Power equipment fault data conflict identification and grading method based on deep learning

Abstract

The invention discloses a power equipment fault data conflict identification and grading method based on deep learning, and the method comprises the steps: generating a state event based on real-time operation data, determining a dynamic threshold value, and constructing a dynamic activation topological graph structure; constructing a space-time state coding library and performing active state predictive coding; dynamic graph features are extracted through a bidirectional adaptive graph Transform; establishing an attention weight cross-time window comparison sample space, and generating a graph comparison sample pair; performing real-time data conflict code identification on the training feature encoder by using the graph comparison sample; constructing a three-level associated hierarchical topology conflict mode library and updating in real time; and actively matching and outputting a fault conflict grading result in real time by utilizing a grading topology mode library. According to the method, the accuracy and real-time performance of power equipment fault data conflict identification are improved, and the fault studying and judging efficiency and reliability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault diagnosis, and in particular to a method for identifying and grading power equipment fault data conflicts based on deep learning. Background Art

[0002] Power equipment fault diagnosis technology is an important foundation for ensuring the stable operation of power grids. With the continuous expansion of power grids in recent years, power equipment fault data has become multi-source, heterogeneous, time-series complex, and highly dynamic. Traditional power equipment fault diagnosis methods mostly rely on artificial rules or empirical knowledge, analyzing and judging fault data by presetting fixed thresholds or establishing static models. For example, expert system-based diagnostic solutions typically identify and classify fault types through the manual construction of knowledge bases. However, due to the limitations and static nature of empirical rules, they are difficult to adapt to the dynamic and complex trends of fault data. Furthermore, diagnostic methods based on traditional machine learning algorithms typically adopt predetermined feature extraction strategies and use classical classifiers for pattern recognition. While these methods improve some automation levels, they lack sensitivity to complex, multi-dimensional, and dynamically changing data, making it difficult to effectively handle conflicts between different data sources.

[0003] In recent years, with the rapid development of artificial intelligence technology, especially deep learning algorithms, some studies have proposed methods for intelligent analysis of power equipment fault data using deep neural networks, and have achieved certain results. Among them, graph neural networks and Transformer models have gradually been introduced into the field of power equipment fault diagnosis due to their ability to effectively capture the structural characteristics and global dependencies of data, and have initially achieved more accurate analysis of fault data. However, most existing technical solutions based on graph neural networks and Transformer algorithms only focus on the extraction of single static topological structure features, ignoring the spatiotemporal correlation relationship and potential conflicts of equipment data under dynamic state changes, resulting in insufficient recognition accuracy and real-time performance in actual complex scenarios. In addition, the existing technology lacks an active dynamic prediction mechanism for data conflict features, making it difficult to achieve keen recognition and accurate classification of subtle differences, resulting in a significant reduction in the efficiency and accuracy of fault data conflict identification.

[0004] Therefore, how to provide a method for identifying and classifying power equipment fault data conflicts based on deep learning is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] One purpose of the present invention is to propose a method for identifying and grading power equipment fault data conflicts based on deep learning. The present invention significantly improves the recognition accuracy and real-time performance of power equipment fault data conflicts, and improves the efficiency and reliability of fault data conflict diagnosis and grading.

[0006] According to an embodiment of the present invention, a method for identifying and grading power equipment fault data conflicts based on deep learning includes the following steps:

[0007] S1. Generate status events based on the operating status data of the power equipment, use the status events to determine the dynamic threshold, use the dynamic threshold as the activation condition of the edge, and construct a dynamically activated power equipment topology data structure;

[0008] S2. Based on the dynamically activated power equipment topology data structure, a spatiotemporal state coding library of nodes and edges is constructed. Active state prediction coding is performed for each node and edge that newly enters the system through the coding library.

[0009] S3. Based on the dynamic graph data structure after active state prediction coding, the forward and reverse dynamic graph features are interactively extracted through the bidirectional adaptive graph Transformer network to obtain the forward and reverse bidirectional dynamic graph feature vectors;

[0010] S4. Based on the forward and inverse bidirectional dynamic graph feature vectors, a cross-time window comparison sample space based on attention weight is established. By constructing the correlation and comparison relationship of graph features across the time window, graph comparison sample pairs are generated;

[0011] S5. Using the feature encoder trained with the graph comparison sample, perform data conflict coding recognition on the forward and reverse bidirectional dynamic graph feature vectors to obtain an initial recognition result;

[0012] S6. Construct a hierarchical association topology analysis network based on three levels of association: device, region, and global. Import the initial identification results into the three-level association network to form a hierarchical topology conflict pattern library.

[0013] S7. By using the updated hierarchical topology conflict pattern library and adopting an active matching conflict classification processing strategy, when a new data conflict event occurs, the power equipment fault conflict classification result is actively output.

[0014] Optionally, the S1 specifically includes:

[0015] S11. Extract voltage fluctuation period difference value, current instantaneous jump duration, power load dynamic balance conversion interval and frequency stability fluctuation interval value from power equipment operation data;

[0016] S12. Based on the extracted voltage fluctuation period difference value and current transient jump duration, a comparison benchmark for the timing characteristics of the equipment operation status is established, and the degree of nonlinear difference in the timing characteristics change is calculated;

[0017] S13. Based on the calculated nonlinear difference degree of the time series characteristics, a step-by-step adaptive determination mechanism is used to determine the state event triggering threshold, thereby obtaining an adaptive dynamic state event determination standard;

[0018] S14. Monitor the equipment operation data based on the dynamic state event judgment standard, generate a state event sequence, and record the occurrence frequency, duration, and associated equipment number of each state event sequence;

[0019] S15. Calculate the dynamic correlation strength between each associated device and its adjacent devices based on the occurrence frequency and duration of the state event sequence, and use the dynamic correlation strength to determine the activation condition of the connection edge in the topology graph;

[0020] S16, applying the activation condition to the connection edge of the power equipment topology graph, actively judging the edge activation or failure of the topology graph according to the edge activation condition, and recording the edge activation and failure times;

[0021] S17. Actively adjust the equipment topology structure based on the updated edge activation and failure judgment results to form an updated and dynamically activated power equipment topology data structure.

[0022] Optionally, the S2 specifically includes:

[0023] S21. Based on the dynamically activated power equipment topology data structure, collect and record the absolute timestamp of the first activation of each node, calculate the difference between the first activation timestamps of any adjacent nodes, and determine the relative timing difference of the node activations;

[0024] S22. Determine the temporal position of edge activation based on the first activation time of the edge connecting each node, calculate the absolute difference between the first activation time of each edge and the first activation time of the connected node, and determine the edge activation timing correlation difference;

[0025] S23. Using the relative timing difference of node activation, a timing difference clustering method is used to determine the timing state feature classification of the node, and a timing classification label is assigned to each node;

[0026] S24. Determine the activation order characteristics of edge connections based on the edge activation timing correlation difference, construct a timing connection strength benchmark that reflects the stability level of the edge connection, and assign a connection strength level label to each edge.

