A transformer health management method and device based on multi-agent cooperation and a knowledge graph

Abstract

The application relates to a transformer health management method and device based on multi-agent cooperation and a knowledge graph, comprising the following steps: constructing a vector knowledge base, a knowledge graph and a retrieval engine; constructing a transformer fault diagnosis model; constructing a management agent, a state perception agent, a health evaluation agent, a fault diagnosis agent and a state overhauling agent; and performing a closed-loop health management task from state perception, health evaluation, fault diagnosis to state overhauling strategy generation. The application realizes unified fusion and correlation management of multi-source heterogeneous data such as transformer evaluation guidelines, general regulations and fault cases, can comprehensively analyze multi-parameter information such as oil chromatography, acoustic fingerprints, partial discharge and infrared thermal images, improves equipment state perception capability and data utilization efficiency, can improve the accuracy, comprehensiveness and interpretability of health evaluation and fault diagnosis in a complex fault scenario, and is favorable for improving the intelligent level and the automation degree of transformer operation and maintenance management.

Description

Technical Field

[0001] This invention relates to the fields of power equipment health management and artificial intelligence technology, and in particular to a transformer health management method and device based on multi-agent collaboration and knowledge graph. Background Technology

[0002] With the continuous expansion of power system scale and the increasing complexity of operating environment, the safe and stable operation of power equipment is of great significance for ensuring the quality of power supply. As a key piece of equipment in the power system, the operating status of transformers directly affects the safety and stability of the power grid and the reliability of power supply. Existing transformer operation and maintenance methods mainly rely on periodic inspections and manual experience-based judgment, which is not only costly and inefficient, but also results in highly subjective diagnoses that fail to comprehensively and accurately reflect the actual health status of the equipment. Therefore, conducting research on transformer health management has significant technical value for improving equipment operational reliability, extending service life, and reducing operational risks.

[0003] However, existing power equipment health management systems still have areas for improvement. On one hand, transformer health is influenced by various parameters such as oil temperature, oil chromatography, acoustic signals, current, voltage, and environmental factors. Existing methods typically analyze data based on only a single parameter or a single type of data, making it difficult to achieve comprehensive correlation analysis of multi-dimensional data, thus limiting the accuracy and reliability of health assessments. On the other hand, traditional diagnostic models are often designed for specific fault types, such as partial discharge detection or oil chromatography analysis, lacking comprehensive diagnostic capabilities across fault types and failing to meet the needs of equipment health management in complex operating environments. Furthermore, while large-scale models have shown outstanding performance in natural language processing and knowledge reasoning in recent years, their direct application to power equipment health management still suffers from the illusion problem—the model may generate diagnostic results that do not match the actual equipment condition or professional experience, lacking sufficient interpretability and reliability. Summary of the Invention

[0004] To address the current shortcomings in transformer health management, such as insufficient multi-parameter analysis, weak knowledge collaboration, and weak closed-loop control capabilities, the primary objective of this invention is to provide a transformer health management method based on multi-agent collaboration and knowledge graphs. This method can comprehensively analyze multi-parameter information such as oil chromatography, acoustic signature, partial discharge, and infrared thermography, thereby improving equipment status perception and data utilization efficiency, and realizing a closed-loop health management process from status perception, health assessment, fault diagnosis to maintenance strategy generation.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a transformer health management method based on multi-agent collaboration and knowledge graph, the method comprising the following sequential steps:

[0006] (1) Construct a vector knowledge base, knowledge graph and retrieval engine to store structured and unstructured knowledge and provide knowledge retrieval and related reasoning functions;

[0007] (2) Construct a transformer fault diagnosis model for analyzing multiple parameters of the transformer;

[0008] (3) Construct management intelligent agents, state perception intelligent agents, health assessment intelligent agents, fault diagnosis intelligent agents and condition maintenance intelligent agents; management intelligent agents, state perception intelligent agents, health assessment intelligent agents, fault diagnosis intelligent agents and condition maintenance intelligent agents execute closed-loop health management tasks from state perception, health assessment, fault diagnosis to condition maintenance strategy generation by calling vector knowledge base, knowledge graph and transformer fault diagnosis model.

