Ultra-high-voltage transformer fault diagnosis method and system based on knowledge graph
By using a knowledge graph-based approach, a mapping table between fault types and feature quantities is established. The fault probability is calculated using the BERT embedding model and graph convolutional neural network, which solves the problems of accuracy and efficiency in fault diagnosis of UHV transformers and realizes intelligent operation and maintenance management.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- NORTH CHINA BRANCH OF STATE GRID CORPORATION OF CHINA
- Filing Date
- 2025-12-05
- Publication Date
- 2026-06-11
Smart Images
Figure CN2025140335_11062026_PF_FP_ABST
Abstract
Description
A Knowledge Graph-Based Fault Diagnosis Method and System for UHV Transformers
[0001] This application claims priority to Chinese Patent Application No. 202411774214.4, filed with the Chinese Patent Office on December 5, 2024, the entire contents of which are incorporated herein by reference. Technical Field
[0002] This application relates to a method and system for assessing the condition of power equipment, such as a method and system for diagnosing faults in ultra-high voltage transformers. Background Technology
[0003] Ultra-high voltage (UHV) transformers are among the most expensive pieces of equipment in power systems, responsible for voltage transformation and power transmission. Their operational status directly affects the safe and stable operation of the power system; a failure can lead to enormous economic losses. Due to various reasons such as poor manufacturing quality during transformer design and insulation aging, failures are frequent and complex. Furthermore, the fault characteristics are not obvious, making it difficult to accurately determine the type and location of the fault. The large capacity of UHV transformers also means that failures cause greater damage. Therefore, conducting condition assessment and fault diagnosis research on UHV transformers is essential.
[0004] For transformer fault diagnosis, maintenance personnel typically use dissolved gas analysis (DGA) to diagnose potential latent faults. The types, contents, components, and proportions of dissolved gases in the oil vary significantly depending on the type and stage of the fault. Therefore, the thermal decomposition nature of the oil around the fault point can be reflected by the dominant and subordinate changes in the composition and content of various combustible gases dissolved in the oil, thereby revealing the fault mode and severity.
[0005] With the continuous growth of the scale of ultra-high voltage equipment, many problems have been exposed in the operation and maintenance mode of related technologies. For example, the regular maintenance mode has significant problems of over-maintenance, low defect detection rate during power outage maintenance, and persistent safety risks from frequent power outages, resulting in low maintenance efficiency.
[0006] Therefore, it is desirable to provide a new method for diagnosing faults in ultra-high voltage equipment to solve the above problems. Summary of the Invention
[0007] One of the objectives of this application is to provide a knowledge graph-based fault diagnosis method for ultra-high voltage transformers. This method calculates the conditional probability of each feature quantity for each fault mode of the equipment, thereby achieving more robust, reliable, and accurate fault diagnosis of power transformers and assisting power equipment operation and maintenance personnel in their operation and maintenance decisions.
[0008] In accordance with the aforementioned purpose, this application proposes a knowledge graph-based fault diagnosis method for ultra-high voltage transformers, comprising the following steps:
[0009] 100: Establish a mapping table of fault types, corresponding state variables, and characteristic variables for UHV transformers;
[0010] 200: Construct a knowledge graph based on the mapping table;
[0011] 300: The node information in the knowledge graph is converted into node embedding vectors, and the structure of the knowledge graph is modeled using the adjacency matrix;
[0012] 400: Based on the adjacency matrix and the node embedding vector, extract the fault node feature vector and the feature quantity node feature vector of the knowledge graph;
[0013] 500: Based on the fault node feature vector and the feature quantity node feature vector, calculate the conditional probability of the fault mode corresponding to each feature quantity occurring, and based on the conditional probability of the fault mode corresponding to each feature quantity occurring, obtain the total probability of the fault mode occurring.
[0014] Furthermore, in step 300 of the ultra-high voltage transformer fault diagnosis method described in this application, the node information in the knowledge graph is converted into node embedding vectors using the BERT embedding model.
[0015] Furthermore, in step 400 of the UHV transformer fault diagnosis method described in this application, the adjacency matrix and the node are embedded into a graph convolutional neural network, and the graph convolutional neural network is used to extract the fault node feature vector and the feature quantity node feature vector of the knowledge graph.
