] Based on the topological connection relationship of power transmission and distribution equipment, establish a set of nodes for line sections, and establish topological connection edges between equipment nodes and line section nodes according to the power supply direction;

[0023] Based on the ownership relationship of the protected area sections, a set of protection device nodes is established, and protection ownership edges are established between the protection device nodes and the corresponding coverage equipment nodes and the corresponding line section nodes.

[0024] Based on the correlation of fault events, a set of fault event nodes is established, which includes alarm event nodes, historical fault nodes, and maintenance record nodes. Event association edges are established between various types of fault event nodes and corresponding equipment nodes.

[0025] The set of equipment nodes, the set of line segment nodes, the set of protection device nodes, the set of fault event nodes, and various edge relationships are combined according to the node type identifier, edge type identifier, and time identifier to generate a heterogeneous fault diagram of equipment.

[0026] Optionally, the generation of the device fault embedding representation, alarm time difference direction embedding, and fault propagation boundary matrix includes:

[0027] An improved meta-path aggregation graph neural network is constructed, which includes a fault mechanism dynamic meta-path generation layer, an alarm time difference direction aggregation layer, and a protected area propagation boundary discrimination layer.

[0028] Read the set of device nodes, line segment nodes, protection device nodes and fault event nodes in the heterogeneous fault diagram of the equipment, perform embedding mapping and dimension alignment on the node attributes of different types of nodes, and generate the initial embedding of heterogeneous nodes.

[0029] The fault mechanism dynamic meta-path generation layer reads the initial embedding of heterogeneous nodes, equipment type, fault type, protection zone affiliation, alarm event type, and maintenance record type, and generates a set of fault mechanism meta-paths according to thermal defects, insulation degradation, protection linkage, grounding short circuit propagation, and environmental common causes.

[0030] The alarm time difference direction aggregation layer reads the fault mechanism meta-path set and the initial embedding of heterogeneous nodes, calculates the time difference between the equipment status abnormal time, protection action time, alarm trigger time and maintenance confirmation time, and performs direction aggregation on the fault mechanism meta-path instance in combination with the power supply direction to generate the equipment fault embedding representation and alarm time difference direction embedding.

[0031] The protection zone propagation boundary discrimination layer reads the equipment fault embedding representation, alarm time difference direction embedding, protection zone affiliation, protection action status and switch on / off status, and determines the propagable status, restricted propagation status and blocked propagation status between equipment nodes, and generates a fault propagation boundary matrix, with rows corresponding to risk source equipment, columns corresponding to affected equipment, and elements corresponding to the propagation boundary status between equipment nodes.

[0032] The improved metapath aggregation graph neural network was trained by combining the fault classification error corresponding to the equipment fault embedding representation, the timing direction error corresponding to the alarm time difference direction embedding, and the boundary discrimination error corresponding to the fault propagation boundary matrix as the joint optimization objective. The network parameters of the fault mechanism dynamic metapath generation layer, the alarm time difference direction aggregation layer, and the protection zone propagation boundary discrimination layer were continuously updated. When the change of the joint loss value in five consecutive training rounds was less than 0.001, the improved metapath aggregation graph neural network was determined to have completed convergence training.

[0033] Optionally, the generated device diagnostic evaluation data includes:

[0034] The device fault embedding representation and alarm time difference direction embedding are associated according to the device number to generate joint diagnostic features of the device;

[0035] Based on the joint diagnostic features of the equipment, the category mapping value corresponding to each fault category is calculated, and the category mapping value is processed by normalization index to generate the fault category probability corresponding to each power transmission and distribution equipment.

[0036] The equipment risk score is calculated based on the abnormal category probability in the fault category probability, the abnormal offset in the equipment joint diagnostic features, and the directional consistency parameter in the alarm time difference direction embedding.

[0037] The diagnostic confidence level is calculated based on the maximum category probability, the dispersion of the fault category probability distribution, and the directional consistency parameter in the fault category probability.

[0038] The fault category probability, equipment risk score, and diagnostic confidence are summarized according to the equipment number to generate equipment diagnostic evaluation data.

[0039] Optionally, generating the optimized diagnostic parameter set includes:

[0040] Read the equipment diagnostic evaluation data, alarm time difference direction embedding and fault propagation boundary matrix, establish a set of diagnostic parameters to be optimized. The set of diagnostic parameters to be optimized includes fault category judgment threshold, equipment risk score correction coefficient, propagation boundary state correction coefficient, fault source equipment screening threshold, affected equipment screening threshold and maintenance priority ranking coefficient, and encode the set of diagnostic parameters to be optimized as Crowned porcupine candidate individuals.

[0041] Based on the fault category probability, equipment risk score and diagnostic confidence in the equipment diagnostic evaluation data, the fault identification evaluation value, fault missed evaluation value, normal false alarm evaluation value and diagnostic confidence evaluation value of the candidate individuals of the Crown Porcupine were calculated.

[0042] Based on the fault propagation boundary matrix, the propagation boundary consistency evaluation value of candidate individuals of Crowned Porcupine is calculated;

[0043] The alarm time difference interval is divided according to the direction of alarm time difference embedding. The alarm time difference interval includes the time difference from abnormal status to alarm trigger, the time difference from protection action to alarm trigger, and the time difference from alarm trigger to maintenance confirmation. The candidate individual update method corresponding to visual defense, sound defense, odor defense and physical attack is switched according to the alarm time difference interval to update the crown porcupine candidate individuals.

[0044] In the collaborative optimization process of fault source tracing, device pairs are generated based on the fault propagation boundary matrix. Device pairs with propagation boundary states that are blocked are screened out. Based on the device risk score, diagnostic confidence and propagation boundary state of the retained device pairs, the fault source candidate value, the affected candidate value and the propagation direction candidate value are calculated to generate the fault source tracing evaluation value.

[0045] The evaluation values ​​of fault identification, fault omission, normal false alarm, diagnosis confidence, propagation boundary consistency, and fault source tracing are combined into a multi-objective evaluation vector. Based on the multi-objective evaluation vector, non-dominated sorting and crowding distance sorting are performed on the candidate individuals of Crown Porcupine to generate a Pareto candidate diagnostic parameter set.

[0046] The candidate diagnostic parameters with the first level of non-dominated status in the Pareto candidate diagnostic parameter set are sorted from largest to smallest according to the crowding distance, and the sorted candidate diagnostic parameters are written into the external archive to generate the optimized diagnostic parameter set.

[0047] Optionally, the generation of candidate fault source devices, candidate affected devices, and candidate fault propagation chains includes:

[0048] Read the optimized diagnostic parameter set, and perform joint correction on the fault category probability, equipment risk score and fault propagation boundary matrix based on the optimized diagnostic parameter set to generate the corrected fault category probability, corrected equipment risk score and corrected fault propagation boundary matrix;

[0049] Based on the corrected fault propagation boundary matrix, the device node connection relationships corresponding to the blocked propagation state are filtered out, while the device node connection relationships corresponding to the propagable state and the restricted propagation state are retained.

[0050] Based on the calibration equipment risk score, calibration failure category probability, failure source equipment screening threshold, and affected equipment screening threshold, candidate failure source equipment and candidate affected equipment are screened.

[0051] Connect the candidate fault source device and the candidate affected device according to the preserved device node connection relationship to generate a candidate fault propagation chain.

[0052] Optionally, generating fault diagnosis results for power transmission and distribution equipment includes:

[0053] Read the maintenance priority ranking coefficients from the candidate fault source equipment, candidate affected equipment, candidate fault propagation chain and optimized diagnostic parameter set, and calculate the maintenance priority score of the candidate maintenance object;

[0054] Candidate maintenance objects are sorted according to their maintenance priority scores to generate a maintenance priority sequence;

[0055] Based on the maintenance priority sequence, candidate fault source equipment, candidate affected equipment, and candidate fault propagation chain, generate fault diagnosis results for power transmission and distribution equipment.

[0056] Receive maintenance feedback data, update the fault labels, alarm labels and maintenance labels in the standardized equipment status dataset based on the maintenance feedback data, and update the set of fault event nodes and event association edges in the heterogeneous fault graph of the equipment.

[0057] The beneficial effects of this invention are:

[0058] This invention proposes a deep learning-based fault diagnosis method for power transmission and distribution equipment. By constructing a standardized equipment status dataset, a heterogeneous fault graph, an improved meta-path aggregation graph neural network, and a multi-objective optimization algorithm, it achieves unified modeling and fault feature extraction of multi-source operating data for power transmission and distribution equipment. Compared to traditional fault diagnosis methods that rely on single-class monitoring data, manual rules, or fixed graph structures, this invention can uniformly associate electrical operating data, equipment status monitoring data, protection action data, alarm data, historical fault data, and maintenance record data into the heterogeneous fault graph. This reduces diagnostic errors caused by inconsistent data formats, inconsistent time granularities, and scattered correlations, improving the completeness and accuracy of fault feature representation for power transmission and distribution equipment.

[0059] This invention improves the generation of dynamic meta-paths for fault mechanisms, aggregation of alarm time difference directions, and discrimination of propagation boundaries in protected areas by performing meta-path aggregation graph neural networks. This enables the fault diagnosis process to simultaneously consider equipment topology connections, protected area affiliations, fault event correlations, and alarm timing relationships. Compared to ordinary graph neural networks that use fixed adjacency relationships or fixed meta-path aggregation, this invention can form diagnostic meta-paths based on fault mechanisms such as thermal defects, insulation degradation, protection linkage, grounding short-circuit propagation, and environmental common causes. It also determines the fault propagation direction based on the time difference between equipment abnormality, protection action, alarm triggering, and maintenance confirmation. Furthermore, it identifies propagable, restricted, and blocked propagation states based on protection action status and switch open / closed status, improving the ability to identify fault propagation boundaries and reducing the probability of affected equipment being misidentified as fault source equipment.

