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

Abstract

The invention relates to the technical field of coal mine intellectualization and artificial intelligence, and discloses a cross-mine multi-agent collaboration and semantic alignment system based on a large language model, which comprises a local knowledge extraction agent, a semantic alignment agent and a collaborative decision agent. The local agent accesses underground unstructured data, converts the underground unstructured data into a structured causal triple by using a coal mine large model, and screens and uploads the structured causal triple based on confidence; performing vector coding and similarity calculation on the triple by the semantic alignment agent, and constructing a coal mine domain ontology map; the collaborative decision-making agent converts an abnormal event into a natural language problem, initiates semantic inquiry in a cross-mine manner to obtain a historical strategy, and generates a safety suggestion in combination with a graph. Through cooperation of a large model and multiple agents, efficient alignment and reuse 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.

Description

Technical Field

[0001] This invention relates to the fields of coal mine intelligence and artificial intelligence technology, specifically to a cross-mine multi-agent collaboration and semantic alignment system based on a large language model. Background Technology

[0002] With the deepening of intelligent coal mine construction, utilizing large language models to process complex underground data has become an industry trend. Large language models are a generative artificial intelligence technology based on deep learning, capable of understanding and generating natural language, and possessing powerful logical reasoning and text processing capabilities. Cross-mine multi-agent collaboration and semantic alignment refers to utilizing autonomous agents distributed across different mines to achieve industry knowledge sharing and collaborative decision-making through specific semantic interaction protocols without transmitting raw data. Current intelligent coal mine systems generally employ locally deployed supervised learning models, using traditional natural language processing technologies (such as Word2Vec and BERT) to process sensor logs or safety inspection texts within the mine to assist in equipment fault prediction and disaster early warning within a single mine.

[0003] However, due to significant differences in geological conditions across coal mines and stringent data compliance requirements, raw sensor data or monitoring videos are difficult to share directly between different mines, resulting in data silos. Valuable production experience cannot be effectively transferred through model training, leading to low knowledge reuse rates. Furthermore, influenced by regional customs or equipment manufacturers, different mines exhibit significant heterogeneity in their terminology for the same physical phenomena. For example, "roof subsidence" and "interlayer fracture propagation" are difficult to identify as equivalent concepts using traditional word vector similarity methods. Moreover, existing technologies lack the ability to model semantic consistency in structured causal relationships such as "event-cause-effect," making model generalization difficult. In addition, due to the lack of effective automated knowledge collaboration mechanisms, cross-mine emergency responses often rely on experts manually comparing historical cases, resulting in lengthy decision-making processes that fail to meet the urgent needs for low latency and high accuracy in sudden disaster scenarios.

[0004] Therefore, this invention proposes a cross-mine multi-agent collaboration and semantic alignment system based on a large language model to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a cross-mine multi-agent collaboration and semantic alignment system based on a large language model. This system solves the problems in existing intelligent coal mine scenarios, such as the difficulty in sharing knowledge across mines due to data compliance restrictions, the difficulty in model generalization caused by semantic heterogeneity of multi-source terms, and the low response efficiency caused by excessive reliance on human experience in emergency decision-making.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: a cross-mine multi-agent collaboration and semantic alignment system based on a large language model, including a local knowledge extraction agent, a semantic alignment agent, and a collaborative decision-making agent; The local knowledge extraction agent accesses unstructured underground data, uses a pre-set coal mining industry big language model to transform the unstructured underground data into structured causal triples, and uses a confidence screening mechanism to select structured causal triples that meet quality requirements and upload them to the semantic alignment agent. The semantic alignment agent receives the structured causal triples, performs semantic vector encoding on the structured causal triples, calculates the semantic similarity between triples across mines, and constructs or updates the ontology graph of the coal mine field based on the semantic similarity calculation results. When an abnormal event occurs in the mine, the collaborative decision-making agent converts the abnormal event into a natural language question, initiates semantic knowledge inquiry to the local knowledge extraction agents of other coal mines to obtain historical handling strategies, and generates safety recommendations by combining the coal mine domain ontology graph with a preset weighted fusion model.