[0009] In step (1), building a vector knowledge base specifically means: extracting text from the collected PDF texts and scanned copies of transformer health management-related documents using a text recognition model based on the PaddleOCR engine to obtain complete text documents; the transformer health management-related documents include transformer condition evaluation guidelines and transformer fault case reports;

[0010] This paper reviews relevant documents on transformer health management, clarifies the knowledge scope and classification system, and obtains a knowledge set on transformer health management, which is then defined as follows: :

[0011] ;

[0012] in, This indicates the specific categories of knowledge related to transformer health management; Indicates the number of categories; Indicates the total number of categories; Represents a knowledge entry, each knowledge entry Includes different categories of information;

[0013] ;

[0014] In the formula, For voltage level, For the device name, For the number of years of operation, For manufacturers, The faulty part, Faulty component, This is a fault mode. The fault name and description are provided. For cause analysis, For key data, As a measure to deal with the situation;

[0015] Extracted knowledge entries Preprocessing is performed: first, missing fields are deleted or added; then, values ​​exceeding the threshold are corrected; finally, uniqueness is determined based on field combinations to obtain the preprocessed knowledge entries. ;

[0016] To deploy a localized embedded model using LM Studio, let's define the localized embedded model as a function. The preprocessed knowledge entries Mapping to knowledge vectors :

[0017] ;

[0018] Where d is the knowledge vector Dimensions It is the preprocessed knowledge entries Input Embedded Model Function Obtain knowledge vectors ; For preprocessed knowledge entries The total quantity;

[0019] knowledge vector Store in fassi vector database .

[0020] In step (1), the knowledge graph is constructed in the following way: Let the knowledge graph... For a triple: ,in, It represents a set of entities, including the transformer body, windings, insulation components, fault types, handling measures, and maintenance strategies; Represents a set of relations, describing the semantic relationships between entities; It represents a set of attributes that records the characteristic values ​​of an entity, including transformer model, rated voltage, detection time, and fault probability;

[0021] Extract entities from structured monitoring data, inspection reports, and maintenance records. And identify relationships between entities. Assign attribute values ​​to entities Align and disambiguate the multi-source extraction results to form a knowledge graph; store the constructed knowledge graph in the graph database Neo4j and create an entity index. Relational Index and attribute index Among them, entity index Used to retrieve corresponding entities based on entity keywords; relation index Used to retrieve relationships between entities based on entity pairs; attribute index Used to retrieve a collection of entities that have corresponding attributes based on an attribute field.

[0022] In step (1), the search engine specifically refers to: given the user's query content The retrieval plan is defined as a quadruple. ,in, This indicates a query for the relevant domain; Indicates the subject entity in the query; This represents the set of relations to be filtered, used to exclude irrelevant semantic paths; Indicates the number of hops in the search, controlling the search depth;

[0023] The retrieval process employs iterative neighbor expansion and semantic filtering:

[0024] Use an LLM based on Qwen2.5-14b to parse query content. Output the subject entity And query the domain ;from Start by recursively obtaining... All neighboring entities and relationships:

[0025] Filtering is performed based on semantic similarity and query semantics to obtain results that match the query content. Related sets of relationships:

[0026] The search terminates when one of the following three conditions is met: search depth, no new related relationships to expand the search, or the search results have covered all entities required by the query.

[0027] The search results are returned in the form of subgraphs to support fault tracing and strategy reasoning.

[0028] In step (2), the transformer fault diagnosis model includes one or more of the following: oil chromatography diagnosis model, acoustic fingerprint recognition model, ultra-high frequency partial discharge recognition model, infrared overheating recognition model, and core grounding current analysis model;

[0029] The oil chromatography diagnostic model is based on the three-ratio method to diagnose transformer fault types. The oil chromatography diagnostic model receives the concentration data of characteristic gases such as hydrogen, methane, ethane, ethylene, and acetylene as input, calculates three specific ratios: methane / hydrogen, ethylene / ethane, and acetylene / ethylene. According to the preset ratio coding rules and fault type mapping table, the calculated ratios are converted into standard codes to determine the fault type.