[0060] This invention employs a multi-objective optimization algorithm to jointly optimize equipment diagnostic evaluation data, alarm time difference direction embedding, and fault propagation boundary matrix. It utilizes alarm time difference defense switching and fault source tracing to collaboratively generate an optimized diagnostic parameter set, and jointly corrects fault category probabilities, equipment risk scores, and the fault propagation boundary matrix. Compared to diagnostic methods that only output fault categories or risk scores, this invention can filter out propagation blocking relationships, generate candidate fault source devices, candidate affected devices, and candidate fault propagation chains, and further form maintenance priority ranking results. This enables the fault diagnosis results of power transmission and distribution equipment to simultaneously possess fault identification, risk assessment, propagation chain analysis, and maintenance decision support capabilities, improving the accuracy of fault source location and the efficiency of operation and maintenance in scenarios with multiple devices simultaneously alarming. Attached Figure Description

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0062] Figure 1 This is a flowchart of a deep learning-based fault diagnosis method for power transmission and distribution equipment proposed in this invention;

[0063] Figure 2 This is a schematic diagram of the structure of the improved meta-path aggregation graph neural network of a deep learning-based fault diagnosis method for power transmission and distribution equipment proposed in this invention.

[0064] Figure 3 This is a flowchart illustrating the process of generating an optimized diagnostic parameter set using a multi-objective crowned porcupine optimization algorithm for a deep learning-based fault diagnosis method for power transmission and distribution equipment proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1 , Figure 2 and Figure 3 A deep learning-based fault diagnosis method for power transmission and distribution equipment includes:

[0067] Collect multi-source operation data of power transmission and distribution equipment, and perform preprocessing on the multi-source operation data to generate a standardized equipment status dataset;

[0068] Based on a standardized equipment status dataset, a heterogeneous fault diagram of equipment is constructed according to the topological connection relationship of power transmission and distribution equipment, the ownership relationship of protected sections, and the correlation relationship of fault events.

[0069] An improved meta-path aggregation graph neural network is constructed to perform dynamic meta-path generation of fault mechanisms, alarm time difference direction aggregation, and protection zone propagation boundary discrimination on heterogeneous fault graphs of equipment, generating equipment fault embedding representation, alarm time difference direction embedding, and fault propagation boundary matrix;

[0070] Based on the equipment fault embedding representation and alarm time difference direction embedding, the fault category probability, equipment risk score and diagnostic confidence are calculated to generate equipment diagnostic evaluation data;

[0071] The equipment diagnostic evaluation data, alarm time difference direction embedding, and fault propagation boundary matrix are used as candidate individual evaluation data for the multi-objective crown porcupine optimization algorithm. Alarm time difference defense switching and fault source tracing collaborative optimization are performed to generate an optimized diagnostic parameter set.

[0072] Based on the optimized set of diagnostic parameters, joint correction is performed on the fault category probability, equipment risk score and fault propagation boundary matrix to screen out blocking propagation relationships and generate candidate fault source equipment, candidate affected equipment and candidate fault propagation chain;

[0073] Based on candidate fault source equipment, candidate affected equipment, and candidate fault propagation chain, the power transmission and distribution equipment is prioritized for maintenance, generating fault diagnosis results for the power transmission and distribution equipment. Maintenance feedback data is received to update the standardized equipment status dataset and the heterogeneous fault diagram of the equipment.

[0074] In this embodiment, generating a standardized device status dataset includes:

[0075] Collect multi-source operating data of power transmission and distribution equipment, including electrical operating data, equipment status monitoring data, protection action data, alarm data, historical fault data, and maintenance record data;

[0076] Multi-source operational data are correlated according to equipment number and collection time to generate equipment operational data sequences;

[0077] The equipment operation data sequence is processed to remove outliers, fill in missing values, and delete duplicate records, generating a cleaned equipment operation data sequence.

[0078] The cleaned equipment operation data sequence is processed by timestamp alignment, unit unification and numerical standardization to generate a standardized equipment operation data sequence.

[0079] Standardized equipment operation data sequences are matched with corresponding fault labels, alarm labels, and maintenance labels to generate standardized equipment status datasets.

[0080] In this embodiment, the construction of the heterogeneous fault diagram of the equipment includes:

[0081] Read the device number, device type, collection time, and status features from the standardized device status dataset, establish a set of device nodes, and write the status features into the node attributes of the corresponding device nodes, where:

[0082] Write the state characteristics into the node attributes of the corresponding device node, specifically:

[0083] The status characteristics of the corresponding equipment are read from the standardized equipment status dataset according to the equipment number. The status characteristics include voltage offset value, current fluctuation value, load change value, temperature change value, partial discharge change value, infrared temperature measurement change value, protection action indicator, alarm count, historical fault count, and maintenance status indicator.

[0084] Establish a device node attribute table. The rows of the device node attribute table correspond to the device node number, and the columns correspond to each status feature field. Write the status features corresponding to the same device number and the same collection time into the node attribute field of the corresponding device node in the device node attribute table.

[0085] When the same device node has multiple sets of state features within a continuous acquisition time, a node attribute sequence is formed according to the acquisition time order, and a binding relationship is established between the node attribute sequence and the device node.

[0086] Based on the topological connections of power transmission and distribution equipment, a set of nodes for line segments is established, and topological connection edges between equipment nodes and line segment nodes are established according to the power supply direction, wherein:

[0087] Establish a set of nodes for the line segment, specifically as follows:

[0088] Read the line section number, starting equipment number, ending equipment number, line type, line length, rated capacity, feeder number, and switch status from the power transmission and distribution equipment topology connection relationship;

[0089] Each transmission line, distribution line, cable section, and connecting line with an independent section number is considered as a line section node;

[0090] Establish a line segment node attribute table. The rows of the line segment node attribute table correspond to the line segment node number, and the columns correspond to the line type, line length, rated capacity, feeder number, and switch status.

[0091] Establish a numbering mapping relationship between the line segment nodes and the corresponding starting and ending equipment nodes to generate a set of line segment nodes;

[0092] Establish topological connection edges between device nodes and line segment nodes, specifically as follows:

[0093] Read the starting device number, ending device number, and power supply direction identifier from the set of nodes in the line segment, establish an upstream topology connection edge between the device node corresponding to the starting device number and the node in the line segment, and establish a downstream topology connection edge between the node in the line segment and the device node corresponding to the ending device number.

[0094] Write the line segment number, power supply direction identifier, feeder number, switch status and connection time identifier in the upstream and downstream topology connection edges;

[0095] When the line section is in the interconnection operation state, the direction identifiers of the upstream and downstream topology connection edges are updated according to the current power supply direction to form a topology connection edge with the power supply direction.

[0096] Based on the affiliation of protected sections, a set of protection device nodes is established, and protection affiliation edges are established between the protection device nodes and their corresponding coverage equipment nodes and corresponding line section nodes, wherein:

[0097] Establish a set of protection device nodes, specifically as follows:

[0098] Retrieve the protection device number, protection device type, installed equipment number, protection section number, coverage equipment number, coverage line section number, protection action type, action setting value, and action status identifier from the protection section attribution relationship;

[0099] Each set of relay protection device, feeder protection device, differential protection device, overcurrent protection device, grounding protection device, and reclosing device is treated as a protection device node;

[0100] Establish a protection device node attribute table. The rows of the protection device node attribute table correspond to the protection device node number, and the columns correspond to the protection device type, protection section number, protection action type, action setting value, and action status identifier.

[0101] Establish a numbering mapping relationship between the protection device nodes and the installed equipment nodes, and establish a coverage mapping relationship between the protection device nodes and the coverage equipment nodes and the coverage line segment nodes within the protected section, thereby generating a set of protection device nodes;

[0102] Establish protected ownership edges, specifically as follows:

[0103] Read the protection device number, protection section number, coverage equipment number, coverage line section number, protection action type, action status identifier, and protection action time identifier from the protection device node set;

[0104] Establish a device protection ownership edge between the protection device node and the corresponding coverage equipment node, and establish a line protection ownership edge between the protection device node and the corresponding coverage line segment node;

[0105] Write the protection device number, protection section number, protection action type, action status identifier, and protection action time identifier in the equipment protection ownership edge and the line protection ownership edge;

[0106] When the protection action status flag is updated from not activated to activated, the protection action time flag is synchronously written to the corresponding equipment protection ownership edge and line protection ownership edge to form a protection ownership edge.

[0107] Based on the correlation of fault events, a set of fault event nodes is established, including alarm event nodes, historical fault nodes, and maintenance record nodes. Event association edges are then established between each type of fault event node and its corresponding equipment node.

[0108] Establish a set of fault event nodes, specifically as follows:

[0109] Read alarm number, historical fault number, maintenance record number, associated equipment number, event type, event occurrence time, fault type, and handling status from the fault event correlation.

[0110] Map alarm records to alarm event nodes, historical fault records to historical fault nodes, and maintenance records to maintenance record nodes. Establish alarm event node attributes, historical fault node attributes, and maintenance record node attributes respectively, and write event type, event occurrence time, fault type, associated equipment number, and handling status to generate a set of fault event nodes.

[0111] Establish event-related edges, specifically as follows:

[0112] Read the associated device number, event type, event occurrence time, fault type and handling status from the fault event node set, establish alarm association edges between alarm event nodes and device nodes that triggered alarms, establish fault association edges between historical fault nodes and device nodes that experienced historical faults, and establish maintenance association edges between maintenance record nodes and device nodes corresponding to maintenance objects.

[0113] Write the event type, event occurrence time, fault type, and handling status into the alarm association edge, fault association edge, and maintenance association edge to form an event association edge;

[0114] The equipment node set, line section node set, protection device node set, fault event node set, and various edge relationships are combined according to node type identifier, edge type identifier, and time identifier to generate a heterogeneous equipment fault diagram, wherein:

[0115] Generate a heterogeneous fault diagram for the equipment, specifically:

[0116] Write the set of equipment nodes, the set of line section nodes, the set of protection device nodes, and the set of fault event nodes into the heterogeneous node table, and write a node type identifier and a time identifier for each node.

[0117] Write the topology connection edges, protection ownership edges, and event association edges into the heterogeneous edge table, and write the edge type identifier, start node number, end node number, and time identifier for each edge;

[0118] Establish the index relationship between nodes and edges based on the heterogeneous node table and heterogeneous edge table, and generate a heterogeneous fault diagram of the device.