[0007] Preferably, the unstructured downhole data accessed by the local knowledge extraction agent includes sensor alarm logs, safety inspection report text, and pre-shift meeting voice transcription data; The local knowledge extraction agent performs data preprocessing operations, using a preset natural language template to convert the time-series tuples in the sensor alarm logs into text descriptions with semantic information, mapping the security inspection report text to a predefined text space, and accessing the pre-shift meeting voice transcription data processed by automatic speech recognition technology, uniformly encoding homogeneous and heterogeneous data into a text sequence with context enhancement.

[0008] Preferably, the local knowledge extraction agent uses the coal mining industry big language model to extract the structured causal triples containing abnormal event entities, cause entities, and consequence entities from the context-enhanced text sequence. The confidence screening mechanism is as follows: obtain the self-assessment confidence level when the coal mining industry big language model generates the structured causal triples; when the self-assessment confidence level is higher than a preset confidence threshold, the structured causal triples are determined to be valid and uploaded; when the self-assessment confidence level is not higher than the preset confidence threshold, the structured causal triples are determined to be invalid and discarded.

[0009] Preferably, the semantic alignment agent uses a pre-set semantic encoder to map the structured causal triples into semantic vectors; The semantic alignment agent uses cosine similarity as a metric to calculate the semantic similarity between the cross-mine triples. The semantic similarity is calculated by: calculating the dot product of the semantic vectors corresponding to the two structured causal triples, and dividing the dot product by the product of the magnitudes of the two semantic vectors to obtain the semantic similarity that quantifies the consistency of the semantic space direction.

[0010] Preferably, the semantic alignment agent constructs or updates the coal mine domain ontology map based on the semantic similarity calculation results, including: When the semantic similarity is greater than the preset semantic similarity judgment threshold, it is determined that the two structured causal triples describe the same physical phenomenon, and a term equivalence mapping relationship is established in the coal mining domain ontology map; When the semantic similarity is less than or equal to the preset semantic similarity judgment threshold, it is determined that the two structured causal triples describe different business scenarios, and the corresponding terms are stored as independent nodes in the coal mine domain ontology graph.

[0011] Preferably, the semantic alignment agent constructs or updates the coal mine domain ontology map based on the semantic similarity calculation results, including: When the semantic similarity is greater than the preset semantic similarity judgment threshold, it is determined that the two structured causal triples describe the same physical phenomenon, and a term equivalence mapping relationship is established in the coal mining domain ontology map; When the semantic similarity is less than or equal to the preset semantic similarity judgment threshold, it is determined that the two structured causal triples describe different business scenarios, and the corresponding terms are stored as independent nodes in the coal mine domain ontology graph.

[0012] Preferably, the historical handling strategy includes text of handling measures for similar abnormal events, the success rate of the handling measures in historical applications, and records of failure cases; The collaborative decision-making agent uses the preset weighted fusion model to comprehensively score the historical handling strategies, sorts the historical handling strategies according to the comprehensive score, and selects the historical handling strategy with the highest score to generate the security recommendation.

[0013] Preferably, the preset weighted fusion model calculates the comprehensive score in the following way: Obtain the success rate of the historical handling strategy in similar historical scenarios; Calculate the compliance matching degree between the historical disposal strategy and the pre-stored coal mine safety regulations clauses; The first value is obtained by multiplying the success rate by the preset experience weight coefficient, the second value is obtained by multiplying the compliance matching degree by the complement of the experience weight coefficient, and the comprehensive score is obtained by adding the first value and the second value. The complement of the empirical weight coefficient is one minus the empirical weight coefficient.

