CN-122019745-A - Cross-mine multi-agent collaboration and semantic alignment system based on large language model

CN122019745ACN 122019745 ACN122019745 ACN 122019745ACN-122019745-A

Abstract

The invention relates to the technical field of coal mine intellectualization and artificial intelligence, and discloses a large language model-based cross-mine multi-agent collaborative and semantic alignment system. The method comprises the steps of accessing local agents into underground unstructured data, converting the underground unstructured data into structured causal triples by using a large coal mine model, screening and uploading the structured causal triples based on confidence, carrying out vector coding and similarity calculation on the triples by a semantic alignment agent to construct a coal mine field ontology map, converting abnormal events into natural language problems by a collaborative decision agent, initiating a semantic question acquisition historical strategy across mines, and generating a safety suggestion by combining the map. According to the invention, through cooperation of a large model and multiple agents, efficient alignment and multiplexing of causal knowledge are realized under zero original data transmission, the problem of term isomerism is solved, and the accuracy of cross-mine safety decision is remarkably improved.

Inventors

Song Zhengyifan
YANG LINGKAI
QIN YU
CHENG JIAN
WANG GUANGFU
WANG HAO
ZHU JIE

Assignees

煤炭科学研究总院有限公司
天地科技股份有限公司北京煤炭共性技术研究分公司

Dates

Publication Date: 20260512
Application Date: 20251218

Claims (10)

1. The cross-mine multi-agent collaboration and semantic alignment system based on the large language model is characterized by comprising a local knowledge extraction agent, a semantic alignment agent and a collaborative decision agent; The local knowledge extraction agent is connected with underground unstructured data, the underground unstructured data are converted into structured causal triples by using a preset large language model of coal mine industry, and the structured causal triples meeting quality requirements are screened based on a confidence screening mechanism and uploaded to the semantic alignment agent; The semantic alignment agent receives the structured causal triples, performs semantic vector coding on the structured causal triples, calculates semantic similarity among the cross-mine triples, and constructs or updates a coal mine field ontology graph according to the semantic similarity calculation result; When the collaborative decision-making agent generates an abnormal event in a mine, the abnormal event is converted into a natural language problem, semantic knowledge inquiry is initiated to the local knowledge extraction agent of other coal mines to obtain a history treatment strategy, and a safety suggestion is generated by combining the coal mine field ontology map and a preset weighted fusion model.
2. The large language model-based cross-mine multi-agent collaboration and semantic alignment system of claim 1, wherein the downhole unstructured data accessed by the local knowledge extraction agents comprises sensor alarm logs, security report text and pre-shift meeting voice transcription data; The local knowledge extraction agent performs data preprocessing operation, converts a time sequence tuple in the sensor alarm log into text description with semantic information by using a preset natural language template, maps the security report text to a predefined text space, accesses the pre-class speech transcription data processed by an automatic speech recognition technology, and uniformly encodes homologous heterogeneous data into a text sequence with context enhancement.
3. The large language model based cross-mine multi-agent collaboration and semantic alignment system of claim 2, wherein the local knowledge extraction agent utilizes the coal industry large language model to extract the structured causal triplet comprising an abnormal event entity, a cause entity, and a result entity from the contextually enhanced text sequence; The confidence level screening mechanism comprises the steps of obtaining the self-evaluation confidence level when the large language model of the coal mine industry generates the structured causal triad, judging that the structured causal triad is valid and uploaded when the self-evaluation confidence level is higher than a preset confidence level threshold, and judging that the structured causal triad is invalid and discarded when the self-evaluation confidence level is not higher than the preset confidence level threshold.
4. The large language model-based cross-mine multi-agent collaboration and semantic alignment system of claim 1, wherein the semantic alignment agents map the structured causal triples to semantic vectors using a preset semantic encoder; the semantic alignment agent calculates the semantic similarity between the cross-mine triples by adopting cosine similarity as a measurement standard, wherein the semantic similarity is calculated by calculating dot products of the semantic vectors corresponding to the two structuring causal triples, and dividing the dot products by the product of the modular lengths of the two semantic vectors to obtain the semantic similarity for quantifying the semantic space direction consistency.
5. The large language model-based cross-mine multi-agent collaboration and semantic alignment system of claim 4, wherein the semantic alignment agent building or updating the coal mine field ontology graph according to the semantic similarity calculation result comprises: when the semantic similarity is larger than a preset semantic similarity judging threshold, judging that two structural causal triples describe the same physical phenomenon, and establishing a term equivalent mapping relation in the coal mine field ontology graph; And when the semantic similarity is smaller than or equal to the preset semantic similarity judging threshold, judging that the two structural causal triples describe different service scenes, and storing the corresponding terms as independent nodes in the coal mine field body map.
6. The large language model-based cross-mine multi-agent collaboration and semantic alignment system of claim 1, wherein the collaborative decision-making agent converts structured data in the abnormal event into the natural language questions conforming to human language habits using a preset prompt engineering template; The collaborative decision-making agent performs semantic expansion on the natural language problem by combining the coal mine field body map, retrieves synonymous terms with equivalent mapping relation with key entities in the natural language problem in the coal mine field body map, and packages the natural language problem and the synonymous terms into a structured query request.
7. The large language model based cross-mine multi-agent collaboration and semantic alignment system of claim 6, wherein the historical treatment policy contains treatment text for similar exception events, success rates of the treatments in historical applications, and failure case records; And the collaborative decision-making agent utilizes the preset weighted fusion model to comprehensively score the historical treatment strategies, sorts the historical treatment strategies according to the comprehensive scores, and selects the historical treatment strategy with the highest score to generate the safety suggestion.
8. The large language model-based cross-mine multi-agent collaboration and semantic alignment system of claim 7, wherein the preset weighted fusion model computes the composite score by: acquiring the success proportion of the history treatment strategy in a similar scene of the history; Calculating the compliance matching degree of the historical treatment strategy and prestored coal mine safety regulation clauses; Multiplying the success ratio by a preset experience weight coefficient to obtain a first value, multiplying the compliance matching degree by the complement of the experience weight coefficient to obtain a second value, and adding the first value and the second value to obtain the comprehensive score; the complement of the empirical weight coefficient is one minus the empirical weight coefficient.
9. The large language model-based cross-mine multi-agent collaboration and semantic alignment system of claim 8, wherein the way the collaborative decision-making agent calculates the compliance match is: carrying out semantic comparison on the text content of the historical treatment strategy and the coal mine safety regulation clause, calculating cosine similarity between the text vector of the historical treatment strategy and the vector of the most relevant coal mine safety regulation clause by utilizing a semantic vector model, and taking the cosine similarity as the compliance matching degree; the safety advice generated by the collaborative decision-making agent includes the selected historical treatment policy content, the corresponding success rate, and the matched coal mine safety code clause number.
10. The large language model-based cross-mine multi-agent cooperation and semantic alignment method applied to the large language model-based cross-mine multi-agent cooperation and semantic alignment system as claimed in any one of claims 1 to 9 is characterized by comprising the following steps: Accessing underground unstructured data, converting the underground unstructured data into structured causal triples by using a preset large language model of coal mine industry, and screening and uploading the structured causal triples meeting quality requirements based on a confidence screening mechanism; Receiving the structured causal triples, carrying out semantic vector coding on the structured causal triples, calculating semantic similarity among the cross-mine triples, and constructing or updating a coal mine field ontology graph according to the semantic similarity calculation result; When an abnormal event occurs in a mine, the abnormal event is converted into a natural language problem, semantic knowledge inquiry is initiated to other coal mines to obtain a history handling strategy, and a safety suggestion is generated by combining the coal mine field ontology map and a preset weighted fusion model.

