A Method and System for Coal Mine Safety Risk Assessment Based on Multi-Agent Collaboration and Adversarial Review

By employing a multi-agent collaborative and adversarial review approach, the unreliability problem in coal mine safety risk assessment was solved, achieving highly accurate and reliable comprehensive risk judgment and enhancing the professional collaborative capabilities of coal mine safety risk auxiliary decision-making.

CN123117765APending Publication Date: 2026-11-17CHINA COAL RES INST +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611140796.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-29
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing coal mine safety risk assessment schemes lack an adversarial review mechanism, which can lead to unreliable assessment illusions when information is missing or rules are incomplete. Furthermore, the fixed processes lack continuous absorption of expert experience and historical cases, affecting the accuracy and reliability of safety-assisted decision-making.

Method used

The method employs multi-agent collaboration and adversarial review. Initial conclusions are generated by professional judgment agents and then reviewed in reverse by adversarial review agents. These conclusions are then corrected by combining expert memory. Finally, a comprehensive consultation agent performs multi-domain correlation analysis and priority ranking to form a closed-loop correction of the comprehensive judgment result.

Benefits of technology

It effectively suppresses the analytical illusions generated by a single intelligent agent, improves the accuracy and reliability of auxiliary assessment of coal mine safety risks, and continuously accumulates experience through an expert memory mechanism to achieve comprehensive diagnosis and priority ranking of risks in multiple fields.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117765A_ABST
    Figure CN123117765A_ABST
Patent Text Reader

Abstract

This application proposes a method and system for coal mine safety risk assessment based on multi-agent collaboration and adversarial review. The method includes: receiving coal mine safety risk auxiliary assessment tasks and safety production information, and generating a risk task profile; selecting corresponding professional assessment agents and adversarial review agents for each type of risk in the task based on the profile, and invoking expert memory; generating preliminary risk conclusions through the professional assessment agents combined with expert memory, and conducting reverse review of the preliminary risk conclusions through the adversarial review agents to generate a closed-loop corrected professional assessment result; and performing multi-domain correlation analysis and priority ranking of the professional assessment results and review opinions through a comprehensive consultation agent, outputting a comprehensive assessment result, and updating expert memory based on expert feedback. This method can effectively suppress the illusion of assessment by a single agent, achieve continuous accumulation of experience, and improve the accuracy and reliability of comprehensive coal mine safety risk assessment through multi-domain consultation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of coal mine safety production technology, and in particular to a coal mine safety risk assessment method and system based on multi-agent collaborative and adversarial review. Background Technology

[0002] As a crucial pillar of national energy security, coal mine safety risk assessment involves multiple professional fields, including gas, water hazards, fire, rock bursts, ventilation, roof collapse, and electromechanical transportation, exhibiting high complexity and specialization. Related technologies typically employ risk analysis through a single intelligent agent or a single large-scale language model, generating risk assessment conclusions based on input information.

[0003] However, the risk assessment schemes in related technologies directly adopt the output conclusions of a single agent without introducing an adversarial review mechanism to question the completeness of the basis, logical consistency, and alternative interpretations of the preliminary results. This may lead to a seemingly reasonable but actually unreliable assessment illusion when information is missing or rules are incomplete. Furthermore, the fixed processing procedures lack the ability to continuously absorb expert experience, historical cases, and lessons learned from failures, causing the system to repeat "novice-like" judgments in similar scenarios. In addition, the isolated analysis of risks from different specialties ignores the transmission and superposition effects between risks such as gas, fire, and rock bursts, thus affecting the accuracy and reliability of coal mine safety auxiliary decision-making. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art.

[0005] Therefore, the first objective of this application is to propose a coal mine safety risk assessment method based on multi-agent collaborative and adversarial review.

[0006] The second objective of this application is to propose a coal mine safety risk assessment system based on multi-agent collaborative and adversarial review.

[0007] The third objective of this application is to propose an electronic device.

[0008] The fourth objective of this application is to provide a computer-readable storage medium.

[0009] To achieve the above objectives, the first aspect of this application is to propose a coal mine safety risk assessment method based on multi-agent collaborative and adversarial review, comprising the following steps: Receive coal mine safety risk auxiliary assessment tasks and safety production information related to the assessment tasks, and generate risk task profiles, wherein the risk task profiles are used to provide multi-dimensional basic data for the assessment tasks; Based on the risk task profile, select a corresponding professional assessment agent and adversarial review agent for each type of risk in the task, and call upon multiple expert memories; The preliminary risk conclusion is generated by combining the expert memory with the various professional judgment agents. The preliminary risk conclusion is then reviewed by the adversarial review agent to identify hallucination risks and provide feedback correction, thus generating a closed-loop corrected professional judgment result. The comprehensive consultation intelligent agent performs multi-domain correlation analysis and priority ranking on the professional judgment results and review opinions, outputs a comprehensive judgment result for the task, and updates the expert memory based on expert feedback.