[0014] Preferably, the preset weighted fusion model calculates the comprehensive score in the following way: Obtain the success rate of the historical handling strategy in similar historical scenarios; Calculate the compliance matching degree between the historical disposal strategy and the pre-stored coal mine safety regulations clauses; The first value is obtained by multiplying the success rate by the preset experience weight coefficient, the second value is obtained by multiplying the compliance matching degree by the complement of the experience weight coefficient, and the comprehensive score is obtained by adding the first value and the second value. The complement of the empirical weight coefficient is one minus the empirical weight coefficient.

[0015] This invention also provides a cross-mine multi-agent collaboration and semantic alignment method based on a large language model, comprising the following steps: Access unstructured underground data, use a pre-set coal mining industry big language model to transform the unstructured underground data into structured causal triples, and use a confidence screening mechanism to screen the structured causal triples that meet the quality requirements and upload them. The structured causal triples are received, semantic vector encoding is performed on the structured causal triples, and the semantic similarity between triples across mines is calculated. Based on the semantic similarity calculation results, an ontology map of the coal mining field is constructed or updated. When an abnormal event occurs in the mine, the abnormal event is converted into a natural language question, and semantic knowledge queries are initiated to other coal mines to obtain historical handling strategies. Safety recommendations are then generated by combining the coal mine domain ontology map with a preset weighted fusion model.

[0016] This invention provides a cross-mine multi-agent collaborative and semantic alignment system based on a large language model. It has the following beneficial effects: 1. This invention utilizes a local knowledge extraction agent and a large language model to transform multi-source unstructured data from underground mines into structured causal triples. A confidence-based filtering mechanism is then employed to upload only highly reliable logical knowledge. This mechanism, which exchanges only semantic causal knowledge rather than raw data, effectively breaks down data silos across mines while adhering to the coal mining industry's compliance requirement that data not leave the mine. It achieves secure data storage and efficient flow of industry knowledge, solving the data interoperability problem caused by privacy restrictions in traditional supervised learning models.

[0017] 2. This invention utilizes a semantic alignment agent to perform semantic vector encoding and similarity calculation on structured causal triples, constructing and dynamically updating an ontology graph in the coal mining field that includes terminological equivalence mapping relationships. This mechanism can identify the same physical phenomenon behind different expressions across mines, effectively solving the problem of semantic inconsistency in terminology caused by regional customs or equipment differences. It overcomes the shortcomings of traditional word vector methods in handling the equivalence of structured causal logic, significantly improving the generalization ability and reusability of causal knowledge in heterogeneous scenarios.

[0018] 3. This invention transforms abnormal events into natural language questions through a collaborative decision-making intelligent agent and initiates cross-mine semantic queries. It then combines a coal mine ontology graph with a weighted fusion model to comprehensively evaluate historical handling strategies. This multi-agent collaborative mechanism changes the inefficient traditional model that relies on manual comparison of historical cases. It enables newly built mines lacking historical data to quickly acquire industry-wide validated handling experience, significantly shortening emergency response time while generating highly interpretable and safety-compliant decision recommendations, thus ensuring production safety. Attached Figure Description

[0019] Figure 1 This is a diagram of the cross-mine multi-agent collaboration and semantic alignment system architecture based on a large language model, as presented in this invention. Figure 2 This is a flowchart of the semantic alignment and graph construction of cross-mine causal triples in this invention; Figure 3 This is a flowchart of the cross-mine multi-agent collaboration and semantic alignment method based on a large language model according to the present invention. Among them, 100 is the local knowledge extraction agent; 200 is the semantic alignment agent; and 300 is the collaborative decision-making agent. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See attached document Figure 1This invention provides a cross-mine multi-agent collaborative and semantic alignment system based on a large language model, including a local knowledge extraction agent 100, a semantic alignment agent 200, and a collaborative decision-making agent 300. The local knowledge extraction agent 100 is deployed on coal mine edge computing nodes. The semantic alignment agent 200 is deployed on a cloud platform. Each coal mine edge computing node communicates and connects to the cloud platform via a network. The local knowledge extraction agent 100 is configured to connect to underground data sources and perform causal triple generation tasks. The semantic alignment agent 200 is configured to maintain a coal mine domain ontology graph and perform semantic standardization tasks. The collaborative decision-making agent 300 is configured to connect the local knowledge extraction agent 100 and the semantic alignment agent 200 and perform policy generation tasks.

