A coal mine safety regulation processing method, system, device and readable storage medium

Abstract

The present application provides a kind of coal mine safety regulation processing method, system, equipment and readable storage medium, the present application is aimed at the problem that low computing power equipment cannot process super-long industrial regulation, big model is easy to produce illusion and cause safety production accident, adopts master-slave multi-agent architecture, and the coal mine safety regulation is cut according to professional field physics and constructs multiple independent field vector knowledge base;By master agent, safety query is received and analyzed, is disassembled into retrieval subtask and is issued to corresponding field slave agent;After each slave agent retrieves and generates preliminary reply based on exclusive knowledge base, the master agent generates final safety decision reply by rule engine according to the preset coal mine safety business rule for conflict detection and logical arbitration of multiple preliminary replies.This application can realize complete retrieval of super-long regulation document under limited hardware conditions, effectively avoid big model illusion, improve the reliability and safety of mine intelligent question and answer system.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and natural language processing technology, specifically to a method, system, equipment, and readable storage medium for processing coal mine safety regulations. Background Technology

[0002] In the process of intelligent construction of coal mines, using large language models to provide intelligent question-and-answer services for coal mine safety regulations and related operation guidelines is an important technical means to improve the efficiency of on-site decision-making and reduce violations. Currently, the mainstream implementation solution in the industry adopts retrieval-enhanced generation technology, which inputs the complete coal mine safety regulations into a vector database. When a user asks a question, the model retrieves relevant paragraphs and generates a response based on the large language model. However, underground and edge-side computer rooms in coal mines typically only have limited computing power to deploy equipment, while the full text of the coal mine safety regulations is quite long, covering safety clauses from multiple professional fields. Its text length far exceeds the context window capacity of a large model with limited video memory. Directly feeding the entire regulations into the model would cause video memory overflow, leading to system malfunction. If a text truncation strategy is used to compress the input length, the latter part of the document cannot be effectively retrieved by the model, resulting in the omission of key safety clauses and failing to guarantee the comprehensiveness and accuracy of the question-and-answer results, making it difficult to meet the compliance query needs of underground operations.

[0003] In addition, existing single-model retrieval enhancement generation schemes have inherent security flaws. The underlying logic of large language models is based on a probabilistic generation model, which inherently conflicts with the deterministic red-line regulations required in the coal mine safety field. This makes them highly susceptible to multiple fatal errors when handling complex cross-domain problems. Firstly, there is the problem of information silos: the model only retrieves domain documents directly related to the query action, ignoring restrictive clauses in related domains, leading to compliance loopholes in the output conclusions. Secondly, there is the problem of conflicting answers and misjudgments of priority: when the requirements of different professional fields conflict, especially in complex scenarios involving multiple disasters, a single model cannot determine the correct order of action based on professional safety logic, easily providing incorrect guidance that could lead to secondary accidents. Thirdly, there is the problem of missing key constraints: some special operations have sub-domain prohibitions that can lead to veto power; large models do not pay enough attention to these marginal clauses, often outputting incorrect work permit conclusions. Existing technologies cannot simultaneously solve the two core challenges of low-computing-power deployment and high security and reliability, making it difficult to meet the practical application needs of the coal mine industry. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a coal mine safety procedure processing method, system, equipment, and readable storage medium. This method reduces reliance on video memory by physically cutting up excessively long documents and, by introducing an arbitration mechanism for a main intelligent agent containing industry-mandated rules, forcibly suppresses the logical illusions of large language models, thereby ensuring the absolute reliability of coal mine safety decisions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for processing coal mine safety regulations, running on an edge computing device including a processor, memory, a large language model inference engine, and a rule engine module, the method comprising the following steps:

[0007] S1. Document splitting and knowledge base construction: Using the processor to run a physical splitting algorithm, the long text of coal mine safety regulations is physically divided into multiple independent industry sub-documents according to professional fields, and vectorized by using a vector embedding model to construct and store multiple independent domain vector knowledge bases accordingly.

[0008] S2. Reception and Intent Parsing: The main intelligent agent module receives the input coal mine safety query text request through the user interface, performs semantic analysis through the large language model inference engine, identifies all safety domains involved in the query text request, and uses the task allocation algorithm to decompose it into corresponding retrieval subtasks.

