Power equipment fault diagnosis method and system based on knowledge graph

Abstract

The invention discloses a power equipment fault diagnosis method and system based on knowledge maps, and relates to the technical field of power equipment monitoring, and the method comprises the steps: constructing a first power equipment knowledge map and a second power equipment knowledge map deployed at edge equipment and a cloud processor, the node levels of the first power equipment knowledge map are less than the node levels of the second power equipment knowledge map; collecting equipment operation data of the target power area; the edge device performs preliminary diagnosis based on the first knowledge graph to obtain a first fault positioning result; and the cloud processor receives the result and performs deep diagnosis in combination with a second knowledge graph to obtain a second fault positioning result. And thus, the technical effects of balanced response efficiency and high-precision fault diagnosis are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment monitoring, and in particular to a knowledge graph-based power equipment fault diagnosis method and system. Background Art

[0002] Fault diagnosis for power equipment is crucial for ensuring reliable grid operation. Traditional fault diagnosis methods rely primarily on single-node knowledge processing, such as performing fault reasoning solely in the cloud or on edge devices. This often creates a trade-off between efficiency and accuracy in the complex and ever-changing power operating environment. With the increasing number of power equipment and the explosive growth in operational data, efficiently and accurately locating faults has become a pressing challenge for the power industry.

[0003] During the fault diagnosis process, if we rely entirely on edge devices for knowledge reasoning, although we can respond quickly, we are limited by the storage and computing capabilities of the devices, the knowledge graph is small in scale, and it is difficult to handle complex faults; if it is only processed by the cloud, although the knowledge graph is complete, the data transmission delay and network bandwidth limitations will cause the response time to be too long. Summary of the Invention

[0004] The present invention provides a method and system for diagnosing faults of electric power equipment based on knowledge graphs to solve the technical problem in the prior art that it is difficult to balance the efficiency and accuracy of fault diagnosis, and to achieve the technical effect of balanced response efficiency and high-precision fault diagnosis.

[0005] In a first aspect, the present invention provides a method for diagnosing faults of electric power equipment based on a knowledge graph, wherein the method for diagnosing faults of electric power equipment based on a knowledge graph comprises: Construct a first power equipment knowledge graph and a second power equipment knowledge graph for the target power area, wherein the first power equipment knowledge graph is deployed to an edge device and the second power equipment knowledge graph is deployed to a cloud processor, the edge device and the cloud processor include a data transmission channel, and the number of node levels of the first power equipment knowledge graph is less than the number of node levels of the second power equipment knowledge graph.

[0006] Collect real-time operation data sets of equipment in the target power area.

[0007] The edge device diagnoses the real-time operation data set of the device according to the first power equipment knowledge graph to obtain a first fault location result.

[0008] The cloud-based processor performs a diagnosis on the second power equipment knowledge graph based on the received first fault location result to obtain a second fault location result.

[0009] In a feasible implementation, the edge device diagnoses the real-time operation dataset of the device according to the first power equipment knowledge graph to obtain a first fault location result, including: The status field of each node in the first power equipment knowledge graph is updated according to the real-time operation data set of the equipment to detect abnormal nodes.

[0010] Identify the neighboring nodes of the abnormal node.

[0011] The abnormal nodes and the neighborhood-related nodes are traversed in the first power equipment knowledge graph, and the built-in lightweight fault diagnosis module in the edge device is called to perform a diagnosis to obtain a first fault location result, including candidate fault devices, candidate fault types, and confidence scores of the candidate fault devices.

[0012] In a feasible implementation, the cloud processor performs a diagnosis on the second power equipment knowledge graph based on the received first fault location result to obtain a second fault location result, including: The cloud-based processor locks a candidate fault node in the second power equipment knowledge graph based on the received first fault location result.

[0013] Obtain multi-level neighborhood association nodes of the candidate fault node, and construct a local knowledge graph based on the candidate fault node and the multi-level neighborhood association nodes.

[0014] The local knowledge graph is traversed in the second power equipment knowledge graph, and the deep fault diagnosis module built into the cloud processor is called to perform secondary diagnosis to obtain a second fault location result, including the faulty device, the fault type, and the confidence score of the faulty device.

[0015] In a feasible implementation, a lightweight fault diagnosis module is built into the edge device. The lightweight fault diagnosis module is obtained through training with a lightweight fault operation sample dataset of known lightweight fault events and lightweight fault location labels, including: Acquire historical operating data of the target power area.

[0016] Known lightweight fault events under the first power equipment knowledge graph are extracted from the historical operation data to construct a lightweight fault operation sample data set and a lightweight fault location label.

[0017] A KNN lightweight model structure is defined, the KNN lightweight model structure is trained using the lightweight fault operation sample data set and the lightweight fault location label, and the obtained lightweight fault diagnosis module is downloaded to the edge device.

[0018] In a feasible implementation, a deep fault diagnosis module is built into the cloud processor. The deep fault diagnosis module is trained and acquired through a deep fault operation sample dataset of known deep fault events and deep fault location labels, including: Acquire historical operating data of the target power area.

[0019] Known deep fault events under the second power equipment knowledge graph are extracted from the historical operation data to construct a deep fault operation sample data set and deep fault location labels.

[0020] A GNN deep model structure is defined, the GNN deep model structure is trained using the deep fault operation sample data set and the deep fault location label, and the obtained deep fault diagnosis module is downloaded to the cloud processor.

[0021] In a feasible implementation, the second power equipment knowledge graph is traversed within a preset hierarchical range to obtain multi-level neighborhood associated nodes of the candidate fault node.