[0022] See attached document Figure 3 This invention provides a cross-mine multi-agent collaboration and semantic alignment method based on a large language model, comprising the following steps: S10, Local knowledge extraction agent 100 accesses unstructured underground data, uses a pre-set coal mining industry big language model to transform the unstructured underground data into structured causal triples, and uses a confidence screening mechanism to select structured causal triples that meet the quality requirements and upload them to semantic alignment agent 200. S20, the semantic alignment agent 200 receives structured causal triples, performs semantic vector encoding on the structured causal triples and calculates the semantic similarity between triples across mines, and constructs or updates the ontology graph of the coal mine field based on the semantic similarity calculation results. S30, when an abnormal event occurs in the mine, the collaborative decision-making agent 300 converts the abnormal event into a natural language question, initiates a semantic knowledge inquiry to the local knowledge extraction agent 100 of other coal mines to obtain historical handling strategies, and generates safety suggestions by combining the coal mine domain ontology graph with a preset weighted fusion model.

[0023] See attached document Figure 1 The local knowledge extraction agent 100 is deployed on the edge computing node of the coal mine. Its specific implementation mainly involves three parts: access and standardization processing of unstructured data in the mine, causal triple extraction based on industry large model, and knowledge filtering based on confidence.

[0024] The local knowledge extraction agent 100 first accesses unstructured data from underground. This unstructured data includes at least sensor alarm logs, safety inspection report text, and pre-shift meeting voice transcription data. To bridge the semantic gap between different modalities of data, the local knowledge extraction agent 100 performs data preprocessing, uniformly encoding the aforementioned homogeneous but heterogeneous data into a text sequence with context enhancement.

[0025] For sensor alarm logs, the local knowledge extraction agent 100 parses them into a set of time-series tuples. Let the sensor alarm log set be... Its mathematical expression is defined as: ; in, Indicates the specific time when the alarm occurred; Indicates the sensor type that triggered the alarm; This indicates the specific observation value collected by the sensor at that moment; This represents the total number of data records. The local knowledge extraction agent 100 uses a preset natural language template to transform the aforementioned discrete numerical tuples into textual descriptions with semantic information. For example, for the tuple... The system converts it into text form: "sensor" In time The value is This transformation gives numerical data a clear context.

[0026] For the security inspection report text, the local knowledge extraction agent 100 directly reads the document content, denoted as... ,in Belongs to a predefined text space This text space This encompasses a collection of all characters, words, and syntactic structures used in coal mine safety inspection records. For the pre-shift meeting audio transcription data, a local knowledge extraction agent 100 accesses the text sequence processed by Automatic Speech Recognition (ASR) technology. Let the original pre-shift meeting audio data be... The processed speech-transcribed text is denoted as It is represented as: ; For the specific implementation of automatic speech recognition technology, those skilled in the art can use existing speech-to-text models, which will not be elaborated here.

[0027] After transforming various types of unstructured underground data into text sequences with a unified context, the local knowledge extraction agent 100 uses a pre-defined coal mining industry large-scale language model to perform reasoning analysis on the text sequences. The pre-defined coal mining industry large-scale language model can be a specialized model trained on industry data, such as the Sunstone Mine large-scale model. Based on a general large-scale language model architecture, this model has undergone supervised fine-tuning using a large-scale professional corpus in the coal mining field. During the training phase, this coal mining industry large-scale language model simultaneously optimizes two tasks: Named Entity Recognition (NER) and Causal Relationship Classification, thereby possessing the ability to accurately identify specific entities and their logical relationships from mixed contexts.

[0028] A local knowledge extraction agent 100 uses a large language model specific to the coal mining industry to extract structured causal triples from the input text sequence. The output structured causal triples are... Defined as: ; in, Represents the identified anomalous event entity; Represents the entity that caused the event; This represents the entity that causes the consequences of the event. This structured causal triple constitutes the basic knowledge unit for cross-mine collaboration.