[0027] S25, combining the temporal classification labels of the nodes and the connection strength level labels of the edges to construct a unified spatiotemporal state fusion matrix of the nodes and edges;

[0028] S26. Periodically screen the spatiotemporal state fusion matrix to extract key fusion feature sequences that can characterize the overall topological structure change trend, and establish a high-confidence dynamic spatiotemporal state fusion coding library;

[0029] S27. When a new node or new edge enters the system, the initial activation timing characteristics of the new node or new edge are actively matched with the key fusion feature sequence in the dynamic spatiotemporal state fusion coding library, and the new node or new edge is assigned an active state prediction code based on the matching results.

[0030] Optionally, the S26 specifically includes:

[0031] S261. Based on the data of the spatiotemporal state fusion matrix, collect the state transition paths of the spatiotemporal state fusion codes of nodes and edges within multiple consecutive fixed periods, and construct a state transition path sequence;

[0032] S262, record and calculate the path length, state switching frequency, and path recurrence frequency of each state transition path sequence, and establish a state transition path feature set;

[0033] S263. Determine the path importance of nodes and edges in the state transition path based on the state transition path feature set, and assign a path importance level label to each state transition path;

[0034] S264, combining the path importance level mark with the corresponding state transition path sequence to construct and dynamically update the spatiotemporal state path importance fusion feature matrix;

[0035] S265. Based on the spatiotemporal state path importance fusion feature matrix, identify state transition paths with high path importance and stable occurrence frequency, and determine them as key fusion feature path sequences;

[0036] S266. Based on the key fusion feature path sequence, extract the fusion feature key nodes and key edges that characterize the abnormal change trend of the overall topological structure, and determine the high-confidence coding of the fusion feature key nodes and key edges;

[0037] S267. Establish and dynamically update the determined high-confidence code into a dynamic spatiotemporal state fusion code library with path feature tags.

[0038] Optionally, the S3 specifically includes:

[0039] S31. Based on the dynamic graph data structure after active state prediction coding, collect and record the absolute time point when the state prediction coding of each node and edge is first generated, and determine the initial propagation order of the state prediction coding of each node and edge one by one;

[0040] S32. Define and assign initial propagation direction tags for node and edge state prediction codes based on the initial propagation order of the node and edge state prediction codes;

[0041] S33, extracting the node activation order and edge connection order encoded on the forward propagation path based on the initial propagation direction mark encoded by the node and edge state prediction code, and obtaining a detailed topological sequence of the forward propagation path;

[0042] S34, starting from the end node of the detailed topology sequence of the forward propagation path, gradually backtracking to extract the activation order and connection order of the nodes and edges on the reverse propagation path, to obtain the detailed topology sequence of the reverse propagation path;

[0043] S35. Compare the detailed topological sequences of the forward and reverse propagation paths, and determine the nodes and edges that overlap between the two propagation path sequences, which are defined as key interaction overlapping nodes and edges;

[0044] S36. For key interaction overlapping nodes and edges, calculate the frequency of occurrence, propagation distance, and path density on the forward and reverse propagation paths, and quantify the similarity of the forward and reverse propagation paths;

[0045] S37. Based on the quantified similarity results of key interaction overlapping nodes and edges, the differential features of the forward and reverse propagation paths are integrated to construct and dynamically update the forward and reverse bidirectional dynamic graph feature vectors that simultaneously characterize the interaction trends of the forward and reverse features.

[0046] Optionally, the S4 specifically includes:

[0047] S41, collecting the forward and reverse bidirectional dynamic graph feature vectors, dividing the feature dynamic monitoring time windows of multiple adaptive lengths according to the feature vector fluctuation rate and amplitude, and recording the start and end time of each time window and the feature vector change trend;

[0048] S42, extracting nodes and edges dominated by feature vector change trends within each feature dynamic monitoring time window, and constructing feature vector trend evolution paths for each node and edge within each time window;

[0049] S43, analyzing the eigenvector trend evolution path of each node and edge, calculating the trend dominance metric value of the path based on the eigenvalue change trajectory on the path, and quantifying and determining the node and edge trend path dominance level;

[0050] S44, identifying nodes and edge pairs with the same trend path dominance level and overlapping topological connection positions within the adjacent feature dynamic monitoring time window, defining them as candidate comparison nodes and edge pairs across the time window, and assigning candidate comparison identifiers;

[0051] S45. For each set of candidate comparison nodes and edge pairs across the time window, extract the temporal trajectory of the feature vector change path, and calculate the temporal path difference value between the candidate comparison nodes and edge pairs;

[0052] S46. Compare the temporal path difference values ​​of candidate comparison nodes and edge pairs across the time window, and determine the node and edge pairs whose difference values ​​exceed the dynamically adjusted threshold as high-confidence dynamic graph comparison sample pairs;

[0053] S47. Store and dynamically update high-confidence dynamic graph comparison sample pairs, and construct and maintain cross-time window graph comparison sample pairs including candidate comparison identifiers, trend path dominance levels, and time series path difference values.

[0054] Optionally, the S5 specifically includes:

[0055] S51, obtaining a topological feature change pattern of each node and edge in a cross-time window graph comparison sample pair, and generating a topological feature evolution pattern trajectory of the node and edge in each time window according to the topological feature change pattern;

[0056] S52, calculating the evolution trajectory distance between each node and edge topology feature evolution pattern trajectory one by one, and constructing the node and edge evolution trajectory difference matrix according to the calculation results of the evolution trajectory distance;

[0057] S53, monitoring and obtaining forward and reverse bidirectional dynamic graph feature vectors, calculating a topological feature difference metric value in an evolution trajectory difference matrix for each feature vector, and determining abnormal nodes and edges whose topological feature difference metric values ​​exceed a preset conflict threshold;

[0058] S54, for the identified abnormal nodes and edges, extract the characteristic evolution pattern path of the adjacent topological nodes and edges, and locate the starting node, path node and end node of the abnormal change of the topological characteristics;

[0059] S55, tracking the path of the abnormal topological feature change, recording the path length, the amplitude of the topological feature change, and the duration of the abnormal topological feature change one by one, and defining them as a topological abnormal path event;

[0060] S56. Calculate the comprehensive anomaly intensity of the topological anomaly path event, determine the significance level of the conflict feature of each topological anomaly path event based on the comprehensive anomaly intensity, and assign a conflict identifier;

[0061] S57. Based on the conflict feature significance level and the conflict identifier, an association structure of abnormal nodes and edges of topological features is formed, and an initial identification result of the power equipment topology structure data conflict is generated and updated.

[0062] Optionally, the S6 specifically includes:

[0063] S61. Obtain the location, conflict identifier, and significance level of each topological anomaly node and edge in the initial identification results, and dynamically divide the topological anomaly propagation impact range into three levels: device level, regional level, and global level based on the location and propagation characteristics of the anomaly nodes and edges;

[0064] S62. Based on the impact range of device-level topology anomaly propagation, extract and record the starting node, abnormal topology path length, abnormal propagation speed between nodes, and continuous propagation duration of each device topology anomaly propagation one by one, and construct a device-level anomaly propagation path feature set;

[0065] S63. Based on the impact range of regional topological anomaly propagation, identify the initial time difference, propagation path topological structure differences, and anomaly propagation synchronization degree of topological anomaly cooperative propagation among multiple devices in the region, and construct a regional anomaly propagation cooperative feature set;

[0066] S64. Based on the impact range of global topological anomaly propagation, determine the difference in the starting position of the path of cross-regional topological anomaly propagation, the correlation strength between the cross-regional propagation time series difference and the cross-regional anomaly topological propagation, and construct a global anomaly propagation interaction feature set;

[0067] S65. Calculate the topological anomaly propagation correlation between the device-level anomaly propagation path feature set, the regional-level anomaly propagation collaborative feature set, and the global-level anomaly propagation interaction feature set, and establish a correlation mapping relationship between the device, regional, and global levels of anomaly propagation features.