[0009] S3. Divide and conquer retrieval from agents: The main agent module distributes the retrieval subtasks to the corresponding domain agents through the internal bus; each domain agent only loads its own domain vector knowledge base to perform semantic retrieval, and combines the prompt words parameters of its own domain to generate a preliminary response containing the normative constraints of its domain through the large language model inference engine;

[0010] S4. Rule Fusion and Arbitration Output: The main agent module receives the preliminary responses returned by all called agents and calls the rule engine module to perform semantic feature extraction, conflict detection and logical arbitration on each preliminary response based on the preset coal mine safety business rules, and generates and outputs the final safety decision response.

[0011] Among them, the preset coal mine safety business rules include at least prohibition priority rules and multi-hazard coupling priority rules.

[0012] Preferably, in step S1, the professional field is one or more of the following: ventilation, gas, outburst prevention, fire prevention and extinguishing, water control, rock burst prevention, blasting, transportation, electrical, and mining.

[0013] Preferably, the prohibition priority rule in step S4 refers to the rule engine module extracting text features from all preliminary responses. When any preliminary response is found to contain negative security constraint feature words such as "strictly prohibited," "must not," or "forbidden," the main agent module resets the logical weight of the sentence containing the constraint feature word to the highest, filters out the permissive responses from other agents, and directly outputs the final security decision centered on the negative security constraint condition.

[0014] Preferably, the multi-hazard coupling priority rule in step S4 refers to the following: when multiple disaster response measures given by the intelligent agent conflict in execution timing or logic, the main intelligent agent module calls a preset disaster lethality sequence for arbitration to determine the order of operations; the priority of the disaster lethality sequence from high to low is as follows: gas treatment, outburst prevention, fire prevention and extinguishing, rockburst prevention, water prevention and control, and other routine operations.

[0015] Preferably, the preset coal mine safety business rules also include the strictest standard rules. Specifically, when multiple preliminary responses from intelligent agents contain numerical threshold requirements for the same safety parameter, the rule engine module automatically extracts the numerical boundaries from all preliminary responses. If the safety parameter is an upper limit constraint, the minimum value among all candidate upper limit values ​​is selected as the final safety decision indicator. If the safety parameter is a lower limit constraint, the maximum value among all candidate lower limit values ​​is selected as the final safety decision indicator.

[0016] Preferably, in step S3, each of the domain agents is executed serially on the same computing node using time-sharing multiplexing and memory replacement mechanisms; the main agent module calls the corresponding agents sequentially according to the time-series queue of the retrieval subtasks; when calling the target agent, the corresponding domain vector knowledge base context and current model weights are loaded into the memory; after the target agent generates a preliminary response, the memory space occupied by the target agent and the large language model key-value cache are released, and then the context of the next target agent to be triggered is loaded.

[0017] A coal mine safety procedure processing system based on master-slave multi-agent and long document splitting includes a preprocessing module, a master agent module, a slave agent cluster module and a rule engine module.

[0018] The preprocessing module is used to divide the complete coal mine safety regulations into independent domain slices according to professional fields and construct multiple independent domain vector knowledge bases.

[0019] The main intelligent agent module is deployed on the edge computing device as a scheduling hub, used to receive user questions, identify the relevant professional fields, distribute sub-tasks, and perform forced logical arbitration and text integration on the received opinions of the subordinate intelligent agents through the rule engine module;

[0020] The slave agent cluster module contains multiple independent domain slave agents. Each domain slave agent is bound to and retrieves its corresponding single domain vector knowledge base to process the sub-tasks issued by the master agent module and generate domain-specific constraints to be fed back to the master agent module.

[0021] The rule engine module stores preset arbitration logic based on the hard safety constraints of coal mines, which is used to perform conflict detection and priority filtering on the initial responses from the intelligent agents.

[0022] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the coal mine safety regulation processing method.