Description

Cross-mine multi-agent collaboration and semantic alignment system based on large language model Technical Field The invention relates to the technical field of coal mine intellectualization and artificial intelligence, in particular to a cross-mine multi-agent cooperation and semantic alignment system based on a large language model. Background With the penetration of intelligent construction of coal mines, the processing of underground complex data by using a large language model has become an industry trend. The large language model is a deep learning-based generation type artificial intelligence technology, can understand and generate natural language, and has strong logic reasoning and text processing capability. The cross-mine multi-agent cooperation and semantic alignment refers to the realization of industry knowledge sharing and cooperation decision on the premise of not transmitting original data by utilizing autonomous agents distributed in different mines through a specific semantic interaction protocol. The existing coal mine intelligent system generally adopts a locally deployed supervised learning model, and utilizes the traditional natural language processing technology (such as Word2Vec and BERT) to process the sensor logs or security check texts in the mine so as to assist in equipment fault prediction and disaster early warning in the single mine range. However, due to the significant differences in geological conditions of coal mines and the strict data compliance requirements, raw sensor data or monitoring videos are difficult to directly communicate among different mines, so that each mine forms a data island, precious production experience is difficult to effectively migrate through model training, and the knowledge multiplexing rate is low. Meanwhile, different mines are affected by regional habits or equipment manufacturers, the term descriptions of the same physical phenomenon are severely heterogeneous, for example, roof subsidence and interlayer crack expansion are difficult to identify as equivalent concepts in the traditional method based on word vector similarity, and the prior art lacks semantic consistency modeling capability of structural causal relationships such as event-cause-result, so that model generalization is difficult. In addition, due to the lack of an effective automatic knowledge coordination mechanism, the emergency response of the cross-mine often depends on the manual comparison of historical cases by experts, the decision time is long, and the urgent demands of low delay and high accuracy in the sudden disaster scene are difficult to meet. Therefore, the invention provides a cross-mine multi-agent cooperation and semantic alignment system based on a large language model to solve the defects in the prior art. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a cross-mine multi-agent collaboration and semantic alignment system based on a large language model, which solves the problems of difficult sharing of cross-mine knowledge due to data compliance limitation, difficult model generalization due to multi-source term semantic isomerism and low response efficiency due to excessive dependence on manual experience in the existing coal mine intelligent scene. The large language model-based cross-mine multi-agent collaboration and semantic alignment system comprises a local knowledge extraction agent, a semantic alignment agent and a collaborative decision agent; The local knowledge extraction agent is connected with underground unstructured data, the underground unstructured data are converted into structured causal triples by using a preset large language model of coal mine industry, and the structured causal triples meeting quality requirements are screened based on a confidence screening mechanism and uploaded to the semantic alignment agent; The semantic alignment agent receives the structured causal triples, performs semantic vector coding on the structured causal triples, calculates semantic similarity among the cross-mine triples, and constructs or updates a coal mine field ontology graph according to the semantic similarity calculation result; When the collaborative decision-making agent generates an abnormal event in a mine, the abnormal event is converted into a natural language problem, semantic knowledge inquiry is initiated to the local knowledge extraction agent of other coal mines to obtain a history treatment strategy, and a safety suggestion is generated by combining the coal mine field ontology map and a preset weighted fusion model. Preferably, the underground unstructured data accessed by the local knowledge extraction agent comprises sensor alarm logs, security report texts and pre-class speech transcription data; The local knowledge extraction agent performs data preprocessing operation, converts a time sequence tuple in the sensor alarm log into text description with semantic information by using a pres