A cross-domain collaborative analysis system and method for power equipment faults

Abstract

The present invention is a cross-domain collaborative analysis system and method for power equipment faults, which relates to the field of power grid dispatching technology, including obtaining pre-processed multi-source heterogeneous data of power equipment, constructing a cross-domain knowledge graph based on the topological relationship of the pre-processed data and historical fault data, marking the fault propagation path, and using a graph neural network for embedding representation. Designing a spatiotemporal multi-branch network uses the spatiotemporal multi-branch network to extract spatial, temporal and modal interaction features respectively, and fuse them in the feature fusion layer to obtain fused features and branch weights. Combined with the knowledge graph embedding representation, a collaborative reasoning model is constructed using a Bayesian network, and reasoning decisions are made on the fused features, ultimately obtaining cross-domain collaborative analysis results of power equipment faults. Through the combination of knowledge graph and spatiotemporal multi-branch network, the fusion and efficient reasoning of multi-source heterogeneous data are realized, thereby improving the accuracy and efficiency of fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid management, and in particular to a cross-domain collaborative analysis system and method for power equipment faults. Background Art

[0002] With the development of society, the demand for electricity in various industries is constantly increasing, and the requirements for the stability and security of power grid power supply are also gradually increasing. Power equipment is an important part of the power grid system, and the safe and stable operation of power equipment is an important factor in ensuring power supply reliability.

[0003] During the operation of power equipment, fault diagnosis usually relies on data from a single area or a single type. For example, the existing published Chinese invention patents, application number CN202311396061.X, are a power equipment defect prediction method based on the gray wolf optimization algorithm and LSTM-Attention, and application number CN202411283917.7, are a cloud-edge collaborative smart grid fault processing method, device, equipment and storage medium. The methods involved, such as the gray wolf optimization algorithm, all focus on a single device or local time series characteristics, and cannot effectively process multi-source heterogeneous data features. They lack the ability to dynamically model cross-regional fault propagation paths, and it is difficult to handle coupled fault analysis of multi-region and multi-type equipment in the power grid, affecting the accuracy and timeliness of fault diagnosis and early warning. Summary of the Invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the main purpose of the present invention is to provide a cross-domain collaborative analysis system and method for power equipment faults, which can achieve rapid fault location and root cause analysis through multi-source data fusion, spatiotemporal alignment algorithm and cross-domain knowledge sharing, thereby improving diagnostic accuracy and efficiency.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a cross-domain collaborative analysis method for power equipment faults, comprising the following steps:

[0006] Acquire cross-domain multi-source heterogeneous data of power equipment, perform preprocessing, and obtain preprocessed cross-domain multi-source heterogeneous data;

[0007] Based on the topological relationship of the preprocessed cross-domain multi-source heterogeneous data and historical fault data, a cross-domain knowledge graph is constructed, fault propagation paths are marked, and embedded representations of graph neural networks are supported to obtain a knowledge graph and a knowledge graph embedded representation;

[0008] Designing a spatiotemporal multi-branch network, the spatiotemporal multi-branch network including a parallel spatial topology branch, a temporal evolution branch, and a cross-modal association branch, and a feature fusion layer;

[0009] The spatiotemporal multi-branch network is used to extract the spatial feature matrix, the temporal feature matrix, and the modal interaction feature matrix from the preprocessed cross-domain multi-source heterogeneous data through the parallel spatial topology branch, the time evolution branch, and the cross-modal association branch, and feature fusion is performed in the feature fusion layer to obtain fused features and branch weights;

[0010] Based on the knowledge graph and Bayesian network decision, a collaborative reasoning model is constructed. Combined with the knowledge graph embedding representation, the collaborative reasoning model is used to make reasoning decisions on the fusion features to obtain cross-domain collaborative analysis results of power equipment faults.

[0011] The cross-domain multi-source heterogeneous data includes device attributes, operation data, environmental data, image data and topological relationship data;

[0012] The device attributes of the power equipment include the device type and real-time status label; the operating data of the power equipment include current data, voltage data and temperature data; the environmental data include meteorological data and geographical data of the power equipment; and the image data include infrared thermal imaging data and partial discharge maps of the power equipment;

[0013] The cross-domain collaborative analysis results of the power equipment fault include the fault type, fault location and fault cause of the power equipment.

[0014] The spatial topology branch adopts a heterogeneous graph convolutional network;

[0015] The time evolution branch includes multi-scale temporal coding and dynamic graph learning;

[0016] The cross-modal association branch includes modality alignment and constructing a modality relationship graph;

[0017] The feature fusion layer includes spatiotemporal registration alignment and dynamic gated fusion.