[0010] To achieve the above objectives, the second aspect of this application also proposes a coal mine safety risk assessment system based on multi-agent collaborative and adversarial review, comprising the following modules: The receiving module is used to receive coal mine safety risk auxiliary assessment tasks and safety production information related to the assessment tasks, and generate risk task profiles, wherein the risk task profiles are used to provide multi-dimensional basic data for the assessment tasks; The calling module is used to select the corresponding professional judgment agent and adversarial review agent for each type of risk in the task based on the risk task profile, and to call multiple expert memories; The generation module is used to generate preliminary risk conclusions by combining the expert memory with the various professional judgment agents, and to conduct reverse review of the preliminary risk conclusions by the adversarial review agent to identify hallucination risks and make feedback corrections, thereby generating a closed-loop corrected professional judgment result. The output module is used to perform multi-domain correlation analysis and priority ranking on the professional judgment results and review opinions through the comprehensive consultation intelligent agent, output the comprehensive judgment result for the task, and update the expert memory based on expert feedback.

[0011] To achieve the above objectives, a third aspect of this application also proposes an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the coal mine safety risk assessment method based on multi-agent collaborative and adversarial review as described in any one of the first aspects above.

[0012] To achieve the above objectives, the fourth aspect of this application also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the coal mine safety risk assessment method based on multi-agent collaborative and adversarial review as described in any one of the first aspects above.

[0013] The technical solution provided by the embodiments of this application brings at least the following beneficial effects: This application achieves a closed-loop process in the auxiliary assessment of coal mine safety risks—from professional assessment to adversarial review, then to expert memory, and finally to multi-domain consultation—through the collaborative work of a professional assessment agent, an adversarial review agent, an expert memory update agent, and a comprehensive consultation agent. This application can identify and suppress assessment illusions by the adversarial review mechanism, continuously accumulate expert experience and failure cases through the expert memory mechanism, enabling the system to gradually improve its assessment capabilities in subsequent similar tasks, and discover the inherent connections between risks such as gas, water hazards, fires, rock bursts, ventilation, roof collapse, and electromechanical transportation through the multi-domain consultation mechanism, forming a comprehensive risk judgment and priority handling recommendations. Therefore, this application effectively suppresses assessment illusions generated by a single agent through multi-agent collaborative assessment and adversarial review mechanisms, continuously accumulates experience with the help of expert memory enhancement mechanisms, and simultaneously achieves comprehensive consultation and priority ranking of risks across multiple domains, thereby significantly improving the accuracy, reliability, and professional collaborative capabilities of auxiliary assessment of coal mine safety risks.

[0014] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0015] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a coal mine safety risk assessment method based on multi-agent collaboration and adversarial review proposed in this application embodiment; Figure 2 This is a flowchart illustrating a specific coal mine safety risk assessment method based on multi-agent collaboration and adversarial review proposed in an embodiment of this application. Figure 3 This is a schematic diagram of the structure of a coal mine safety risk assessment system based on multi-agent collaborative and adversarial review proposed in an embodiment of this application. Detailed Implementation

[0016] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0017] The following description, with reference to the accompanying drawings, illustrates a method and system for assessing coal mine safety risks based on multi-agent collaborative and adversarial review, as proposed in an embodiment of this application.

[0018] Example 1 Figure 1 This is a flowchart of a coal mine safety risk assessment method based on multi-agent collaboration and adversarial review proposed in an embodiment of this application, as shown below. Figure 1 As shown, the method includes the following steps: Step S101: Receive coal mine safety risk auxiliary assessment task and related safety production information, and generate risk task profile, wherein the risk task profile is used to provide multi-dimensional basic data for the assessment task.

[0019] Specifically, after receiving the current coal mine safety risk auxiliary assessment task and related safety production information, this application first identifies and analyzes the input task to determine the key attributes such as the risk object involved in the task, the professional field to which it belongs, the abnormal signals or risk clues that trigger the assessment, the analysis intention, and the output form and depth requirements, and generates a risk task profile accordingly.

[0020] Among them, the risk task profile serves as the basis for subsequent agent selection, expert memory retrieval, adversarial review configuration, and multi-domain consultation, providing a unified context and task boundary for the entire collaborative assessment process, and can provide various types of data related to the current coal mine safety risk auxiliary assessment task.

[0021] As one possible implementation, the risk task profile can be represented as a structured object, containing fields such as task number, input information source, risk domain, object being assessed, abnormal signals, analysis intent, confidence level requirements, and output requirements. The risk domain covers coal mine safety-related areas such as gas, water hazards, fire, rock bursts, ventilation, roof collapse, and electromechanical transportation. Analysis intents include types such as risk identification, anomaly judgment, hazard analysis, handling suggestions, or comprehensive assessment. Through this step, the system transforms the raw input information into a task description that can be collaboratively processed by subsequent intelligent agents, ensuring that each agent operates within a unified framework.

[0022] Therefore, by generating a risk task profile, this application can clarify the specific boundaries and requirements of the current assessment task, providing a structured basis for subsequent intelligent agent collaborative orchestration, accurate retrieval of expert memory, and targeted configuration for adversarial review, thereby improving the pertinence and efficiency of risk assessment and avoiding misjudgment or omission due to ambiguity in task information.

[0023] Step S102: Based on the risk task profile, select the corresponding professional assessment agent and adversarial review agent for each type of risk in the task, and call upon multiple expert memories.

[0024] Specifically, based on the obtained risk task profile, the system performs collaborative orchestration of intelligent agents and targeted recall of expert memories. The core of this step lies in automatically identifying the professional risk domains involved in the current task based on the structured information contained in the task profile, such as risk objects, involved fields, abnormal signals, analysis objectives, and output requirements, and matching each type of risk with a corresponding professional judgment intelligent agent and a paired adversarial review intelligent agent.