[0068] S66. Generate device-level, regional-level, and global-level topology anomaly propagation patterns based on the correlation mapping relationship between the three-level anomaly propagation features, quantify the abnormal topology change intensity of each topology anomaly propagation pattern one by one, and assign a unique topology conflict pattern identifier;

[0069] S67. Store and dynamically update the device-level, regional-level, and global-level topology anomaly propagation patterns corresponding to the topology conflict pattern identifiers, and build and continuously maintain a hierarchical topology conflict pattern library containing three-level topology conflict pattern features and associated mapping relationships.

[0070] Optionally, the S7 specifically includes:

[0071] S71, obtaining the abnormal starting position, initial abnormal characteristic value and conflict occurrence time of the topological node and edge of the newly occurring data conflict event, and actively constructing the abnormal propagation topological path of the event in the power equipment topology network;

[0072] S72, performing topological feature tag encoding on the abnormal propagation topological path to form a topological abnormality feature fingerprint sequence representing the uniqueness of the abnormal event, and recording the propagation rate, number of path topological nodes, and total path propagation time of each feature fingerprint;

[0073] S73, calling and querying the hierarchical topology conflict pattern library, matching the topology anomaly feature fingerprint sequence with the device-level, regional-level, and global-level topology conflict pattern feature fingerprint sequences in the pattern library, and determining the matching similarity between the topology anomaly feature fingerprint sequence and the topology conflict patterns at each level;

[0074] S74, analyzing the propagation path topology node deviation, propagation rate difference, and path topology node number difference between the topology anomaly feature fingerprint sequence and the third-level topology conflict pattern with the highest matching similarity, to form an abnormal propagation path topology deviation feature vector;

[0075] S75. Calculate a comprehensive evaluation index value of the topological anomaly path difference based on the abnormal propagation path topological deviation characteristic vector;

[0076] S76. Determine the topology conflict risk level corresponding to the new data conflict event based on the topology anomaly path difference comprehensive evaluation index value and a preset topology anomaly risk level adaptive classification threshold;

[0077] S77. Output the power equipment fault conflict classification processing result including the abnormal starting position of the new data conflict event, the time when the conflict occurs, the matching three-level topology conflict pattern identifier and the topology conflict risk level.

[0078] The beneficial effects of the present invention are:

[0079] (1) The present invention forms a dynamic topology structure based on device status events in real time through a real-time dynamic threshold determination method and a dynamically activated power equipment topology map construction technology, effectively improving the response speed to device status changes and the real-time performance of topology structure updates, and enhancing the sensitivity and accuracy of fault data conflict detection.

[0080] (2) By constructing active spatiotemporal state prediction coding technology for nodes and edges and a bidirectional adaptive graph Transformer network, the present invention can achieve accurate and in-depth extraction of dynamic graph features of power equipment, significantly improving the dynamic adaptability and robustness of feature extraction, and showing better feature expression capabilities and prediction accuracy in complex time series changes and dynamic equipment association environments.

[0081] (3) In terms of dynamic graph data conflict recognition and classification, the present invention effectively solves the problem of difficulty in identifying subtle conflicts with feature differences in the existing technology through the graph feature comparison sample space construction technology across time windows based on attention weights and the real-time data conflict recognition method of the feature encoder, breaks through the limitation of insufficient sensitivity of static models, and realizes the keen detection and real-time classification processing of tiny data conflicts, thereby effectively improving the real-time decision-making efficiency and accuracy of power equipment fault diagnosis.

[0082] (4) The present invention realizes multi-scale and multi-level conflict identification and precise hierarchical management through the construction technology of topological conflict pattern library with three levels of association: device level, regional level and global level, and the real-time conflict risk level assessment method of active matching, effectively optimizing the intelligence level of power equipment fault warning decision-making, and showing higher adaptability and reliability in the actual power system operation environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0084] Figure 1 This is a schematic diagram of the overall process of the deep learning-based power equipment fault data conflict identification and classification method proposed in the present invention;

[0085] Figure 2 This is a flow chart of the dynamic graph feature extraction and graph comparison sample construction of the deep learning-based power equipment fault data conflict identification and classification method proposed in the present invention;

[0086] Figure 3 This is a flowchart of the three-level associated topology conflict classification processing of the deep learning-based power equipment fault data conflict identification and classification method proposed in the present invention. DETAILED DESCRIPTION

[0087] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0088] refer to Figure 1-Figure 3 , a method for identifying and grading power equipment fault data conflicts based on deep learning, comprising the following steps:

[0089] S1. Generate status events based on the operating status data of the power equipment, use the status events to determine the dynamic threshold, use the dynamic threshold as the activation condition of the edge, and construct a dynamically activated power equipment topology data structure;

[0090] S2. Based on the dynamically activated power equipment topology data structure, a spatiotemporal state coding library of nodes and edges is constructed. Active state prediction coding is performed for each node and edge that newly enters the system through the coding library.

[0091] S3. Based on the dynamic graph data structure after active state prediction coding, the forward and reverse dynamic graph features are interactively extracted through the bidirectional adaptive graph Transformer network to obtain the forward and reverse bidirectional dynamic graph feature vectors;

[0092] S4. Based on the forward and inverse bidirectional dynamic graph feature vectors, a cross-time window comparison sample space based on attention weight is established. By constructing the correlation and comparison relationship of graph features across the time window, graph comparison sample pairs are generated;

[0093] S5. Using the feature encoder trained with the graph comparison sample, perform data conflict coding recognition on the forward and reverse bidirectional dynamic graph feature vectors to obtain an initial recognition result;

[0094] S6. Construct a hierarchical association topology analysis network based on three levels of association: device, region, and global. Import the initial identification results into the three-level association network to form a hierarchical topology conflict pattern library.

[0095] S7. By using the updated hierarchical topology conflict pattern library and adopting an active matching conflict classification processing strategy, when a new data conflict event occurs, the power equipment fault conflict classification result is actively output.

[0096] By constructing a real-time dynamic topology graph structure, adopting active state prediction coding and a bidirectional adaptive graph Transformer network, real-time and efficient extraction and in-depth analysis of power equipment fault data features are achieved; through a cross-time window graph feature comparison method based on attention weights, subtle data conflicts can be accurately identified and distinguished; using a three-level association topology analysis network and a real-time updated conflict pattern library, active conflict risk level judgment and precise classification decisions are supported, thereby effectively improving the accuracy and real-time performance of fault data conflict identification and enhancing the reliability and response speed of fault diagnosis.