[0030] The voiceprint recognition model is used to analyze the sound signals generated during transformer operation and identify the acoustic features of mechanical abnormalities and partial discharge. In the preprocessing stage, the Spectrogram algorithm is used to extract feature representations from the original audio signal and enhance the frequency and time domain features related to the fault in the signal.

[0031] The UHF partial discharge identification model is designed for the identification and classification of partial discharges in UHF electromagnetic signals. It uses LS-SVM as the classifier. The partial discharge signals collected by the UHF sensor are processed by feature extraction to form feature vectors. LS-SVM projects the samples into a high-dimensional feature space through nonlinear mapping and constructs the optimal classification hyperplane in this space to distinguish the type of partial discharge.

[0032] The infrared overheating identification model uses the YOLOV8 target detection algorithm to automatically identify and locate overheated areas on the surface of transformer equipment. It receives infrared thermal images as input, extracts multi-scale thermal features through a convolutional neural network, and uses the anchor frame mechanism and feature pyramid structure to detect abnormal hot spots.

[0033] The core grounding current analysis model receives the transformer core grounding current and uses a multi-parameter dynamic threshold judgment method to achieve parameter analysis. It not only monitors the amplitude of the grounding current, but also comprehensively analyzes the waveform characteristics, harmonic content, and trend of the grounding current, and dynamically adjusts the alarm threshold in combination with operating load, oil temperature and other operating condition information.

[0034] In step (3), the management agent receives health management requests, assigns tasks, coordinates the execution order between units, and resolves and integrates the output results of multiple units.

[0035] The state-aware intelligent agent responds to query requests and obtains real-time operating data, historical monitoring data, test records, operation and maintenance files, and historical analysis results of the transformer;

[0036] The health assessment agent calls the large language model and combines it with the matching evaluation guidelines to analyze the state variables of each component of the transformer, calculate the health score and level, and output the abnormal state variables and health assessment results.

[0037] The fault diagnosis intelligent agent calls the transformer fault diagnosis model based on the transformer's real-time operating data and health assessment results, and combines the vector knowledge base and knowledge graph to perform case matching and graph reasoning to generate a fault diagnosis report containing the fault location and cause chain.

[0038] The condition-based maintenance agent generates operation and maintenance or repair strategies based on health assessment results and fault diagnosis reports, combined with maintenance guidelines and knowledge graphs.

[0039] Another object of the present invention is to provide an electronic device comprising:

[0040] Processor; and

[0041] A memory storing computer program instructions, which, when executed by the processor, cause the processor to perform the transformer health management method based on multi-agent cooperation and knowledge graph as described above.

[0042] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the transformer health management method based on multi-agent cooperation and knowledge graph as described above.

[0043] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, by constructing a vector knowledge base and knowledge graph, the present invention achieves unified integration and associated management of multi-source heterogeneous data such as transformer evaluation guidelines, general regulations, and fault cases. It can comprehensively analyze multi-parameter information such as oil chromatography, acoustic signature, partial discharge, and infrared thermography, thereby improving equipment status perception capabilities and data utilization efficiency. Second, by combining multiple transformer fault diagnosis models and knowledge graph reasoning mechanisms, the present invention performs associated analysis on equipment status, fault type, fault location, and cause chain, which can improve the accuracy, comprehensiveness, and interpretability of health assessment and fault diagnosis in complex fault scenarios, avoiding the limitations of single model analysis. Third, by managing the collaborative operation of management intelligent agents, status perception intelligent agents, health assessment intelligent agents, fault diagnosis intelligent agents, and condition-based maintenance intelligent agents, the present invention realizes a closed-loop health management process from status perception, health assessment, fault diagnosis to maintenance strategy generation, which is conducive to improving the intelligence level and automation of transformer operation and maintenance management, and improving equipment operation reliability and operation and maintenance efficiency. Attached Figure Description