[0018] The extraction of the spatial feature matrix comprises:

[0019] Determine a device topology diagram based on the topology relationship data, determine a cause-effect diagram based on the fault propagation path, and determine an environment coupling diagram based on the pre-processed environment data in the cross-domain multi-source heterogeneous data;

[0020] generating a set of heterogeneous adjacency matrices according to the device topology graph, the causal graph, and the environment coupling graph;

[0021] Inputting the heterogeneous adjacency matrix set into the heterogeneous graph convolutional network to obtain a spatial feature matrix, wherein the spatial correlation strength and causal influence of each power device are encoded;

[0022] Extracting the time series feature matrix includes:

[0023] Convolving the preprocessed cross-domain multi-source heterogeneous data using multiple parallel convolution kernels, processing the convolution results using a four-layer encoder and a multi-head attention mechanism to obtain time series features; mapping the time series features into vectors to obtain a time series feature matrix;

[0024] Extracting the modal interaction feature matrix includes:

[0025] The operating data is mapped to the geographic grid obtained by geographic hash coding according to the power equipment, and then spliced ​​with the operating data and environmental data in the unified spatiotemporal grid according to the channel. The fusion feature is obtained by using cross-modal attention fusion;

[0026] ResNet-18 is used to extract infrared image features, and the dimensions are reduced by principal component analysis and associated with the fusion features to obtain alignment features;

[0027] Extracting running nodes, environment nodes, and image nodes based on running features, environment features, and image features, obtaining the Pearson correlation coefficient of the fused features as statistical associations, learning the alignment features as learned associations through an attention mechanism, obtaining edge weights based on the statistical associations and learned relationships, constructing a modal relationship graph based on the running nodes, environment nodes, and image nodes and the edge weights, and obtaining the modal relationship graph and modal interaction feature matrix through a graph convolutional network;

[0028] Performing feature fusion at the feature fusion layer includes:

[0029] Using a sliding time window and geo-hashing algorithm, the pre-processed cross-domain multi-source heterogeneous data is mapped to a unified spatiotemporal grid for feature alignment to obtain spatiotemporal alignment features.

[0030] The spatial feature matrix, the temporal feature matrix and the cross-modal feature matrix are spatiotemporally registered and aligned to obtain aligned features, branch weights are allocated through a gating mechanism, and branch features are fused according to the branch weights to obtain fused features and branch weights.

[0031] The constructing of the collaborative reasoning model includes:

[0032] Mapping nodes in the knowledge graph to nodes in a Bayesian network;

[0033] Mapping the edges in the knowledge graph into directed edges in a Bayesian network;

[0034] Based on the propagation rules of historical fault data and knowledge graphs, a conditional probability table between nodes is set to obtain a Bayesian network structure.

[0035] The Bayesian network structure is combined with the embedding representation of the knowledge graph to construct a collaborative reasoning model.

[0036] The method of using the collaborative reasoning model to reason and make decisions on the fusion features includes the following steps:

[0037] The feature vectors of each spatiotemporal grid are used as graph nodes, and the physical connections and spatiotemporal adjacencies of devices are used as graph edges. A graph structure is constructed, and a multi-layer graph convolutional network is used for message passing. The final layer uses Sigmoid activation to obtain a node-level fault probability matrix.

[0038] The upstream fault in the knowledge graph is used as the parent node, and the real-time power equipment status is used as the child node to construct a conditional probability table. The conditional probability table is updated based on the cross-domain heterogeneous data of the real-time power equipment to obtain the posterior probability matrix.

[0039] According to the posterior probability matrix, candidate probabilistic fault paths are extracted, the cosine similarity between the candidate probabilistic fault paths and the potential rules of the knowledge graph is obtained, the fault paths with a similarity greater than the set threshold are retained, and the cross-domain collaborative analysis results of power equipment faults are obtained by combining the knowledge graph embedding representation.

[0040] A cross-domain collaborative analysis system for power equipment faults, comprising:

[0041] A cross-domain device data processing module is used to obtain cross-domain multi-source heterogeneous data of power equipment, perform preprocessing, and obtain preprocessed cross-domain multi-source heterogeneous data; based on the topological relationship of the preprocessed cross-domain multi-source heterogeneous data and historical fault data, a cross-domain knowledge graph is constructed, fault propagation paths are marked, and embedded representations of graph neural networks are supported to obtain a knowledge graph and a knowledge graph embedded representation;

[0042] A feature data analysis and processing module is used to design a spatiotemporal multi-branch network, which includes parallel spatial topology branches, time evolution branches, and cross-modal association branches, as well as a feature fusion layer; using the spatiotemporal multi-branch network to extract a spatial feature matrix, a temporal feature matrix, and a modal interaction feature matrix from the preprocessed cross-domain multi-source heterogeneous data through the parallel spatial topology branches, time evolution branches, and cross-modal association branches, and perform feature fusion in the feature fusion layer to obtain fused features and branch weights;