[0022] The multi-level neighborhood association nodes include first-level neighboring nodes, second-level neighboring nodes, and third-level neighboring nodes.

[0023] In a feasible implementation, after obtaining the second fault location result, the method further includes: The second fault location result is sent to the edge device for storage to generate fault location feedback data.

[0024] Based on the fault location feedback data, the confidence of the lightweight fault diagnosis module of the first power equipment knowledge graph is optimized, and an updated lightweight fault diagnosis module is output. The edge device performs a diagnosis based on the updated lightweight fault diagnosis module.

[0025] In a feasible implementation, after obtaining the first fault localization result, if the number of candidate fault devices and candidate fault types output by the first fault localization result both return 1, it is determined whether the confidence score of the candidate fault device is greater than a preset confidence score threshold.

[0026] If the confidence score of the candidate faulty device is greater than a preset confidence score threshold, the first fault location result is output as a diagnosis result.

[0027] In a feasible implementation, there is an associated mapping relationship between the first power equipment knowledge graph and the second power equipment knowledge graph.

[0028] The association mapping relationship includes node semantic alignment mapping, topology path mapping and attribute synchronization mapping.

[0029] In a second aspect, the present invention further provides a power equipment fault diagnosis system based on a knowledge graph, wherein the power equipment fault diagnosis system based on a knowledge graph includes: A graph construction module is used to construct a first power equipment knowledge graph and a second power equipment knowledge graph for a target power area, wherein the first power equipment knowledge graph is deployed to an edge device, and the second power equipment knowledge graph is deployed to a cloud processor, the edge device and the cloud processor include a data transmission channel, and the number of node levels of the first power equipment knowledge graph is less than the number of node levels of the second power equipment knowledge graph.

[0030] The data set acquisition module is used to acquire real-time operation data sets of the equipment in the target power area.

[0031] An edge diagnosis module is used for the edge device to diagnose the real-time operation data set of the device according to the first power equipment knowledge graph to obtain a first fault location result.

[0032] The cloud-based diagnostic module is used for the cloud-based processor to perform a diagnosis in the second power equipment knowledge graph based on the received first fault location result to obtain a second fault location result.

[0033] The present invention discloses a method and system for diagnosing faults of power equipment based on a knowledge graph, comprising: in a target power area, respectively constructing an edge-side power equipment knowledge graph and a cloud-side power equipment knowledge graph, wherein the edge-side knowledge graph is deployed on an edge computing node, and the cloud-side knowledge graph is configured on a remote server, a data communication channel is established between the two, and the number of node levels of the edge-side knowledge graph is set to be less than the number of node levels of the cloud-side knowledge graph; collecting operation process data of equipment in the target power area; the edge computing node performs preliminary fault analysis on the collected operation data based on the locally deployed power equipment knowledge graph, and generates first-level fault location information; after receiving the preliminary location result, the cloud server conducts further diagnosis in combination with its more complex hierarchical structure of power equipment knowledge graph, and obtains a second-level fault location result. The method and system for diagnosing faults of power equipment based on a knowledge graph disclosed in the present invention solve the technical problem of the difficulty in balancing fault diagnosis efficiency and accuracy, and achieves the technical effect of balanced response efficiency and high-precision fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flow chart of the power equipment fault diagnosis method based on knowledge graph of the present invention.

[0035] Figure 2 This is a structural diagram of the power equipment fault diagnosis system based on knowledge graph of the present invention.

[0036] Explanation of the accompanying drawings: graph construction module 11, data set acquisition module 12, edge diagnosis module 13, cloud diagnosis module 14. DETAILED DESCRIPTION

[0037] The above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods of the specification to better understand the above technical solution. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments used only to explain the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, it should be noted that, for the convenience of description, only the parts related to the present invention, rather than all, are shown in the drawings.

[0038] Example 1, as Figure 1 The figure is a flow chart of a method for diagnosing faults of electric power equipment based on a knowledge graph according to the present invention, wherein the method for diagnosing faults of electric power equipment based on a knowledge graph comprises: S100: Construct a first power equipment knowledge graph and a second power equipment knowledge graph of the target power area, wherein the first power equipment knowledge graph is deployed to an edge device, and the second power equipment knowledge graph is deployed to a cloud processor, the edge device and the cloud processor include a data transmission channel, and the number of node levels of the first power equipment knowledge graph is less than the number of node levels of the second power equipment knowledge graph.

[0039] Specifically, two knowledge graphs with different complexities are constructed for the target power area to suit different deployment scenarios; among them, the first power equipment knowledge graph is a lightweight graph structure with power equipment as the core, which is deployed on edge devices. It mainly includes key equipment, main fault types and their basic correlation relationships. It has fewer node levels and is convenient for fast loading and real-time processing; the second power equipment knowledge graph is a complete knowledge graph deployed on a cloud processor. It contains richer equipment attributes, operating parameters, historical faults, maintenance records and other multi-level information, with more node levels, which is suitable for in-depth analysis and global optimization.

[0040] Specifically, edge devices refer to local computing units deployed at the power plant site in the target power area, such as smart substation terminals and distribution automation terminals, which possess certain data collection and preliminary processing capabilities. Cloud processors are high-performance computing platforms deployed in remote servers or data centers, supporting complex model execution and large-scale data analysis. Data transmission channels are communication links established between edge devices and cloud processors, used for bidirectional transmission of information such as device status, inference results, and graph updates.