[0044] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0045] like Figure 1 As shown, a transformer health management method based on multi-agent collaboration and knowledge graph is proposed, which includes the following sequential steps:

[0046] (1) Construct a vector knowledge base, knowledge graph and retrieval engine to store structured and unstructured knowledge and provide knowledge retrieval and related reasoning functions;

[0047] (2) Construct a transformer fault diagnosis model for analyzing multiple parameters of the transformer;

[0048] (3) Construct management intelligent agents, state perception intelligent agents, health assessment intelligent agents, fault diagnosis intelligent agents and condition maintenance intelligent agents; management intelligent agents, state perception intelligent agents, health assessment intelligent agents, fault diagnosis intelligent agents and condition maintenance intelligent agents execute closed-loop health management tasks from state perception, health assessment, fault diagnosis to condition maintenance strategy generation by calling vector knowledge base, knowledge graph and transformer fault diagnosis model.

[0049] In step (1), building a vector knowledge base specifically means: extracting text from the collected PDF texts and scanned copies of transformer health management-related documents using a text recognition model based on the PaddleOCR engine to obtain complete text documents; the transformer health management-related documents include transformer condition evaluation guidelines and transformer fault case reports;

[0050] This paper reviews relevant documents on transformer health management, clarifies the knowledge scope and classification system, and obtains a knowledge set on transformer health management, which is then defined as follows: :

[0051] ;

[0052] in, This indicates the specific categories of knowledge related to transformer health management; Indicates the number of categories; Indicates the total number of categories; Represents a knowledge entry, each knowledge entry Includes different categories of information;

[0053] ;

[0054] In the formula, For voltage level, For the device name, For the number of years of operation, For manufacturers, The faulty part, Faulty component, This is a fault mode. The fault name and description are provided. For cause analysis, For key data, As a measure to deal with the situation;

[0055] Extracted knowledge entries Preprocessing is performed: first, missing fields are deleted or added; then, values ​​exceeding the threshold are corrected; finally, uniqueness is determined based on field combinations to obtain the preprocessed knowledge entries. ;

[0056] To deploy a localized embedded model using LM Studio, let's define the localized embedded model as a function. The preprocessed knowledge entries Mapping to knowledge vectors :

[0057] ;

[0058] Where d is the knowledge vector Dimensions It is the preprocessed knowledge entries Input Embedded Model Function Obtain knowledge vectors ; For preprocessed knowledge entries The total quantity;

[0059] knowledge vector Store in fassi vector database .

[0060] In step (1), the knowledge graph is constructed in the following way: Let the knowledge graph... For a triple: ,in, It represents a set of entities, including the transformer body, windings, insulation components, fault types, handling measures, and maintenance strategies; Represents a set of relations, describing the semantic relationships between entities; It represents a set of attributes that records the characteristic values ​​of an entity, including transformer model, rated voltage, detection time, and fault probability;

[0061] Extract entities from structured monitoring data, inspection reports, and maintenance records. And identify relationships between entities. Assign attribute values ​​to entities Align and disambiguate the multi-source extraction results to form a knowledge graph; store the constructed knowledge graph in the graph database Neo4j and create an entity index. Relational Index and attribute index Among them, entity index Used to retrieve corresponding entities based on entity keywords; relation index Used to retrieve relationships between entities based on entity pairs; attribute index Used to retrieve a collection of entities that have corresponding attributes based on an attribute field.

[0062] In step (1), the search engine specifically refers to: given the user's query content The retrieval plan is defined as a quadruple. ,in, This indicates a query for the relevant domain; Indicates the subject entity in the query; This represents the set of relations to be filtered, used to exclude irrelevant semantic paths; Indicates the number of hops in the search, controlling the search depth;

[0063] The retrieval process employs iterative neighbor expansion and semantic filtering:

[0064] Use an LLM based on Qwen2.5-14b to parse query content. Output the subject entity And query the domain ;from Start by recursively obtaining... All neighboring entities and relationships:

[0065] Filtering is performed based on semantic similarity and query semantics to obtain results that match the query content. Related sets of relationships:

[0066] The search terminates when one of the following three conditions is met: search depth, no new related relationships to expand the search, or the search results have covered all entities required by the query.