[0043] The collaborative reasoning module is used to build a collaborative reasoning model based on the knowledge graph and Bayesian network decision, combine the knowledge graph embedding representation, use the collaborative reasoning model to make reasoning decisions on the fusion features, and obtain cross-domain collaborative analysis results of power equipment faults.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention utilizes a combination of graph neural networks and spatiotemporal multi-branch networks, and combines them with Bayesian networks for decision reasoning, which can achieve accurate fault diagnosis and cross-domain collaborative analysis of power equipment. By integrating multi-source heterogeneous data, the spatiotemporal multi-branch network can accurately capture the correlation between different types of data, significantly improving the accuracy and efficiency of fault diagnosis, and enhancing the fault diagnosis capability of power equipment. In addition, by combining the spatiotemporal multi-branch network with the embedded representation of the graph neural network, the model can process complex spatiotemporal dynamics and cross-modal information, thereby showing strong adaptability and robustness when facing different scenarios and environments. In order to improve the cross-domain data fusion effect, the deep fusion of cross-domain multi-source heterogeneous data effectively enhances the complementarity of different types of data, thereby achieving more accurate power equipment status monitoring and fault analysis. In addition, in order to optimize the integration and reasoning of cross-domain information, a collaborative reasoning model is constructed, and the information in the knowledge graph is combined with the real-time data of power equipment through the reasoning capability of the Bayesian network, which can make more accurate predictions and decisions under uncertain conditions. By discovering potential fault paths in advance and predicting faults early, the risk of power equipment failure can be significantly reduced and the reliability and operating efficiency of the equipment can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings described herein are used to provide further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0046] Figure 1 It is the process framework of the present invention;

[0047] Figure 2 It is a schematic flow chart of the present invention;

[0048] Figure 3 It is a schematic diagram of the process structure of the spatiotemporal multi-branch network of the present invention;

[0049] Figure 4 This is a schematic diagram of the association between nodes in Example 2 of the present invention;

[0050] Figure 5 This is a schematic diagram of the fault propagation logic chain in Example 2 of the present invention. DETAILED DESCRIPTION

[0051] During the operation of power equipment, fault diagnosis usually relies on data from a single area or a single type, focusing on a single device or local time series characteristics. It is unable to effectively process the characteristics of multi-source heterogeneous data, lacks the ability to dynamically model cross-regional fault propagation paths, and is difficult to handle coupled fault analysis of multi-region and multi-type equipment in the power grid, affecting the accuracy and timeliness of fault diagnosis and early warning.

[0052] In order to overcome the shortcomings of the above-mentioned prior art, the main purpose of the present invention is to provide a cross-domain collaborative analysis system and method for power equipment faults, which can integrate multi-source heterogeneous data, realize cross-domain collaborative reasoning, and realize root cause analysis and rapid location of cross-regional faults through multi-modal data spatiotemporal alignment and causal-driven dynamic evolution modeling, and further carry out immediate active prevention and control. Figures 1 to 3 , specifically including the following steps:

[0053] Acquire cross-domain multi-source heterogeneous data of power equipment, perform preprocessing, and obtain preprocessed cross-domain multi-source heterogeneous data;

[0054] Based on the topological relationship of pre-processed cross-domain multi-source heterogeneous data and historical fault data, a cross-domain knowledge graph is constructed, fault propagation paths are marked, and embedded representations of graph neural networks are supported to obtain knowledge graphs and knowledge graph embedding representations.

[0055] Design a spatiotemporal multi-branch network, which includes parallel spatial topology branches, temporal evolution branches, cross-modal association branches, and a feature fusion layer;

[0056] Use spatiotemporal multi-branch networks to process pre-processed cross-domain multi-source heterogeneous data:

[0057] The spatial feature matrix is ​​extracted through the spatial topology branch: the equipment topology map is determined based on the topological relationship data, the causal graph is determined based on the fault propagation path, and the environmental coupling map is determined based on the environmental features in the preprocessed cross-domain multi-source heterogeneous data; a set of heterogeneous adjacency matrices is generated based on the equipment topology map, the causal graph, and the environmental coupling map; the set of heterogeneous adjacency matrices is input into the heterogeneous graph convolutional network to obtain the spatial feature matrix, in which the spatial correlation strength and causal influence of each power device are encoded.

[0058] Extract the time series feature matrix through the time evolution branch: convolve the preprocessed cross-domain multi-source heterogeneous data using three different scales of convolution kernels, and use a four-layer encoder and a multi-head attention mechanism to obtain time series features; map the time series features into vectors to obtain the time series feature matrix.