[0023] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the coal mine safety regulation processing method described above.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] By physically splitting long documents and constructing vector knowledge bases by domain, the global long text retrieval task is decomposed into multiple short text domain retrieval tasks. This significantly reduces the hardware memory requirements of a single round of reasoning, enabling complete and untrunculated semantic retrieval of ultra-long procedure documents on low-computing-power edge devices, covering all security clauses, and effectively reducing the hardware deployment cost of intelligent question-answering systems in coal mining enterprises.

[0026] By employing a hierarchical architecture of master-slave multi-agent systems and a hard arbitration mechanism of the rule engine, this approach breaks away from the traditional single-model approach that relies entirely on the reasoning of a large language model. The master agent verifies the responses of all slave agents using a prohibition-first rule. If a prohibition clause exists in any domain, a prohibition conclusion is directly output. This eliminates the risk of overlooking critical safety constraints at the system architecture level, ensuring that the output strictly complies with the safety red line requirements of coal mines.

[0027] By setting multi-hazard coupling priority rules, disaster response measures with temporal or logical conflicts in different fields are uniformly arbitrated and sorted, clarifying the logic of response sequence in multi-hazard concurrent scenarios, solving the problem of chaotic model decision-making in complex disaster scenarios, and ensuring that the output response plan conforms to the real rescue safety logic of coal mine sites.

[0028] By setting the strictest standard rules, the threshold values ​​of similar security parameters in responses from different fields are uniformly converged, and the most stringent numerical standard is automatically selected as the final decision-making basis, eliminating standard differences across fields and avoiding security risks caused by improper standard selection.

[0029] By employing a time-sharing and memory replacement mechanism for agents, agents from different domains are loaded and run sequentially on the same computing node. Each round loads only the knowledge base and model context corresponding to the current task, and memory resources are released immediately after the task is completed. This further reduces the hardware computing power threshold for system operation and improves the operational stability of low-computing-power devices. Attached Figure Description

[0030] Figure 1 This is a diagram of the overall architecture of the present invention;

[0031] Figure 2 This is a flowchart illustrating the overall processing of the method of the present invention;

[0032] Figure 3 This is a flowchart of the arbitration logic of the present invention;

[0033] Figure 4 This is a timing diagram of time-sharing multiplexing of the intelligent agent's memory in this invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1 To be continued Figure 4 The invention will be further described in detail with reference to specific embodiments and examples. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0035] A coal mine safety procedure processing method based on master-slave multi-agent and long document segmentation, running on an edge computing device including a processor, memory, a large language model inference engine, and a rule engine module, includes the following steps:

[0036] S1. Document splitting and knowledge base construction: Using the processor to run a physical splitting algorithm, the long text "Coal Mine Safety Regulations" is physically divided into multiple independent industry sub-documents according to professional fields, and vectorized by using a vector embedding model to construct and store multiple independent domain vector knowledge bases accordingly.

[0037] The professional fields are divided into one or more combinations of ventilation, gas, outburst prevention, fire prevention and extinguishing, water control, rock burst prevention, blasting, transportation, electrical, and mining fields;

[0038] S2. Reception and Intent Parsing: The main intelligent agent module receives the input coal mine safety query text request through the user interface, performs semantic analysis through the large language model inference engine, identifies all potential safety areas involved in the query text request, and uses the task allocation algorithm to decompose it into corresponding retrieval subtasks.

[0039] S3. Divide and conquer retrieval from agents: The main agent module distributes the retrieval subtasks to the corresponding domain agents through the internal bus; each domain agent only loads its own domain vector knowledge base to perform semantic retrieval, and combines the prompt words parameters of its own domain to generate a preliminary response containing the normative constraints of its domain through the large language model inference engine;

[0040] Each domain agent is executed serially on the same computing node using time-sharing multiplexing and memory replacement mechanisms: the main agent module calls the corresponding slave agents in sequence according to the time-series queue of the retrieval subtasks; when calling the target slave agent, the corresponding domain vector knowledge base context and current model weights are loaded into the memory; after the target slave agent generates a preliminary response, the memory space occupied by the target slave agent and the large language model key-value cache KVCache are released, and then the context of the next target slave agent to be triggered is loaded.