[0067] The search results are returned in the form of subgraphs to support fault tracing and strategy reasoning.

[0068] In step (2), the transformer fault diagnosis model includes one or more of the following: oil chromatography diagnosis model, acoustic fingerprint recognition model, ultra-high frequency partial discharge recognition model, infrared overheating recognition model, and core grounding current analysis model;

[0069] The oil chromatography diagnostic model is based on the three-ratio method to diagnose transformer fault types. The oil chromatography diagnostic model receives the concentration data of characteristic gases such as hydrogen, methane, ethane, ethylene, and acetylene as input, calculates three specific ratios: methane / hydrogen, ethylene / ethane, and acetylene / ethylene. According to the preset ratio coding rules and fault type mapping table, the calculated ratios are converted into standard codes to determine the fault type.

[0070] The voiceprint recognition model is used to analyze the sound signals generated during transformer operation and identify the acoustic features of mechanical abnormalities and partial discharge. In the preprocessing stage, the Spectrogram algorithm is used to extract feature representations from the original audio signal and enhance the frequency and time domain features related to the fault in the signal.

[0071] The UHF partial discharge identification model is designed for the identification and classification of partial discharges in UHF electromagnetic signals. It uses LS-SVM as the classifier. The partial discharge signals collected by the UHF sensor are processed by feature extraction to form feature vectors. LS-SVM projects the samples into a high-dimensional feature space through nonlinear mapping and constructs the optimal classification hyperplane in this space to distinguish the type of partial discharge.

[0072] The infrared overheating identification model uses the YOLOV8 target detection algorithm to automatically identify and locate overheated areas on the surface of transformer equipment. It receives infrared thermal images as input, extracts multi-scale thermal features through a convolutional neural network, and uses the anchor frame mechanism and feature pyramid structure to detect abnormal hot spots.

[0073] The core grounding current analysis model receives the transformer core grounding current and uses a multi-parameter dynamic threshold judgment method to achieve parameter analysis. It not only monitors the amplitude of the grounding current, but also comprehensively analyzes the waveform characteristics, harmonic content, and trend of the grounding current, and dynamically adjusts the alarm threshold in combination with operating load, oil temperature and other operating condition information.

[0074] In step (3), the management agent receives health management requests, assigns tasks, coordinates the execution order between units, and resolves and integrates the output results of multiple units.

[0075] The state-aware intelligent agent responds to query requests and obtains real-time operating data, historical monitoring data, test records, operation and maintenance files, and historical analysis results of the transformer;

[0076] The health assessment agent calls the large language model and combines it with the matching evaluation guidelines to analyze the state variables of each component of the transformer, calculate the health score and level, and output the abnormal state variables and health assessment results.

[0077] The fault diagnosis intelligent agent calls the transformer fault diagnosis model based on the transformer's real-time operating data and health assessment results, and combines the vector knowledge base and knowledge graph to perform case matching and graph reasoning to generate a fault diagnosis report containing the fault location and cause chain.

[0078] The condition-based maintenance agent generates operation and maintenance or repair strategies based on health assessment results and fault diagnosis reports, combined with maintenance guidelines and knowledge graphs.

[0079] In summary, this invention, by constructing a vector knowledge base and knowledge graph, achieves unified integration and associated management of multi-source heterogeneous data such as transformer evaluation guidelines, general regulations, and fault cases. It enables comprehensive analysis of multi-parameter information including oil chromatography, acoustic signatures, partial discharge, and infrared thermography, improving equipment status perception and data utilization efficiency. Furthermore, by combining multiple transformer fault diagnosis models and knowledge graph reasoning mechanisms, this invention performs associated analysis on equipment status, fault type, fault location, and causal chain, improving the accuracy, comprehensiveness, and interpretability of health assessment and fault diagnosis in complex fault scenarios, avoiding the limitations of single-model analysis. Finally, through the collaborative operation of management agents, status perception agents, health assessment agents, fault diagnosis agents, and condition-based maintenance agents, this invention achieves a closed-loop health management process from status perception, health assessment, fault diagnosis to maintenance strategy generation, which is beneficial for improving the intelligence and automation level of transformer operation and maintenance management, and enhancing equipment reliability and maintenance efficiency.