[0059] The modal interaction feature matrix is ​​extracted through cross-modal association branches: the operating data is mapped to the geographic grid obtained by geographic hash coding according to the power equipment, and then spliced ​​with the operating data and environmental data in the unified spatiotemporal grid by channel, and fused using cross-modal attention to obtain fusion features; ResNet-18 is used to extract infrared image features, and the dimension is reduced by principal component analysis and associated with the fusion features to obtain alignment features; operating nodes, environmental nodes, and image nodes are extracted based on the operating features, environmental features, and image features, and the Pearson correlation coefficient of the fusion features is obtained as the statistical association. The alignment features are learned as the learning association through the attention mechanism. The edge weights are obtained based on the statistical association and the learning association. The modal relationship graph is constructed based on the operating nodes, environmental nodes, image nodes and edge weights. The modal relationship graph and the modal interaction feature matrix are obtained through the graph convolutional network.

[0060] Feature fusion is performed at the feature fusion layer: a sliding time window and geo-hashing algorithm are used to map the pre-processed cross-domain multi-source heterogeneous data to a unified spatiotemporal grid for feature alignment. The spatial feature matrix, temporal feature matrix, and cross-modal feature matrix are spatiotemporally registered and aligned to obtain aligned features. Branch weights are assigned through a gating mechanism, and branch features are fused according to the branch weights to obtain fused features and branch weights.

[0061] Based on the knowledge graph and Bayesian network decision, a collaborative reasoning model is constructed. Combined with the knowledge graph embedding representation, the collaborative reasoning model is used to make reasoning decisions on the fusion features to obtain cross-domain collaborative analysis results of power equipment faults.

[0062] Cross-domain multi-source heterogeneous data, including equipment attributes, operating data, environmental data, image data and topological relationship data; equipment attributes of power equipment include equipment type and real-time status label, operating data of power equipment includes current data, voltage data and temperature data, environmental data includes meteorological data and geographic data of power equipment, image data includes infrared thermal imaging data and partial discharge maps of power equipment; cross-domain collaborative analysis results of power equipment faults include fault type, fault location and fault cause of power equipment.

[0063] Build a collaborative reasoning model, including:

[0064] Map the device nodes, state nodes, and edges between them in the knowledge graph into nodes and directed edges in the Bayesian network;

[0065] Based on the fault propagation path and historical fault data in the knowledge graph, the causal relationship between nodes is defined, and the conditional probability table between nodes is set to obtain the Bayesian network structure;

[0066] The Bayesian network structure is combined with the embedding representation of the knowledge graph to construct a collaborative reasoning model.

[0067] The collaborative reasoning model is used to reason and make decisions based on the fused features, which includes the following steps:

[0068] The feature vectors of each spatiotemporal grid are used as graph nodes, and the physical connections and spatiotemporal adjacencies of devices are used as graph edges. A graph structure is constructed, and a multi-layer graph convolutional network is used for message passing. The final layer uses Sigmoid activation to obtain a node-level fault probability matrix.

[0069] The device status nodes in the knowledge graph are used as nodes in the Bayesian network. The conditional probability table between nodes in the Bayesian network is updated based on the cross-domain heterogeneous data of real-time power equipment to obtain the posterior probability matrix.

[0070] The candidate probabilistic fault paths are extracted according to the posterior probability matrix, and the cosine similarity between the candidate probabilistic fault paths and the fault propagation paths in the knowledge graph is calculated. The fault paths with a similarity greater than the set similarity threshold are retained. Combined with the knowledge graph embedding representation, the cross-domain collaborative analysis results of power equipment faults are obtained.

[0071] Example 1

[0072] At a coastal substation (Geohash code: ws3x9y), during a thunderstorm, transformer TR-001 experienced insulation breakdown, resulting in abnormal current harmonics on the associated line Line-005. Using the method presented in this paper, cross-domain collaborative analysis enabled rapid fault location and generated a response strategy.

[0073] This embodiment collects the following five types of data for 110kV substations:

[0074] Equipment attribute data includes the type and real-time status label (normal operation / warning / fault) of equipment such as the main transformer (SFSZ11-31500 / 110) and circuit breaker (LW25-126).

[0075] The operating data include current (fundamental frequency 2.3kA, harmonic content) and voltage (126kV) data collected at a frequency of 4kHz by the PMU device, and pre-processed using wavelet packet denoising and FFT harmonic decomposition.

[0076] Environmental data include temperature (32.5°C), humidity (78%), and lightning location (distance 8.7 km) data collected by a micrometeorological station at a frequency of 1 Hz, and Kalman filtering and geographic interpolation were used for spatiotemporal alignment.