[0041] Specifically, a knowledge graph of the target power area is constructed: first, the structural information, operating data, fault history and maintenance records of all power equipment in the target power area are collected, and entity extraction and relationship modeling are performed; then, based on the extracted entity and relationship modeling results, a second power equipment knowledge graph is constructed on the cloud processor. The graph contains a multi-level node structure, such as equipment-component-fault type-handling solution-historical case and other five-layer structures, which supports complex reasoning and in-depth analysis.

[0042] Furthermore, in order to meet the real-time response requirements on the edge side, key nodes and high-frequency relationships are extracted from the second graph to construct the first power equipment knowledge graph. For example, only the three-layer structure of equipment-fault type-processing suggestion is retained, or only the knowledge graph corresponding to the equipment associated with the edge device is retained as the first power equipment knowledge graph and deployed to the edge device.

[0043] Through this process, a two-layer knowledge graph system is constructed, leveraging the advantages of both edge and cloud computing. The knowledge graph on the edge device can quickly respond, promptly processing field data and performing preliminary fault location, effectively reducing data transmission latency. Meanwhile, the knowledge graph on the cloud leverages the cloud's powerful computing power and storage capabilities to conduct in-depth analysis of preliminary diagnostic results, providing more accurate fault location and diagnostic recommendations.

[0044] In some embodiments, there is an associative mapping relationship between the first power equipment knowledge graph and the second power equipment knowledge graph; the associative mapping relationship includes node semantic alignment mapping, topological path mapping and attribute synchronization mapping.

[0045] Optionally, a multi-dimensional mapping relationship including node semantic alignment mapping, topological path mapping and attribute synchronization mapping is constructed between the first power equipment knowledge graph and the second power equipment knowledge graph, which is used to ensure that the input, output and other data of the first power equipment knowledge graph transmitted through the data transmission channel can be smoothly transferred to the second power equipment knowledge graph, and at the same time ensure that the edge device and the cloud processor can achieve consistent correction of the fault node location when performing fault diagnosis.

[0046] Specifically, node semantic alignment mapping is the alignment of the semantic information of the knowledge graph nodes in the edge device and the cloud processor, which is used to ensure the consistency of the description of the same power equipment or fault mode; topology path mapping is the correspondence established between the knowledge graphs in the edge device and the cloud processor on the device connection relationship and the fault propagation path, so that accurate correspondence can be achieved when locating the fault; attribute synchronization mapping refers to the synchronization of the attribute information of the knowledge graph nodes in the edge device and the cloud processor, such as device parameters, fault thresholds, etc., to ensure the consistency of fault diagnosis.

[0047] By establishing a stable and precise knowledge mapping mechanism between edge devices and cloud processors, we can address the semantic drift, structural differences, and data inconsistencies associated with layered graph deployment. This not only helps improve the accuracy of subsequent collaboration between initial edge diagnostic results and in-depth cloud analysis, but also supports consistent correction of fault node location, ensuring intelligent diagnostic capabilities and reliability in a multi-tiered operating environment.

[0048] S200: Collecting a real-time operating data set of equipment in the target power area.

[0049] Specifically, the real-time equipment operation dataset is a collection of data collected in real time from various types of power equipment within the target power area, reflecting the current operating status of the equipment. This data includes multi-dimensional data such as voltage, current, power, frequency, temperature, vibration, oil level, gas concentration, and switch status. The specific data types or dimensions included are determined by the equipment type and monitoring requirements within the target power area.

[0050] Optionally, the real-time operation data set of the device is obtained by the edge device or monitoring terminal through locally deployed sensors and intelligent acquisition units (such as smart meters, online monitoring devices, and SCADA systems), and is preliminarily processed and stored in a structured or semi-structured form, thereby providing a data basis for subsequent knowledge graph reasoning and fault diagnosis.

[0051] For example, a variety of data acquisition devices are pre-deployed in the target power area, such as temperature sensors and online oil chromatographs installed on transformers, current transformers and switch status sensors on circuit breakers, and conductor temperature monitoring and icing detection devices on transmission lines. Data from all of these devices is aggregated through edge computing nodes. For example, for a transformer, its operating data set might include: winding temperature of 85°C, normal oil level, load current of 380A, gas content (H2) of 120ppm, and vibration frequency of 3Hz.

[0052] Optionally, edge devices perform preliminary cleaning of the collected raw data (e.g., removing outliers and filling missing values) and timestamp it, outputting a real-time data set representing the equipment's operation. This process enables real-time awareness of the operating status of power equipment, providing high-quality, dynamically updated data support for subsequent knowledge graph-driven diagnosis.

[0053] S300: The edge device diagnoses the real-time operation data set of the device according to the first power equipment knowledge graph to obtain a first fault location result.

[0054] Specifically, based on current real-time operating data (device real-time operating data sets), and with the support of the local knowledge graph, a preliminary diagnosis of possible faults is made through rule reasoning, path matching, or lightweight machine learning models. The first fault location result is the output of the above diagnosis, including the suspected faulty device and the corresponding fault type.

[0055] In some embodiments, the edge device diagnoses the real-time operation dataset of the device according to the first power equipment knowledge graph to obtain a first fault location result, including: The status fields of each node in the first power equipment knowledge graph are updated according to the real-time operation data set of the device to detect abnormal nodes; the neighborhood-related nodes of the abnormal nodes are identified; the abnormal nodes and the neighborhood-related nodes are traversed in the first power equipment knowledge graph, and the lightweight fault diagnosis module built into the edge device is called to perform a diagnosis to obtain a first fault location result, including candidate fault devices, candidate fault types and confidence scores of the candidate fault devices.