[0119] In this embodiment, the generation of the device fault embedding representation, alarm time difference direction embedding, and fault propagation boundary matrix includes:

[0120] An improved meta-path aggregation graph neural network is constructed, comprising a fault mechanism dynamic meta-path generation layer, an alarm time difference direction aggregation layer, and a protected area propagation boundary discrimination layer, wherein:

[0121] Construct an improved meta-path aggregation graph neural network, specifically as follows:

[0122] In the traditional meta-path aggregation graph neural network, the network structure includes node content transformation, internal meta-path aggregation, and inter-meta-path aggregation. Before the traditional internal meta-path aggregation, a fault mechanism dynamic meta-path generation layer is added. The fault mechanism dynamic meta-path generation layer generates a set of fault mechanism meta-paths and inputs the set of fault mechanism meta-paths into the internal meta-path aggregation. An alarm time difference direction aggregation layer is added to the traditional internal meta-path aggregation. The alarm time difference direction aggregation layer performs direction aggregation on meta-path instances to generate device fault embedding representation and alarm time difference direction embedding. After the traditional inter-meta-path aggregation, a protected area propagation boundary discrimination layer is added. The protected area propagation boundary discrimination layer generates a fault propagation boundary matrix. The fault mechanism dynamic meta-path generation layer, the alarm time difference direction aggregation layer, and the protected area propagation boundary discrimination layer are connected to the traditional meta-path aggregation graph neural network in the order of meta-path generation, meta-path aggregation, and boundary discrimination to obtain the improved meta-path aggregation graph neural network.

[0123] Read the device node set, line segment node set, protection device node set, and fault event node set from the heterogeneous fault diagram. Perform embedding mapping and dimension alignment on the node attributes of different types of nodes to generate the initial embedding of heterogeneous nodes, where:

[0124] Generate the initial embedding of heterogeneous nodes, specifically as follows:

[0125] The equipment node attributes, line section node attributes, protection device node attributes, and fault event node attributes are converted into corresponding node feature vectors. Dimension alignment is performed on the feature vectors of different types of nodes to make the feature vectors of different types of nodes have the same embedding dimension.

[0126] Write the dimension-aligned node feature vectors into the heterogeneous node embedding table according to the node number to form the initial heterogeneous node embedding.

[0127] The fault mechanism dynamic meta-path generation layer reads the initial embedding of heterogeneous nodes, equipment type, fault type, protected area affiliation, alarm event type, and maintenance record type. It then generates a set of fault mechanism meta-paths based on thermal defects, insulation degradation, protection linkage, grounding short-circuit propagation, and environmental common causes.

[0128] The fault mechanism dynamic meta-path generation layer includes:

[0129] Device Type Index Table: Stores the device type identifier corresponding to the device node;

[0130] Fault Type Matching Table: Fault type identifiers corresponding to storage thermal defects, insulation degradation, protection linkage, ground short circuit propagation, and environmental common causes;

[0131] Fault Mechanism Path Template Library: Stores path templates for thermal defect path, insulation degradation path, protection linkage path, grounding short circuit propagation path, and environmental common cause path;

[0132] Protection Segment Attribution Register: Stores the attribution relationships between protection device nodes, protection segment numbers, coverage equipment nodes, and coverage line segment nodes;

[0133] Alarm event queue: Stores alarm event nodes and corresponding alarm event types according to alarm trigger time;

[0134] Maintenance record matching table: stores the maintenance object number and fault confirmation type corresponding to the maintenance record node;

[0135] Metapath instance generator: Generates metapath instances based on device type, fault type, protected area, alarm event, and maintenance record;

[0136] Meta-path set buffer: Stores meta-path instances according to fault mechanism category, and outputs a meta-path set of fault mechanisms;

[0137] In the dynamic meta-path generation layer of fault mechanism, the equipment type index table and the fault type matching table output the equipment type identifier and the fault type identifier, and together with the path templates in the fault mechanism meta-path template library, input the meta-path instance generator. The protected area attribution register, alarm event queue and maintenance record matching table provide the meta-path instance generator with the protected area attribution relationship, alarm event type and maintenance record matching information, respectively. The meta-path instance generator generates meta-path instances corresponding to different fault mechanisms and writes the meta-path instances into the meta-path set buffer. The meta-path set buffer outputs the fault mechanism meta-path set according to the fault mechanism category.

[0138] The set of fault mechanism meta-paths is generated as follows:

[0139] Based on the initial embedding of heterogeneous nodes and the node number, device type, fault type, alarm event type, and maintenance record type in the corresponding node attributes, the fault mechanism category corresponding to the current device node is determined, specifically as follows:

[0140] Input the equipment type and fault type into the fault type matching table to obtain the initial fault mechanism category corresponding to the current equipment node;

[0141] The fault type matching table records the matching fields corresponding to thermal defects, insulation degradation, protection linkage, ground short circuit propagation, and environmental common causes. Among them, thermal defects correspond to temperature rise alarms, infrared temperature measurement alarms, contact abnormalities, and heating defect handling records; insulation degradation corresponds to partial discharge alarms, insulation alarms, humidity abnormality alarms, and insulation defect handling records; protection linkage corresponds to protection action alarms, trip alarms, reclosing alarms, and protection setting verification records; ground short circuit propagation corresponds to overcurrent alarms, grounding alarms, zero-sequence current alarms, and line test records; and environmental common causes correspond to ambient temperature alarms, humidity alarms, lightning strike alarms, and synchronous alarms of multiple devices in the same area.

[0142] When the device type, fault type, alarm event type, and maintenance record type corresponding to the current device node match the same fault mechanism matching field in the fault type matching table, the fault mechanism to which the matching field belongs is determined as the fault mechanism category of the current device node.

[0143] When the current device node matches multiple fault mechanism categories at the same time, the primary fault mechanism category is determined according to the fault type matching result, the alarm event occurrence time, and the maintenance record confirmation time, and the remaining fault mechanism categories are written as auxiliary fault mechanism categories into the fault mechanism identifier of the current device node;

[0144] The fault mechanism meta-path templates corresponding to the fault mechanism categories are retrieved from the fault mechanism meta-path template library, and fault mechanism meta-path instances are generated by combining alarm event types and maintenance record types. Among them, the thermal defect meta-path template includes equipment nodes and corresponding status characteristics, alarm event nodes and maintenance record nodes; the insulation degradation meta-path template includes equipment nodes and corresponding status characteristics, alarm event nodes, historical fault nodes and maintenance record nodes; the protection linkage meta-path template includes equipment nodes, protection device nodes, protection ownership side and upstream and downstream power supply equipment nodes; the ground short circuit propagation meta-path template includes equipment nodes, line section nodes, protection action status and adjacent equipment nodes; and the environmental common cause meta-path template includes equipment nodes, equipment nodes in the same area, alarm event nodes and maintenance record nodes.

[0145] Based on the protected area affiliation and power supply direction, the fault mechanism element path instances are filtered out. Fault mechanism element path instances that are not associated with the current equipment node in terms of both protected area affiliation and power supply are deleted, while fault mechanism element path instances that belong to the same protected area, adjacent protected areas, or have upstream and downstream power supply relationships are retained.

[0146] The retained fault mechanism meta-path instances are written into the starting device node number, intermediate node number, ending device node number, fault mechanism category, protection section identifier, power supply direction identifier, and time identifier, and are classified and summarized according to thermal defects, insulation degradation, protection linkage, grounding short circuit propagation, and environmental common causes to generate a fault mechanism meta-path set.

[0147] The alarm time difference direction aggregation layer reads the fault mechanism meta-path set and the initial embedding of heterogeneous nodes, calculates the time difference between the equipment status abnormality time, protection action time, alarm trigger time, and maintenance confirmation time, and performs direction aggregation on the fault mechanism meta-path instances in conjunction with the power supply direction to generate the equipment fault embedding representation and the alarm time difference direction embedding, wherein:

[0148] The alarm time difference direction aggregation layer includes:

[0149] Event time register group: storage device status abnormality time, protection action time, alarm trigger time, and maintenance confirmation time;

[0150] Time Difference Calculator: Calculates the time difference between the time of equipment status abnormality and the time of protection action, the time of protection action and the time of alarm triggering, and the time of alarm triggering and maintenance confirmation.

[0151] Power supply direction index table: upstream relationships, downstream relationships, and adjacent segment relationships between storage device nodes;

[0152] Fault mechanism meta-path instance queue: Stores fault mechanism meta-path instances according to fault mechanism category and time identifier;

[0153] Directional consistency discriminator: Based on the power supply direction index table and the node order in the fault mechanism element path instance, it determines the consistency relationship between the propagation direction of the fault mechanism element path instance and the power supply direction, and generates directional consistency parameters.

[0154] Time Difference Direction Embedding Calculator: Generates alarm time difference direction embedding based on time difference, direction consistency parameters, and initial embedding of heterogeneous nodes;

[0155] Meta-path aggregation cache: Stores the aggregation results of multiple fault mechanism meta-path instances corresponding to the same device node;

[0156] Fault Embedding Output: Performs a summary on the aggregation results and outputs a device fault embedding representation and an alarm time difference direction embedding;

[0157] In the alarm time difference direction aggregation layer, the abnormal device status time, protection action time, alarm trigger time, and maintenance confirmation time recorded by the event time register group are input into the time difference calculator. The time difference calculator generates the time difference between each event. The upstream relationship, downstream relationship, and adjacent segment relationship recorded by the power supply direction index table, together with the fault mechanism meta-path instances in the fault mechanism meta-path instance queue, are input into the direction consistency discriminator. The direction consistency discriminator generates direction consistency parameters. The time difference, direction consistency parameters, and heterogeneous node initial embedding are input into the time difference direction embedding calculator to generate alarm time difference direction embedding. The alarm time difference direction embedding and the aggregation result of the fault mechanism meta-path instances are written into the meta-path aggregation buffer. The fault embedding outputter reads the aggregation result in the meta-path aggregation buffer and performs summarization, outputting the device fault embedding representation and the alarm time difference direction embedding.

[0158] The time difference between the time of equipment status abnormality, the time of protection action, the time of alarm triggering, and the time of maintenance confirmation is calculated as follows:

[0159] Read the abnormal device status time, protection action time, alarm trigger time and maintenance confirmation time corresponding to the same fault mechanism meta-path instance from the event time register group, and establish an event time sequence according to the device node number, fault mechanism category and time identifier;

[0160] Using the time of equipment status abnormality as the starting time, calculate the first time difference from the time of equipment status abnormality to the time of protection action, the second time difference from the time of protection action to the time of alarm triggering, and the third time difference from the time of alarm triggering to the time of maintenance confirmation.