[0077] The image data includes the surface temperature field of the device (such as the local hot spot of 85°C) collected by the infrared camera at 30fps, and the 512x512 image is compressed into a 128-dimensional feature vector through PCA dimensionality reduction.

[0078] Topological relationship data includes the device connection relationship (such as busbar-circuit breaker-line) provided by the SCADA system, and uses geographic hash coding for spatial alignment.

[0079] Based on the above data, we construct a cross-domain knowledge graph that includes device entities, fault types, and propagation paths. The key steps are as follows:

[0080] The first step is entity extraction, where equipment, fault types, and environmental factors are extracted as graph nodes. Equipment includes main transformers and circuit breakers; fault types include lightning strikes and overheating; and environmental factors include lightning and humidity.

[0081] Next, define the relationship. Define the device connection and select busbar-circuit breaker; for fault propagation, select lightning strike for insulation breakdown; for environmental impact, select increased humidity and increased flicker probability as the relationship edge.

[0082] Then, knowledge fusion is performed, injecting historical fault records and the arrester damage caused by a lightning strike in July 2022 into the graph as prior knowledge. A graph neural network (TransE) is used to map the graph nodes into a 128-dimensional vector space to support subsequent calculations.

[0083] Regarding the spatiotemporal multi-branch network design, in the present invention, a spatiotemporal multi-branch network with three branches is designed, specifically including:

[0084] The spatial topology branch builds a graph structure based on the physical connection relationship of devices and uses a two-layer graph convolutional network to extract the spatial feature matrix.

[0085] In the time evolution branch, an LSTM network is used to perform time series modeling on the historical 72-hour operation data and extract the time series feature matrix.

[0086] The cross-modal association branch fuses infrared images and operational data through an attention mechanism to extract the modal interaction feature matrix.

[0087] The feature fusion layer uses a gated attention mechanism and combines branch weights, specifically 0.35 for space, 0.42 for time, and 0.23 for modality, to generate fused features.

[0088] Subsequently, the knowledge graph is combined with the Bayesian network to construct a collaborative reasoning model, in which graph nodes, such as "main transformer status", are mapped to Bayesian network nodes, and edges, such as "lightning strike → main transformer failure", are mapped to directed edges.

[0089] Based on historical fault data, such as the main transformer failure probability of 0.67 during lightning strike, a joint conditional probability table is established.

[0090] The fused features output by the spatiotemporal multi-branch network are used as evidence, combined with the knowledge graph embedding representation, and reasoning is performed in the Bayesian network to calculate the posterior probability of equipment failure.

[0091] The graph convolutional network with Sigmoid activation outputs the failure probability matrix of each device, where the main transformer failure probability is 0.92.

[0092] According to real-time data, such as the current lightning intensity in this embodiment, the Bayesian network conditional probability table is updated.

[0093] The cosine similarity between the fault probability and the potential rules of the graph is calculated. In this embodiment, the threshold is set to 0.75, and the fault paths with a similarity greater than 0.75 are selected to screen out the fault paths with high matching degree.

[0094] Combined with the knowledge graph embedding representation, the output fault location is "#2 main transformer high-voltage side bushing", the fault type is "lightning strike + oil overheating", and maintenance suggestions are given.

[0095] This solution was deployed at a 110kV substation in a certain region of East China. The monitoring period was the thunderstorm season from June to August 2023. The results are as follows:

[0096] The accuracy of correlated diagnosis for power equipment increased from 78% to 94.7%, while the false alarm rate decreased by 68%. Cross-domain data utilization increased from 62% to 89%. Single fault diagnosis time was shortened from 2.3 hours to 21 minutes. The user complaint rate decreased by 78%, and the mean time-to-market (MTBF) of key equipment increased to 32,500 hours.

[0097] A cross-domain collaborative analysis method for power equipment faults, through the deep integration of knowledge graphs and spatiotemporal multi-branch networks, enables accurate diagnosis of complex faults such as lightning strikes. This method has been proven effective in actual power grids and provides important technical support for the safe operation of smart grids.

[0098] Example 2

[0099] This example simulates the process of diagnosing a fault on the main transformer of a 500kV smart substation at the dispatching and control center of a provincial power grid company, utilizing cross-domain, multi-source, heterogeneous data, combined with a knowledge graph and a spatiotemporal multi-branch network. The scenario is set during a typhoon, with the main transformer experiencing abnormal operating conditions.