[0097] In this embodiment, S1 specifically includes:

[0098] S11. Extract voltage fluctuation period difference value, current instantaneous jump duration, power load dynamic balance conversion interval and frequency stability fluctuation interval value from power equipment operation data;

[0099] S12. Based on the extracted voltage fluctuation period difference value and current transient jump duration, a comparison benchmark for the timing characteristics of the equipment operation status is established, and the degree of nonlinear difference in the timing characteristics change is calculated;

[0100] S13. Based on the calculated nonlinear difference degree of the time series characteristics, a step-by-step adaptive determination mechanism is used to determine the state event triggering threshold, thereby obtaining an adaptive dynamic state event determination standard;

[0101] S14. Monitor the equipment operation data based on the dynamic state event judgment standard, generate a state event sequence, and record the occurrence frequency, duration, and associated equipment number of each state event sequence;

[0102] S15. Calculate the dynamic correlation strength between each associated device and its adjacent devices based on the occurrence frequency and duration of the state event sequence, and use the dynamic correlation strength to determine the activation condition of the connection edge in the topology graph;

[0103] S16, applying the activation condition to the connection edge of the power equipment topology graph, actively judging the edge activation or failure of the topology graph according to the edge activation condition, and recording the edge activation and failure times;

[0104] S17. Actively adjust the equipment topology structure based on the updated edge activation and failure judgment results to form an updated and dynamically activated power equipment topology data structure.

[0105] By extracting the changing characteristics of parameters such as voltage, current, power and frequency in the operation data of power equipment in real time, an adaptive dynamic state event judgment standard is constructed, and the dynamic correlation strength of nodes and edges is calculated in real time to determine the real-time activation conditions, thereby actively realizing the real-time update and adaptive adjustment of the topological graph structure, effectively improving the dynamic adaptability and accuracy of the topological structure, and enhancing the sensitivity and real-time response speed of equipment fault status identification.

[0106] In this embodiment, S2 specifically includes:

[0107] S21. Based on the dynamically activated power equipment topology data structure, collect and record the absolute timestamp of the first activation of each node, calculate the difference between the first activation timestamps of any adjacent nodes, and determine the relative timing difference of the node activations;

[0108] S22. Determine the temporal position of edge activation based on the first activation time of the edge connecting each node, calculate the absolute difference between the first activation time of each edge and the first activation time of the connected node, and determine the edge activation timing correlation difference;

[0109] S23. Using the relative timing difference of node activation, a timing difference clustering method is used to determine the timing state feature classification of the node, and a timing classification label is assigned to each node;

[0110] S24. Determine the activation order characteristics of edge connections based on the edge activation timing correlation difference, construct a timing connection strength benchmark that reflects the stability level of the edge connection, and assign a connection strength level label to each edge.

[0111] S25, combining the temporal classification labels of the nodes and the connection strength level labels of the edges to construct a unified spatiotemporal state fusion matrix of the nodes and edges;

[0112] S26. Periodically screen the spatiotemporal state fusion matrix to extract key fusion feature sequences that can characterize the overall topological structure change trend, and establish a high-confidence dynamic spatiotemporal state fusion coding library;

[0113] S27. When a new node or new edge enters the system, the initial activation timing characteristics of the new node or new edge are actively matched with the key fusion feature sequence in the dynamic spatiotemporal state fusion coding library, and the new node or new edge is assigned an active state prediction code based on the matching results.

[0114] By extracting and analyzing the first activation timing features of nodes and edges in real time, using the timing difference clustering method to determine the timing state classification of nodes, and constructing a timing connection strength benchmark that reflects the stability of edge connections, we further construct and periodically screen out key fusion feature sequences to establish a high-confidence dynamic spatiotemporal state fusion coding library, realizing active state prediction coding for new nodes or new edges, effectively improving the dynamic adaptability and state prediction accuracy of the topological structure of power equipment, and enhancing the sensitivity to dynamic data conflicts and real-time recognition capabilities.

[0115] In this embodiment, the S26 specifically includes:

[0116] S261. Based on the data of the spatiotemporal state fusion matrix, collect the state transition paths of the spatiotemporal state fusion codes of nodes and edges within multiple consecutive fixed periods, and construct a state transition path sequence;

[0117] S262, record and calculate the path length, state switching frequency, and path recurrence frequency of each state transition path sequence, and establish a state transition path feature set;

[0118] S263. Determine the path importance of nodes and edges in the state transition path based on the state transition path feature set, and assign a path importance level label to each state transition path;

[0119] S264, combining the path importance level mark with the corresponding state transition path sequence to construct and dynamically update the spatiotemporal state path importance fusion feature matrix;

[0120] S265. Based on the spatiotemporal state path importance fusion feature matrix, identify state transition paths with high path importance and stable occurrence frequency, and determine them as key fusion feature path sequences;

[0121] S266. Based on the key fusion feature path sequence, extract the fusion feature key nodes and key edges that characterize the abnormal change trend of the overall topological structure, and determine the high-confidence coding of the fusion feature key nodes and key edges;

[0122] S267. Establish and dynamically update the determined high-confidence code into a dynamic spatiotemporal state fusion code library with path feature tags.

[0123] By constructing and analyzing the state transition path sequences of nodes and edges, quantifying the path length, state switching frequency and recurrence frequency, and dynamically screening the key fusion feature path sequences according to the path importance level, the key fusion feature nodes and edges that can represent the abnormal change trend of the topological structure are extracted, achieving a more accurate and high-confidence dynamic spatiotemporal state fusion coding library construction, effectively improving the sensitivity and recognition accuracy of abnormal changes in the topological structure, and enhancing the accuracy and real-time response capability of power equipment fault data conflict detection and early warning.

[0124] In this embodiment, S3 specifically includes:

[0125] S31. Based on the dynamic graph data structure after active state prediction coding, collect and record the absolute time point when the state prediction coding of each node and edge is first generated, and determine the initial propagation order of the state prediction coding of each node and edge one by one;

[0126] S32. Define and assign initial propagation direction tags for node and edge state prediction codes based on the initial propagation order of the node and edge state prediction codes;

[0127] S33, extracting the node activation order and edge connection order encoded on the forward propagation path based on the initial propagation direction mark encoded by the node and edge state prediction code, and obtaining a detailed topological sequence of the forward propagation path;

[0128] S34, starting from the end node of the detailed topology sequence of the forward propagation path, gradually backtracking to extract the activation order and connection order of the nodes and edges on the reverse propagation path, to obtain the detailed topology sequence of the reverse propagation path;

[0129] S35. Compare the detailed topological sequences of the forward and reverse propagation paths, and determine the nodes and edges that overlap between the two propagation path sequences, which are defined as key interaction overlapping nodes and edges;

[0130] S36. For key interaction overlapping nodes and edges, calculate the frequency of occurrence, propagation distance, and path density on the forward and reverse propagation paths, and quantify the similarity of the forward and reverse propagation paths;

[0131] S37. Based on the quantified similarity results of key interaction overlapping nodes and edges, the differential features of the forward and reverse propagation paths are integrated to construct and dynamically update the forward and reverse bidirectional dynamic graph feature vectors that simultaneously characterize the interaction trends of the forward and reverse features.

