CN-122020088-A - Transformer state diagnosis method based on large model

CN122020088ACN 122020088 ACN122020088 ACN 122020088ACN-122020088-A

Abstract

The invention relates to the technical field of power equipment state evaluation, in particular to a transformer state diagnosis method based on a large model, which provides a transformer comprehensive state evaluation algorithm fused with multi-mode monitoring data, and effectively solves the industry bottleneck that the traditional single-dimension monitoring cannot accurately identify composite faults; the intelligent agent interaction question-answering system is provided, a man-machine interaction fault diagnosis question-answering mechanism is established, hidden faults are converted into accurate diagnosis conclusions through continuous interaction of natural language, and comprehensive and intelligent diagnosis assistants are provided for transformer fault diagnosis.

Inventors

HUANG JUNHUA
LIU YONGXUAN
TANG WEIJIE
LI JIAN
ZHANG ZHIQIANG
LI XUELING
WANG YICHENG
WANG PEILUN
Jin Jizhen
QI WEIFENG
LIU TIANCONG

Assignees

山东电工电气集团科学技术研究有限公司
山东电工电气集团有限公司

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (10)

1. The transformer state diagnosis method based on the large model is characterized by comprising the following steps of: S01, constructing a transformer professional knowledge base, firstly carrying out knowledge acquisition and cleaning, then slicing and manually marking cleaned data, converting the sliced data into low-dimensional dense vectors and storing the low-dimensional dense vectors into the transformer professional knowledge base; S02, constructing a knowledge graph, wherein the knowledge graph comprises entities, attributes and relations, using an ontology editing tool to complete definition of classes, attributes, relations and constraints, then carrying out knowledge extraction and filling, completing relation identification between the entities based on a BERT relation classification model, and importing extracted, cleaned and converted entity, attribute and relation data into a selected storage engine; S03, constructing a state comprehensive evaluation model based on multiple characteristic parameters of the transformer, firstly, establishing an evaluation factor set of the multiple state parameters and an evaluation set formed by multiple evaluation results, calculating membership of each factor in the evaluation set through a fuzzy membership function to obtain a single factor evaluation matrix, multiplying factors in the single factor evaluation matrix by weights of corresponding evaluation factors to obtain a fuzzy comprehensive evaluation result of a second-level evaluation factor, and adopting a weighting method to the fuzzy comprehensive evaluation result to obtain a state evaluation result of the transformer state; S04, constructing an intelligent question-answering system based on an intelligent agent, wherein the intelligent agent comprises a monitoring data analysis module, a state comprehensive evaluation model, a transformer professional knowledge base and a knowledge graph, the monitoring data analysis module analyzes and extracts transformer monitoring data and transmits the transformer monitoring data to the state comprehensive evaluation model to obtain a comprehensive state evaluation result of the transformer, detail information required by fault diagnosis is obtained through multiple rounds of conversations, reasoning and fusing the detail information to obtain a fault diagnosis list, the transformer professional knowledge base and the knowledge graph are called to search the similarity of the fault diagnosis list to obtain detailed fault reasons and fault countermeasure results, the comprehensive state evaluation results, the detailed fault reasons and the fault countermeasure results are fed back to a large model together to carry out knowledge fusion and language organization, and the transformer state diagnosis result is generated.
2. The method for diagnosing the transformer state based on the large model of claim 1, wherein in the step S01, a transformer expertise base is built based on a Rag frame, a sliced data set is vectorized and built by adopting a BGE-M3 model, the BGE-M3 model converts texts into low-dimensional dense vectors with fixed dimensions through a dense retrieval function and stores the low-dimensional dense vectors into transformer expertise data, and processing and storage of the transformer expertise data are completed.
3. The transformer state diagnosis method based on the large model of claim 2, wherein the BGE-M3 model adopts a weighted mixed search mode based on dense search, vocabulary search and multi-vector search, calculates correlation scores of the dense search, the vocabulary search and the multi-vector search respectively, performs weighted summation on the three correlation scores to obtain a mixed calculated score, and the mixed calculated score represents similarity between a query vector and vectors in a transformer professional knowledge base and a knowledge graph, and screens a candidate document list with high similarity.
4. The transformer state diagnosis method based on the large model as set forth in claim 1, wherein the data collected in step S01 comprises a product manual, a technical document, a customer service record, a conference summary, a research report, an industry standard, and a policy rule, text detection and recognition of text data are completed by PaddleOCR for unstructured data, layout elements are extracted, the layout elements comprise text blocks, titles, paragraphs, pictures, and tables, the content is structurally output as a MarkDown document, and a Whisper model is embedded to realize voice recognition, semantic text conversion, translation, and language detection of the audio file.
5. The method of claim 1, wherein in step S01, the method is implemented by configuring an interface to set a slice mode, wherein the slice mode supports a semantic block mode based on the LLM large model according to a fixed length, a paragraph and punctuation mark.
6. The transformer state diagnosis method based on large model as set forth in claim 1, wherein in step S02, ETL script is written to extract entity, attribute, relation example from database, CSV, excel, json file for structured/semi-structured data, auxiliary mapping is completed by using database Schema, entity in text is recognized by NER model in BGE-M3 model for unstructured data, and relation recognition between entities is completed by relation classification model based on BERT, and extracted, cleaned and converted entity, attribute and relation data are imported into selected storage engine.
7. The method for diagnosing transformer states based on large model as recited in claim 6, wherein the knowledge graph uses transformer model number, fault type name, state quantity name and time stamp as query indexes.
8. The transformer state diagnosis method based on the large model of claim 1, wherein the newly added fault data is processed in real time by adopting a streaming data processing architecture Kafka+ SPARK STREAMING, and knowledge graph nodes and side relations are dynamically updated through an incremental learning mechanism, so that the professional knowledge base and knowledge graph basic data are completed and continuously updated.
9. The method for diagnosing a transformer state based on a large model as recited in claim 1, wherein the set of evaluation factors includes a set of dissolved gases in oil, a set of thermodynamic parameters, a set of electrical fault characteristics, a set of mechanical characteristics, and the set of evaluation factors includes 、、 The thermodynamic parameter set includes a top layer oil temperature and a hot spot temperature, the electrical fault characteristic parameter set includes partial discharge and ground current, and the mechanical characteristic parameter set includes vibration and noise.
10. The transformer state diagnosis method based on the large model as set forth in claim 1, wherein the multiple rounds of dialogue of step S04 introduce a reinforcement learning-based dynamic problem generation strategy, and the reinforcement learning strategy is adopted to dynamically generate a chain of additional questions based on the consistency of the current diagnosis confidence and the multi-modal data.