[0100] In order to fully characterize the status of the main transformer and its associated environment, the following data needs to be collected from multiple systems: power equipment attribute data, power equipment operation data, environmental data that affects power equipment operation, image data representing the power equipment status, and the location and connection of the power equipment in the power grid. The details are shown in Table 1:

[0101] Table 1 Acquired multi-source heterogeneous data

[0102]

[0103] The acquired multi-source heterogeneous data was cleaned, converted, and feature extracted to obtain preprocessed data suitable for subsequent analysis. This process included signal denoising, image processing, and geocoding. Specifically, the highly fluctuating Phase A current data was denoised using the db4 wavelet basis to eliminate high-frequency noise. The infrared thermal images were then scaled to a standard size of 224×224 pixels, and a pretrained ResNet-18 model was used to extract 512-dimensional feature vectors to capture key patterns in the image. Finally, the geohash algorithm was used to map the GPS coordinates (30.12°N, 120.45°E) to the string wx4g0, facilitating spatial correlation analysis.

[0104] Based on the pre-processed data and historical fault records, a knowledge graph is constructed to reflect the relationship between devices, environmental coupling, and fault propagation. The nodes defined in this embodiment mainly include the main transformer, 500kV busbar, circuit breaker QF12, line L23, typhoon weather, and historical fault events such as 2023-07 bushing overheating. It is also necessary to define the association between nodes, such as Figure 4 As shown, it includes "connection", "control", "environmental coupling", "historical faults", etc.

[0105] Combined with domain knowledge, possible fault propagation logic chains are marked in the graph, such as Figure 5 As shown, in this example, the sequence is: high humidity during a typhoon → damp bushing → partial discharge → increased oil temperature → insulation degradation. In this example, the graph neural network model uses GraphSAGE and other tools to map nodes in the knowledge graph, such as the main transformer, into low-dimensional dense vectors (128 dimensions), such as the main transformer node → [0.12, -0.45, ..., 0.67]. These embedding vectors capture the semantic information and relational features of the nodes, providing support for subsequent graph neural network modeling.

[0106] A spatiotemporal multi-branch network is designed to extract spatiotemporal multimodal features from different perspectives and enhance the representation capability through dynamic fusion. It includes spatial topology branch, temporal evolution branch, and time evolution branch, as well as fusion part. The details are shown in Table 2:

[0107] Table 2 Spatiotemporal multi-branch network design table

[0108]

[0109] The features extracted from the spatiotemporal multi-branch network are combined with knowledge graphs and Bayesian reasoning for fault diagnosis.

[0110] Bayesian network construction involves mapping nodes and obtaining a conditional probability table (CPT). Node mapping involves mapping key nodes in the knowledge graph, such as the main transformer, humidity, and discharge, to Bayesian network nodes. The conditional probability table is based on historical fault data training or expert experience. For example, P(fault | humidity > 80%, discharge > 200pC) = 0.92. This is shown in Table 3:

[0111] Table 3 Humidity and discharge failure probability

[0112]

[0113] The specific fault reasoning process of the power equipment in this embodiment is as follows:

[0114] The final layer of the spatiotemporal multi-branch network, in this example, uses graph convolution, full connectivity, and sigmoid functions. This layer outputs a probability matrix for a main transformer failure, assuming a preliminary probability of 0.89. A Bayesian posterior update is then obtained by inputting real-time observed data (humidity 85%, discharge 210 pC) into the Bayesian network to update the conditional probabilities. Based on the humidity and discharge failure probability table, the posterior probability increases to 0.94.

[0115] Based on the candidate fault causes derived from Bayesian reasoning, the corresponding propagation path, "High humidity causes discharge," was searched in the knowledge graph. The cosine similarity between this candidate path and the typical fault rule pre-defined in the knowledge graph, "Humidity + discharge causes overheating," was calculated to be 0.87. Fault paths with a similarity greater than a threshold of 0.75 were retained. Combining the graph neural network's embedded representation, capturing device state and relationship information with the Bayesian reasoning results, the fault diagnosis conclusion for the power equipment was ultimately determined, as shown in Table 4.

[0116] Table 4 Fault diagnosis conclusion

[0117]

[0118] In this example, actual maintenance confirmed the diagnostic results, revealing aging of the main transformer phase A bushing seal and moisture-damaged insulation, consistent with the analysis. Furthermore, the system took only 3.68 seconds from data collection completion to generating the final diagnostic report, meeting real-time monitoring requirements.

[0119] Traditional alarm methods based on a single threshold, such as monitoring only the oil temperature, failed to provide a timely warning of this fault because the oil temperature, although elevated, did not completely exceed the absolute threshold. This method, however, successfully captured the precursors to the fault through multi-source data fusion and correlation analysis.