[0016] Furthermore, in step 500 of the UHV transformer fault diagnosis method described in this application, the fault node feature vector and the feature quantity node feature vector are concatenated and input into a fully connected neural network to calculate the conditional probability of the fault mode corresponding to each feature quantity.
[0017] Furthermore, in step 500 of the ultra-high voltage transformer fault diagnosis method described in this application, the conditional probability of a fault mode occurring for each characteristic quantity is calculated based on the following formula:
[0018] ;
[0019] in, Represents the j-th feature quantity When an anomaly occurs, the i-th fault mode The conditional probability of a failure occurring. This represents the weight matrix of a fully connected neural network. This represents the bias of a fully connected neural network. This indicates the Sigmoid activation function.
[0020] Accordingly, another objective of this application is to provide a knowledge graph-based fault diagnosis system for ultra-high voltage transformers, which can achieve more robust, reliable, and accurate fault diagnosis of power transformers by calculating the conditional probability of each feature quantity for each fault mode of the equipment.
[0021] In accordance with the aforementioned purpose, this application proposes a knowledge graph-based fault diagnosis system for ultra-high voltage transformers, comprising:
[0022] The data module stores a table showing the mapping relationship between the fault types of UHV transformers, the corresponding state variables, and the characteristic variables.
[0023] The knowledge graph module constructs a knowledge graph based on the mapping table.
[0024] The conversion module converts the node information in the knowledge graph into node embedding vectors and uses the adjacency matrix to model the structure of the knowledge graph.
[0025] The feature extraction module extracts the fault node feature vector and the feature quantity node feature vector of the knowledge graph based on the adjacency matrix and the node embedding vector.
[0026] The judgment module calculates the conditional probability of a fault mode occurring for each feature quantity based on the feature vector of the fault node and the feature vector of the feature quantity node, and obtains the total probability of a fault mode occurring based on the conditional probability of a fault mode occurring for each feature quantity.
[0027] Furthermore, in the UHV transformer fault diagnosis system described in this application, the conversion module uses the BERT embedding model to convert the node information in the knowledge graph into node embedding vectors.
[0028] Furthermore, in the UHV transformer fault diagnosis system described in this application, the feature extraction module embeds the adjacency matrix and the nodes into a vector input graph convolutional neural network, and uses the graph convolutional neural network to extract the fault node feature vector and feature quantity node feature vector of the knowledge graph.
[0029] Furthermore, in the UHV transformer fault diagnosis system described in this application, the judgment module concatenates the fault node feature vector and the feature quantity node feature vector and inputs them into a fully connected neural network to calculate the conditional probability of the fault mode corresponding to each feature quantity occurring.
[0030] Furthermore, in the UHV transformer fault diagnosis system described in this application, the judgment module calculates the conditional probability of a fault mode occurring for each characteristic quantity based on the following formula:
[0031] ;
[0032] in, Represents the j-th feature quantity When an anomaly occurs, the i-th fault mode The conditional probability of a failure occurring. This represents the weight matrix of a fully connected neural network. This represents the bias of a fully connected neural network. This indicates the Sigmoid activation function.
[0033] The knowledge graph-based fault diagnosis method and system for ultra-high voltage transformers described in this application have the following advantages and beneficial effects:
[0034] The method and system described in this application can calculate the conditional probability of each characteristic quantity for each fault mode of the equipment, thereby achieving more robust, reliable and accurate fault diagnosis of power transformers.
[0035] The methods and systems described in this application can provide support for the intelligent operation, maintenance, and production management of equipment in the new smart power grid operation and maintenance system. They are the foundation for realizing intelligent equipment, intelligent operation and maintenance, intelligent maintenance, and intelligent production management, improving equipment status control and operation and maintenance management penetration, and realizing data-driven innovation and efficiency improvement in operation and maintenance business. They have important engineering application prospects. Attached Figure Description
[0036] Figure 1 schematically illustrates the steps of the ultra-high voltage transformer fault diagnosis method described in this application in one embodiment;
[0037] Figure 2 schematically shows the knowledge graph constructed in one embodiment of the UHV transformer fault diagnosis method described in this application;
[0038] Figure 3 schematically illustrates some steps of the ultra-high voltage transformer fault diagnosis method described in this application under one embodiment;
[0039] Figure 4 schematically shows the system architecture of the UHV transformer fault diagnosis system described in this application in one embodiment. Detailed Implementation
[0040] The following will further explain and illustrate the knowledge graph-based fault diagnosis method and system for ultra-high voltage transformers with reference to the accompanying drawings and specific embodiments. However, this explanation and illustration do not constitute an undue limitation on the technical solution of this application.