[0056] Specifically, first, the data content (such as numerical values) of the real-time operation data set of the equipment is mapped to the first power equipment knowledge graph, that is, the status field of each node in the first power equipment knowledge graph is updated as mentioned above; then the updated status field is compared with the preset normal range in the knowledge graph to detect nodes with abnormal status (exceeding the preset normal range); for example, when the oil temperature status field of a transformer node exceeds the preset normal range after being updated, the node is marked as an abnormal node.

[0057] Specifically, the abnormal node's neighboring nodes are identified. For example, nodes corresponding to equipment such as circuit breakers and busbars directly connected to the abnormal transformer are included in the analysis, as they may be affected by the fault or related to the fault's cause. The abnormal node and its neighboring nodes are then traversed within the first power equipment knowledge graph to construct a subgraph containing node states and attribute relationships. This subgraph then calls the lightweight fault diagnosis module built into the edge device to perform a diagnosis.

[0058] Specifically, the lightweight fault diagnosis module analyzes the involved nodes and their relationships (i.e., the aforementioned subgraph) based on embedded fault patterns and diagnostic rules, inferring possible faulty devices, fault types, and their confidence scores to generate a preliminary fault location result. For example, by analyzing the correlation between abnormal transformer oil temperature and circuit breaker current changes, a preliminary fault location result is obtained, including a candidate faulty device (e.g., transformer), a candidate fault type (e.g., cooling system failure), and a confidence score (e.g., 0.75).

[0059] Through the above process, edge devices can quickly complete preliminary fault diagnosis locally, promptly screening possible faulty devices and types, and quantifying the confidence level of the diagnostic results. The R&D department leverages the low latency of edge computing to enable immediate fault detection and initial location, reducing reliance on cloud processors and improving system response speed. Furthermore, the introduction of confidence scores facilitates further screening and correction of diagnostic results in subsequent steps.

[0060] In some implementations, a lightweight fault diagnosis module is built into the edge device. The lightweight fault diagnosis module is trained using a lightweight fault operation sample dataset of known lightweight fault events and lightweight fault location labels, including: Acquire historical operating data of the target power area; extract known lightweight fault events under the first power equipment knowledge graph from the historical operating data, and construct a lightweight fault operation sample data set and lightweight fault location labels; define a KNN lightweight model structure, use the lightweight fault operation sample data set and lightweight fault location labels to train the KNN lightweight model structure, and download the obtained lightweight fault diagnosis module to the edge device.

[0061] Specifically, the lightweight fault diagnosis module is an efficient diagnostic model designed specifically for edge devices, with the characteristics of low computing resource usage and fast response. Among them, lightweight fault events refer to fault types that have little impact on the operation of the system or equipment in the target power area, obvious symptoms, and relatively simple diagnostic logic, such as fan failure, capacitor aging, local overheating, etc. Correspondingly, the lightweight fault operation sample data set refers to multiple groups of data fragments related to lightweight faults extracted from historical data, which, for example, include sensor readings, equipment status, environmental parameters and other features. The lightweight fault location label is the known fault device and fault type corresponding to the sample data, which is used to supervise model training.

[0062] Specifically, we first extract samples related to known minor fault events from the historical operating data of the target power area. For example, based on a pre-defined minor fault table, we filter out events such as abnormal fan speed and excessive capacitor temperature rise from the substation's operating data over the past two years. We then extract the corresponding operating data (such as current, voltage, temperature, and speed) as features to construct a dataset of minor fault operating samples. Furthermore, we label these samples with corresponding minor fault location tags, such as fan failure and capacitor aging.

[0063] Next, an appropriate K value is selected to define the KNN model structure. The sample dataset and labels are used for supervised training and model performance evaluation. The operational data from the sample serves as the training input, while the labels provide supervision. The trained KNN model is capable of rapidly matching similarities and classifying faults for new input data, and can be output as a lightweight fault diagnosis module. Finally, this lightweight fault diagnosis module is downloaded and deployed to edge devices.

[0064] Through the above process, we achieve localized training and edge deployment of lightweight fault diagnosis models, helping to improve the autonomous diagnostic capabilities of edge devices. The KNN model is simple to train, fast to infer, and easy to update, making it suitable for deployment in edge environments with limited computing resources.

[0065] S400: The cloud-based processor performs a diagnosis on the second power equipment knowledge graph based on the received first fault location result to obtain a second fault location result.

[0066] Specifically, a cloud processor refers to high-performance computing resources deployed on a cloud platform or data center. The second fault location result is the final fault identification result output by the cloud processor after receiving the initial edge diagnosis results and performing in-depth analysis and verification based on the second knowledge graph. This result is typically more accurate and reliable.

[0067] In some embodiments, the cloud processor performs a diagnosis on the second power equipment knowledge graph based on the received first fault location result to obtain a second fault location result, including: The cloud-based processor locks a candidate fault node in the second power equipment knowledge graph based on the received first fault location result; obtains multi-level neighborhood association nodes of the candidate fault node, and constructs a local knowledge graph based on the candidate fault node and the multi-level neighborhood association nodes; traverses the local knowledge graph in the second power equipment knowledge graph, calls the deep fault diagnosis module built into the cloud-based processor for secondary diagnosis, and obtains a second fault location result, including the fault device, the fault type, and the confidence score of the fault device.