[0120] This example fully demonstrates how to apply a cross-domain collaborative analysis approach to a specific 500kV substation main transformer failure scenario. By integrating multi-source heterogeneous data, constructing a knowledge graph, extracting features using a spatiotemporal multi-branch network, and incorporating Bayesian reasoning, this approach achieves accurate and rapid diagnosis of complex faults, validating the feasibility and superiority of this approach in actual power system operations and maintenance.

[0121] It should be noted that, in the present invention, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises", or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article, or apparatus.

[0122] The above embodiments are merely examples of the present invention and do not limit the scope of protection of the present invention. Any designs that are identical or similar to the present invention fall within the scope of protection of the present invention.

Claims

1. A cross-domain collaborative analysis method for power equipment faults, characterized in that: The following steps are involved: Acquire cross-domain multi-source heterogeneous data of power equipment, perform preprocessing, and obtain preprocessed cross-domain multi-source heterogeneous data; Based on the topological relationship of the preprocessed cross-domain multi-source heterogeneous data and historical fault data, a cross-domain knowledge graph is constructed, fault propagation paths are marked, and embedded representations of graph neural networks are supported to obtain a knowledge graph and a knowledge graph embedded representation; Designing a spatiotemporal multi-branch network, the spatiotemporal multi-branch network including a parallel spatial topology branch, a temporal evolution branch, and a cross-modal association branch, and a feature fusion layer; The spatiotemporal multi-branch network is used to extract the spatial feature matrix, the temporal feature matrix, and the modal interaction feature matrix from the preprocessed cross-domain multi-source heterogeneous data through the parallel spatial topology branch, the time evolution branch, and the cross-modal association branch, and feature fusion is performed at the feature fusion layer to obtain fusion features and branch weights; wherein, feature fusion is performed at the feature fusion layer, including: Using a sliding time window and geo-hashing algorithm, the pre-processed cross-domain multi-source heterogeneous data is mapped to a unified spatiotemporal grid for feature alignment to obtain spatiotemporal alignment features. The spatial feature matrix, the temporal feature matrix, and the cross-modal feature matrix are spatiotemporally registered and aligned to obtain aligned features. Branch weights are assigned through a gating mechanism, and the concatenated vector of each branch feature is input. The original weights are generated through a two-layer perceptron. Softmax normalization is used to obtain branch weights, and branch features are weighted and fused according to the branch weights to obtain fused features and branch weights. Based on the knowledge graph and Bayesian network decision, a collaborative reasoning model is constructed, wherein the collaborative reasoning model is constructed including: Mapping nodes in the knowledge graph to nodes in a Bayesian network; Mapping the edges in the knowledge graph into directed edges in a Bayesian network; Based on historical fault data and the propagation rules in the knowledge graph, multiple propagation paths in the knowledge graph are mapped into the chain structure of the Bayesian network, and a joint conditional probability table is set to obtain the Bayesian network structure; Combining the Bayesian network structure with the embedded representation of the knowledge graph to build a collaborative reasoning model; Combined with the knowledge graph embedding representation, the collaborative reasoning model is used to make reasoning decisions on the fusion features to obtain cross-domain collaborative analysis results of power equipment faults.

2. The cross-domain collaborative analysis method for power equipment faults according to claim 1, characterized in that: The cross-domain multi-source heterogeneous data includes device attributes, operation data, environmental data, image data and topological relationship data; The device attributes of the power equipment include the device type and real-time status label; the operating data of the power equipment include current data, voltage data and temperature data; the environmental data include meteorological data and geographical data of the power equipment; and the image data include infrared thermal imaging data and partial discharge maps of the power equipment; The cross-domain collaborative analysis results of the power equipment fault include the fault type, fault location and fault cause of the power equipment.

3. The cross-domain collaborative analysis method for power equipment faults according to claim 2, characterized in that: The spatial topology branch adopts a heterogeneous graph convolutional network; The time evolution branch includes multi-scale temporal coding and dynamic graph learning; The cross-modal association branch includes modality alignment and constructing a modality relationship graph; The feature fusion layer includes spatiotemporal registration alignment and dynamic gated fusion.