[0041] Figure 1 schematically illustrates the steps of one implementation of the knowledge graph-based fault diagnosis method for ultra-high voltage transformers described in this application.
[0042] In some embodiments of this application, as shown in FIG1, the knowledge graph-based fault diagnosis method for ultra-high voltage transformers proposed in this application may include the following steps:
[0043] 100: Establish a mapping table of fault types, corresponding state variables, and characteristic variables of UHV transformers.
[0044] In some implementations, the fault types of ultra-high voltage transformers can be structurally classified into three categories: large, medium, and small. In a more specific example, the large category can be classified according to equipment components, such as 13 types including winding and lead wire faults and insulating oil performance faults; the medium category can be further refined from these large categories, such as including 57 types including current-conducting circuit faults and insulation faults; the small category can be further refined, such as including 93 types including overheating caused by poor welding or crimping of conductors and leads, and moisture absorption of solid insulation.
[0045] In addition, based on the specific fault type, statistical analysis is performed on the state variables involved, such as the content of various gases, winding frequency response, and winding DC resistance.
[0046] In addition, based on the state quantity, the characteristic quantities involved are statistically analyzed. For example, the state quantity of total hydrocarbon content involves two characteristic quantities: the total hydrocarbon content value and the relative weekly growth rate.
[0047] Table 1 provides an exemplary partial view of the mapping relationship between fault types, corresponding state variables, and characteristic variables of an ultra-high voltage transformer in a specific instance.
[0048] Table 1.
[0049]
[0050]
[0051] 200: Construct a knowledge graph based on the mapping table.
[0052] In some more specific implementations, as shown in Figure 2, the content in the mapping table can be converted into a knowledge graph. Specifically, the various fault types, state variables, and feature variables in the mapping table are transformed into nodes in the knowledge graph; the table headers "Major Category" and "Medium Category" are then transformed into edges connecting the knowledge graph nodes.
[0053] 300: The node information in the knowledge graph is converted into node embedding vectors, and the structure of the knowledge graph is modeled using the adjacency matrix.
[0054] The meanings represented by nodes in a knowledge graph are expressed in natural language. Since algorithms cannot directly understand natural language, this application first converts the node information in the knowledge graph into node embedding vectors.
[0055] In some more specific implementations, the BERT embedding model (Bidirectional Encoder Representations from Transformers) can be used to transform the names (i.e., node information) of each node in the knowledge graph into node embedding vectors, as shown in the following formula:
[0056] ; (1)
[0057] In the formula, This represents the BERT embedding model. The j-th character in the name of the i-th graph node. The length of the name of the i-th graph node. Let be the node embedding vector of the i-th node.
[0058] In this way, the node embedding matrix of the knowledge graph can be obtained. ,in for The i-th row in The number of nodes in the knowledge graph.
[0059] Furthermore, to enable the model to better handle the structural information of knowledge graphs, an adjacency matrix is used. Model the structure of the graph, where the value in the i-th row and j-th column of the adjacency matrix is... Let be the number of edges between the i-th node and the j-th node in the knowledge graph.
[0060] 400: Based on the adjacency matrix and the node embedding vector, extract the fault node feature vector and the feature quantity node feature vector of the knowledge graph.
[0061] In some more specific implementations, the adjacency matrix and the nodes can be embedded in a vector input graph convolutional neural network (GCN), and the graph convolutional neural network can be used to extract the fault node feature vector and the feature quantity node feature vector of the knowledge graph.
[0062] Specifically, the knowledge graph node embedding matrix and adjacency matrix can be input into a graph convolutional neural network to calculate the feature vector of each node in the knowledge graph, as shown in the following formula:
[0063] ; (2)
[0064] In the formula, For activation functions; ,in It is the identity matrix. It is an adjacency matrix; This is a degree matrix, where all elements except the main diagonal are 0, and the value of the i-th main diagonal is... The sum of the values in the i-th row of the adjacency matrix ; This refers to the trainable convolutional kernel matrix in GCN; Let be the feature vector matrix of the knowledge graph, where the i-th row... Let be the feature vector of the i-th node.