[0068] Specifically, a candidate fault node refers to the physical node included in the first fault location result uploaded by the edge device, representing a specific piece of power equipment (such as a transformer, circuit breaker, cooling fan, etc.). Multi-level neighborhood association nodes refer to other nodes in the knowledge graph that have a direct or indirect association with the candidate fault node, involving multiple levels of relationships such as physical connection, logical dependency, and functional coupling.

[0069] Specifically, the local knowledge graph is a subgraph structure centered on the candidate fault node and combined with its multi-level neighborhood nodes, which is used to focus on the fault propagation path and causal chain in a specific area.

[0070] Specifically, the deep fault diagnosis module is a high-performance model deployed in the cloud. It can be constructed using graph neural networks (GNNs), attention mechanisms, or causal reasoning models, and has the ability to model and reason about fault modes in complex topological structures.

[0071] Specifically, after receiving the first fault location result, the cloud processor first locates the corresponding candidate fault node in the second power equipment knowledge graph. Then, with this node as the center, it expands outward to obtain its multi-level neighboring nodes, such as the temperature control system node, transformer node, power supply line node, and protection device node connected to it, forming a structured local knowledge graph. This local knowledge graph contains the upstream and downstream dependencies of the candidate fault node in the power system and the possible fault propagation paths.

[0072] Furthermore, the cloud-based processor traverses the local knowledge graph to identify potential high-risk nodes and their associated paths, and invokes the deep fault diagnosis module for modeling and analysis. For example, a graph neural network is used to encode node characteristics (such as real-time current, voltage, temperature, and historical fault records) into the network structure. Graph convolution operations are then used to identify possible root cause nodes, ultimately outputting a second fault location result. This second fault location result includes the final faulty device, fault type, and confidence score.

[0073] Through this process, a structured, hierarchical fault diagnosis mechanism, powered by a knowledge graph, is implemented. By locking in candidate nodes and constructing local subgraphs, we can focus on key areas relevant to preliminary diagnosis, avoiding the computational overhead of traversing the entire graph. Furthermore, reasoning and analyzing these subgraphs using deep models improves the ability to identify complex fault mechanisms.

[0074] In some implementations, the second power equipment knowledge graph is traversed within a preset hierarchical range to obtain multi-level neighborhood association nodes of the candidate fault node; the multi-level neighborhood association nodes include first-level adjacent nodes, second-level adjacent nodes, and third-level adjacent nodes.

[0075] Specifically, the preset hierarchical range refers to the pre-set graph hierarchical depth limit during the knowledge graph traversal process to limit the search scope and control the computational complexity. It is used to define the maximum adjacency level that can be reached by expanding outward from the candidate fault node.

[0076] Optionally, the multi-level neighborhood association nodes of the candidate fault node include first-level adjacent nodes directly connected to the candidate fault node, second-level adjacent nodes directly connected to the first-level adjacent nodes, and third-level adjacent nodes directly connected to the second-level adjacent nodes.

[0077] Specifically, after receiving a candidate fault node, the cloud processor will traverse the second power equipment knowledge graph based on a preset hierarchy (e.g., three levels) starting from the candidate fault node. Optionally, the traversal process uses a breadth-first search (BFS) strategy, expanding outward layer by layer and recording the associated nodes at each layer.

[0078] For example, if the candidate node is cooling fan #5, the first-level adjacent nodes might be transformer #5 and the fan control module; the second-level adjacent nodes might include the transformer power input, temperature sensor, and protection relay; and the third-level adjacent nodes might involve the main transformer busbar and line protection devices. These three levels of nodes have different types of edge relationships with the candidate node, such as power supply, control, and thermal coupling, forming a multi-level causal chain. This multi-level neighborhood structure allows the construction of a local knowledge graph, providing a structural foundation for in-depth diagnosis.

[0079] In some implementations, a deep fault diagnosis module is built into the cloud processor, and the deep fault diagnosis module is trained using a deep fault operation sample dataset of known deep fault events and deep fault location labels, including: Acquire historical operating data of the target power area; extract known deep fault events under the second power equipment knowledge graph from the historical operating data, and construct a deep fault operation sample data set and deep fault location labels; define a GNN deep model structure, train the GNN deep model structure using the deep fault operation sample data set and deep fault location labels, and download the obtained deep fault diagnosis module to the cloud processor.

[0080] Specifically, a deep fault diagnosis module refers to a complex fault identification system deployed on the cloud or high-performance computing platform. It usually adopts models such as deep learning (such as CNN, LSTM, Transformer), and has the ability to process multi-dimensional, large-scale data and identify complex fault modes.

[0081] Specifically, the deep fault operation sample dataset is a collection of samples containing known deep fault events extracted from historical power system operation data. Each sample includes device state parameters (such as current, voltage, temperature, and frequency), time series characteristics, and device topology information. Deep fault location labels are the annotation information corresponding to these samples, used to clearly identify the ultimately confirmed faulty device and its fault type (such as winding short circuit, insulation breakdown, and control failure) corresponding to each sample.

[0082] Specifically, the same method and idea as the above-mentioned lightweight fault diagnosis module is adopted. First, historical operation data is retrieved from the target power area. This data is then analyzed to identify deep fault events that have occurred, such as short circuits between turns of the main transformer winding, partial discharges in the GIS, and circuit breaker refusal to operate. Subsequently, the corresponding operating parameters are extracted as input features to construct a deep fault operation sample dataset. At the same time, combined with operation and maintenance records or expert annotations, the actual fault equipment and type corresponding to each sample are used as labels to construct deep fault location labels. Subsequently, a GNN model adapted to the knowledge graph structure of power equipment is defined, such as the Graph Attention Network (GAT) structure, which introduces an attention mechanism to distinguish the influence weights of different adjacent nodes on fault propagation.