4. The cross-domain collaborative analysis method for power equipment faults according to claim 3, characterized in that: The extraction of the spatial feature matrix comprises: Determine a device topology diagram based on the topology relationship data, determine a cause-effect diagram based on the fault propagation path, and determine an environment coupling diagram based on the pre-processed environment data in the cross-domain multi-source heterogeneous data; generating a set of heterogeneous adjacency matrices according to the device topology graph, the causal graph, and the environment coupling graph; Inputting the heterogeneous adjacency matrix set into the heterogeneous graph convolutional network to obtain a spatial feature matrix; Extracting the time series feature matrix includes: The preprocessed cross-domain multi-source heterogeneous data is convolved using multiple parallel convolution kernels, and the convolution results are processed using a four-layer encoder and a multi-head attention mechanism to obtain time series features; the time series features are mapped into embedding vectors, and a time-varying adjacency matrix is ​​calculated to obtain a time series feature matrix; Extracting the modal interaction feature matrix includes: The operating data is mapped to the geographic grid obtained by geographic hash coding according to the power equipment, and then spliced ​​with the operating data and environmental data in the unified spatiotemporal grid according to the channel. The fusion feature is obtained by using cross-modal attention fusion; ResNet-18 is used to extract infrared image features, and the dimensions are reduced by principal component analysis and associated with the fusion features to obtain alignment features; According to the running features, environmental features and image features, the running nodes, environmental nodes and image nodes are extracted, the Pearson correlation coefficient of the fusion features is obtained as the statistical association, the alignment features are learned as the learning association through the attention mechanism, and the edge weights are obtained according to the statistical association and the learning relationship. According to the running nodes, environmental nodes and image nodes and the edge weights, a modal relationship graph is constructed, and the modal relationship graph and the modal interaction feature matrix are obtained through the graph convolutional network.

5. The cross-domain collaborative analysis method for power equipment faults according to claim 1, characterized in that: The method of using the collaborative reasoning model to make reasoning decisions on the fusion features includes the following steps: The feature vectors of each spatiotemporal grid are used as graph nodes, and the physical connections and spatiotemporal adjacencies of devices are used as graph edges. A graph structure is constructed, and a multi-layer graph convolutional network is used for message passing. The final layer uses Sigmoid activation to obtain a fault probability matrix. The upstream fault in the knowledge graph is used as the parent node, and the real-time power equipment status is used as the child node to construct a conditional probability table. The conditional probability table is updated based on the cross-domain heterogeneous data of the real-time power equipment to obtain the posterior probability matrix. According to the posterior probability matrix, candidate probabilistic fault paths are extracted, the cosine similarity between the candidate probabilistic fault paths and the potential rules of the knowledge graph is obtained, the fault paths with a similarity greater than the set threshold are retained, and the cross-domain collaborative analysis results of power equipment faults are obtained by combining the knowledge graph embedding representation.

6. A system using the cross-domain collaborative analysis method for power equipment faults according to any one of claims 1 to 5, characterized in that: include: A cross-domain device data processing module is used to obtain cross-domain multi-source heterogeneous data of power equipment, perform preprocessing, and obtain preprocessed cross-domain multi-source heterogeneous data; Based on the topological relationship of the preprocessed cross-domain multi-source heterogeneous data and historical fault data, a cross-domain knowledge graph is constructed, fault propagation paths are marked, and embedded representations of graph neural networks are supported to obtain a knowledge graph and a knowledge graph embedded representation; A feature data analysis and processing module is used to design a spatiotemporal multi-branch network, which includes parallel spatial topology branches, time evolution branches, and cross-modal association branches, as well as a feature fusion layer; the spatiotemporal multi-branch network is used to extract spatial feature matrices, temporal feature matrices, and modal interaction feature matrices from the pre-processed cross-domain multi-source heterogeneous data through parallel spatial topology branches, time evolution branches, and cross-modal association branches, and feature fusion is performed in the feature fusion layer to obtain fusion features and branch weights; wherein, feature fusion is performed in the feature fusion layer, including using a sliding time window and a geographic hashing algorithm to map the pre-processed cross-domain multi-source heterogeneous data to a unified spatiotemporal grid for feature alignment to obtain spatiotemporal alignment features; the spatial feature matrix, the temporal feature matrix, and the cross-modal feature matrix are spatiotemporally registered and aligned to obtain aligned features, branch weights are allocated through a gating mechanism, the feature splicing vectors of each branch are input, and the original weights are generated through a two-layer perceptron; Softmax normalization is used to obtain branch weights, and branch features are weighted and fused according to the branch weights to obtain fusion features and branch weights; The collaborative reasoning module is used to build a collaborative reasoning model based on the knowledge graph and Bayesian network decision. The construction of the collaborative reasoning model includes mapping the nodes in the knowledge graph to the nodes in the Bayesian network; mapping the edges in the knowledge graph to the directed edges in the Bayesian network; based on the historical fault data and the propagation rules in the knowledge graph, mapping the multiple propagation paths in the knowledge graph to the chain structure of the Bayesian network, and setting a joint conditional probability table to obtain the Bayesian network structure; combining the Bayesian network structure with the embedded representation of the knowledge graph to build a collaborative reasoning model; combining the knowledge graph embedded representation, using the collaborative reasoning model to make inference decisions on the fusion features to obtain cross-domain collaborative analysis results of power equipment faults.

7. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

8. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 5 is implemented.

Images