[0025] The system employs a specialized analytical agent to generate preliminary risk conclusions based on domain knowledge and input information, while an adversarial review agent is dedicated to challenging and verifying these preliminary conclusions, forming a collaborative verification relationship. Simultaneously, based on the identified risk domain, the system retrieves relevant historical experience from a pre-built expert memory database. This database encompasses various types of expert memories, including rule memories, case memories, risk pattern memories, and failure case memories. This allows the system to reuse validated expert experience, historical cases, and lessons learned, avoiding the need to restart from a fixed process for each task.

[0026] Thus, through the above arrangement and invocation, the system completes the division of labor and configuration of intelligent agents and the loading of expert experience before the risk assessment is initiated, providing targeted knowledge support and verification basis for subsequent professional assessment and adversarial review.

[0027] As one possible implementation, collaborative orchestration and memory recall can be represented as: AO={risk_domain,judgement_agent,review_agent,memory_set,review_focus,collaboration_mode}, where each field corresponds to the risk domain, the selected professional judgment agent, the corresponding adversarial review agent, the expert memory set invoked for the current task, the review focus, and the collaboration mode.

[0028] Therefore, this step achieves precise allocation of agent resources and effective reuse of historical experience by dynamically selecting agents based on task profiles and calling expert memories in a targeted manner. This improves the pertinence and efficiency of risk assessment and provides the necessary knowledge base and verification basis for the subsequent adversarial review process.

[0029] Step S103: A preliminary risk conclusion is generated by combining the memory of experts with the judgment of various professional intelligence agents. The preliminary risk conclusion is then reviewed in reverse by an adversarial review intelligence agent to identify the risk of hallucination and make feedback corrections, thereby generating a closed-loop corrected professional judgment result.

[0030] Specifically, after receiving the risk task profile and completing the collaborative orchestration of intelligent agents and the retrieval of expert memories, the system enters a closed-loop processing stage of professional assessment and adversarial review. Each professional assessment agent, based on its assigned sub-task, combines current task information with rule memories, case memories, risk pattern memories, and failure case memories retrieved from expert memories to independently generate preliminary risk conclusions for its respective professional field. These preliminary conclusions cover risk item identification, risk level judgment, risk formation cause analysis, assessment basis citations, and preliminary handling suggestions. The generation process fully utilizes historical expert experience, avoiding starting from fixed rules for each task.

[0031] Subsequently, the system invokes adversarial review agents paired with the professional assessment agents to conduct a reverse review of the aforementioned preliminary risk conclusions. The adversarial review agents focus on checking whether the preliminary conclusions are generated without evidence, have insufficient evidence, jump in causal relationships, ignore key conditions, have biased risk levels, misjudge correlation as causation, or fail to rule out alternative explanations, and based on this, propose challenging opinions, requests for supplementary information, or suggestions for adjusting the risk level.

[0032] Furthermore, when significant doubts are discovered, the system feeds back the review opinions to the original professional assessment agent or related professional agents, prompting them to re-analyze and revise their conclusions, thus forming a closed-loop iterative process of "assessment—questioning—revision." After one or more rounds of closed-loop revision, the system outputs a professional assessment result verified by adversarial review. This result includes the revised risk conclusions, review opinions, and the final confidence level.

[0033] As one possible implementation, in coal mine safety risk scenarios, the gas assessment agent can focus on analyzing abnormal gas concentrations and trends exceeding limits; the water prevention and control assessment agent can analyze water hazard risks and water exploration and release measures; the fire assessment agent can analyze temperature and carbon monoxide anomalies; and the rockburst assessment agent can analyze ground pressure monitoring and mining layout. Each professional agent forms preliminary conclusions based on its own professional rules, and the corresponding adversarial review agent reviews the completeness of the basis, logical consistency, and possibility of alternative interpretations, triggering a correction cycle when necessary.

[0034] Therefore, through the closed-loop mechanism of professional assessment and adversarial review, this application can effectively suppress the risk of a single intelligent agent directly generating erroneous or unfounded conclusions, significantly improve the accuracy, reliability and traceability of preliminary risk conclusions, reduce the illusion of judgment caused by missing information, logical jumps or excessive inference, and provide a verified high-quality professional assessment basis for subsequent multi-domain comprehensive consultation.

[0035] Step S104: The comprehensive consultation intelligent agent performs multi-domain correlation analysis and priority ranking on the professional judgment results and review opinions, outputs the comprehensive judgment results for the task, and updates the expert memory based on expert feedback.

[0036] Specifically, after receiving the closed-loop revised assessment results from various professional fields and the corresponding adversarial review opinions, the integrated consultation agent performs multi-domain correlation analysis and priority ranking on all input information. The agent first identifies the inherent connections between different professional risks, including but not limited to causal relationships, transmission relationships, superposition relationships, conflict relationships, and handling dependencies, thereby integrating the originally scattered professional judgments into a unified view of the overall safety status of the coal mine.

[0037] Based on this, the comprehensive consultation intelligent agent merges, classifies, and prioritizes the identified risks according to their severity, urgency, scope of impact, and order of handling, identifies key or major risks, and generates a suggested order of handling.

[0038] Ultimately, the comprehensive consultation intelligent agent outputs a comprehensive assessment result that includes a risk summary, professional assessment opinions, adversarial review opinions, multi-domain risk correlations, a risk priority list, final auxiliary assessment conclusions, and treatment recommendations.