[0041] S4. Rule Fusion and Arbitration Output: The main agent module receives the preliminary responses returned by all called agents and calls the rule engine module to perform semantic feature extraction, conflict detection and logical arbitration on each preliminary response based on the preset coal mine safety business rules, and generates and outputs the final safety decision response.

[0042] The prohibition priority rule means that the rule engine module extracts text features from all preliminary responses. When it identifies any preliminary response containing negative security constraint feature words such as "strictly prohibited", "must not", or "prohibited", the main agent module resets the logical weight of the sentence containing the constraint feature word to the highest, filters out other permissive responses from the slave agents, and directly outputs the final security decision of "prohibition" with the negative security constraint condition as the core.

[0043] The multi-hazard coupling priority rule refers to the following: when multiple disaster response measures given by the intelligent agent conflict in execution timing or logic, the main intelligent agent module calls a preset disaster fatality sequence for arbitration to determine the order of operations; the priority of the disaster fatality sequence is as follows: gas treatment > outburst prevention > fire prevention and extinguishing > rock burst prevention > water control > other routine operations;

[0044] The preset coal mine safety business rules include at least: prohibition priority rules and multi-hazard coupling priority rules; and also include the strictest standard rules, characterized in that: when multiple preliminary responses from intelligent agents contain numerical threshold requirements for the same safety parameter, the rule engine module automatically extracts the numerical boundaries from all preliminary responses; wherein, if the safety parameter is an upper limit constraint, the minimum value among all candidate upper limit values ​​is selected as the final safety decision indicator; if the safety parameter is a lower limit constraint, the maximum value among all candidate lower limit values ​​is selected as the final safety decision indicator.

[0045] A coal mine safety procedure processing system based on master-slave multi-agent and long document splitting includes:

[0046] Preprocessing module: used to divide the complete "Coal Mine Safety Regulations" into independent domain slices according to profession, and build multiple independent domain vector knowledge bases;

[0047] The main intelligent agent module is deployed on the edge computing device as a scheduling hub. It is used to receive user questions, identify the relevant professional fields, distribute sub-tasks, and perform mandatory logical arbitration and text integration on the received opinions of the subordinate intelligent agents through the rule engine module.

[0048] The agent cluster module contains multiple independent domain agents. Each domain agent is bound to and retrieves its corresponding single domain vector knowledge base. It is used to process sub-tasks issued by the main agent module and generate domain-specific constraints to be fed back to the main agent module.

[0049] Rule Engine Module: Stores preset arbitration logic based on hard safety constraints in coal mines, used for conflict detection and priority filtering of the initial responses from intelligent agents.

[0050] A computer device includes a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of a coal mine safety procedure processing method based on master-slave multi-agent and long document splitting.

[0051] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of a coal mine safety procedure processing method based on master-slave multi-agent and long document splitting are disclosed.

[0052] Example 1: System Hardware and Model Initialization Deployment

[0053] This embodiment is specifically designed for industrial edge computing scenarios with low computing power and high security requirements. In this embodiment, the system is physically deployed on an edge server equipped with a single NVIDIA RTX 4090 (24GB VRAM) graphics card.

[0054] Document preprocessing and physical segmentation:

[0055] First, the complete "Coal Mine Safety Regulations" (taking the underground coal mine section as an example, approximately 225 pages in total) was physically segmented by chapter. It was rigidly divided into: ventilation sub-documents, gas sub-documents, outburst prevention sub-documents, fire prevention and extinguishing sub-documents, water control sub-documents, rockburst sub-documents, blasting sub-documents, and mining sub-documents. The system used a lightweight vector embedding model (such as BGE-M3, with approximately 567 million model parameters) to perform text vectorization and chunking on each sub-document, setting the chunk size to 512 characters and the overlap size to 50 characters. The vectorized data was then stored in multiple physically isolated ChromaDB local vector databases.

[0056] Time-sharing of video memory and CUDA cache replacement mechanism:

[0057] Because 24GB of video memory cannot simultaneously accommodate the context retrieval overhead of eight domain knowledge bases and the runtime state of a large language model exceeding 14 bytes, this system designs a time-sharing scheduling mechanism for video memory. When the main agent starts the inference process:

[0058] 1. The main agent resides in video memory (occupies approximately 12GB of video memory, deploying a large model with 14B parameters for fine-tuning).