[0083] Furthermore, the GNN model is trained using the aforementioned deep fault operation sample dataset and deep fault location labels, enabling it to automatically identify fault sources from the graph structure. After training, the GNN model is exported and deployed to a cloud processor as a deep fault diagnosis module for subsequent online diagnostic calls. It should be noted that for the sake of brevity, the same methodological approach as used in the training and verification of the lightweight fault diagnosis module will not be further elaborated.

[0084] Through the above process, a deep fault diagnosis module with self-learning capabilities was constructed, capable of learning the fault propagation patterns and structural dependencies between devices from historical data. The construction and training of the GNN model can more effectively model nonlinear fault modes in complex topologies, helping to improve the accuracy and robustness of diagnosis.

[0085] In some embodiments, after obtaining the second fault location result, the method further includes: The second fault location result is sent to the edge device for storage to generate fault location feedback data; based on the fault location feedback data, the lightweight fault diagnosis module of the first power equipment knowledge graph is confidence optimized, and the updated lightweight fault diagnosis module is output, and the edge device performs a diagnosis based on the updated lightweight fault diagnosis module.

[0086] Specifically, fault location feedback data is the data record generated after the second fault location result is transmitted back to the edge device. It is used to guide the optimization and correction of edge-side models. The high-precision diagnostic results (i.e., the second fault location result) in the cloud can be used to correct the prediction confidence of the edge model, retrain it, or fine-tune its parameters, thereby improving the accuracy and reliability of the edge model.

[0087] Specifically, the second fault location result generated in the cloud is packaged into structured feedback data (e.g., in JSON format) and sent back to the edge device via the communication link between the edge node and the cloud. After receiving it, the edge device stores it locally as historical diagnostic feedback for subsequent model optimization.

[0088] For example, if the initial diagnosis result of the edge device is that device A is suspected of overheating (confidence 0.6), and the second fault location result returned by the cloud is that device B has insulation breakdown (confidence 0.95), then the deviation is recorded and included in the optimized sample library of the lightweight fault diagnosis module.

[0089] Furthermore, the edge device uses this feedback data as new sample data to optimize the confidence of the local lightweight model. Optimization methods may include retraining model parameters, updating sample weights, adjusting threshold strategies, etc. For example, an incremental learning method is used to incorporate cloud results as pseudo-labels into the training set, and fine-tune the edge model to make it closer to the cloud results in subsequent diagnoses.

[0090] Through the above process, the lightweight fault diagnosis module can be optimized and updated to achieve continuous adaptive upgrades of the model; the edge device uses the updated lightweight fault diagnosis module to perform a diagnosis on the newly connected data to achieve higher diagnostic accuracy.

[0091] In some embodiments, after obtaining the first fault location result, if the number of candidate fault devices and candidate fault types output by the first fault location result both return 1, it is determined whether the confidence score of the candidate fault device is greater than a preset confidence score threshold; if the confidence score of the candidate fault device is greater than the preset confidence score threshold, the first fault location result is output as a diagnosis result.

[0092] Specifically, if the first fault location result has been defined for a candidate fault device, the fault type of the candidate fault device has been determined, and the confidence score of the candidate fault device meets the preset requirements, then it can be considered that accurate and complete fault location information has been obtained through the first fault location result. Furthermore, the first fault location result can be directly output as a diagnostic result without further transmitting the first fault location result to the cloud for redundant repeated judgment and positioning.

[0093] Specifically, the confidence score is a quantitative indicator used to measure the credibility of a diagnosis. It can be derived from the diagnostic algorithm's internal probability calculations or from historical data accuracy statistics. The preset confidence score threshold is a pre-set standard value used to determine the reliability of a diagnosis. When the confidence score exceeds this threshold, the diagnosis is considered to have a high degree of credibility.

[0094] The above process avoids repeated judgments on simple, intuitive fault location tasks, helping to improve the overall response efficiency and effectiveness of fault location and diagnosis, and avoiding the waste of resources caused by inefficient repeated judgments. Furthermore, confidence scoring ensures that only reliable results are directly adopted, avoiding incorrect decisions caused by the uncertainty of preliminary diagnostic results.

[0095] In summary, the knowledge graph-based power equipment fault diagnosis method provided by the present invention has the following technical effects: By constructing an edge-side power equipment knowledge graph and a cloud-side power equipment knowledge graph in the target power area, respectively, wherein the edge-side knowledge graph is deployed on the edge computing node and the cloud-side knowledge graph is configured on the remote server side, a data communication channel is established between the two, and the number of node levels of the edge-side knowledge graph is set to be less than the number of node levels of the cloud-side knowledge graph; the operation process data of the equipment in the target power area are collected; the edge computing node performs preliminary fault analysis on the collected operation data based on the locally deployed power equipment knowledge graph, and generates first-level fault location information; after receiving the preliminary location result, the cloud server combines it with the power equipment knowledge graph with a more complex hierarchical structure to carry out further diagnosis and obtain the second-level fault location result, thereby achieving the technical effect of balanced response efficiency and high-precision fault diagnosis.

[0096] Example 2, as Figure 2 This is a schematic diagram of the structure of the power equipment fault diagnosis system based on the knowledge graph of the present invention. For example, Figure 1 The flow chart of the power equipment fault diagnosis method based on knowledge graph in the present invention can be shown as follows: Figure 2 The structure shown is implemented.