[0039] Furthermore, after human experts confirm, modify, or supplement the comprehensive judgment results output by the system, the comprehensive consultation intelligent agent or its collaborative memory update mechanism can analyze the expert feedback content, extract the risk correction conclusions, judgment basis, handling experience, professional rules, risk patterns, and error cases contained therein, and update the expert memory accordingly. The expert memory includes rule memory, case memory, risk pattern memory, and failure case memory.

[0040] As an example, the objects of comprehensive consultation and priority review can be represented as: SR={professional_results,review_results,relation_items,coupled_risks,conflict_items,priority_order,major_risk,disposal_order,manual_review_items}, the final output object can be expressed as JR={task_s ummary,professional_results,review_results,hallucination_check_result,multi_domain_result,priority_list,final_judgement,suggestions,manual_review_items,memory_update}.

[0041] Therefore, this step, through multi-domain correlation analysis and prioritization, can overcome the limitations of a single professional perspective, discover the coupling and transmission relationships between different risks, and form a comprehensive judgment on the overall safety status of the coal mine, avoiding decision-making biases caused by isolated analysis. Simultaneously, based on expert feedback, the system continuously updates expert memory, enabling it to accumulate verified and effective expert experience and failure cases. This historical knowledge can be reused in subsequent similar tasks, gradually improving the accuracy and reliability of the judgment, and achieving a closed-loop optimization from one-time judgment to continuous learning and evolution.

[0042] The following describes in detail the specific implementation process of each step in the coal mine safety risk assessment method based on multi-agent collaboration and adversarial review proposed in this application, using several specific embodiments.

[0043] Example 2 Based on the above embodiments, this embodiment provides a detailed description of the specific implementation of "receiving coal mine safety risk auxiliary assessment tasks and safety production information related to the assessment tasks, and generating risk task profiles" in step S101.

[0044] In this embodiment, the process of generating a risk task profile specifically includes the following steps: First, the system receives coal mine safety risk auxiliary assessment tasks and various related safety production information, including but not limited to gas monitoring data, water control information, fire-related data, rockburst monitoring information, ventilation status information, roof management information, electromechanical transportation status, hazard management records, on-site handling situations, historical cases, and expert review opinions. The system identifies the input task, determines the risk assessment task number (which uniquely identifies the current task), determines the source of the input information to trace the data origin, identifies the relevant risk professional fields (including gas, water hazards, fire, rockburst, ventilation, roof, or electromechanical transportation), determines the object to be assessed (which can be a region, equipment, working face, system, or hazard object), extracts abnormal signals or risk clues that trigger the assessment, identifies the task intent (including risk identification, anomaly judgment, hazard analysis, handling suggestions, or comprehensive assessment), sets the confidence level requirements for the reliability of the assessment conclusions, and clarifies the output format and assessment depth requirements.

[0045] Based on the identified and determined information, the system generates a structured risk task profile, which is represented as: RT={task_id,input_source,risk_domain,risk_object,abnormal_signal,analysis_intent,confidence_requirement,output_requirement}.

[0046] In this framework, `task_id` represents the risk assessment task number; `input_source` represents the source of the input information; `risk_domain` represents the relevant risk domain; `risk_object` represents the object being assessed; `abnormal_signal` represents the abnormal signal or risk clue that triggers the assessment; `analysis_intent` represents the task intent; `confidence_requirement` represents the reliability requirements for the assessment conclusion; and `output_requirement` represents the output format and assessment depth requirements. This risk task profile serves as the foundation for subsequent agent selection, expert memory retrieval, adversarial review configuration, and multi-domain consultation, ensuring that subsequent steps can work collaboratively based on a unified and clear task description.

[0047] Therefore, this embodiment can transform the input raw task information into a structured risk task profile. This profile not only clarifies the basic attributes and analysis objectives of the task, but also provides precise input parameters for the accurate selection of agents and the targeted recall of expert memories, thereby improving the pertinence and efficiency of the entire multi-agent collaborative judgment process.

[0048] Example 3 Based on the above embodiments, this embodiment provides a detailed description of the specific implementation of step S102, "selecting a corresponding professional assessment agent and adversarial review agent for each type of risk in the task according to the risk task profile, and calling multiple expert memories".

[0049] In this embodiment, the specific implementation method for selecting the corresponding professional judgment agent and adversarial review agent for each type of risk based on the risk task profile and calling the expert memory is as follows.

[0050] First, the system takes the risk task profile RT generated in step S101 as input. This profile includes a risk_domain field, which identifies the risk domain involved in the current task, such as gas, water hazards, fire, rock bursts, ventilation, roof support, or electromechanical transportation. Based on this field, the system identifies one or more domains involved in the current task. For each identified domain, the system selects a corresponding professional assessment agent from a pre-configured agent library, such as a gas assessment agent, a water prevention assessment agent, or a fire assessment agent. Simultaneously, it selects an adversarial review agent paired with this professional assessment agent. This adversarial review agent is specifically used to challenge and verify the reliability of the preliminary conclusions in the corresponding professional domain.

[0051] Simultaneously, based on the identified risk areas, the system retrieves the relevant expert memory set (memory_set) from the expert memory bank. This memory set includes rule memories, case memories, risk pattern memories, and failure case memories, enabling the reuse of historical expert experience in this assessment. Building upon this, the system further determines the review focus (review_focus) for this collaborative orchestration. This focus includes one or more of the following: completeness of evidence, logical consistency, reasonableness of risk level, and alternative explanations or missing information. This guides the focus of the adversarial review agent in subsequent reviews. The system also determines the collaboration mode (collaboration_mode), which includes one-on-one review, multiple-on-multiple review, cross-review, or joint consultation review, defining the collaboration methods between the professional assessment agent and the adversarial review agent.