[0080] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A transformer health management method based on multi-agent collaboration and knowledge graph, characterized in that: The method includes the following steps in sequence: (1) Construct a vector knowledge base, knowledge graph and retrieval engine to store structured and unstructured knowledge and provide knowledge retrieval and related reasoning functions; (2) Construct a transformer fault diagnosis model for analyzing multiple parameters of the transformer; (3) Construct management intelligent agents, state perception intelligent agents, health assessment intelligent agents, fault diagnosis intelligent agents and condition maintenance intelligent agents; management intelligent agents, state perception intelligent agents, health assessment intelligent agents, fault diagnosis intelligent agents and condition maintenance intelligent agents execute closed-loop health management tasks from state perception, health assessment, fault diagnosis to condition maintenance strategy generation by calling vector knowledge base, knowledge graph and transformer fault diagnosis model.

2. The transformer health management method based on multi-agent collaboration and knowledge graph as described in claim 1, characterized in that: In step (1), building a vector knowledge base specifically means: extracting text from the collected PDF texts and scanned copies of documents related to transformer health management using a text recognition model based on the PaddleOCR engine to obtain complete text documents; The relevant documents on transformer health management include guidelines for transformer condition assessment and transformer failure case reports; This paper reviews relevant documents on transformer health management, clarifies the knowledge scope and classification system, and obtains a knowledge set on transformer health management, which is then defined as follows: : ； in, This indicates the specific categories of knowledge related to transformer health management; Indicates the number of categories; Indicates the total number of categories; Represents a knowledge entry, each knowledge entry Includes different categories of information; ； In the formula, For voltage level, For the device name, For the number of years of operation, For manufacturers, The faulty part, Faulty component, This is a fault mode. The fault name and description are provided. For cause analysis, For key data, As a measure to deal with the situation; Extracted knowledge entries Preprocessing is performed: first, missing fields are deleted or added; then, values ​​exceeding the threshold are corrected; finally, uniqueness is determined based on field combinations to obtain the preprocessed knowledge entries. ; To deploy a localized embedded model using LM Studio, let's define the localized embedded model as a function. The preprocessed knowledge entries Mapping to knowledge vectors : ； Where d is the knowledge vector Dimensions It is the preprocessed knowledge entries Input Embedded Model Function Obtain knowledge vectors ; For preprocessed knowledge entries The total quantity; knowledge vector Store in fassi vector database .

3. The transformer health management method based on multi-agent collaboration and knowledge graph as described in claim 1, characterized in that: In step (1), the knowledge graph is constructed in the following way: Let the knowledge graph... For a triple: ,in, It represents a set of entities, including the transformer body, windings, insulation components, fault types, handling measures, and maintenance strategies; Represents a set of relations, describing the semantic relationships between entities; It represents a set of attributes that records the characteristic values ​​of an entity, including transformer model, rated voltage, detection time, and fault probability; Extract entities from structured monitoring data, inspection reports, and maintenance records. And identify relationships between entities. Assign attribute values ​​to entities Align and disambiguate the multi-source extraction results to form a knowledge graph; store the constructed knowledge graph in the graph database Neo4j and create an entity index. Relational Index and attribute index Among them, entity index Used to retrieve corresponding entities based on entity keywords; relation index Used to retrieve relationships between entities based on entity pairs; attribute index Used to retrieve a collection of entities that have corresponding attributes based on an attribute field.