[0059] 2. When the "anti-sneak attack agent" needs to be invoked, the system dynamically loads the "anti-sneak attack vector database retriever" and the current inference context into the video memory through the PyTorch interface;

[0060] 3. Once the "anti-spy agent" completes its specific library search and generates an initial response text, the system memory manager immediately releases the memory context object occupied by the agent by executing `del`, and forcibly calls Python's garbage collection mechanism `importgc;gc.collect()` and the CUDA memory cleanup command `torch.cuda.empty_cache()` to clear the KVCache cache in the memory.

[0061] 4. Then, load the next triggered domain from the agent's data.

[0062] Through this mechanism, the system runs stably under the 24G graphics card limit, successfully avoiding the video memory overflow (OOM) problem.

[0063] Example 2: Main Agent Prompt and Rule Engine Arbitration Logic

[0064] The main intelligent agent adopts a large open-source language model with relatively large parameters (such as a fine-tuned model with 14B or 30B parameters) to give it a global perspective. Its core central logic is double-verified through the following structured Prompt combined with a hardware rule engine:

[0065] The structure of the main intelligent agent SystemPrompt is as follows:

[0066] “

[0067] You are the global lead AI agent for answering questions about coal mine safety regulations. Your task execution must strictly follow these steps:

[0068]

Step 1

[0069]

Step 2

[0070]

Step 3

[0071] - Rule A (Prohibition Priority Rule): Check all returned results. If any response from any domain contains restrictive words such as 'strictly prohibited', 'must not', or 'forbidden', you must veto it regardless of how relaxed the conditions are in other domains. The final output should be 'absolutely prohibited'.

[0072] - Rule B (Strictest Standard Rule): When different agents give different numerical thresholds (such as gas concentration ≤1.0% and ≤1.5%), you must automatically select the most stringent value (such as ≤1.0%) as the sole execution standard.

[0073] - Rule C (Multi-hazard Priority Rule): If there is a time-series conflict in the processing suggestions of various agents, the sorting and guidance shall be enforced in the order of [Gas > Outburst > Fire Prevention and Extinguishing > Rockburst > Water Control].

[0074]

Step 4

[0075] "

[0076] In the specific implementation of the system, the main agent first extracts JSON structured data from the outputs of each slave agent using a large language model, for example:

[0077] json

[0078] {

[0079] "agent_name":"Agent_anti-surprise",

[0080] "permission":"prohibited",

[0081] "reason": "Hydraulic mining is strictly prohibited in coal mines prone to geological disasters".

[0082] "numerical_constraints":[

[0083] }

[0084] After extraction, the system does not directly rely on the probability generation of a large model for output. Instead, it inputs the structured data into a rigid rule engine module. The rule engine performs physical-level filtering and coverage through simple conditional logic (If-Else) and numerical comparison algorithms (such as min() or max()), thereby achieving "zero illusion" decision output.

[0085] Example 3: Comparative Experiment and Effect Verification

[0086] To verify the substantial progress of the present invention, the applicant conducted comparative experiments with the present invention system (master-slave multi-agent + memory replacement + rule engine) and the existing single-model RAG system (loading a single model under 24g memory for full procedure retrieval) in three typical multi-hazard coupling scenarios.

[0087] Experimental Scenario 1: Resolving the "Omitted Critical Constraints" (Conflict between Hydraulic Coal Mining and Outburst Prevention)

[0088] A user asked: "Is it possible to use hydraulic mining in outburst-prone coal seams?"

[0089] The existing technology responds as follows: During single-model retrieval, the attention mechanism is attracted to "hydraulic coal mining," matching Article 139 of the mining regulations, and outputting the conclusion: "Hydraulic coal mining can be carried out, just follow the regulations." (Danger! This seriously misses the prohibited clauses in the anti-outburst rules).