[0029] To ensure the accuracy of uploaded data and filter out low-quality semantic noise, the local knowledge extraction agent 100 executes a confidence-based filtering mechanism. This mechanism judges the confidence level based on the self-assessed confidence level of the coal mining industry large language model when generating structured causal triples. Let's assume the coal mining industry large language model evaluates the generated... A structured causal triple The self-assessment credibility is The system's confidence threshold is set at 75%. The specific filtering logic is as follows: ; When the self-rated credibility of the structured causal triples generated by the large language model of the coal mining industry is... When the self-assessed confidence level is above 75%, the local knowledge extraction agent 100 determines the triple to be valid and uploads it to the semantic alignment agent 200; when the self-assessed confidence level is not higher than 75%, the local knowledge extraction agent 100 determines the triple to be invalid and discards it. Through this mechanism, the local knowledge extraction agent 100 can share only high-quality structured knowledge without uploading the original data.

[0030] See attached document Figure 1 and Figure 2 The semantic alignment agent 200 is deployed on a cloud platform. Its specific implementation mainly involves semantic vector encoding of structured causal triples, semantic similarity calculation between triples across mines, and construction and updating of ontology graphs in the coal mining field.

[0031] The semantic alignment agent 200 receives set of structured causal triples uploaded by edge computing nodes in various coal mines via the network. In actual coal mine production scenarios, different mines often have inconsistent descriptions of the same physical phenomenon due to differences in equipment manufacturers, regional customs, etc. For example, for the same type of geological change in roof collapse disasters, Mine A may describe it as "roof subsidence," while Mine B may describe it as "interlayer separation." To solve the problem of knowledge reusability caused by the above-mentioned inconsistency in terminology and to provide a unified semantic benchmark for subsequent collaborative decision-making, the semantic alignment agent 200 performs semantic vector encoding on the received structured causal triples.

[0032] The semantic alignment agent 200 uses a pre-built semantic encoder to map structured causal triples in text form into high-dimensional real vectors. For any structured causal triple... The system converts it into a corresponding semantic vector. This process transforms discrete symbolic information into a continuous vector space representation, enabling computers to calculate the semantic distance between different terms. For the specific network structure of the semantic encoder, those skilled in the art can implement it using existing deep learning language models (such as BERT or the Transformer architecture), which are well-known technologies in the field and will not be elaborated upon here.

[0033] Based on the generated semantic vectors, the semantic alignment agent 200 calculates the semantic similarity between cross-mine triples. The system uses cosine similarity as the metric. Let two structured causal triples from different mines be... and Their corresponding semantic vectors are respectively and Then the semantic similarity between the two The calculation formula is as follows: ; in, Represents the dot product operation of vectors; This represents the magnitude of the vector. This formula quantifies the directional similarity of two triples in the semantic space; the closer the output value is to 1, the more semantically similar they are.

[0034] The semantic alignment agent 200 constructs or updates a coal mine ontology graph based on the calculated semantic similarity results between cross-mine triples. The coal mine ontology graph is a structured knowledge base storing standard terminology entities and their semantic relationships within the coal mine safety production field. Nodes represent specific business terms (such as equipment names, disaster types, and operational actions), and edges represent logical relationships between terms (such as causal relationships and equivalence relationships). The system sets a semantic similarity threshold. The similarity score is 0.85. The system compares the newly received structured causal triples with the existing knowledge nodes in the coal mining domain ontology graph, and performs graph construction or update operations based on the comparison results.