Description

Transformer state diagnosis method based on large model Technical Field The invention relates to the technical field of power equipment state evaluation, in particular to a transformer state diagnosis method realized by combining multi-factor decision and fuzzy mathematics such as fusion temperature, vibration, partial discharge, oil chromatography, voiceprint and the like and combining a large model and an intelligent agent arrangement technology. Background The transformer is a core device of the power system, plays an irreplaceable role in power transmission, distribution and utilization, and is a core hub for maintaining the stable operation of the power system. With the development of large-scale grid connection of renewable energy sources and intelligent power grids, transformers become key interfaces for reliably connecting intermittent power sources such as wind energy, photovoltaics and the like with a main network. From traditional electric power to modern industry, rail transit, data center and even new energy fields, the economy and the safety of the whole energy system are directly determined by the high efficiency and the reliability of the transformer. Whether the traditional power grid or the renewable energy source is connected, the transformer is a key for ensuring the stable operation of the voltage class conversion and the power grid, and supports the power supply and the technical development of the modern society. However, with the continuous development of the power system, the running environment and the load condition of the transformer are increasingly complex, and the fault probability of insulation aging, mechanical deformation, partial discharge and the like is also increased under the working conditions of high voltage, high current, complex electromagnetic environment and temperature change for a long time, so that equipment damage and even power grid paralysis can be caused in serious cases. Therefore, the method has extremely important significance in real-time monitoring and intelligent diagnosis of the running state of the transformer. With the wide application and rapid development of information technology, the scale and complexity of various information systems are continuously increased, and operation and maintenance management work becomes increasingly heavy and complex. The traditional operation and maintenance fault diagnosis processing method mainly depends on experience judgment of operation and maintenance personnel, and is low in efficiency and easy to make mistakes. In addition, on one hand, the existing operation and maintenance fault management system mostly adopts a database or file form to store historical operation and maintenance fault diagnosis processing cases, but lacks potential value mining and systematic management of the historical operation and maintenance fault diagnosis processing cases, so that precious fault diagnosis processing experiences are difficult to effectively integrate and reuse, and knowledge resources are wasted. On the other hand, the systems generally lack intelligent fault diagnosis and processing mechanisms, can not quickly identify the fault source and automatically recommend an accurate solution, and further increase the operation and maintenance management cost. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a transformer state diagnosis method based on a large model, which realizes the rapid evaluation and accurate diagnosis of the transformer state, reduces the occurrence probability of faults and prolongs the service life of the transformer. In order to solve the technical problems, the technical scheme adopted by the invention is that the transformer state diagnosis method based on the large model comprises the following steps: S01, constructing a transformer professional knowledge base, firstly carrying out knowledge acquisition and cleaning, then slicing and manually marking cleaned data, converting the sliced data into low-dimensional dense vectors and storing the low-dimensional dense vectors into the transformer professional knowledge base; S02, constructing a knowledge graph, wherein the knowledge graph comprises entities, attributes and relations, using an ontology editing tool to complete definition of classes, attributes, relations and constraints, then carrying out knowledge extraction and filling, completing relation identification between the entities based on a BERT relation classification model, and importing extracted, cleaned and converted entity, attribute and relation data into a selected storage engine; S03, constructing a state comprehensive evaluation model based on multiple characteristic parameters of the transformer, firstly, establishing an evaluation factor set of the multiple state parameters and an evaluation set formed by multiple evaluation results, calculating membership of each factor in the evaluation set through a fuzzy membership function to obtai