[0052] Finally, the system combines the above information to generate a Co-Organization and Memory Recall Object (AO), which is represented as: .

[0053] Here, `risk_domain` represents the risk expertise domain, `judgement_agent` represents the selected expert judgment agent, `review_agent` represents the corresponding adversarial review agent, `memory_set` represents the set of expert memories invoked, `review_focus` represents the review focus, and `collaboration_mode` represents the collaboration mode. This AO object serves as the configuration basis for the expert judgment agent to generate preliminary risk conclusions and for the adversarial review agent to perform closed-loop corrections in subsequent step S3, ensuring the accuracy and traceability of agent division of labor and expert experience loading.

[0054] Therefore, this embodiment realizes the dynamic arrangement of intelligent agent combinations and the invocation of expert memory based on risk task profiles, and clarifies the review focus and collaboration mode, so that the subsequent judgment and review process can be differentiated according to different risk areas and task characteristics, thereby improving the pertinence and efficiency of multi-agent collaborative judgment, and providing a structured collaborative foundation for closed-loop correction and multi-domain consultation.

[0055] Example 4 Based on the above embodiments, this embodiment provides a detailed description of the specific implementation of step S103, which involves "generating preliminary risk conclusions by combining the expert memory with the professional judgment agents, and conducting reverse review of the preliminary risk conclusions by the adversarial review agents to identify hallucination risks and provide feedback corrections to generate a closed-loop corrected professional judgment result."

[0056] In this embodiment, the specific implementation method of generating preliminary risk conclusions by combining expert memory with the various professional judgment agents is as follows.

[0057] First, each professional judgment agent receives a subtask assigned by the collaborative orchestration module in step S102. The subtask includes the risk object, the field involved, the abnormal signal, the analysis target and output requirements in the risk task profile, as well as the expert memory set related to the current task called from the expert memory module. This set includes rule memory, case memory, risk pattern memory and failure case memory.

[0058] Then, the specialized analytical agent, using current task information (such as gas monitoring data, water control information, fire-related data, etc.) and expert memory as input, independently performs risk analysis within its professional field. During the analysis, the specialized analytical agent performs logical reasoning based on rules from expert memory, compares similar scenarios in case memory, refers to typical abnormal combinations in risk pattern memory, and avoids erroneous paths from failed case memory.

[0059] For example, when analyzing abnormal gas concentrations, the gas analysis agent will call the ventilation rule memory to determine whether it is caused by sensor drift, call the historical case memory to compare similar over-limit trends, and call the risk pattern memory to identify the coupling pattern between gas and ventilation abnormalities, while avoiding the misjudgment logic that has appeared in failed cases.

[0060] Subsequently, after each specialized analytical agent completes its analysis, it generates a specialized analytical result containing all fields, which is represented as follows: AR={agent_id,domain_type,risk_item,risk_level,risk_reason,memory_used,basis_reference,suggestion,confidence} in, A unique identifier for professional analysis of intelligent agents; This indicates the professional field to which the intelligent agent belongs, such as gas, water control, fire prevention, etc. For specific risk items identified, such as "the trend of gas concentration exceeding the limit continues to rise"; The initial risk level can be divided into low, medium, high, and major levels. For the reasons for the formation of the risk or the logic behind the judgment, describe in detail which monitoring data, rules or cases were used to reach the conclusion; Record the expert memory content actually used in this assessment, including specific rule numbers, case identifiers, or risk pattern names; List the types of basis for judgment, such as rule-based basis, historical case basis, monitoring anomaly basis, or expert experience basis; Preliminary recommendations, such as "It is recommended to immediately cease mining operations and intensify gas drainage"; To professionally assess the agent's confidence level in this conclusion, the assessment is expressed numerically or in rating scales. This professional assessment result serves as input for subsequent reverse review by the adversarial review agent and is not directly used as the final output.

[0061] Therefore, through the above specific implementation methods, each professional judgment agent can conduct independent analysis based on expert memory, generate structured professional judgment results, and provide accurate and traceable input data for subsequent adversarial review and multi-domain consultation, thereby improving the professionalism and reliability of coal mine safety risk assessment.

[0062] In this embodiment, the specific implementation method of the adversarial review agent to reverse review the preliminary risk conclusion, identify the risk of illusion and provide feedback for correction, forming a closed-loop corrected professional judgment result is as follows.

[0063] First, the adversarial review agent receives a preliminary risk conclusion from the professional assessment agent. This conclusion includes the professional assessment agent's identifier, professional field, identified risk items, preliminary risk level, cause of risk or judgment logic, accessed expert memory, assessment basis, preliminary handling suggestions, and confidence level. Using this preliminary risk conclusion as input, the adversarial review agent initiates a reverse review process.