[0035] When the calculated semantic similarity When the value is greater than the threshold of 0.85, the semantic alignment agent 200 determines the triple. and Both describe the same physical phenomenon and are semantically equivalent. In this case, the semantic alignment agent 200 performs a graph update operation, establishing terminological equivalence mappings in the coal mine ontology graph. For example, when the system detects that the semantic similarity between "initial support force" and "hydraulic support pressure" exceeds 0.85, or that the semantic similarity between "abnormal gas concentration" and "exceeding borehole monitoring standards" exceeds 0.85, the system establishes a bidirectional mapping link between these terms in the coal mine ontology graph (i.e., "initial support force"). "Hydraulic support pressure"), thereby achieving cross-mine semantic standardization.

[0036] When the calculated semantic similarity When the threshold value is less than or equal to 0.85, the semantic alignment agent 200 determines that the two terms describe different business scenarios. The semantic alignment agent 200 then performs a graph expansion operation, storing the corresponding terms as independent nodes in the coal mining domain ontology graph. Through this continuous comparison and mapping process, the semantic alignment agent 200 gradually constructs and maintains a coal mining domain ontology graph containing standard terms and equivalent mapping relationships. This solved the problem of semantic heterogeneity of data across mines.

[0037] See attached document Figure 1 The collaborative decision-making intelligent agent 300 is deployed on a cloud platform or core scheduling node. Its specific implementation mainly involves semantic inquiry of abnormal events, cross-mine acquisition of historical handling strategies, and generation of security suggestions based on a preset weighted fusion model.

[0038] The collaborative decision-making agent 300 monitors and receives data in real time. When an abnormal event occurs in a single coal mine and the local knowledge base lacks a corresponding contingency plan, the collaborative decision-making agent 300 initiates a cross-mine collaborative decision-making process. The collaborative decision-making agent 300 first converts the received structured abnormal event data into natural language questions. This process utilizes pre-set prompting templates to transform structured triples containing time, location, and numerical values ​​of the phenomenon into interrogative sentences that conform to human language habits. For example, regarding the event of "a sudden increase in gas concentration in the return airway," the system converts it into a natural language question: "In the event of a sudden increase in gas concentration in the return airway, what effective contingency measures have been implemented historically?"

[0039] To ensure that natural language questions can be accurately understood by coal mine systems with different terminology systems, the collaborative decision-making agent 300 semantically expands the natural language questions by combining them with the coal mine domain ontology graph. The collaborative decision-making agent 300 searches the coal mine domain ontology graph to identify synonyms or equivalent terms for key entities in the natural language questions (e.g., expanding "gas surge" to "abnormal gas outburst" or "methane exceedance"). The collaborative decision-making agent 300 encapsulates the natural language questions and expanded terms into a structured query request and initiates a semantic knowledge inquiry to the local knowledge extraction agents 100 of all other coal mines in the network.

[0040] The collaborative decision-making agent 300 receives historical handling strategies from local knowledge extraction agents 100 in other coal mines. These historical handling strategies include textual descriptions of specific handling measures for similar abnormal events, the success rate of those measures in historical applications, and records of related failure cases.

[0041] To select the most suitable and compliant solution for the current scenario from multiple candidate historical handling strategies, the collaborative decision-making agent 300 uses a pre-defined weighted fusion model to comprehensively score each submitted historical handling strategy. The pre-defined weighted fusion model is a mathematical evaluation algorithm that quantifies and ranks the feasibility of strategies based on multi-dimensional indicators. It balances the relationship between historical data and safety regulatory constraints by introducing empirical weight coefficients.

[0042] Let the first The historical handling strategy for each candidate is Its overall evaluation score The calculation formula is as follows: ; in, This strategy The success rate in similar historical scenarios, this value is derived from statistics compiled by the source coal mine that provided feedback on this strategy based on historical operation and maintenance records, and the range is [missing value]. ; This strategy The degree of matching with the current "Coal Mine Safety Regulations" is within the range of [value range missing]. ; This refers to the empirical weighting coefficient. In this embodiment, the empirical weighting coefficient... The value is set to 0.7, which is the optimal parameter derived from a large number of field tests. This indicates that the system assigns 70% of the weight to verified historical buying experience when making decisions, while reserving 30% of the weight to ensure the compliance of the strategy.