[0090] Processing flow of this invention:

[0091] The main intelligent agent identified the keywords: "outburst coal seam" and "hydraulic coal mining";

[0092] Tasks are distributed to Agent-Mining and Agent-Extrusion;

[0093] Agent-Mining replied: "Hydraulic coal mining is permitted (in accordance with Article 139)";

[0094] Agent-highlighted the response after searching the special volume on coal mine gas outburst prevention: "The regulations clearly stipulate that hydraulic mining is strictly prohibited in coal mines prone to gas outbursts."

[0095] Upon receiving the above preliminary response, the main agent triggers Rule A: Prohibition takes precedence, granting Agent-prominent "Strict Prohibition" veto power and blocking Agent-mining permission;

[0096] The final safety decision output was: "Absolutely prohibited! Although general mining regulations allow hydraulic mining, according to regulations for preventing coal outbursts, hydraulic mining is strictly prohibited in outburst-prone coal seams. The final decision is that it is not allowed." (Successfully avoided violating regulations).

[0097] Experimental Scenario 2: Resolving "Misjudgments in Disaster Response Priorities and Logical Conflicts"

[0098] A user asked: "If signs of spontaneous combustion are suddenly detected in a rockburst hazard zone, which should be addressed first?"

[0099] Current technical response: Lacking professional timing logic, it outputs arbitrary guesses: "Rockbursts are extremely destructive, and drilling should be prioritized to relieve pressure and prevent rockbursts." (Fatal error, drilling will supply oxygen to the fire zone, leading to a large-scale gas explosion).

[0100] Processing flow of this invention:

[0101] The main agent distributes tasks to Agent-Impact Burst (responds: "Depressurize and prevent impact") and Agent-Fire Prevention (responds: "Inject grouting to extinguish fire immediately, and strictly prevent air leakage");

[0102] There is a timing conflict between the two actions during on-site execution. The main intelligent agent rule engine forcibly triggers [Rule C: Multi-hazard Priority], and finds in the database that "fire prevention and extinguishing priority (rank 3) is higher than rockburst prevention (rank 4)";

[0103] The final safety decision output is: "Spontaneous fires should be dealt with first. Since fire prevention and extinguishing take priority over rockburst prevention, drilling operations to prevent rockburst and pressure relief are strictly prohibited at this stage to prevent air leakage from expanding the fire zone and causing a gas explosion. Please take grouting fire extinguishing measures immediately." (This conforms to the ironclad rules of real-world rescue).

[0104] Experimental Scenario 3: Solving the "most stringent standard for cross-domain numerical requirements"

[0105] A user asked: "Can a fire-resistant sealed wall also serve as a drainage wall?"

[0106] The existing technology response is: "Yes, just set a drain hole." (This will cause the fire to reignite due to oxygen supply issues).

[0107] Processing flow of this invention:

[0108] Agent-Firefighting replied: "The sealed walls in the fire zone must be absolutely airtight and must not have any holes."

[0109] Agent-Water Prevention replied: "Drainage walls need to have drainage holes."

[0110] The main intelligent agent's rule engine identified a direct conflict between "no holes allowed" and "leave drainage holes," and the sealed wall falls under the "strictly prohibited" category for fire prevention. Rule A / B was triggered.

[0111] The final safety decision output: "It cannot be used for both purposes. Fire prevention and extinguishing regulations require that the fire area be tightly sealed, and leaving drainage holes will cause air leakage and reignition in the fire area. The fire prevention and extinguishing standards must be followed, and the drainage wall should be treated separately."

[0112] This invention achieves safe and zero-illusion question-and-answer application of coal mine safety regulations under extremely low computing power conditions through the above-mentioned multi-agent physical segmentation and rigid injection of rule engine.