[0065] 500: Based on the fault node feature vector and the feature quantity node feature vector, calculate the conditional probability of the fault mode corresponding to each feature quantity occurring, and based on the conditional probability of the fault mode corresponding to each feature quantity occurring, obtain the total probability of the fault mode occurring.
[0066] In some more specific implementations, the fault node feature vector and the feature quantity node feature vector are concatenated and input into a fully connected neural network to calculate the conditional probability of the fault mode corresponding to each feature quantity occurring.
[0067] To compute the j-th feature in a knowledge graph When an anomaly occurs, the i-th fault mode Conditional probability of failure It is necessary to extract the i-th row from the eigenvector matrix. and the jth row Extract them. These two vectors are respectively and The feature vector of a node.
[0068] Then and The input is fed into a fully connected neural network, and the final conditional probability is calculated as shown in the following formula:
[0069] ; (3)
[0070] in, Represents the j-th feature quantity When an anomaly occurs, the i-th fault mode The conditional probability of a failure occurring. This represents the weight matrix of a fully connected neural network. This represents the bias of a fully connected neural network. This represents the Sigmoid activation function.
[0071] According to the law of total probability, when the equipment malfunctions... mode failure probability The sum of all conditional probabilities associated with it is shown in the following formula:
[0072] ; (4)
[0073] Ultimately, the above method can be used to calculate the probability of various faults occurring in ultra-high voltage transformers, which can then be used to determine the fault. For example, if the fault probability is higher than a set threshold, then the fault is determined to have occurred.
[0074] As shown in Figure 4, in another embodiment of this application, a knowledge graph-based fault diagnosis system for ultra-high voltage transformers is also proposed, which includes:
[0075] Data module 800 stores a mapping table of fault types, corresponding state quantities, and characteristic quantities of ultra-high voltage transformers.
[0076] The knowledge graph module 802 constructs a knowledge graph based on the mapping relationship table.
[0077] The conversion module 804 converts the node information in the knowledge graph into node embedding vectors and uses the adjacency matrix to model the structure of the knowledge graph.
[0078] The feature extraction module 806 extracts the fault node feature vector and the feature quantity node feature vector of the knowledge graph based on the adjacency matrix and the node embedding vector.
[0079] The judgment module 808 calculates the conditional probability of a fault mode occurring for each feature quantity based on the fault node feature vector and the feature quantity node feature vector, and obtains the total probability of a fault mode occurring based on the conditional probability of the fault mode occurring for each feature quantity.
[0080] In some more specific implementations, the conversion module 804 may use the BERT embedding model to convert the node information in the knowledge graph into node embedding vectors.
[0081] In some more specific implementations, the feature extraction module 806 embeds the adjacency matrix and the nodes into a graph convolutional neural network, and uses the graph convolutional neural network to extract the fault node feature vector and the feature quantity node feature vector of the knowledge graph.
[0082] In some more specific implementations, the judgment module 808 concatenates the fault node feature vector and the feature quantity node feature vector and inputs them into a fully connected neural network to calculate the conditional probability of the fault mode corresponding to each feature quantity occurring.
[0083] In some more specific implementations, the determination module 808 calculates the conditional probability of a failure mode occurring for each feature based on the following formula:
[0084] ; (3)
[0085] in, Represents the j-th feature quantity When an anomaly occurs, the i-th fault mode The conditional probability of a failure occurring. This represents the weight matrix of a fully connected neural network. This represents the bias of a fully connected neural network. This represents the Sigmoid activation function.
[0086] Therefore, the method and system described in this application can achieve more robust, reliable and accurate fault diagnosis of power transformers by calculating the conditional probability of each characteristic quantity for each fault mode of the equipment.
[0087] It should be noted that the prior art portion of the protection scope of this application is not limited to the embodiments given in this application document. All prior art that does not contradict the solution of this application, including but not limited to prior patent documents, prior publications, prior public uses, etc., can be included in the protection scope of this application.
[0088] Furthermore, the combination of the technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.