[0161] When there are multiple alarm trigger times for the same device node, the alarm trigger times that are located in the same fault mechanism meta-path instance as the device abnormality time and whose time sequence is continuous are selected for time difference calculation.

[0162] When the same fault mechanism path instance lacks maintenance confirmation time, the alarm trigger time is taken as the end event time, and the third time difference is marked as the time difference to be confirmed.

[0163] Write the first time difference, the second time difference, and the third time difference into the time difference calculation result table according to the fault mechanism metapath instance number to generate the time difference;

[0164] Combined with the power supply direction, the fault mechanism meta-path instance is aggregated in terms of direction, specifically as follows:

[0165] Read fault mechanism path instances from the fault mechanism path instance queue, and read the upstream, downstream and adjacent segment relationships between adjacent device nodes in the fault mechanism path instances from the power supply direction index table;

[0166] Based on the node arrangement order in the fault mechanism meta-path instance, the propagation direction corresponding to the fault mechanism meta-path instance is determined, and the propagation direction is compared with the upstream and downstream relationships in the power supply direction index table to generate direction consistency parameters.

[0167] When the propagation direction of the fault mechanism path instance is consistent with the power supply direction, the direction consistency parameter is marked as the forward propagation parameter. When the propagation direction of the fault mechanism path instance is opposite to the power supply direction, the direction consistency parameter is marked as the reverse propagation parameter. When the device nodes in the fault mechanism path instance belong to the adjacent segment relationship, the direction consistency parameter is marked as the adjacent propagation parameter.

[0168] The heterogeneous nodes corresponding to each node in the fault mechanism metapath instance are initially embedded into the input time difference direction embedding calculator, and the heterogeneous node initial embedding of each node in the fault mechanism metapath instance is unified into a 128-dimensional vector, and the metapath node embedding sequence is formed according to the node arrangement order.

[0169] The time difference is converted into minutes and normalized to generate a time difference normalization vector. The forward propagation parameter, backward propagation parameter, and adjacent propagation parameter are mapped to 1, -1, and 0.5 respectively to generate direction identification values.

[0170] The time difference normalized vector and the direction identifier value are concatenated into the meta-path node embedding sequence to form the time difference direction enhanced embedding sequence. A linear mapping is performed on the time difference direction enhanced embedding sequence to generate the fault mechanism meta-path instance aggregation result.

[0171] The aggregation results of multiple fault mechanism meta-path instances are written into the meta-path aggregation cache according to the device node number, and the aggregation results corresponding to the same device node are summarized to generate device fault embedding representation and alarm time difference direction embedding.

[0172] The protected area propagation boundary discrimination layer reads the device fault embedding representation, alarm time difference direction embedding, protected area attribution, protection action status, and switch on / off status. It then determines the propagable, restricted, and blocked propagation states between device nodes, generating a fault propagation boundary matrix. Rows correspond to the risk source devices, columns correspond to the affected devices, and elements correspond to the propagation boundary states between device nodes.

[0173] The propagation boundary discrimination layer of the protected area includes:

[0174] Protection Segment Attribution Register: Stores the attribution relationships between protection device nodes, protection segment numbers, coverage equipment nodes, and coverage line segment nodes;

[0175] Protection Action Status Table: Stores the protection action type, protection action status, and protection action time identifier corresponding to the protection device node;

[0176] Switch Open / Close Status Table: Stores the open / close status of circuit breakers, sectionalizing switches, tie switches, and disconnecting switches;

[0177] Propagation Path Index Table: Stores the line segment nodes and protection device nodes between the risk source device node and the affected device node;

[0178] Propagation status discriminator: Based on the protection section affiliation, protection action status, switch on / off status, and propagation path index table, it determines the propagation status, restricted propagation status, and blocked propagation status between equipment nodes;

[0179] Propagation Boundary Matrix Writer: The risk source device node is used as the matrix row, the affected device node is used as the matrix column, and the propagation boundary state is written into the corresponding matrix element to generate the fault propagation boundary matrix.

[0180] In the protection zone propagation boundary discrimination layer, the protection zone attribution register, protection action status table, and switch on / off status table transmit protection attribution information, protection action information, and switch status information to the propagation status discriminator, respectively. The propagation path index table synchronously inputs the path information between the risk source device node and the affected device node into the propagation status discriminator. The propagation status discriminator generates the propagation boundary status based on the input information and transmits it to the propagation boundary matrix writer. The propagation boundary matrix writer writes matrix elements according to the correspondence between the risk source device node and the affected device node, generating the fault propagation boundary matrix.

[0181] The process of determining the propagational, restricted, and blocked propagation states between device nodes is as follows:

[0182] Read the path information from the risk source device node to the affected device node in the propagation path index table. The path information includes line segment nodes, protection device nodes, switch nodes, and power supply direction identifiers.

[0183] Read the protection attribution information in the protection zone attribution register to determine whether the risk source device node and the affected device node are located in the same protection zone, adjacent protection zones, or the same power supply path;

[0184] Read the protection action status table and the switch open / closed status table, determine whether the protection device node in the propagation path is in the activated state, and determine whether the circuit breaker, sectionalizing switch, tie switch and disconnecting switch in the propagation path are in the open state.

[0185] When the risk source device node and the affected device node are located on the same power supply path, the protection device node in the propagation path is not in the activated state, and all the switch nodes in the propagation path are in the closed state, the propagation boundary state between the corresponding device nodes is determined to be a propagable state.

[0186] When the risk source device node and the affected device node are located in adjacent protection zones, and there are nodes with activated protection devices or disconnected switches in the propagation path, and the power supply path still has a bypass connection, the propagation boundary state between the corresponding device nodes is determined to be a restricted propagation state.

[0187] When there is no power supply path between the risk source device node and the affected device node, or when the propagation path is cut off by the activated protection device node and the disconnect switch node, the propagation boundary state between the corresponding device nodes is determined to be the propagation blocking state.

[0188] The improved meta-path aggregation graph neural network is trained by combining the fault classification error corresponding to the equipment fault embedding representation, the temporal direction error corresponding to the alarm time difference direction embedding, and the boundary discrimination error corresponding to the fault propagation boundary matrix as the joint optimization objective. The network parameters of the fault mechanism dynamic meta-path generation layer, the alarm time difference direction aggregation layer, and the protected area propagation boundary discrimination layer are continuously updated. When the change in the joint loss value in five consecutive training rounds is less than 0.001, the improved meta-path aggregation graph neural network is considered to have completed convergence training.

[0189] The improved meta-path aggregation graph neural network is trained as follows:

[0190] Read the heterogeneous fault graph, fault label, alarm time difference label and propagation boundary label from the training samples. Input the heterogeneous fault graph into the improved metapath aggregation graph neural network. Output the device fault embedding representation, alarm time difference direction embedding and fault propagation boundary matrix. Map the device fault embedding representation to the fault category prediction result, map the alarm time difference direction embedding to the time sequence direction prediction result, and map the fault propagation boundary matrix to the propagation boundary prediction result.

[0191] The fault classification error is obtained by averaging the squared differences between the fault category prediction result and the fault label. The time-series direction prediction error is obtained by averaging the squared differences between the time-series direction prediction result and the alarm time difference label. The boundary discrimination error is obtained by averaging the squared differences between the propagation boundary prediction result and the propagation boundary label. The fault classification error, time-series direction error, and boundary discrimination error are multiplied by 0.40, 0.30, and 0.30, and then summed to generate the joint loss value.

[0192] Read the rate of change of the joint loss value for each network parameter to obtain the gradient value for each network parameter. Multiply the gradient value by the learning rate of 0.001 and subtract it from the current network parameter value to obtain the updated network parameter value. Then rewrite it into the fault mechanism dynamic meta-path generation layer, alarm time difference direction aggregation layer and protection zone propagation boundary discrimination layer. Repeat the forward calculation, error calculation and parameter update according to the batch size of 64. When the change of the joint loss value in five consecutive training rounds is less than 0.001, it is determined that the improved meta-path aggregation graph neural network has completed convergence training.

[0193] In this embodiment, the generated device diagnostic evaluation data includes:

[0194] The device fault embedding representation and alarm time difference direction embedding are associated according to the device number to generate joint diagnostic features of the device;

[0195] Based on the joint diagnostic features of the equipment, the category mapping value corresponding to each fault category is calculated, and the category mapping value is subjected to normalized exponential processing to generate the fault category probability corresponding to each power transmission and distribution equipment, where:

[0196] The fault category probabilities corresponding to each power transmission and distribution equipment are generated as follows:

[0197] Read the joint diagnostic features of the equipment and determine the joint diagnostic features of each power transmission and distribution equipment according to the equipment number;

[0198] Based on the fault labels obtained by matching historical fault data and maintenance record data in the standardized equipment status dataset, the joint diagnostic features of the equipment are divided into category feature groups corresponding to normal status, thermal defects, insulation degradation, protection linkage, grounding short circuit propagation, and environmental common causes.

[0199] Calculate the category center feature vector for each category feature group. The value of each dimension in the category center feature vector is the average value of the joint diagnostic features of all devices in the same category feature group on the corresponding feature dimension.

[0200] Read the joint diagnostic features of the current power transmission and distribution equipment, and calculate the feature distance between the joint diagnostic features of the equipment and the feature vector of each category center respectively. The feature distance is the value obtained by summing the squares of the differences of each corresponding feature dimension and then taking the square root.

[0201] Take the inverse of the feature distance corresponding to each fault category to generate the category mapping value of the corresponding fault category, and perform normalization exponential processing on the category mapping value to obtain the probability of a single fault category;

[0202] The probabilities of a single fault category are summarized according to the equipment number and fault category to generate the fault category probabilities corresponding to each power transmission and distribution equipment.

[0203] Based on the abnormal category probability in the fault category probability, the abnormal offset in the equipment joint diagnostic features, and the directional consistency parameter in the alarm time difference direction embedding, the equipment risk score is calculated, where:

[0204] The risk assessment of computing devices is as follows:

[0205] Read the probability of thermal defects, insulation degradation, protection linkage, grounding short circuit propagation, and environmental common causes from the fault category probability, and sum the above probabilities to generate abnormal category probability;

[0206] Read the joint diagnostic features of the device and the category center feature vector corresponding to the normal state, calculate the feature distance between the joint diagnostic features of the device and the category center feature vector of the normal state, and divide the feature distance by the maximum feature distance in the same diagnostic batch to generate the abnormal offset.