[0043] Regarding the compliance matching degree in the formula The computation of the collaborative decision-making intelligent agent 300 will incorporate historical handling strategies. The system performs a semantic comparison between the text content and a pre-stored database of standard clauses in the "Coal Mine Safety Regulations". It uses a semantic vector model to calculate the cosine similarity between the strategy text vector and the vector of the most relevant regulation clause, and uses this similarity value as the basis for its calculation. The value. This step ensures that the recommended strategy is not only historically effective but also strictly complies with safety regulations, avoiding the risk of unauthorized command.

[0044] The collaborative decision-making agent 300 calculates a comprehensive evaluation score. All candidate historical handling strategies are ranked, and the strategy with the highest score is selected as the basis for generating a safety recommendation. The safety recommendation is an interpretable text containing the decision-making basis. The collaborative decision-making agent 300 combines the selected strategy content, corresponding historical success rate data, and matching clause numbers from the *Coal Mine Safety Regulations* to output the final safety recommendation. For example, the output safety recommendation is: "Recommendation: Immediately increase the density of hydraulic supports to..." 0.8m / frame. Reason: Similar cases at Mine B have a 90% success rate and comply with Article 12.3 of the Regulations. The Collaborative Decision-Making Intelligent Agent 300 will send this safety recommendation to the requesting console to assist on-site commanders in making scientific decisions.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cross-mine multi-agent collaborative and semantic alignment system based on a large language model, characterized in that, This includes local knowledge extraction agents, semantic alignment agents, and collaborative decision-making agents; The local knowledge extraction agent accesses unstructured underground data, uses a pre-set coal mining industry big language model to transform the unstructured underground data into structured causal triples, and uses a confidence screening mechanism to select structured causal triples that meet quality requirements and upload them to the semantic alignment agent. The semantic alignment agent receives the structured causal triples, performs semantic vector encoding on the structured causal triples, calculates the semantic similarity between triples across mines, and constructs or updates the ontology graph of the coal mine field based on the semantic similarity calculation results. When an abnormal event occurs in the mine, the collaborative decision-making agent converts the abnormal event into a natural language question, initiates semantic knowledge inquiry to the local knowledge extraction agents of other coal mines to obtain historical handling strategies, and generates safety recommendations by combining the coal mine domain ontology graph with a preset weighted fusion model.

2. The cross-mine multi-agent collaborative and semantic alignment system based on a large language model according to claim 1, characterized in that, The unstructured downhole data accessed by the local knowledge extraction agent includes sensor alarm logs, safety inspection report text, and pre-shift meeting voice transcription data. The local knowledge extraction agent performs data preprocessing operations, using a preset natural language template to convert the time-series tuples in the sensor alarm logs into text descriptions with semantic information, mapping the security inspection report text to a predefined text space, and accessing the pre-shift meeting voice transcription data processed by automatic speech recognition technology, uniformly encoding homogeneous and heterogeneous data into a text sequence with context enhancement.

3. The cross-mine multi-agent collaborative and semantic alignment system based on a large language model according to claim 2, characterized in that, The local knowledge extraction agent uses the coal mining industry big language model to extract the structured causal triples containing abnormal event entities, cause entities, and consequence entities from the context-enhanced text sequence. The confidence screening mechanism is as follows: obtain the self-assessment confidence level when the coal mining industry big language model generates the structured causal triples; when the self-assessment confidence level is higher than a preset confidence threshold, the structured causal triples are determined to be valid and uploaded; when the self-assessment confidence level is not higher than the preset confidence threshold, the structured causal triples are determined to be invalid and discarded.

4. The cross-mine multi-agent collaborative and semantic alignment system based on a large language model according to claim 1, characterized in that, The semantic alignment agent uses a pre-set semantic encoder to map the structured causal triples into semantic vectors; The semantic alignment agent uses cosine similarity as a metric to calculate the semantic similarity between the cross-mine triples. The semantic similarity is calculated by: calculating the dot product of the semantic vectors corresponding to the two structured causal triples, and dividing the dot product by the product of the magnitudes of the two semantic vectors to obtain the semantic similarity that quantifies the consistency of the semantic space direction.