[0089] It should also be noted that the embodiments listed above are merely specific embodiments of this application. Obviously, this application is not limited to the above embodiments, and similar changes or modifications made thereto are those that can be directly derived or easily conceived by those skilled in the art from the content disclosed in this application, and should all fall within the protection scope of this application.
Claims
1. A fault diagnosis method for ultra-high voltage transformers based on knowledge graphs, comprising the following steps: 100: Establish a mapping table of fault types, corresponding state variables, and characteristic variables for UHV transformers; 200: Construct a knowledge graph based on the mapping table; 300: The node information in the knowledge graph is converted into node embedding vectors, and the structure of the knowledge graph is modeled using the adjacency matrix; 400: Based on the adjacency matrix and the node embedding vector, extract the fault node feature vector and the feature quantity node feature vector of the knowledge graph; 500: Based on the fault node feature vector and the feature quantity node feature vector, calculate the conditional probability of the fault mode corresponding to each feature quantity occurring, and based on the conditional probability of the fault mode corresponding to each feature quantity occurring, obtain the total probability of the fault mode occurring.
2. The fault diagnosis method for ultra-high voltage transformers as described in claim 1, wherein, In step 300, the BERT embedding model is used to convert the node information in the knowledge graph into node embedding vectors.
3. The fault diagnosis method for ultra-high voltage transformers as described in claim 1, wherein, In step 400, the adjacency matrix and the node embedding vectors are input into a graph convolutional neural network, and the graph convolutional neural network is used to extract the fault node feature vector and the feature quantity node feature vector of the knowledge graph.
4. The fault diagnosis method for ultra-high voltage transformers as described in claim 1, wherein, In step 500, the fault node feature vector and the feature quantity node feature vector are concatenated and input into a fully connected neural network to calculate the conditional probability of the fault mode corresponding to each feature quantity.
5. The fault diagnosis method for ultra-high voltage transformers as described in claim 4, wherein, In step 500, the conditional probability of failure occurring for each feature mode is calculated based on the following formula: ； in, Represents the j-th feature quantity When an anomaly occurs, the i-th fault mode The conditional probability of a failure occurring. This represents the weight matrix of a fully connected neural network. This represents the bias of a fully connected neural network. This indicates the Sigmoid activation function.
6. A knowledge graph-based fault diagnosis system for ultra-high voltage transformers, comprising: The data module stores a table showing the mapping relationship between the fault types of UHV transformers, the corresponding state variables, and the characteristic variables. The knowledge graph module constructs a knowledge graph based on the mapping table. The conversion module converts the node information in the knowledge graph into node embedding vectors and uses the adjacency matrix to model the structure of the knowledge graph. The feature extraction module extracts the fault node feature vector and the feature quantity node feature vector of the knowledge graph based on the adjacency matrix and the node embedding vector. The judgment module calculates the conditional probability of a fault mode occurring for each feature quantity based on the feature vector of the fault node and the feature vector of the feature quantity node, and obtains the total probability of a fault mode occurring based on the conditional probability of a fault mode occurring for each feature quantity.
7. The ultra-high voltage transformer fault diagnosis system as described in claim 6, wherein, The conversion module uses the BERT embedding model to convert the node information in the knowledge graph into node embedding vectors.
8. The ultra-high voltage transformer fault diagnosis system as described in claim 6, wherein, The feature extraction module embeds the adjacency matrix and the nodes into a graph convolutional neural network, and uses the graph convolutional neural network to extract the fault node feature vector and the feature quantity node feature vector of the knowledge graph.
9. The ultra-high voltage transformer fault diagnosis system as described in claim 6, wherein, The judgment module concatenates the feature vectors of the fault nodes and the feature vectors of the feature quantity nodes and inputs them into the fully connected neural network to calculate the conditional probability of the fault mode corresponding to each feature quantity.
10. The ultra-high voltage transformer fault diagnosis system as described in claim 6, wherein, The judgment module calculates the conditional probability of a failure mode occurring for each feature quantity based on the following formula: ； in, Represents the j-th feature quantity When an anomaly occurs, the i-th fault mode The conditional probability of a failure occurring. This represents the weight matrix of a fully connected neural network. This represents the bias of a fully connected neural network. This represents the Sigmoid activation function.