[0207] Read the direction consistency parameter embedded in the alarm time difference direction, convert the forward propagation parameter to 1, the adjacent propagation parameter to 0.5, and the backward propagation parameter to 0;

[0208] The abnormal category probability is multiplied by 0.45, the abnormal offset by 0.35, and the directional consistency parameter by 0.20, and then summed to generate the equipment risk score.

[0209] When the equipment risk score is greater than 1, the equipment risk score is corrected to 1; when the equipment risk score is less than 0, the equipment risk score is corrected to 0.

[0210] The diagnostic confidence is calculated based on the maximum category probability, the dispersion of the fault category probability distribution, and the directional consistency parameter, where:

[0211] The diagnostic confidence level is calculated as follows:

[0212] Read the fault category probability corresponding to each power transmission and distribution equipment, and select the fault category probability with the largest value from the fault category probabilities corresponding to normal state, thermal defects, insulation deterioration, protection linkage, ground short circuit propagation and environmental common causes, as the maximum category probability;

[0213] Calculate the average probability of each fault category, sum the squared differences between the probability of each fault category and the average, and divide the sum by the number of fault categories to obtain the dispersion of the fault category probability distribution.

[0214] Read the direction consistency parameter embedded in the alarm time difference direction, convert the forward propagation parameter to 1, the adjacent propagation parameter to 0.5, and the backward propagation parameter to 0;

[0215] The diagnostic confidence score is generated by multiplying the maximum category probability by 0.50, the dispersion of the fault category probability distribution by 0.30, and the direction consistency parameter by 0.20.

[0216] When the diagnostic confidence level is greater than 1, the diagnostic confidence level is adjusted to 1; when the diagnostic confidence level is less than 0, the diagnostic confidence level is adjusted to 0.

[0217] The fault category probability, equipment risk score, and diagnostic confidence are summarized according to the equipment number to generate equipment diagnostic evaluation data.

[0218] In this embodiment, generating the optimized diagnostic parameter set includes:

[0219] Read equipment diagnostic evaluation data, alarm time difference direction embedding, and fault propagation boundary matrix to establish a set of diagnostic parameters to be optimized. This set includes fault category determination thresholds, equipment risk score correction coefficients, propagation boundary state correction coefficients, fault source equipment screening thresholds, affected equipment screening thresholds, and maintenance priority ranking coefficients. The set of diagnostic parameters to be optimized is then encoded as candidate individuals for *Corydalis porcupine* species.

[0220] Establish a set of diagnostic parameters to be optimized, specifically as follows:

[0221] Read the fault category probability, equipment risk score and diagnostic confidence in the equipment diagnostic evaluation data, read the directional consistency parameter in the alarm time difference direction embedding, and read the propagation boundary state in the fault propagation boundary matrix;

[0222] The fault category determination threshold is determined based on the distribution range of the probability of each fault category. The initial value of the fault category determination threshold is set as the midpoint between the average probability of each fault category and the maximum category probability. The range of the fault category determination threshold is set to 0.50 to 0.90.

[0223] The equipment risk score correction coefficient is determined based on the mean and standard deviation of the equipment risk score. The initial value of the equipment risk score correction coefficient is set to 1.00, and the range of the equipment risk score correction coefficient is set to 0.80 to 1.20.

[0224] The propagation boundary state correction coefficients are determined based on the propagable state, restricted propagation state, and blocked propagation state in the fault propagation boundary matrix. The initial value of the propagable state correction coefficient is set to 1.00, the initial value of the restricted propagation state correction coefficient is set to 0.60, and the initial value of the blocked propagation state correction coefficient is set to 0.00.

[0225] The screening thresholds for fault source equipment and affected equipment are determined based on equipment risk scores and diagnostic confidence levels. The initial value of the screening threshold for fault source equipment is set as the product of the average equipment risk score and the average diagnostic confidence level, and the initial value of the screening threshold for affected equipment is set as 0.70 times the screening threshold for fault source equipment.

[0226] Based on the ranking requirements of candidate fault source equipment, candidate affected equipment, and candidate fault propagation chain, maintenance priority ranking coefficients are established. The maintenance priority ranking coefficients include risk score ranking coefficient, fault source ranking coefficient, propagation chain length ranking coefficient, and diagnostic confidence ranking coefficient. The initial values ​​of the four ranking coefficients are set to 0.35, 0.30, 0.20, and 0.15, respectively.

[0227] The fault category determination threshold, equipment risk score correction coefficient, propagation boundary state correction coefficient, fault source equipment screening threshold, affected equipment screening threshold and maintenance priority ranking coefficient are summarized to form a set of diagnostic parameters to be optimized.

[0228] The set of diagnostic parameters to be optimized is encoded into candidate individuals of crowned porcupines, specifically:

[0229] The diagnostic parameters to be optimized are arranged in the following order: fault category determination threshold, equipment risk score correction coefficient, propagation boundary state correction coefficient, fault source equipment screening threshold, affected equipment screening threshold, and maintenance priority ranking coefficient, and a diagnostic parameter vector is generated.

[0230] Numerical normalization is performed on each parameter in the diagnostic parameter vector, mapping each parameter to the interval between 0 and 1, and generating a normalized diagnostic parameter vector.

[0231] Each parameter in the normalized diagnostic parameter vector is used as a positional component of a candidate individual of the Crowned Porcupine and written into the Crowned Porcupine candidate individual coding table according to the parameter arrangement order to generate candidate individuals of the Crowned Porcupine.

[0232] Based on the fault category probability, equipment risk score, and diagnostic confidence level in the equipment diagnostic evaluation data, the fault identification evaluation value, fault missed detection evaluation value, normal false alarm evaluation value, and diagnostic confidence evaluation value of the candidate individuals of the Crown Porcupine are calculated, where:

[0233] The fault identification evaluation value of candidate individuals of the crowned porcupine is calculated as follows:

[0234] Read the fault category determination threshold in the candidate individuals of the crowned porcupine and read the fault category probability in the equipment diagnostic evaluation data;

[0235] The fault category corresponding to the fault category with the highest numerical fault category probability in each power transmission and distribution equipment is used as the predicted fault category;

[0236] When the probability of the predicted fault category reaches the fault category determination threshold, and the predicted fault category is consistent with the corresponding fault label in the standardized equipment status dataset, the corresponding power transmission and distribution equipment is marked as the correctly identified equipment.

[0237] The ratio of the number of correctly identified devices to the total number of devices participating in the evaluation is used to generate a fault identification evaluation value;

[0238] The fault miss rate evaluation value of candidate individuals of the crowned porcupine is calculated as follows:

[0239] Read the standardized equipment status data and identify the transmission and distribution equipment with fault labels of thermal defects, insulation deterioration, protection linkage, ground short circuit propagation and environmental common causes as the actual faulty equipment.

[0240] When the probability of the abnormal category corresponding to the actual faulty equipment does not reach the fault category judgment threshold, or the equipment risk score corresponding to the actual faulty equipment does not reach the fault source equipment screening threshold, the corresponding actual faulty equipment will be marked as a missed judgment equipment.

[0241] The ratio of the number of missed faults to the total number of actual faulty equipment is used to generate a fault missed evaluation value.

[0242] The normal false alarm rating for crowned porcupine candidate individuals is calculated as follows:

[0243] Read the power transmission and distribution equipment whose fault labels are in the normal state from the standardized equipment status dataset and classify them as normal equipment;

[0244] When the probability of abnormal category corresponding to normal equipment reaches the fault category judgment threshold, or the equipment risk score corresponding to normal equipment reaches the fault source equipment screening threshold, the corresponding normal equipment is marked as a false alarm equipment; the ratio of the number of false alarm equipment to the total number of normal equipment is calculated to generate a normal false alarm evaluation value;

[0245] The diagnostic confidence score for candidate individuals of the crowned porcupine is calculated as follows:

[0246] Read the diagnostic confidence level from the equipment diagnostic evaluation data, and associate the diagnostic confidence level with the predicted fault category according to the equipment number;

[0247] Statistically predict the power transmission and distribution equipment whose fault categories reach the fault category determination threshold, calculate the average diagnostic confidence level of the corresponding power transmission and distribution equipment, and generate a diagnostic confidence evaluation value;

[0248] Based on the fault propagation boundary matrix, the propagation boundary consistency evaluation value of candidate individuals of the crowned porcupine is calculated as follows:

[0249] Read the propagation boundary state in the fault propagation boundary matrix and read the propagation boundary state correction coefficient in the crowned porcupine candidate individuals;

[0250] Multiply the matrix elements corresponding to the propagable states in the fault propagation boundary matrix by the propagable state correction coefficient, multiply the matrix elements corresponding to the restricted propagation states by the restricted propagation state correction coefficient, and multiply the matrix elements corresponding to the blocked propagation states by the blocked propagation state correction coefficient to generate the corrected propagation boundary matrix.

[0251] Read the maintenance and fault labels obtained by matching historical fault data and maintenance record data in the standardized equipment status dataset. Mark the equipment pairs whose fault labels and maintenance labels both point to the same fault propagation chain as real propagation equipment pairs, and mark the equipment pairs whose fault labels and maintenance labels do not both point to the same fault propagation chain as non-propagation equipment pairs.

[0252] The matrix elements corresponding to the actual propagation device pairs in the modified propagation boundary matrix are compared with the propagation state for consistency. The matrix elements corresponding to the non-propagation device pairs in the modified propagation boundary matrix are compared with the propagation blocking state for consistency. The ratio of the number of consistent device pairs to the total number of device pairs participating in the comparison is calculated to generate a propagation boundary consistency evaluation value.

[0253] Alarm time difference intervals are defined based on the direction of the alarm time difference embedding. These intervals include the time difference between an abnormal state and alarm triggering, the time difference between protection action and alarm triggering, and the time difference between alarm triggering and maintenance confirmation. The candidate individual update method for visual defense, sound defense, odor defense, and physical attack is switched according to the alarm time difference interval, updating the candidate individuals for the Crowned Porcupine.