[0064] During the review process, the adversarial review agent meticulously examines each preliminary risk conclusion to determine if it is generated without sufficient evidence. This includes assessing whether the risk items, causes, or proposed solutions are supported by clear monitoring data, rules, or case studies; checking for insufficient evidence, ensuring the supporting evidence is adequate and that key monitoring indicators or historical data are not omitted; identifying causal jumps, ensuring the logical chain from the abnormal signal to the risk conclusion is complete and that no intermediate steps are missing; ignoring key conditions, ensuring the conclusion does not consider current operational status, geological conditions, equipment status, or environmental parameters; verifying risk level discrepancies, ensuring the preliminary risk level aligns with the severity of the abnormal signal, risk levels in historical cases, or industry standards; misjudging correlation as causation, ensuring two simultaneous anomalies are not incorrectly identified as causally related; and failing to exclude alternative explanations, ensuring the preliminary conclusion does not exclude or discuss other reasonable explanations for the current anomaly. For each check, the adversarial review agent generates corresponding points of contention and records potential illusory risk items.

[0065] Furthermore, the adversarial review agent identifies missing information or conditions that need to be supplemented, such as additional monitoring data, geological data, historical cases, or expert experience, and generates possible alternative interpretations. Based on the above analysis, the adversarial review agent generates risk level adjustment suggestions, such as suggesting that the initial risk level be adjusted from high to medium, or from medium to high.

[0066] Subsequently, the adversarial review agent constructs an adversarial review and closed-loop correction object, which includes the adversarial review agent identifier, the professional judgment agent identifier being reviewed, the judgment conclusion being reviewed, potential illusion risk items, key points of questioning, missing information or conditions that need to be supplemented, potential alternative interpretations, risk level adjustment suggestions, the judgment result after feedback correction, and the review conclusion.

[0067] The review conclusions include: passed, requires correction, requires supplementary information, controversial, or requires expert review. When the review conclusion requires correction or supplementary information, the adversarial review agent provides feedback to the original professional judgment agent or related professional agent, requesting them to re-analyze. Based on the feedback, the original professional judgment agent or related professional agent supplements the missing information, corrects the risk level, adjusts the judgment logic, or eliminates alternative interpretations, generating a corrected judgment result, which is then resubmitted to the adversarial review agent for review. If the review is successful, the review conclusion is updated to "passed," and the corrected judgment result is output as the closed-loop corrected professional judgment result. If problems remain after the review, feedback continues until the review conclusion is "passed" or requires expert review. When the review conclusion is "controversial" or requires expert review, the system marks the matter as requiring further manual review and outputs it to subsequent steps.

[0068] Therefore, through the aforementioned adversarial review and closed-loop correction mechanism, this embodiment can effectively identify and suppress the illusion risks that may arise when a single professional judgment agent generates preliminary risk conclusions, such as unfounded generation, insufficient evidence, causal relationship jumps, ignoring key conditions, risk level deviations, misjudging correlation as causality, or failing to exclude alternative explanations. This significantly improves the accuracy and reliability of coal mine safety risk assessment conclusions and reduces the risk of safety management decision-making errors due to erroneous conclusions.

[0069] Example 5 Based on the above embodiments, this embodiment provides a detailed description of the specific implementation of step S104, "using a comprehensive consultation intelligent agent to perform multi-domain correlation analysis and priority ranking on the professional judgment results and review opinions, outputting a comprehensive judgment result for the task, and updating the expert memory based on expert feedback."

[0070] In this embodiment, the specific implementation process of the comprehensive consultation intelligent agent performing multi-domain correlation analysis and priority ranking of all professional judgment results and review opinions is as follows.

[0071] First, the integrated consultation agent receives the professional assessment results output by each professional assessment agent in step S103, as well as the review opinions, disputed items, and revised results output by each adversarial review agent in step four, as input data for unified analysis. The professional assessment results include the risk items identified in each professional field, their risk levels, causes, assessment basis, and confidence levels; the review opinions include the adversarial review agents' key points of contention regarding the preliminary conclusions, missing information, alternative interpretations, level adjustment suggestions, and the revised assessment results.

[0072] Then, the integrated consultation agent performs structured analysis on these input data, extracting risk items and their attributes for each professional field, and establishing a cross-field risk correlation matrix. When identifying the intrinsic connections between risks from different professional fields, the integrated consultation agent analyzes, based on preset association rules and risk patterns remembered by experts, whether there are causal relationships (e.g., gas anomalies caused by ventilation anomalies), transmission relationships (e.g., roof pressure changes leading to rockburst risk), superposition relationships (e.g., risk amplification when gas exceedances and fire hazards coexist), conflict relationships (e.g., inconsistent judgments on the same risk level by different professions), and treatment dependency relationships (e.g., ventilation problems must be addressed before gas risks can be addressed). For the identified correlations, the integrated consultation agent records them as relational items and marks risk items with coupled or superimposed effects.

[0073] Based on this, the integrated consultation agent merges all risk items, combining risks with the same root cause or those that are interrelated into comprehensive risk groups. Then, it ranks and prioritizes these risks according to their risk level, urgency, scope of impact, and treatment dependencies, generating a risk priority ranking result and identifying key or major risks. Simultaneously, the integrated consultation agent identifies conflicting items that are not yet fully agreed upon by multiple agents and lists them separately.