5. The cross-mine multi-agent collaborative and semantic alignment system based on a large language model according to claim 4, characterized in that, The semantic alignment agent constructs or updates the coal mine domain ontology map based on the semantic similarity calculation results, including: When the semantic similarity is greater than the preset semantic similarity judgment threshold, it is determined that the two structured causal triples describe the same physical phenomenon, and a term equivalence mapping relationship is established in the coal mining domain ontology map; When the semantic similarity is less than or equal to the preset semantic similarity judgment threshold, it is determined that the two structured causal triples describe different business scenarios, and the corresponding terms are stored as independent nodes in the coal mine domain ontology graph.

6. The cross-mine multi-agent collaborative and semantic alignment system based on a large language model according to claim 1, characterized in that, The collaborative decision-making agent uses a pre-set prompting engineering template to transform the structured data in the abnormal event into a natural language question that conforms to human language habits. The collaborative decision-making agent combines the coal mining domain ontology graph to semantically expand the natural language problem, retrieves synonyms in the coal mining domain ontology graph that have an equivalent mapping relationship with the key entities in the natural language problem, and encapsulates the natural language problem and the synonyms into a structured query request.

7. The cross-mine multi-agent collaborative and semantic alignment system based on a large language model according to claim 6, characterized in that, The historical handling strategy includes textual descriptions of handling measures for similar abnormal events, the success rate of the handling measures in historical applications, and records of failure cases. The collaborative decision-making agent uses the preset weighted fusion model to comprehensively score the historical handling strategies, sorts the historical handling strategies according to the comprehensive score, and selects the historical handling strategy with the highest score to generate the security recommendation.

8. The cross-mine multi-agent collaborative and semantic alignment system based on a large language model according to claim 7, characterized in that, The preset weighted fusion model calculates the comprehensive score in the following way: Obtain the success rate of the historical handling strategy in similar historical scenarios; Calculate the compliance matching degree between the historical disposal strategy and the pre-stored coal mine safety regulations clauses; The first value is obtained by multiplying the success rate by the preset experience weight coefficient, the second value is obtained by multiplying the compliance matching degree by the complement of the experience weight coefficient, and the comprehensive score is obtained by adding the first value and the second value. The complement of the empirical weight coefficient is one minus the empirical weight coefficient.

9. The cross-mine multi-agent collaborative and semantic alignment system based on a large language model according to claim 8, characterized in that, The collaborative decision-making agent calculates the compliance matching degree in the following way: The text content of the historical disposal strategy is semantically compared with the clauses of the coal mine safety regulations. The cosine similarity between the text vector of the historical disposal strategy and the vector of the most relevant clause of the coal mine safety regulations is calculated using a semantic vector model. The cosine similarity is used as the compliance matching degree. The safety recommendations generated by the collaborative decision-making agent include the selected historical handling strategies, the corresponding success rates, and the matching coal mine safety regulation clause numbers.

10. A cross-mine multi-agent collaboration and semantic alignment method based on a large language model, applied to the cross-mine multi-agent collaboration and semantic alignment system based on a large language model as described in any one of claims 1-9, characterized in that, Includes the following steps: Access unstructured underground data, use a pre-set coal mining industry big language model to transform the unstructured underground data into structured causal triples, and use a confidence screening mechanism to screen the structured causal triples that meet the quality requirements and upload them. The structured causal triples are received, semantic vector encoding is performed on the structured causal triples, and the semantic similarity between triples across mines is calculated. Based on the semantic similarity calculation results, an ontology map of the coal mining field is constructed or updated. When an abnormal event occurs in the mine, the abnormal event is converted into a natural language question, and semantic knowledge queries are initiated to other coal mines to obtain historical handling strategies. Safety recommendations are then generated by combining the coal mine domain ontology map with a preset weighted fusion model.

Images