[0254] The alarm time difference intervals are divided as follows:

[0255] Read the time difference from state abnormality to alarm trigger, the time difference from protection action to alarm trigger, and the time difference from alarm trigger to maintenance confirmation embedded in the alarm time difference direction, and convert each time difference into minutes;

[0256] When the time difference of a single item is greater than or equal to 0 minutes and less than or equal to 5 minutes, the corresponding time difference is divided into a concentrated time difference interval;

[0257] When the time difference of a single item is greater than 5 minutes and less than or equal to 15 minutes, the corresponding time difference of the single item will be divided into a time difference delay interval;

[0258] When the time difference of a single item is greater than 15 minutes, the corresponding time difference will be divided into discrete time difference intervals;

[0259] Based on the interval results corresponding to the time difference from abnormal status to alarm trigger, the time difference from protection action to alarm trigger, and the time difference from alarm trigger to maintenance confirmation, generate alarm time difference interval identifiers;

[0260] The method for updating candidate individuals is switched as follows:

[0261] Read the alarm time difference interval identifier and count the interval types corresponding to the time difference from abnormal status to alarm trigger, the time difference from protection action to alarm trigger, and the time difference from alarm trigger to maintenance confirmation.

[0262] When the number of concentrated time difference intervals in the alarm time difference interval identifier is greater than the number of delayed time difference intervals and the number of discrete time difference intervals, the candidate individual update method corresponding to odor defense and physical attack is selected, and the fault category judgment threshold, fault source device screening threshold and affected device screening threshold are locally adjusted.

[0263] When the number of delayed time difference intervals in the alarm time difference interval identifier is greater than the number of centralized time difference intervals and the number of discrete time difference intervals, the candidate individual update method corresponding to the sound defense is selected, and the device risk score correction coefficient and the propagation boundary state correction coefficient are subjected to neighborhood perturbation.

[0264] When the number of discrete time difference intervals in the alarm time difference interval identifier is greater than the number of concentrated time difference intervals and the number of delayed time difference intervals, the candidate individual update method corresponding to visual defense is selected, and all parameter components in the Crowned Porcupine candidate individuals are globally searched and updated.

[0265] When the number of centralized time difference intervals, the number of delayed time difference intervals, and the number of discrete time difference intervals have the same maximum value, the candidate individual update method is determined according to the interval type corresponding to the time difference from the state abnormality to the alarm trigger.

[0266] The updated candidate individuals for the crowned porcupine are as follows:

[0267] Read the fault category determination threshold, equipment risk score correction coefficient, propagation boundary state correction coefficient, fault source equipment screening threshold, affected equipment screening threshold, and maintenance priority ranking coefficient from the candidate individuals of the crown porcupine;

[0268] When the candidate individual update method is odor defense, a local perturbation value with a value range of [-0.02, 0.02] is superimposed on the current parameter components to update the fault category determination threshold, fault source device screening threshold, and affected device screening threshold;

[0269] When the candidate individual update method is physical attack, read the corresponding parameter component in the current non-dominated candidate individual, add 0.50 times the difference between the current parameter component and the corresponding parameter component of the non-dominated candidate individual to the current parameter component, and update the fault category judgment threshold, fault source device screening threshold and affected device screening threshold.

[0270] When the candidate individual update method is sound defense, a neighborhood perturbation value with a value range of [-0.05, 0.05] is superimposed on the current parameter components to update the device risk score correction coefficient and the propagation boundary state correction coefficient;

[0271] When the candidate individual update method is visual defense, a global perturbation value with a value range of [-0.10, 0.10] is superimposed on the current parameter components to update all parameter components in the Crowned Porcupine candidate individual;

[0272] The updated parameter components are limited to the range [0,1]. When a parameter component is greater than 1, it is corrected to 1. When a parameter component is less than 0, it is corrected to 0. The corrected parameter components are then rewritten into the coding table of the Crowned Porcupine candidate individuals to obtain the updated Crowned Porcupine candidate individuals.

[0273] In the collaborative optimization process of fault source tracing, device pairs are generated based on the fault propagation boundary matrix. Device pairs with propagation boundary states that are blocked from propagation are eliminated. Based on the device risk score, diagnostic confidence, and propagation boundary state of the retained device pairs, candidate values ​​for fault sources, affected candidate values, and propagation direction candidate values ​​are calculated to generate fault source tracing evaluation values, where:

[0274] Generate device pairs, specifically:

[0275] Read the row devices, column devices, and corresponding matrix elements in the fault propagation boundary matrix, and take the row devices as the risk source device nodes and the column devices as the affected device nodes;

[0276] Combine the risk-source device node and the affected device node into a device pair, and write the propagation boundary state in the corresponding matrix element into the device pair attribute;

[0277] Delete device pairs where the risk source device node is the same as the affected device node, and retain device pairs consisting of different device nodes;

[0278] The devices that screen out and block the transmission of the virus are specifically:

[0279] Read the propagation boundary state from the device pair attributes, and remove the device pairs whose propagation boundary state is blocked from the device pair set.

[0280] The device pairs with propagation boundary states of propagable state and restricted propagation state are retained, and a set of retained device pairs is established according to the device node number of the risk source and the device node number of the affected device.

[0281] The following parameters are used to generate fault source tracing evaluation values:

[0282] Read the device risk score and diagnostic confidence level corresponding to the risk source device node in the reserved device pair, multiply the device risk score by 0.60 and the diagnostic confidence level by 0.40, and then sum them to generate the candidate value of the fault source.

[0283] Read the device risk score, diagnostic confidence, and propagation boundary status corresponding to the affected device node in the reserved device pair, map the propagable state to 1, map the restricted propagation state to 0.6, and multiply the device risk score by 0.35, the diagnostic confidence by 0.25, and the propagation boundary status mapping value by 0.40, and then sum them up to generate the affected candidate value;

[0284] Based on the direction consistency parameter, the forward propagation parameter is mapped to 1, the adjacent propagation parameter is mapped to 0.5, and the backward propagation parameter is mapped to 0. The direction consistency parameter mapping value is multiplied by the propagation boundary state mapping value to generate the propagation direction candidate value.

[0285] The fault source candidate value is multiplied by 0.40, the affected candidate value is multiplied by 0.30, and the propagation direction candidate value is multiplied by 0.30, and then summed to generate the fault source tracking evaluation value.

[0286] The evaluation values ​​of fault identification, fault omission, normal false alarm, diagnosis confidence, propagation boundary consistency, and fault source tracing are combined into a multi-objective evaluation vector. Based on the multi-objective evaluation vector, non-dominated sorting and crowding distance sorting are performed on the candidate individuals of Crown Porcupine to generate a Pareto candidate diagnostic parameter set.

[0287] The candidate diagnostic parameters with the first level of non-dominated status in the Pareto candidate diagnostic parameter set are sorted from largest to smallest according to the crowding distance, and the sorted candidate diagnostic parameters are written into the external archive to generate the optimized diagnostic parameter set.

[0288] In this embodiment, generating candidate fault source devices, candidate affected devices, and candidate fault propagation chains includes:

[0289] Read the optimized diagnostic parameter set, and perform joint correction on the fault category probability, equipment risk score, and fault propagation boundary matrix based on the optimized diagnostic parameter set to generate corrected fault category probability, corrected equipment risk score, and corrected fault propagation boundary matrix, where:

[0290] Joint correction is performed on the fault category probability, equipment risk score, and fault propagation boundary matrix, specifically as follows:

[0291] Read the fault category judgment threshold, equipment risk score correction coefficient and propagation boundary state correction coefficient from the optimized diagnostic parameter set, and read the corresponding fault category probability, equipment risk score and fault propagation boundary matrix according to the equipment number;

[0292] The fault category probability corresponding to each power transmission and distribution equipment is compared with the fault category judgment threshold. The fault category probabilities that reach the fault category judgment threshold are retained, and the fault category probabilities that do not reach the fault category judgment threshold are written as 0 to generate the corrected fault category probability.

[0293] The equipment risk score is multiplied by the equipment risk score correction factor to obtain the corrected equipment risk score. When the corrected equipment risk score is greater than 1, the corrected equipment risk score is corrected to 1. When the corrected equipment risk score is less than 0, the corrected equipment risk score is corrected to 0.

[0294] Read the propagation boundary state corresponding to each matrix element in the fault propagation boundary matrix, multiply the matrix element corresponding to the propagable state by the propagable state correction coefficient, multiply the matrix element corresponding to the restricted propagation state by the restricted propagation state correction coefficient, and multiply the matrix element corresponding to the blocked propagation state by the blocked propagation state correction coefficient to generate the corrected fault propagation boundary matrix.

[0295] Based on the corrected fault propagation boundary matrix, the device node connection relationships corresponding to the blocked propagation state are filtered out, while the device node connection relationships corresponding to the propagable state and the restricted propagation state are retained.

[0296] Based on the calibration equipment risk score, calibration failure category probability, failure source equipment screening threshold, and affected equipment screening threshold, candidate failure source equipment and candidate affected equipment are screened.

[0297] Connect the candidate fault source device and the candidate affected device according to the preserved device node connection relationship to generate a candidate fault propagation chain.

[0298] In this embodiment, generating fault diagnosis results for power transmission and distribution equipment includes:

[0299] Read the maintenance priority ranking coefficients from the candidate fault source equipment, candidate affected equipment, candidate fault propagation chain, and optimized diagnostic parameter set, and calculate the maintenance priority score for the candidate maintenance objects, where:

[0300] The maintenance priority score for candidate maintenance objects is calculated as follows:

[0301] Read the candidate fault source device, candidate affected device and candidate fault propagation chain, merge the candidate fault source device and candidate affected device into a candidate maintenance object set, and write the device number, device role, candidate fault propagation chain to which it belongs and propagation chain position for each candidate maintenance object;

[0302] Read the maintenance priority ranking coefficients from the optimized diagnostic parameter set. The maintenance priority ranking coefficients include risk score ranking coefficients, fault source ranking coefficients, propagation chain location ranking coefficients, and diagnostic confidence ranking coefficients.