[0074] Finally, the integrated consultation agent outputs the integrated consultation and priority review object, which is represented as: SR={professional_results,review_results,relation_items,coupled_risks,conflict_items,priority_order,major_risk,disposal_order,manual_review_items} The process comprises the following components: professional_results (containing outputs from various professional assessment agents); review_results (containing outputs from adversarial review agents); relation_items (recording the relationships between different risks); coupled_risks (listing risk items with coupled or overlapping effects); conflict_items (listing conflicting items among multiple agents that are not yet fully agreed upon); priority_order (ranking the risks by priority); major_risk (identifying key or major risks); disposal_order (providing the suggested order of action); and manual_review_items (listing items recommended for further manual review). Through this process, the integrated consultation agent elevates single-specialty judgments to multi-domain consultations and comprehensive decision-making, providing a structured and comprehensive assessment basis for subsequent expert feedback and memory updates.

[0075] Therefore, this specific implementation method, by establishing a cross-domain risk correlation matrix and a priority ranking mechanism, achieves systematic integration and comprehensive judgment of risks from multiple disciplines, avoids the limitations of single-discipline analysis, and enhances the comprehensiveness and decision support capabilities of auxiliary judgment of coal mine safety risks.

[0076] In this embodiment, the specific implementation process of updating the expert memory based on expert feedback is as follows. The expert opinion understanding and memory updating agent first receives confirmation, modification, or supplementary opinions from human experts on the comprehensive judgment results, which serve as the input source. The agent parses the expert opinions, extracting the risk correction conclusions, judgment basis, handling experience, professional rules, risk patterns, and erroneous cases contained therein. Specifically, for rules confirmed as valid by experts, the agent writes them into the rule memory; for typical risk scenarios, it writes them into the case memory; for recurring abnormal combinations, it writes them into the risk pattern memory; and for erroneous conclusions rejected by experts, it writes them into the failure case memory. In this way, the system transforms a one-time judgment task into a closed-loop process of continuous learning and evolution.

[0077] Through the aforementioned expert opinion understanding and memory update mechanism, the system can continuously accumulate expert experience and failure cases, enabling it to draw on historical experience in subsequent similar tasks, avoid repeating mistakes, gradually improve judgment capabilities, and thus enhance the accuracy and reliability of auxiliary judgment on coal mine safety risks.

[0078] Example 6 Based on the above embodiments, this embodiment will describe in detail the complete implementation of the coal mine safety risk assessment method based on multi-agent collaboration and adversarial review. This embodiment constructs a collaborative assessment system composed of a professional assessment agent, an adversarial review agent, an expert memory module, a comprehensive consultation agent, and an expert opinion understanding and memory update agent. Through risk task profile generation, expert memory enhancement, professional agent assessment, adversarial review verification, multi-domain comprehensive consultation, and expert feedback updates, a closed-loop process for coal mine safety risk assessment is achieved, from professional identification, reverse review, comprehensive judgment to continuous optimization. The implementation process of the safety risk assessment method in this embodiment is as follows: Figure 2 As shown.

[0079] First, the system receives coal mine safety risk auxiliary assessment tasks and related safety production information. This information includes, but is not limited to, gas monitoring data, water control information, fire-related data, rockburst monitoring information, ventilation status information, roof management information, electromechanical and transportation status, hazard mitigation records, on-site handling situations, historical cases, and expert review opinions. The system identifies the input task, determines the risk object, involved areas, abnormal signals, analysis objectives, credibility requirements, and output requirements, and generates a risk task profile.

[0080] Specifically, the risk task profile is represented as follows: RT={task_id,input_source,risk_domain,risk_object,abnormal\signal,analysis_intent,confidence_requirement,output_requirement}.

[0081] In this framework, `task_id` represents the risk assessment task number; `input_source` represents the source of the input information; `risk_domain` represents the relevant risk domain, including gas, water hazards, fire, rock bursts, ventilation, roof support, electromechanical transportation, etc.; `risk_object` represents the area, equipment, work surface, system, or potential hazard being assessed; `abnormal_signal` represents the abnormal signal or risk clue that triggers the assessment; `analysis_intent` represents the task intent, including risk identification, anomaly judgment, hazard analysis, handling suggestions, or comprehensive assessment; `confidence_requirement` represents the reliability requirements for the assessment conclusions; and `output_requirement` represents the output format and assessment depth requirements. This risk task profile serves as the foundation for subsequent agent selection, expert memory retrieval, adversarial review configuration, and multi-domain consultation.

[0082] Next, based on the risk task profile, the execution engine identifies the professional fields involved in the current task and selects corresponding professional assessment agents and adversarial review agents for each type of risk. Professional assessment agents include those for gas assessment, water control, fire assessment, rockburst assessment, ventilation assessment, roof assessment, and electromechanical transportation assessment. Adversarial review agents are paired with professional assessment agents to conduct reverse questioning and reliability reviews of the corresponding professional conclusions. Simultaneously, the system invokes expert memories based on the risk field, including rule memories, case memories, risk pattern memories, handling experience memories, and failure case memories, enabling the reuse of historical expert experience in this assessment. The collaborative orchestration and memory retrieval objects are represented as follows: AO={risk_domain,judgement_agent,review_agent,memory_set,review_focus,collaboration_mode}.

[0083] Here, `risk_domain` represents the risk expertise domain; `judgement_agent` represents the selected expert judgment agent; `review_agent` represents the corresponding adversarial review agent; `memory_set` represents the set of expert memories invoked for the current task; `review_focus` represents the review focus, including the completeness of the basis, logical consistency, reasonableness of the risk level, alternative explanations, and missing information; and `collaboration_mode` represents the collaboration mode, including one-on-one review, multiple-on-multiple review, cross-review, or joint consultation review. Through this step, the system completes the agent division of labor and expert experience loading before risk assessment begins, avoiding starting from a fixed process every time.