[0132] By recording and analyzing the initial propagation order and propagation path of the node and edge state prediction coding in real time, and adopting a bidirectional path detailed topology sequence extraction method, the key interactive overlapping nodes and edges of the forward and reverse paths are accurately identified, and their features are fused. The feature vector with a bidirectional dynamic interaction trend is constructed and updated in real time, thereby effectively improving the comprehensiveness and accuracy of the dynamic graph feature expression, and significantly enhancing the real-time recognition and accurate judgment capabilities of conflicting features of power equipment fault data.

[0133] In this embodiment, the S4 specifically includes:

[0134] S41, collecting the forward and reverse bidirectional dynamic graph feature vectors, dividing the feature dynamic monitoring time windows of multiple adaptive lengths according to the feature vector fluctuation rate and amplitude, and recording the start and end time of each time window and the feature vector change trend;

[0135] S42, extracting nodes and edges dominated by feature vector change trends within each feature dynamic monitoring time window, and constructing feature vector trend evolution paths for each node and edge within each time window;

[0136] S43, analyzing the eigenvector trend evolution path of each node and edge, calculating the trend dominance metric value of the path based on the eigenvalue change trajectory on the path, and quantifying and determining the node and edge trend path dominance level;

[0137] S44, identifying nodes and edge pairs with the same trend path dominance level and overlapping topological connection positions within the adjacent feature dynamic monitoring time window, defining them as candidate comparison nodes and edge pairs across the time window, and assigning candidate comparison identifiers;

[0138] S45. For each set of candidate comparison nodes and edge pairs across the time window, extract the temporal trajectory of the feature vector change path, and calculate the temporal path difference value between the candidate comparison nodes and edge pairs;

[0139] S46. Compare the temporal path difference values ​​of candidate comparison nodes and edge pairs across the time window, and determine the node and edge pairs whose difference values ​​exceed the dynamically adjusted threshold as high-confidence dynamic graph comparison sample pairs;

[0140] S47. Store and dynamically update high-confidence dynamic graph comparison sample pairs, and construct and maintain cross-time window graph comparison sample pairs including candidate comparison identifiers, trend path dominance levels, and time series path difference values.

[0141] By dividing the bidirectional dynamic graph feature vectors into adaptive dynamic monitoring time windows in real time, accurately extracting the trend evolution paths of the feature vectors of nodes and edges, quantitatively determining the degree of path trend dominance and the timing path differences of candidate comparison nodes and edges across time windows, dynamically screening and determining high-confidence dynamic graph comparison sample pairs, and establishing and updating the dynamic graph feature comparison sample space with clear identification in real time, the accuracy of capturing dynamic graph data feature changes and the ability to identify abnormal differences are effectively improved, and the accuracy and real-time response capability of power equipment fault data conflict analysis are significantly enhanced.

[0142] In this embodiment, the S5 specifically includes:

[0143] S51, obtaining a topological feature change pattern of each node and edge in a cross-time window graph comparison sample pair, and generating a topological feature evolution pattern trajectory of the node and edge in each time window according to the topological feature change pattern;

[0144] S52, calculating the evolution trajectory distance between each node and edge topology feature evolution pattern trajectory one by one, and constructing the node and edge evolution trajectory difference matrix according to the calculation results of the evolution trajectory distance;

[0145] S53, monitoring and obtaining forward and reverse bidirectional dynamic graph feature vectors, calculating a topological feature difference metric value in an evolution trajectory difference matrix for each feature vector, and determining abnormal nodes and edges whose topological feature difference metric values ​​exceed a preset conflict threshold;

[0146] S54, for the identified abnormal nodes and edges, extract the characteristic evolution pattern path of the adjacent topological nodes and edges, and locate the starting node, path node and end node of the abnormal change of the topological characteristics;

[0147] S55, tracking the path of the abnormal topological feature change, recording the path length, the amplitude of the topological feature change, and the duration of the abnormal topological feature change one by one, and defining them as a topological abnormal path event;

[0148] S56. Calculate the comprehensive anomaly intensity of the topological anomaly path event, determine the significance level of the conflict feature of each topological anomaly path event based on the comprehensive anomaly intensity, and assign a conflict identifier;

[0149] S57. Based on the conflict feature significance level and the conflict identifier, an association structure of abnormal nodes and edges of topological features is formed, and an initial identification result of the power equipment topology structure data conflict is generated and updated.

[0150] By extracting the trajectory of topological feature change patterns across time windows, constructing the topological feature evolution trajectory difference matrix in real time and determining abnormal nodes and edges, accurately recording the starting node, path length, feature change amplitude and duration of topological abnormal path events, and calculating the comprehensive abnormality intensity in real time to determine the real-time conflict identification and significance level, the initial identification results of power equipment topological structure data conflicts are dynamically generated and updated, thereby effectively improving the precision of power equipment fault data anomaly identification and the accuracy of topological anomaly propagation path positioning, and enhancing the real-time and reliability of data conflict diagnosis and processing.

[0151] In this embodiment, S6 specifically includes:

[0152] S61. Obtain the location, conflict identifier, and significance level of each topological anomaly node and edge in the initial identification results, and dynamically divide the topological anomaly propagation impact range into three levels: device level, regional level, and global level based on the location and propagation characteristics of the anomaly nodes and edges;

[0153] S62. Based on the impact range of device-level topology anomaly propagation, extract and record the starting node, abnormal topology path length, abnormal propagation speed between nodes, and continuous propagation duration of each device topology anomaly propagation one by one, and construct a device-level anomaly propagation path feature set;

[0154] S63. Based on the impact range of regional topological anomaly propagation, identify the initial time difference, propagation path topological structure differences, and anomaly propagation synchronization degree of topological anomaly cooperative propagation among multiple devices in the region, and construct a regional anomaly propagation cooperative feature set;

[0155] S64. Based on the impact range of global topological anomaly propagation, determine the difference in the starting position of the path of cross-regional topological anomaly propagation, the correlation strength between the cross-regional propagation time series difference and the cross-regional anomaly topological propagation, and construct a global anomaly propagation interaction feature set;

[0156] S65. Calculate the topological anomaly propagation correlation between the device-level anomaly propagation path feature set, the regional-level anomaly propagation collaborative feature set, and the global-level anomaly propagation interaction feature set, and establish a correlation mapping relationship between the device, regional, and global levels of anomaly propagation features.

[0157] S66. Generate device-level, regional-level, and global-level topology anomaly propagation patterns based on the correlation mapping relationship between the three-level anomaly propagation features, quantify the abnormal topology change intensity of each topology anomaly propagation pattern one by one, and assign a unique topology conflict pattern identifier;

[0158] S67. Store and dynamically update the device-level, regional-level, and global-level topology anomaly propagation patterns corresponding to the topology conflict pattern identifiers, and build and continuously maintain a hierarchical topology conflict pattern library containing three-level topology conflict pattern features and associated mapping relationships.

[0159] By dynamically dividing the impact range of topological anomaly propagation at the device level, regional level and global level, constructing the device-level anomaly propagation path feature set, the regional-level anomaly collaborative propagation feature set and the global-level anomaly propagation interaction feature set, and quantifying the correlation mapping relationship between the three-level propagation features, a three-level topological conflict pattern library is established and updated in real time, forming a unique topological conflict pattern identification, thereby effectively improving the correlation analysis capability of the abnormal propagation characteristics of power equipment faults, enhancing the accuracy of abnormal propagation pattern recognition, and significantly optimizing the real-time early warning and precise hierarchical management of data conflict events.