[0097] Based on the same concept as the power equipment fault diagnosis method based on the knowledge graph in the embodiment, the present invention also provides a power equipment fault diagnosis system based on the knowledge graph, including: The graph construction module 11 is used to construct a first power equipment knowledge graph and a second power equipment knowledge graph for a target power area, wherein the first power equipment knowledge graph is deployed to an edge device, and the second power equipment knowledge graph is deployed to a cloud processor, the edge device and the cloud processor include a data transmission channel, and the number of node levels of the first power equipment knowledge graph is less than the number of node levels of the second power equipment knowledge graph.

[0098] The data set collection module 12 is used to collect real-time operation data sets of the equipment in the target power area.

[0099] The edge diagnosis module 13 is used for the edge device to diagnose the real-time operation data set of the device according to the first power equipment knowledge graph to obtain a first fault location result.

[0100] The cloud diagnosis module 14 is used for the cloud processor to perform a diagnosis in the second power equipment knowledge graph according to the received first fault location result to obtain a second fault location result.

[0101] In some embodiments, the edge diagnostic module 13 includes: The node status update and anomaly detection unit is used to update the status field of each node in the first power equipment knowledge graph according to the real-time operation data set of the equipment and detect abnormal nodes.

[0102] The neighborhood associated node identification unit is used to identify the neighborhood associated nodes of the abnormal node.

[0103] A primary diagnosis and fault location result acquisition unit is used to traverse the abnormal node and the neighborhood-related nodes in the first power equipment knowledge graph, call the lightweight fault diagnosis module built into the edge device to perform a primary diagnosis, and obtain a first fault location result, including candidate fault devices, candidate fault types, and confidence scores of the candidate fault devices.

[0104] In some embodiments, the cloud diagnostic module 14 includes: A candidate fault node locking unit is used for the cloud-based processor to lock a candidate fault node in the second power equipment knowledge graph according to the received first fault location result.

[0105] The local knowledge graph construction unit is used to obtain the multi-level neighborhood association nodes of the candidate fault node and construct a local knowledge graph based on the candidate fault node and the multi-level neighborhood association nodes.

[0106] The secondary diagnosis and fault location result acquisition unit is used to traverse the local knowledge graph in the second power equipment knowledge graph, call the deep fault diagnosis module built into the cloud processor to perform secondary diagnosis, and obtain the second fault location result, including the faulty device, the fault type, and the confidence score of the faulty device.

[0107] In some embodiments, the lightweight fault diagnosis module is obtained by training a lightweight fault operation sample dataset of known lightweight fault events and lightweight fault location labels, and the execution steps include: Acquire historical operating data of the target power area; extract known lightweight fault events under the first power equipment knowledge graph from the historical operating data, and construct a lightweight fault operation sample data set and lightweight fault location labels; define a KNN lightweight model structure, use the lightweight fault operation sample data set and lightweight fault location labels to train the KNN lightweight model structure, and download the obtained lightweight fault diagnosis module to the edge device.

[0108] In some embodiments, the deep fault diagnosis module is trained and acquired through a deep fault operation sample dataset of known deep fault events and deep fault location labels, and the execution steps include: Acquire historical operating data of the target power area; extract known deep fault events under the second power equipment knowledge graph from the historical operating data, and construct a deep fault operation sample data set and deep fault location labels; define a GNN deep model structure, train the GNN deep model structure using the deep fault operation sample data set and deep fault location labels, and download the obtained deep fault diagnosis module to the cloud processor.

[0109] In some embodiments, the local knowledge graph construction unit also includes an association analysis and association extraction sub-unit, which is used to traverse the second power equipment knowledge graph within a preset hierarchical range to obtain multi-level neighborhood association nodes of the candidate fault node; the multi-level neighborhood association nodes include first-level adjacent nodes, second-level adjacent nodes and third-level adjacent nodes.

[0110] In some embodiments, the cloud diagnostic module 14 further includes: The fault location feedback data generating unit is configured to send the second fault location result to the edge device for storage, and generate fault location feedback data.

[0111] A confidence optimization unit is used to perform confidence optimization on the lightweight fault diagnosis module of the first power equipment knowledge graph based on the fault location feedback data, output an updated lightweight fault diagnosis module, and the edge device performs a diagnosis based on the updated lightweight fault diagnosis module.

[0112] In some embodiments, the edge diagnosis module 13 also includes a judgment output unit, which is used to: if the number of candidate fault devices and candidate fault types output by the first fault location result are both returned as 1, determine whether the confidence score of the candidate fault device is greater than a preset confidence score threshold; if the confidence score of the candidate fault device is greater than the preset confidence score threshold, output the first fault location result as a diagnosis result.

[0113] In some implementations, there is an associative mapping relationship between the first power equipment knowledge graph and the second power equipment knowledge graph; the associative mapping relationship includes node semantic alignment mapping, topological path mapping and attribute synchronization mapping.

[0114] It should be understood that the embodiments mentioned in this specification focus on their differences from other embodiments. The specific embodiments in the aforementioned embodiment one are also applicable to the knowledge graph-based power equipment fault diagnosis system described in embodiment two. For the sake of brevity of the specification, they will not be further elaborated here.

[0115] It should be understood that the embodiments disclosed in the present invention and the above description can enable those skilled in the art to use the present invention to implement the present invention. At the same time, the present invention is not limited to the embodiments mentioned above. It should be understood that those skilled in the art can still modify the technical solutions described in the above embodiments or replace some of the technical features therein with equivalents; and such modifications or replacements do not deviate from the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention and are all included in the scope of protection of the present invention.