[0303] Read the risk score, diagnostic confidence and propagation chain position of the calibration equipment corresponding to the candidate maintenance object, set the equipment role value of the candidate fault source equipment to 1, set the equipment role value of the candidate affected equipment to 0.6, set the propagation chain start position value to 1, and decrease it by 0.1 according to the order of the candidate maintenance object in the candidate fault propagation chain;

[0304] The calibration equipment risk score is multiplied by the risk score ranking coefficient, the equipment role value is multiplied by the fault source ranking coefficient, the propagation chain position value is multiplied by the propagation chain position ranking coefficient, and the diagnostic confidence is multiplied by the diagnostic confidence ranking coefficient. The product results are then summed to generate the maintenance priority score of the candidate maintenance object.

[0305] Candidate maintenance objects are sorted according to their maintenance priority scores to generate a maintenance priority sequence;

[0306] Based on the maintenance priority sequence, candidate fault source equipment, candidate affected equipment, and candidate fault propagation chain, generate fault diagnosis results for power transmission and distribution equipment.

[0307] Receive maintenance feedback data, update the fault labels, alarm labels and maintenance labels in the standardized equipment status dataset based on the maintenance feedback data, and update the set of fault event nodes and event association edges in the heterogeneous fault graph of the equipment.

[0308] Example 1: To verify the feasibility of this invention in practice, it was applied to a fault diagnosis scenario in a power transmission and distribution maintenance area. During a continuous power transmission and distribution maintenance diagnosis cycle, this method was used to diagnose faults in power transmission and distribution equipment within the maintenance area. The equipment involved in the diagnosis included 16 distribution transformers, 48 ​​switchgear units, 32 circuit breakers, 18 ring main units, 26 cable line sections, 41 protection devices, and 96 online monitoring terminals. A total of 864,000 multi-source operational data points were collected, including 438,000 electrical operation data points, 216,000 equipment status monitoring data points, 12,600 protection action data points, 158,000 alarm data points, 14,200 historical fault data points, and 25,200 maintenance record data points. The original data exhibited inconsistencies in equipment number formats, inconsistent sampling intervals, duplicate alarm records, and missing status monitoring data. The missing equipment status monitoring data rate was 4.8%, and the duplicate alarm rate was 7.2%.

[0309] After the data enters the processing flow, preprocessing is performed on the multi-source operational data. Data from different sources are associated according to equipment number and acquisition window, with a unified sampling granularity of 1 minute. Missing status values ​​are imputed using adjacent operational windows, duplicate alarm records are merged, and voltage, current, temperature, partial discharge, and infrared thermometry data are standardized. The resulting standardized equipment status dataset contains 63,240 sets of equipment operational data sequences. Each sequence records the equipment number, operational characteristics, protection action identifier, alarm tag, fault tag, and maintenance tag. After preprocessing, the data missing rate decreased to 0.6%, the duplicate record rate decreased to 0.4%, and the continuity rate of adjacent window data increased from 88.1% to 97.3%.

[0310] Based on a standardized equipment status dataset, a heterogeneous fault graph was constructed according to equipment nodes, line segment nodes, protection device nodes, and fault event nodes. The resulting graph contains 140 equipment nodes, 26 line segment nodes, 41 protection device nodes, and 1846 fault event nodes. 326 topology connection edges were established according to power supply direction, 412 protection attribution edges were established according to protection segment affiliation, and 7386 event association edges were established according to alarm events, historical faults, and maintenance records. The heterogeneous fault graph contains a total of 2053 nodes and 8124 edges, with each edge recording its type and time identifier.

[0311] After inputting the heterogeneous fault graph of the equipment into the improved meta-path aggregation graph neural network, embedding mapping and dimension alignment are first performed on the node attributes of different types of nodes to form an initial embedding of 128-dimensional heterogeneous nodes. The fault mechanism dynamic meta-path generation layer generates a set of fault mechanism meta-paths based on equipment type, fault type, protected area affiliation, alarm event type, and maintenance record type, including 428 thermal defect meta-paths, 316 insulation degradation meta-paths, 392 protection linkage meta-paths, 274 grounding short circuit propagation meta-paths, and 186 environmental common cause meta-paths. Compared with the fixed meta-path method, the proportion of invalid meta-paths decreased from 31.5% to 9.8% after dynamic generation.

[0312] In a ground fault diagnosis sample, cable section L-12 experienced a sudden increase in zero-sequence current over three consecutive operating windows, with the normalized value of the zero-sequence current rising from 0.22 to 0.87, and the voltage drop reaching 0.19. Protection device P-07 activated 2 minutes after the abnormal status occurred. Downstream ring main unit R-04, switchgear K-18, and distribution transformer T-09 triggered voltage abnormality alarms within 1, 3, and 4 minutes after the protection activation. The alarm time difference directional aggregation layer calculated the time difference from the abnormal status to the protection activation to be 2 minutes, the time difference from the protection activation to the alarm triggering to be 1 minute, 3 minutes, and 4 minutes respectively, and the time difference from the alarm triggering to maintenance confirmation to be 46 minutes. Directional aggregation was then performed on the original path instances in conjunction with the power supply direction. After aggregation, the average directional embedding value for the alarm time difference of L-12 was 0.74, the average directional embedding value for the affected downstream equipment was 0.61, and the average directional embedding value for the background equipment was 0.18.

[0313] The protected area propagation boundary discrimination layer continues to read the equipment fault embedding representation, alarm time difference direction embedding, protected area affiliation, protection action status, and switch on / off status to generate a fault propagation boundary matrix. In the matrix, elements corresponding to L-12 to R-04 are determined to be propagable, elements corresponding to L-12 to K-18 are determined to be restricted propagation, and elements corresponding to L-12 to the adjacent connecting line L-19 are determined to be blocked propagation. After propagation boundary discrimination, there are 38 groups of propagable equipment pairs, 17 groups of restricted propagation equipment pairs, and 44 groups of blocked propagation equipment pairs.

[0314] Based on the equipment fault embedding representation and alarm time difference direction embedding, equipment diagnostic evaluation data were calculated. The ground fault propagation category probability for L-12 is 0.92, the equipment risk score is 0.89, and the diagnostic confidence level is 0.86; the fault category probability for R-04 is 0.23, the equipment risk score is 0.55, and the diagnostic confidence level is 0.78; the fault category probability for K-18 is 0.27, the equipment risk score is 0.58, and the diagnostic confidence level is 0.76; and the fault category probability for T-09 is 0.19, the equipment risk score is 0.49, and the diagnostic confidence level is 0.73. These data indicate that L-12 better matches the characteristics of the fault source, and the downstream equipment better matches the characteristics of the affected equipment.

[0315] Equipment diagnostic evaluation data, alarm time difference direction embedding, and fault propagation boundary matrix were used as candidate individual evaluation data for the multi-objective Crowned Porcupine optimization algorithm. The number of candidate individuals was set to 50. The set of diagnostic parameters to be optimized included fault category judgment threshold, equipment risk score correction coefficient, propagation boundary state correction coefficient, fault source equipment screening threshold, affected equipment screening threshold, and maintenance priority ranking coefficient. Alarm time difference intervals were divided based on alarm time difference direction embedding. When the time difference between protection action and alarm trigger was concentrated within 5 minutes, the candidate individual update method corresponding to odor defense and physical attack was adopted; when the alarm time difference was dispersed beyond 12 minutes, the candidate individual update method corresponding to visual defense and sound defense was adopted. After 42 iterations, 8 sets of candidate diagnostic parameters with a non-dominant level of first grade were retained from the external archive, ultimately generating the optimized diagnostic parameter set.

[0316] Based on the optimized diagnostic parameter set, joint corrections were performed on the fault category probability, equipment risk score, and fault propagation boundary matrix. After correction, the equipment risk score of L-12 increased to 0.93, the equipment risk score of R-04 was corrected to 0.51, the equipment risk score of K-18 was corrected to 0.54, and the equipment risk score of T-09 was corrected to 0.46. Subsequently, the equipment node connection relationships corresponding to the blocked propagation state were screened out, while the equipment node connection relationships corresponding to the propagable and restricted propagation states were retained. L-12 was identified as the candidate fault source device, and R-04, K-18, and T-09 were identified as candidate affected devices. Candidate fault propagation chains L-12 to R-04, L-12 to K-18, and L-12 to T-09 were generated.

[0317] Based on candidate fault source devices, candidate affected devices, and candidate fault propagation chains, the maintenance priorities of power transmission and distribution equipment were ranked. In the ranking results, L-12 was ranked first, protection device P-07 verification was ranked second, R-04 retest was ranked third, K-18 retest was ranked fourth, and T-09 retest was ranked fifth. Maintenance feedback data indicated that there were traces of grounding discharge in the cable section corresponding to L-12, the P-07 protection action record was consistent with the fault section, no inherent defects were found in R-04, K-18, and T-09, and the retested voltage returned to normal. The maintenance feedback data was written into the standardized equipment status dataset, and the fault event node set and event association edges in the heterogeneous fault graph were updated synchronously.

[0318] In the comparative experiments, the number of training samples was 48,000 sets, and the number of test samples was 12,000 sets. The fault category identification accuracy of the traditional threshold combined with ordinary graph neural network method was 85.6%, while that of our method was 94.2%; the false negative rate of high-risk faults of the traditional method was 10.4%, while that of our method was 3.5%; the false alarm rate of normal equipment of the traditional method was 12.1%, while that of our method was 5.6%; the fault source localization accuracy of the traditional method was 76.8%, while that of our method was 92.1%; the accuracy of affected equipment identification of the traditional method was 73.4%, while that of our method was 89.8%; the accuracy of blocking propagation relationships screening of the traditional method was 68.9%, while that of our method was 91.0%; the hit rate of the top three maintenance priorities of the traditional method was 75.7%, while that of our method was 90.6%; the average single diagnosis time of the traditional method was 36.4 seconds, while that of our method was 24.7 seconds. As can be seen from this embodiment, the method has a clear data change process in terms of multi-source operation data fusion, fault propagation boundary discrimination, fault source tracing and maintenance priority ranking. It can reduce the false alarm rate and false negative rate in scenarios where multiple devices alarm simultaneously, and verify the feasibility and effectiveness of the method in the fault diagnosis scenario of power transmission and distribution equipment.