[0160] In this embodiment, the S7 specifically includes:

[0161] S71, obtaining the abnormal starting position, initial abnormal characteristic value and conflict occurrence time of the topological node and edge of the newly occurring data conflict event, and actively constructing the abnormal propagation topological path of the event in the power equipment topology network;

[0162] S72, performing topological feature tag encoding on the abnormal propagation topological path to form a topological abnormality feature fingerprint sequence representing the uniqueness of the abnormal event, and recording the propagation rate, number of path topological nodes, and total path propagation time of each feature fingerprint;

[0163] S73, calling and querying the hierarchical topology conflict pattern library, matching the topology anomaly feature fingerprint sequence with the device-level, regional-level, and global-level topology conflict pattern feature fingerprint sequences in the pattern library, and determining the matching similarity between the topology anomaly feature fingerprint sequence and the topology conflict patterns at each level;

[0164] S74, analyzing the propagation path topology node deviation, propagation rate difference, and path topology node number difference between the topology anomaly feature fingerprint sequence and the third-level topology conflict pattern with the highest matching similarity, to form an abnormal propagation path topology deviation feature vector;

[0165] S75. Based on the topological deviation feature vector of the abnormal propagation path, the comprehensive evaluation index value D of the topological abnormal path difference is calculated. t :

[0166] D t =λ1·D n +λ2·D s +λ3·D l ;

[0167] Among them, D t represents the comprehensive evaluation index value of topological abnormal path difference, D n Indicates the path topology node deviation, D s Denotes the difference in propagation rate, D lRepresents the difference in the number of path topology nodes. The weight coefficients λ1, λ2, and λ3 are determined based on the influence of the device-level, regional-level, and global-level characteristic paths, and satisfy the condition λ1+λ2+λ3=1.

[0168] In the formula, the comprehensive evaluation index value D of topological abnormal path difference t By weighted comprehensive path topology node deviation D n , the difference in propagation rate D s and the difference in the number of path topology nodes D l To achieve the difference measurement of the topological anomaly event path, the weight coefficients λ1, λ2, and λ3 respectively reflect the influence of the characteristic paths at the device level, regional level, and global level. By adjusting the weight coefficients, accurate difference evaluation can be performed on topological feature changes at different levels, thereby achieving a comprehensive and quantitative assessment of the difference characteristics of topological anomaly paths, effectively supporting the accurate classification of abnormal events and determination of risk levels.

[0169] S76, based on the comprehensive evaluation index value D of the topological abnormal path difference t ,Combined with the preset topology anomaly risk level adaptive grading threshold, determine the topology conflict risk level corresponding to the new data conflict event;

[0170] S77. Output the power equipment fault conflict classification processing result including the abnormal starting position of the new data conflict event, the time when the conflict occurs, the matching three-level topology conflict pattern identifier and the topology conflict risk level.

[0171] By acquiring the topological anomaly feature fingerprint sequence of new data conflict events in real time, actively matching the most similar topological anomaly pattern in the three-level topological conflict pattern library and quantifying the path topological node deviation, propagation rate difference and path node number difference, the comprehensive evaluation index of topological anomaly path difference is calculated in real time, the topological conflict risk level is accurately determined and the graded processing results including the abnormal starting position, precise occurrence time and risk level are output in real time, thereby effectively improving the precise matching and risk assessment capabilities of topological anomaly events of power equipment and enhancing the accuracy and real-time performance of abnormal diagnosis decisions.

[0172] Example 1:

[0173] To verify the feasibility of the present invention in practice, the method was applied to a real-time status monitoring and fault warning analysis system for power equipment in a provincial power grid. Specifically, real-time monitoring of the operating status of multiple types of power equipment (such as transformers, circuit breakers, and transmission lines) in the grid was performed, along with fault data conflict analysis. Due to the complex operating environment of the power grid, the status data generated by various types of equipment is significantly heterogeneous and dynamic. Traditional methods based on threshold alarms and static rules have difficulty effectively identifying and processing dynamic conflicts between equipment fault data, resulting in low efficiency and accuracy in fault analysis, which seriously affects equipment operation and maintenance decisions.

[0174] In this implementation scenario, the method first collects real-time operational status data from equipment such as transformers, circuit breakers, and transmission lines, such as voltage fluctuation cycle differences, current transient jump durations, power load transition intervals, and frequency stability fluctuation ranges. It then generates dynamic thresholds and constructs a dynamic topology map of power equipment in real time to accurately visualize the real-time operational status relationships between devices. Within this real-time topology, the system actively predicts the state of all nodes and edges, comprehensively capturing the changing spatiotemporal characteristics of power equipment operation.

[0175] The system then leverages a bidirectional adaptive graph transformer network to simultaneously extract forward and reverse dynamic graph features of device operation, enabling in-depth analysis of dynamic operational characteristics within complex topologies. Based on these dynamic graph features, it generates graph feature comparison sample pairs across time windows, enabling precise identification of subtle data differences and conflicting features between devices. Furthermore, by establishing a library of topological conflict patterns at the device, regional, and global levels, the system proactively matches and outputs the risk level of device failure conflicts in real time, enabling rapid identification of risk levels and implementation of targeted response strategies.

[0176] To further verify the practical application effect of the method of the present invention, statistics were collected over three consecutive months on the time required to identify equipment fault data conflicts and complete analysis in a power grid system using both the traditional method and the method of the present invention. The fault data identification accuracy, false alarm rate, and average response time are shown in Table 1.

[0177] Table 1: Comparison of power equipment fault data conflict identification and processing performance

[0178]

[0179]

[0180] As shown in Table 1, it can be clearly seen from the actual data that the method of the present invention shortens the fault data conflict identification time by about 80.9% compared with the traditional method, greatly improving the real-time processing capability; in terms of the fault data identification accuracy, it is improved from 83.2% of the traditional method to 96.5%, an increase of about 16%, significantly improving the reliability of fault data diagnosis; in terms of the fault data false alarm rate, it is significantly reduced by about 81.9%, significantly reducing the processing burden and risk decision-making pressure of operation and maintenance personnel; in terms of the average response decision time, it is significantly reduced from 480 seconds of the traditional method to 75 seconds, an increase of about 84.4%, effectively ensuring the safety and stability of power equipment operation.

[0181] In addition, to verify the real-time processing capability and classification accuracy of the method of the present invention, a detailed comparison was made between actual data statistics and the traditional method in a typical equipment failure data conflict event, as shown in Table 2:

[0182] Table 2: Comparison of typical equipment failure data conflict event processing effects

[0183] Technical indicators Results of traditional methods Results of the method of the present invention Improvement Data conflict detection delay (seconds) 75 8 89.3％ Conflict classification accuracy (%) 79 98 24.1％ Total event processing time (seconds) 190 22 88.4％

[0184] As shown in Table 2, compared with traditional methods, the method of the present invention reduces the data conflict discovery delay time by about 89.3%, improves the conflict classification accuracy by 24.1%, and shortens the total event processing time by about 88.4%, demonstrating the significant advantages of the method of the present invention in real-time fault conflict identification and hierarchical processing capabilities.