Claims

1. A power equipment fault diagnosis method based on knowledge graph, characterized in that: include: Constructing a first power equipment knowledge graph and a second power equipment knowledge graph for a target power area, wherein the first power equipment knowledge graph is deployed to an edge device, and the second power equipment knowledge graph is deployed to a cloud processor, the edge device and the cloud processor include a data transmission channel, and the number of node levels in the first power equipment knowledge graph is less than the number of node levels in the second power equipment knowledge graph; Collecting real-time operation data sets of equipment in the target power area; The edge device diagnoses the real-time operation data set of the device according to the first power equipment knowledge graph to obtain a first fault location result; The cloud-based processor performs a diagnosis on the second power equipment knowledge graph based on the received first fault location result to obtain a second fault location result.

2. The method according to claim 1, wherein The edge device diagnoses the real-time operation dataset of the device according to the first power equipment knowledge graph to obtain a first fault location result, including: Update the status field of each node in the first power equipment knowledge graph according to the real-time operation data set of the equipment to detect abnormal nodes; Identifying neighboring nodes of the abnormal node; The abnormal nodes and the neighborhood-related nodes are traversed in the first power equipment knowledge graph, and the built-in lightweight fault diagnosis module in the edge device is called to perform a diagnosis to obtain a first fault location result, including candidate fault devices, candidate fault types, and confidence scores of the candidate fault devices.

3. The method according to claim 2, wherein The cloud-based processor performs a diagnosis on the second power equipment knowledge graph based on the received first fault location result to obtain a second fault location result, including: The cloud-based processor locks a candidate fault node in the second power equipment knowledge graph according to the received first fault location result; Acquire multi-level neighborhood associated nodes of the candidate fault node, and construct a local knowledge graph based on the candidate fault node and the multi-level neighborhood associated nodes; The local knowledge graph is traversed in the second power equipment knowledge graph, and the deep fault diagnosis module built into the cloud processor is called to perform secondary diagnosis to obtain a second fault location result, including the faulty device, the fault type, and the confidence score of the faulty device.

4. The method according to claim 2, wherein A lightweight fault diagnosis module is built into the edge device. The lightweight fault diagnosis module is obtained by training a lightweight fault operation sample dataset of known lightweight fault events and lightweight fault location labels, including: Acquiring historical operating data of the target power area; Extracting known lightweight fault events under the first power equipment knowledge graph from the historical operation data, and constructing a lightweight fault operation sample data set and a lightweight fault location label; A KNN lightweight model structure is defined, the KNN lightweight model structure is trained using the lightweight fault operation sample data set and the lightweight fault location label, and the obtained lightweight fault diagnosis module is downloaded to the edge device.

5. The method according to claim 3, wherein A deep fault diagnosis module is built into the cloud processor. The deep fault diagnosis module is obtained through deep fault operation sample data sets and deep fault location label training of known deep fault events, including: Acquiring historical operating data of the target power area; Extracting known deep fault events in the second power equipment knowledge graph from the historical operation data, and constructing a deep fault operation sample dataset and a deep fault location label; A GNN deep model structure is defined, the GNN deep model structure is trained using the deep fault operation sample data set and the deep fault location label, and the obtained deep fault diagnosis module is downloaded to the cloud processor.

6. The method according to claim 3, wherein Traversing the second electric power equipment knowledge graph within a preset hierarchical range to obtain multi-level neighborhood associated nodes of the candidate fault node; The multi-level neighborhood association nodes include first-level neighboring nodes, second-level neighboring nodes, and third-level neighboring nodes.

7. The method according to claim 1, wherein After obtaining the second fault location result, the method further includes: Sending the second fault location result to the edge device for storage to generate fault location feedback data; Based on the fault location feedback data, the confidence of the lightweight fault diagnosis module of the first power equipment knowledge graph is optimized, and an updated lightweight fault diagnosis module is output. The edge device performs a diagnosis based on the updated lightweight fault diagnosis module.

8. The method according to claim 2, wherein After obtaining the first fault location result, if the number of candidate fault devices and candidate fault types outputted by the first fault location result both return 1, determining whether the confidence score of the candidate fault device is greater than a preset confidence score threshold; If the confidence score of the candidate faulty device is greater than a preset confidence score threshold, the first fault location result is output as a diagnosis result.

9. The method according to claim 1, wherein There is an associated mapping relationship between the first power equipment knowledge graph and the second power equipment knowledge graph; The association mapping relationship includes node semantic alignment mapping, topology path mapping and attribute synchronization mapping.

10. The power equipment fault diagnosis system based on knowledge graph is characterized by: A method for diagnosing faults of electric power equipment based on a knowledge graph, for implementing any one of claims 1 to 9, comprising: A graph construction module, configured to construct a first electric power equipment knowledge graph and a second electric power equipment knowledge graph for a target electric power area, wherein the first electric power equipment knowledge graph is deployed to an edge device, and the second electric power equipment knowledge graph is deployed to a cloud processor, the edge device and the cloud processor include a data transmission channel, and the number of node levels in the first electric power equipment knowledge graph is smaller than the number of node levels in the second electric power equipment knowledge graph; A data set acquisition module, configured to acquire real-time operation data sets of equipment in the target power area; An edge diagnosis module is configured to enable the edge device to diagnose the real-time operation data set of the device according to the first power equipment knowledge graph to obtain a first fault location result; The cloud-based diagnostic module is used for the cloud-based processor to perform a diagnosis in the second power equipment knowledge graph based on the received first fault location result to obtain a second fault location result.

Images