[0319] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fault diagnosis method for power transmission and distribution equipment based on deep learning, characterized in that, include: Collect multi-source operation data of power transmission and distribution equipment, and perform preprocessing on the multi-source operation data to generate a standardized equipment status dataset; Based on a standardized equipment status dataset, a heterogeneous fault diagram of equipment is constructed according to the topological connection relationship of power transmission and distribution equipment, the ownership relationship of protected sections, and the correlation relationship of fault events. An improved meta-path aggregation graph neural network is constructed to perform dynamic meta-path generation of fault mechanisms, alarm time difference direction aggregation, and protection zone propagation boundary discrimination on heterogeneous fault graphs of equipment, generating equipment fault embedding representation, alarm time difference direction embedding, and fault propagation boundary matrix; Based on the equipment fault embedding representation and alarm time difference direction embedding, the fault category probability, equipment risk score and diagnostic confidence are calculated to generate equipment diagnostic evaluation data; The equipment diagnostic evaluation data, alarm time difference direction embedding, and fault propagation boundary matrix are used as candidate individual evaluation data for the multi-objective crown porcupine optimization algorithm. Alarm time difference defense switching and fault source tracing collaborative optimization are performed to generate an optimized diagnostic parameter set. Based on the optimized set of diagnostic parameters, joint correction is performed on the fault category probability, equipment risk score and fault propagation boundary matrix to screen out blocking propagation relationships and generate candidate fault source equipment, candidate affected equipment and candidate fault propagation chain; Based on candidate fault source equipment, candidate affected equipment, and candidate fault propagation chain, the power transmission and distribution equipment is prioritized for maintenance, generating fault diagnosis results for the power transmission and distribution equipment. Maintenance feedback data is received to update the standardized equipment status dataset and the heterogeneous fault diagram of the equipment.

2. The method for fault diagnosis of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, The generation of the standardized device status dataset includes: Collect multi-source operating data of power transmission and distribution equipment, including electrical operating data, equipment status monitoring data, protection action data, alarm data, historical fault data, and maintenance record data; Multi-source operational data are correlated according to equipment number and collection time to generate equipment operational data sequences; The equipment operation data sequence is processed to remove outliers, fill in missing values, and delete duplicate records, generating a cleaned equipment operation data sequence. The cleaned equipment operation data sequence is processed by timestamp alignment, unit unification and numerical standardization to generate a standardized equipment operation data sequence. Standardized equipment operation data sequences are matched with corresponding fault labels, alarm labels, and maintenance labels to generate standardized equipment status datasets.

3. The method for fault diagnosis of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, The constructed device heterogeneous fault diagram includes: Read the device number, device type, collection time and status characteristics from the standardized device status dataset, establish a set of device nodes, and write the status characteristics into the node attributes of the corresponding device nodes; Based on the topological connection relationship of power transmission and distribution equipment, establish a set of nodes for line sections, and establish topological connection edges between equipment nodes and line section nodes according to the power supply direction; Based on the ownership relationship of the protected area sections, a set of protection device nodes is established, and protection ownership edges are established between the protection device nodes and the corresponding coverage equipment nodes and the corresponding line section nodes. Based on the correlation of fault events, a set of fault event nodes is established, which includes alarm event nodes, historical fault nodes, and maintenance record nodes. Event association edges are established between various types of fault event nodes and corresponding equipment nodes. The set of equipment nodes, the set of line segment nodes, the set of protection device nodes, the set of fault event nodes, and various edge relationships are combined according to the node type identifier, edge type identifier, and time identifier to generate a heterogeneous fault diagram of equipment.

4. The method for fault diagnosis of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, The generated device fault embedding representation, alarm time difference direction embedding, and fault propagation boundary matrix include: An improved meta-path aggregation graph neural network is constructed, which includes a fault mechanism dynamic meta-path generation layer, an alarm time difference direction aggregation layer, and a protected area propagation boundary discrimination layer. Read the set of device nodes, line segment nodes, protection device nodes and fault event nodes in the heterogeneous fault diagram of the equipment, perform embedding mapping and dimension alignment on the node attributes of different types of nodes, and generate the initial embedding of heterogeneous nodes. The fault mechanism dynamic meta-path generation layer reads the initial embedding of heterogeneous nodes, equipment type, fault type, protection zone affiliation, alarm event type, and maintenance record type, and generates a set of fault mechanism meta-paths according to thermal defects, insulation degradation, protection linkage, grounding short circuit propagation, and environmental common causes. The alarm time difference direction aggregation layer reads the fault mechanism meta-path set and the initial embedding of heterogeneous nodes, calculates the time difference between the equipment status abnormal time, protection action time, alarm trigger time and maintenance confirmation time, and performs direction aggregation on the fault mechanism meta-path instance in combination with the power supply direction to generate the equipment fault embedding representation and alarm time difference direction embedding. The protection zone propagation boundary discrimination layer reads the equipment fault embedding representation, alarm time difference direction embedding, protection zone affiliation, protection action status and switch on / off status, and determines the propagable status, restricted propagation status and blocked propagation status between equipment nodes, and generates a fault propagation boundary matrix, with rows corresponding to risk source equipment, columns corresponding to affected equipment, and elements corresponding to the propagation boundary status between equipment nodes. The improved metapath aggregation graph neural network was trained by combining the fault classification error corresponding to the equipment fault embedding representation, the timing direction error corresponding to the alarm time difference direction embedding, and the boundary discrimination error corresponding to the fault propagation boundary matrix as the joint optimization objective. The network parameters of the fault mechanism dynamic metapath generation layer, the alarm time difference direction aggregation layer, and the protection zone propagation boundary discrimination layer were continuously updated. When the change of the joint loss value in five consecutive training rounds was less than 0.001, the improved metapath aggregation graph neural network was determined to have completed convergence training.

5. The method for fault diagnosis of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, The generated device diagnostic evaluation data includes: The device fault embedding representation and alarm time difference direction embedding are associated according to the device number to generate joint diagnostic features of the device; Based on the joint diagnostic features of the equipment, the category mapping value corresponding to each fault category is calculated, and the category mapping value is processed by normalization index to generate the fault category probability corresponding to each power transmission and distribution equipment. The equipment risk score is calculated based on the abnormal category probability in the fault category probability, the abnormal offset in the equipment joint diagnostic features, and the directional consistency parameter in the alarm time difference direction embedding. The diagnostic confidence level is calculated based on the maximum category probability, the dispersion of the fault category probability distribution, and the directional consistency parameter in the fault category probability. The fault category probability, equipment risk score, and diagnostic confidence are summarized according to the equipment number to generate equipment diagnostic evaluation data.

6. The method for fault diagnosis of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, The generated optimized diagnostic parameter set includes: Read the equipment diagnostic evaluation data, alarm time difference direction embedding and fault propagation boundary matrix, establish a set of diagnostic parameters to be optimized. The set of diagnostic parameters to be optimized includes fault category judgment threshold, equipment risk score correction coefficient, propagation boundary state correction coefficient, fault source equipment screening threshold, affected equipment screening threshold and maintenance priority ranking coefficient, and encode the set of diagnostic parameters to be optimized as Crowned porcupine candidate individuals. Based on the fault category probability, equipment risk score and diagnostic confidence in the equipment diagnostic evaluation data, the fault identification evaluation value, fault missed evaluation value, normal false alarm evaluation value and diagnostic confidence evaluation value of the candidate individuals of the Crown Porcupine were calculated. Based on the fault propagation boundary matrix, the propagation boundary consistency evaluation value of candidate individuals of Crowned Porcupine is calculated; The alarm time difference interval is divided according to the direction of alarm time difference embedding. The alarm time difference interval includes the time difference from abnormal status to alarm trigger, the time difference from protection action to alarm trigger, and the time difference from alarm trigger to maintenance confirmation. The candidate individual update method corresponding to visual defense, sound defense, odor defense and physical attack is switched according to the alarm time difference interval to update the crown porcupine candidate individuals. In the collaborative optimization process of fault source tracing, device pairs are generated based on the fault propagation boundary matrix. Device pairs with propagation boundary states that are blocked are screened out. Based on the device risk score, diagnostic confidence and propagation boundary state of the retained device pairs, the fault source candidate value, the affected candidate value and the propagation direction candidate value are calculated to generate the fault source tracing evaluation value. The evaluation values ​​of fault identification, fault omission, normal false alarm, diagnosis confidence, propagation boundary consistency, and fault source tracing are combined into a multi-objective evaluation vector. Based on the multi-objective evaluation vector, non-dominated sorting and crowding distance sorting are performed on the candidate individuals of Crown Porcupine to generate a Pareto candidate diagnostic parameter set. The candidate diagnostic parameters with the first level of non-dominated status in the Pareto candidate diagnostic parameter set are sorted from largest to smallest according to the crowding distance, and the sorted candidate diagnostic parameters are written into the external archive to generate the optimized diagnostic parameter set.

7. The method for fault diagnosis of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, The generation of candidate fault source devices, candidate affected devices, and candidate fault propagation chains includes: Read the optimized diagnostic parameter set, and perform joint correction on the fault category probability, equipment risk score and fault propagation boundary matrix based on the optimized diagnostic parameter set to generate the corrected fault category probability, corrected equipment risk score and corrected fault propagation boundary matrix; Based on the corrected fault propagation boundary matrix, the device node connection relationships corresponding to the blocked propagation state are filtered out, while the device node connection relationships corresponding to the propagable state and the restricted propagation state are retained. Based on the calibration equipment risk score, calibration failure category probability, failure source equipment screening threshold, and affected equipment screening threshold, candidate failure source equipment and candidate affected equipment are screened. Connect the candidate fault source device and the candidate affected device according to the preserved device node connection relationship to generate a candidate fault propagation chain.

8. The method for fault diagnosis of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, The generation of fault diagnosis results for power transmission and distribution equipment includes: Read the maintenance priority ranking coefficients from the candidate fault source equipment, candidate affected equipment, candidate fault propagation chain and optimized diagnostic parameter set, and calculate the maintenance priority score of the candidate maintenance object; Candidate maintenance objects are sorted according to their maintenance priority scores to generate a maintenance priority sequence; Based on the maintenance priority sequence, candidate fault source equipment, candidate affected equipment, and candidate fault propagation chain, generate fault diagnosis results for power transmission and distribution equipment. Receive maintenance feedback data, update the fault labels, alarm labels and maintenance labels in the standardized equipment status dataset based on the maintenance feedback data, and update the set of fault event nodes and event association edges in the heterogeneous fault graph of the equipment.

Images