4. The transformer health management method based on multi-agent collaboration and knowledge graph as described in claim 1, characterized in that: In step (1), the search engine specifically refers to: given the user's query content The retrieval plan is defined as a quadruple. ,in, This indicates a query for the relevant domain; Indicates the subject entity in the query; This represents the set of relations to be filtered, used to exclude irrelevant semantic paths; Indicates the number of hops in the search, controlling the search depth; The retrieval process employs iterative neighbor expansion and semantic filtering: Use an LLM based on Qwen2.5-14b to parse query content. Output the subject entity And query the domain ;from Start by recursively obtaining... All neighboring entities and relationships: Filtering is performed based on semantic similarity and query semantics to obtain results that match the query content. Related sets of relationships: The search terminates when one of the following three conditions is met: search depth, no new related relationships to expand the search, or the search results have covered all entities required by the query. The search results are returned in the form of subgraphs to support fault tracing and strategy reasoning.

5. The transformer health management method based on multi-agent collaboration and knowledge graph as described in claim 1, characterized in that: In step (2), the transformer fault diagnosis model includes one or more of the following: oil chromatography diagnosis model, acoustic fingerprint recognition model, ultra-high frequency partial discharge recognition model, infrared overheating recognition model, and core grounding current analysis model; The oil chromatography diagnostic model is based on the three-ratio method to diagnose transformer fault types. The oil chromatography diagnostic model receives the concentration data of characteristic gases such as hydrogen, methane, ethane, ethylene, and acetylene as input, calculates three specific ratios: methane / hydrogen, ethylene / ethane, and acetylene / ethylene. According to the preset ratio coding rules and fault type mapping table, the calculated ratios are converted into standard codes to determine the fault type. Voiceprint recognition models are used to analyze sound signals generated during transformer operation and identify acoustic features of mechanical abnormalities and partial discharge. In the preprocessing stage, the Spectrogram algorithm is used to extract feature representations from the original audio signal to enhance the frequency and time domain features related to the fault in the signal. The UHF partial discharge identification model is designed for the identification and classification of partial discharges in UHF electromagnetic signals. It uses LS-SVM as the classifier. The partial discharge signals collected by the UHF sensor are processed by feature extraction to form feature vectors. LS-SVM projects the samples into a high-dimensional feature space through nonlinear mapping and constructs the optimal classification hyperplane in this space to distinguish the type of partial discharge. The infrared overheating identification model uses the YOLOV8 target detection algorithm to automatically identify and locate overheated areas on the surface of transformer equipment. It receives infrared thermal images as input, extracts multi-scale thermal features through a convolutional neural network, and uses the anchor frame mechanism and feature pyramid structure to detect abnormal hot spots. The core grounding current analysis model receives the transformer core grounding current and uses a multi-parameter dynamic threshold judgment method to achieve parameter analysis. It not only monitors the amplitude of the grounding current, but also comprehensively analyzes the waveform characteristics, harmonic content, and trend of the grounding current, and dynamically adjusts the alarm threshold in combination with operating load, oil temperature and other operating condition information.

6. The transformer health management method based on multi-agent collaboration and knowledge graph as described in claim 1, characterized in that: In step (3), the management agent receives health management requests, assigns tasks, coordinates the execution order between units, and resolves and integrates the output results of multiple units. The state-aware intelligent agent responds to query requests and obtains real-time operating data, historical monitoring data, test records, operation and maintenance files, and historical analysis results of the transformer; The health assessment agent calls the large language model and combines it with the matching evaluation guidelines to analyze the state variables of each component of the transformer, calculate the health score and level, and output the abnormal state variables and health assessment results. The fault diagnosis intelligent agent calls the transformer fault diagnosis model based on the transformer's real-time operating data and health assessment results, and combines the vector knowledge base and knowledge graph to perform case matching and graph reasoning to generate a fault diagnosis report containing the fault location and cause chain. The condition-based maintenance agent generates operation and maintenance or repair strategies based on health assessment results and fault diagnosis reports, combined with maintenance guidelines and knowledge graphs.

7. An electronic device, comprising: processor; as well as A memory storing computer program instructions, which, when executed by the processor, cause the processor to perform the transformer health management method based on multi-agent cooperation and knowledge graph as described in any one of claims 1-6.

8. A computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the transformer health management method based on multi-agent cooperation and knowledge graph as described in any one of claims 1-6.

Images