[0185] The detailed application data of this embodiment fully demonstrates that the power equipment fault data conflict identification and classification method based on the graph Transformer algorithm combined with the attention-enhanced contrastive learning algorithm proposed in the present invention has achieved outstanding results in the accuracy of real-time dynamic conflict identification, the timeliness of data analysis and the accuracy of conflict risk classification. It is particularly suitable for complex, multi-dimensional, dynamic data interaction and frequent power system operating environments, and can effectively improve the efficiency of power grid fault diagnosis and the level of equipment safety management.

[0186] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for identifying and grading power equipment fault data conflicts based on deep learning, characterized in that: The steps include: S1. Generate status events based on the operating status data of the power equipment, use the status events to determine the dynamic threshold, use the dynamic threshold as the activation condition of the edge, and construct a dynamically activated power equipment topology data structure; S2. Based on the dynamically activated power equipment topology data structure, a spatiotemporal state coding library of nodes and edges is constructed. Active state prediction coding is performed for each node and edge that newly enters the system through the coding library. S3. Based on the dynamic graph data structure after active state prediction coding, the forward and reverse dynamic graph features are interactively extracted through the bidirectional adaptive graph Transformer network to obtain the forward and reverse bidirectional dynamic graph feature vectors; S4. Based on the forward and inverse bidirectional dynamic graph feature vectors, a cross-time window comparison sample space based on attention weight is established. By constructing the correlation and comparison relationship of graph features across the time window, graph comparison sample pairs are generated; S5. Using the feature encoder trained with the graph comparison sample, perform data conflict coding recognition on the forward and reverse bidirectional dynamic graph feature vectors to obtain an initial recognition result; S6. Construct a hierarchical association topology analysis network based on three levels of association: device, region, and global. Import the initial identification results into the three-level association network to form a hierarchical topology conflict pattern library. S7. By using the updated hierarchical topology conflict pattern library and adopting an active matching conflict classification processing strategy, when a new data conflict event occurs, the power equipment fault conflict classification result is actively output.

2. The method for identifying and grading power equipment fault data conflicts based on deep learning according to claim 1, characterized in that: Said S1 specifically includes: S11. Extract voltage fluctuation period difference value, current instantaneous jump duration, power load dynamic balance conversion interval and frequency stability fluctuation interval value from power equipment operation data; S12. Based on the extracted voltage fluctuation period difference value and current transient jump duration, a comparison benchmark for the timing characteristics of the equipment operation status is established, and the degree of nonlinear difference in the timing characteristics change is calculated; S13. Based on the calculated nonlinear difference degree of the time series characteristics, a step-by-step adaptive determination mechanism is used to determine the state event triggering threshold, thereby obtaining an adaptive dynamic state event determination standard; S14. Monitor the equipment operation data based on the dynamic state event judgment standard, generate a state event sequence, and record the occurrence frequency, duration, and associated equipment number of each state event sequence; S15. Calculate the dynamic correlation strength between each associated device and its adjacent devices based on the occurrence frequency and duration of the state event sequence, and use the dynamic correlation strength to determine the activation condition of the connection edge in the topology graph; S16, applying the activation condition to the connection edge of the power equipment topology graph, actively judging the edge activation or failure of the topology graph according to the edge activation condition, and recording the edge activation and failure times; S17. Actively adjust the equipment topology structure based on the updated edge activation and failure judgment results to form an updated and dynamically activated power equipment topology data structure.

3. The method for identifying and grading power equipment fault data conflicts based on deep learning according to claim 1, characterized in that: The S2 specifically includes: S21. Based on the dynamically activated power equipment topology data structure, collect and record the absolute timestamp of the first activation of each node, calculate the difference between the first activation timestamps of any adjacent nodes, and determine the relative timing difference of the node activations; S22. Determine the temporal position of edge activation based on the first activation time of the edge connecting each node, calculate the absolute difference between the first activation time of each edge and the first activation time of the connected node, and determine the edge activation timing correlation difference; S23. Using the relative timing difference of node activation, a timing difference clustering method is used to determine the timing state feature classification of the node, and a timing classification label is assigned to each node; S24. Determine the activation order characteristics of edge connections based on the edge activation timing correlation difference, construct a timing connection strength benchmark that reflects the stability level of the edge connection, and assign a connection strength level label to each edge. S25, combining the temporal classification labels of the nodes and the connection strength level labels of the edges to construct a unified spatiotemporal state fusion matrix of the nodes and edges; S26. Periodically screen the spatiotemporal state fusion matrix to extract key fusion feature sequences that can characterize the overall topological structure change trend, and establish a high-confidence dynamic spatiotemporal state fusion coding library; S27. When a new node or new edge enters the system, the initial activation timing characteristics of the new node or new edge are actively matched with the key fusion feature sequence in the dynamic spatiotemporal state fusion coding library, and the new node or new edge is assigned an active state prediction code based on the matching results.

4. The method for identifying and grading power equipment fault data conflicts based on deep learning according to claim 3, characterized in that: The S26 specifically includes: S261. Based on the data of the spatiotemporal state fusion matrix, collect the state transition paths of the spatiotemporal state fusion codes of nodes and edges within multiple consecutive fixed periods, and construct a state transition path sequence; S262, record and calculate the path length, state switching frequency, and path recurrence frequency of each state transition path sequence, and establish a state transition path feature set; S263. Determine the path importance of nodes and edges in the state transition path based on the state transition path feature set, and assign a path importance level label to each state transition path; S264, combining the path importance level mark with the corresponding state transition path sequence to construct and dynamically update the spatiotemporal state path importance fusion feature matrix; S265. Based on the spatiotemporal state path importance fusion feature matrix, identify state transition paths with high path importance and stable occurrence frequency, and determine them as key fusion feature path sequences; S266. Based on the key fusion feature path sequence, extract the fusion feature key nodes and key edges that characterize the abnormal change trend of the overall topological structure, and determine the high-confidence coding of the fusion feature key nodes and key edges; S267. Establish and dynamically update the determined high-confidence code into a dynamic spatiotemporal state fusion code library with path feature tags.

5. The method for identifying and grading power equipment fault data conflicts based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31. Based on the dynamic graph data structure after active state prediction coding, collect and record the absolute time point when the state prediction coding of each node and edge is first generated, and determine the initial propagation order of the state prediction coding of each node and edge one by one; S32. Define and assign initial propagation direction tags for node and edge state prediction codes based on the initial propagation order of the node and edge state prediction codes; S33, extracting the node activation order and edge connection order encoded on the forward propagation path based on the initial propagation direction mark encoded by the node and edge state prediction code, and obtaining a detailed topological sequence of the forward propagation path; S34, starting from the end node of the detailed topology sequence of the forward propagation path, gradually backtracking to extract the activation order and connection order of the nodes and edges on the reverse propagation path, to obtain the detailed topology sequence of the reverse propagation path; S35. Compare the detailed topological sequences of the forward and reverse propagation paths, and determine the nodes and edges that overlap between the two propagation path sequences, which are defined as key interaction overlapping nodes and edges; S36. For key interaction overlapping nodes and edges, calculate the frequency of occurrence, propagation distance, and path density on the forward and reverse propagation paths, and quantify the similarity of the forward and reverse propagation paths; S37. Based on the quantified similarity results of key interaction overlapping nodes and edges, the differential features of the forward and reverse propagation paths are integrated to construct and dynamically update the forward and reverse bidirectional dynamic graph feature vectors that simultaneously characterize the interaction trends of the forward and reverse features.