Claims

1. A method for processing coal mine safety regulations, running on an edge computing device including a processor, memory, a large language model inference engine, and a rule engine module, characterized in that, Includes the following steps: S1. Document splitting and knowledge base construction: Using the processor to run a physical splitting algorithm, the long text of coal mine safety regulations is physically divided into multiple independent industry sub-documents according to professional fields, and vectorized by using a vector embedding model to construct and store multiple independent domain vector knowledge bases accordingly. S2. Reception and Intent Parsing: The main intelligent agent module receives the input coal mine safety query text request through the user interface, performs semantic analysis through the large language model inference engine, identifies all safety domains involved in the query text request, and uses the task allocation algorithm to decompose it into corresponding retrieval subtasks. S3. Divide and conquer retrieval from agents: The master agent module distributes the retrieval subtasks to the corresponding domain agents through the internal bus; Each domain loads only its own domain vector knowledge base from the intelligent agent for semantic retrieval, and combines the prompt word parameters of its own domain to generate a preliminary response containing the normative constraints of its domain through the large language model inference engine; S4. Rule Fusion and Arbitration Output: The main agent module receives the preliminary responses returned by all called agents and calls the rule engine module to perform semantic feature extraction, conflict detection and logical arbitration on each preliminary response based on the preset coal mine safety business rules, and generates and outputs the final safety decision response. Among them, the preset coal mine safety business rules include at least prohibition priority rules and multi-hazard coupling priority rules.

2. The coal mine safety regulation processing method as described in claim 1, characterized in that, In step S1, the professional field is one or more of the following: ventilation, gas, outburst prevention, fire prevention and extinguishing, water control, rock burst prevention, blasting, transportation, electrical, and mining.

3. The coal mine safety regulation processing method as described in claim 1, characterized in that, The prohibition priority rule in step S4 refers to the rule engine module extracting text features from all preliminary responses. When any preliminary response is found to contain negative security constraint feature words such as "strictly prohibited," "must not," or "forbidden," the main agent module resets the logical weight of the sentence containing the constraint feature word to the highest, filters out the permissive responses from other agents, and directly outputs the final security decision centered on the negative security constraint condition.

4. The coal mine safety regulation processing method as described in claim 1, characterized in that, The multi-hazard coupling priority rule in step S4 refers to the following: when multiple disaster response measures given by the intelligent agent conflict in execution timing or logic, the main intelligent agent module calls a preset disaster lethality sequence for arbitration to determine the order of operations; the priority of the disaster lethality sequence from high to low is as follows: gas treatment, outburst prevention, fire prevention and extinguishing, rockburst prevention, water prevention and control, and other routine operations.

5. The coal mine safety regulation processing method as described in claim 1, characterized in that, The preset coal mine safety business rules also include the strictest standard rules. Specifically, when multiple preliminary responses from intelligent agents contain numerical threshold requirements for the same safety parameter, the rule engine module automatically extracts the numerical boundaries from all preliminary responses. If the safety parameter is an upper limit constraint, the minimum value among all candidate upper limit values ​​is selected as the final safety decision indicator. If the safety parameter is a lower limit constraint, the maximum value among all candidate lower limit values ​​is selected as the final safety decision indicator.

6. The coal mine safety regulation processing method as described in claim 1, characterized in that, In step S3, each of the domain slave agents is executed serially on the same computing node using time-sharing multiplexing and memory replacement mechanisms; the master agent module calls the corresponding slave agents in sequence according to the time-series queue of the retrieval subtasks; when calling the target slave agent, the corresponding domain vector knowledge base context and the current model weights are loaded into the memory. After the target agent generates an initial response, the GPU memory space occupied by the target agent and the large language model key-value cache are released, and then the context of the next target agent to be triggered is loaded.

7. A coal mine safety procedure processing system based on master-slave multi-agent and long document splitting, used to execute the method according to any one of claims 1 to 6, characterized in that, It includes a preprocessing module, a master agent module, a slave agent cluster module, and a rule engine module; The preprocessing module is used to divide the complete coal mine safety regulations into independent domain slices according to professional fields and construct multiple independent domain vector knowledge bases. The main intelligent agent module is deployed on the edge computing device as a scheduling hub, used to receive user questions, identify the relevant professional fields, distribute sub-tasks, and perform forced logical arbitration and text integration on the received opinions of the subordinate intelligent agents through the rule engine module; The slave agent cluster module contains multiple independent domain slave agents. Each domain slave agent is bound to and retrieves its corresponding single domain vector knowledge base to process the sub-tasks issued by the master agent module and generate domain-specific constraints to be fed back to the master agent module. The rule engine module stores preset arbitration logic based on the hard safety constraints of coal mines, which is used to perform conflict detection and priority filtering on the initial responses from the intelligent agents.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images