6. The method for identifying and grading power equipment fault data conflicts based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41, collecting the forward and reverse bidirectional dynamic graph feature vectors, dividing the feature dynamic monitoring time windows of multiple adaptive lengths according to the feature vector fluctuation rate and amplitude, and recording the start and end time of each time window and the feature vector change trend; S42, extracting nodes and edges dominated by feature vector change trends within each feature dynamic monitoring time window, and constructing feature vector trend evolution paths for each node and edge within each time window; S43, analyzing the eigenvector trend evolution path of each node and edge, calculating the trend dominance metric value of the path based on the eigenvalue change trajectory on the path, and quantifying and determining the node and edge trend path dominance level; S44, identifying nodes and edge pairs with the same trend path dominance level and overlapping topological connection positions within the adjacent feature dynamic monitoring time window, defining them as candidate comparison nodes and edge pairs across the time window, and assigning candidate comparison identifiers; S45. For each set of candidate comparison nodes and edge pairs across the time window, extract the temporal trajectory of the feature vector change path, and calculate the temporal path difference value between the candidate comparison nodes and edge pairs; S46. Compare the temporal path difference values ​​of candidate comparison nodes and edge pairs across the time window, and determine the node and edge pairs whose difference values ​​exceed the dynamically adjusted threshold as high-confidence dynamic graph comparison sample pairs; S47. Store and dynamically update high-confidence dynamic graph comparison sample pairs, and construct and maintain cross-time window graph comparison sample pairs including candidate comparison identifiers, trend path dominance levels, and time series path difference values.

7. The method for identifying and grading power equipment fault data conflicts based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51, obtaining a topological feature change pattern of each node and edge in a cross-time window graph comparison sample pair, and generating a topological feature evolution pattern trajectory of the node and edge in each time window according to the topological feature change pattern; S52, calculating the evolution trajectory distance between each node and edge topology feature evolution pattern trajectory one by one, and constructing the node and edge evolution trajectory difference matrix according to the calculation results of the evolution trajectory distance; S53, monitoring and obtaining forward and reverse bidirectional dynamic graph feature vectors, calculating a topological feature difference metric value in an evolution trajectory difference matrix for each feature vector, and determining abnormal nodes and edges whose topological feature difference metric values ​​exceed a preset conflict threshold; S54, for the identified abnormal nodes and edges, extract the characteristic evolution pattern path of the adjacent topological nodes and edges, and locate the starting node, path node and end node of the abnormal change of the topological characteristics; S55, tracking the path of the abnormal topological feature change, recording the path length, the amplitude of the topological feature change, and the duration of the abnormal topological feature change one by one, and defining them as a topological abnormal path event; S56. Calculate the comprehensive anomaly intensity of the topological anomaly path event, determine the significance level of the conflict feature of each topological anomaly path event based on the comprehensive anomaly intensity, and assign a conflict identifier; S57. Based on the conflict feature significance level and the conflict identifier, an association structure of abnormal nodes and edges of topological features is formed, and an initial identification result of the power equipment topology structure data conflict is generated and updated.

8. The method for identifying and grading power equipment fault data conflicts based on deep learning according to claim 1, characterized in that: The S6 specifically includes: S61. Obtain the location, conflict identifier, and significance level of each topological anomaly node and edge in the initial identification results, and dynamically divide the topological anomaly propagation impact range into three levels: device level, regional level, and global level based on the location and propagation characteristics of the anomaly nodes and edges; S62. Based on the impact range of device-level topology anomaly propagation, extract and record the starting node, abnormal topology path length, abnormal propagation speed between nodes, and continuous propagation duration of each device topology anomaly propagation one by one, and construct a device-level anomaly propagation path feature set; S63. Based on the impact range of regional topological anomaly propagation, identify the initial time difference, propagation path topological structure differences, and anomaly propagation synchronization degree of topological anomaly cooperative propagation among multiple devices in the region, and construct a regional anomaly propagation cooperative feature set; S64. Based on the impact range of global topological anomaly propagation, determine the difference in the starting position of the path of cross-regional topological anomaly propagation, the correlation strength between the cross-regional propagation time series difference and the cross-regional anomaly topological propagation, and construct a global anomaly propagation interaction feature set; S65. Calculate the topological anomaly propagation correlation between the device-level anomaly propagation path feature set, the regional-level anomaly propagation collaborative feature set, and the global-level anomaly propagation interaction feature set, and establish a correlation mapping relationship between the device, regional, and global levels of anomaly propagation features. S66. Generate device-level, regional-level, and global-level topology anomaly propagation patterns based on the correlation mapping relationship between the three-level anomaly propagation features, quantify the abnormal topology change intensity of each topology anomaly propagation pattern one by one, and assign a unique topology conflict pattern identifier; S67. Store and dynamically update the device-level, regional-level, and global-level topology anomaly propagation patterns corresponding to the topology conflict pattern identifiers, and build and continuously maintain a hierarchical topology conflict pattern library containing three-level topology conflict pattern features and associated mapping relationships.

9. The method for identifying and grading power equipment fault data conflicts based on deep learning according to claim 1, characterized in that: The S7 specifically includes: S71, obtaining the abnormal starting position, initial abnormal characteristic value and conflict occurrence time of the topological node and edge of the newly occurring data conflict event, and actively constructing the abnormal propagation topological path of the event in the power equipment topology network; S72, performing topological feature tag encoding on the abnormal propagation topological path to form a topological abnormality feature fingerprint sequence representing the uniqueness of the abnormal event, and recording the propagation rate, number of path topological nodes, and total path propagation time of each feature fingerprint; S73, calling and querying the hierarchical topology conflict pattern library, matching the topology anomaly feature fingerprint sequence with the device-level, regional-level, and global-level topology conflict pattern feature fingerprint sequences in the pattern library, and determining the matching similarity between the topology anomaly feature fingerprint sequence and the topology conflict patterns at each level; S74, analyzing the propagation path topology node deviation, propagation rate difference, and path topology node number difference between the topology anomaly feature fingerprint sequence and the third-level topology conflict pattern with the highest matching similarity, to form an abnormal propagation path topology deviation feature vector; S75. Calculate a comprehensive evaluation index value of the topological anomaly path difference based on the abnormal propagation path topological deviation characteristic vector; S76. Determine the topology conflict risk level corresponding to the new data conflict event based on the topology anomaly path difference comprehensive evaluation index value and a preset topology anomaly risk level adaptive classification threshold; S77. Output the power equipment fault conflict classification processing result including the abnormal starting position of the new data conflict event, the time when the conflict occurs, the matching three-level topology conflict pattern identifier and the topology conflict risk level.

Images