[0084] Then, each specialized assessment agent, based on its assigned sub-tasks and combining current task information with expert memory, independently generates preliminary conclusions on professional risks. The gas assessment agent focuses on analyzing abnormal gas concentrations, gas exceedance trends, ventilation impacts, and response measures; the water control assessment agent focuses on analyzing water hazard risks, water exploration and release measures, geological and hydrological conditions, and the risk of abnormal water inrush; the fire assessment agent focuses on analyzing temperature, carbon monoxide, coal spontaneous combustion tendency, and fire prevention and extinguishing measures; the rockburst assessment agent focuses on analyzing ground pressure monitoring, mining layout, stress concentration, and rockburst hazard; other specialized assessment agents form corresponding conclusions according to their respective professional rules. The professional assessment results are expressed as follows: AR={agent_id,domain_type,risk_item,risk\_level,risk_reason,memory_used,basis_reference,suggestion,confidence}.

[0085] Here, `agent_id` represents the identifier of the professional assessment agent; `domain_type` represents the professional field; `risk_item` represents the identified risk item; `risk_level` represents the preliminary risk level; `risk_reason` represents the cause of the risk or the judgment logic; `memory_used` represents the expert memory content invoked in this assessment; `basis_reference` represents the assessment basis, including rule-based basis, historical case basis, monitoring anomaly basis, or expert experience basis; `suggestion` represents the preliminary handling suggestion; and `confidence` represents the degree of confidence of the professional assessment agent in this conclusion. This preliminary conclusion is not directly used as the final output but enters the adversarial review stage.

[0086] For the preliminary risk conclusions generated by the professional assessment agent, the system invokes the corresponding adversarial review agent for reverse review. The adversarial review agent focuses on checking whether the preliminary conclusions are generated without evidence, have insufficient evidence, jumps in causal relationships, ignore key conditions, have biased risk levels, misjudge relevance as causation, or fail to rule out alternative explanations, and proposes challenging opinions, requests for supplementary information, or suggestions for level adjustment. When significant doubts are found, the system feeds back the review opinions to the original professional assessment agent or a related professional agent for re-analysis, forming a closed loop of "assessment—questioning—correction." The adversarial review and closed-loop correction objects are represented as follows: RV={ review_agent_id,target_agent_id,target_conclusion,hallucination_risk_items,challenge_items,missing_information,alternative_explanation,level_adjustment,revised_result,review_result} Here, `review_agent_id` represents the identifier of the adversarial review agent; `target_agent_id` represents the professional assessment agent being reviewed; `target_conclusion` represents the assessment conclusion being reviewed; `hallucination_risk_items` represents potential hallucination risk items; `challenge_items` represents the key points challenging the conclusion; `missing_information` represents missing information or conditions that need to be supplemented; `alternative_explanation` represents possible alternative explanations; `level_adjustment` represents the risk level adjustment suggestion; `revised_result` represents the assessment result after feedback and correction; and `review_result` represents the review conclusion, including pass, need for correction, need for supplementary information, disputed, or need for expert review. Through this step, the system reduces the risk of a single agent directly generating erroneous or unfounded conclusions.

[0087] After the professional assessments and adversarial reviews are completed, the integrated consultation intelligent agent performs a unified analysis of all professional conclusions, review opinions, disputed items, and revised results. The integrated consultation intelligent agent identifies the inherent connections between different professional risks, including causal relationships, transmission relationships, superposition relationships, conflict relationships, and treatment dependencies. For example, gas anomalies may be related to ventilation anomalies, fire risks may be related to carbon monoxide anomalies and ventilation status, and rockburst risks may be related to roof pressure changes and mining disturbances. Based on this, the system merges, classifies, and prioritizes the risks, forming a comprehensive assessment result for the overall safety status of the coal mine. The objects of the integrated consultation and priority review are represented as follows: SR={professional_results,review_results,relation_items,coupled_risks,conflict_items,priority_order,major_risk,disposal_order,manual_review_items}.

[0088] Among them, professional_results represents the output results of each professional assessment agent; review_results represents the output results of each adversarial review agent; relation_items represents the correlation between different risks; coupled_risks represents risk items with coupled or overlapping effects; conflict_items represents conflict items that are not yet fully agreed upon among multiple agents; priority_order represents the risk priority ranking results; major_risk represents key or major risks; disposal_order represents the recommended order of handling; and manual_review_items represents items that are recommended for further manual review. Through this step, the system can move from single-professional judgment to multi-domain consultation and comprehensive decision-making.

[0089] Finally, after human experts confirm, modify, or supplement the risk assessment results output by the system, the expert opinion understanding and memory update agent analyzes the expert opinions, extracting the risk correction conclusions, judgment basis, handling experience, professional rules, risk patterns, and erroneous cases contained therein. For rules confirmed as valid by experts, the system writes them into rule memory; for typical risk scenarios, the system writes them into case memory; for recurring abnormal combinations, the system writes them into risk pattern memory; and for erroneous conclusions rejected by experts, the system writes them into failure case memory, which is used to remind the corresponding agent to increase its level of caution in subsequent similar tasks. The system also outputs the final risk auxiliary assessment results, including a risk summary, professional assessment opinions, adversarial review opinions, hallucination risk check results, multi-domain risk correlations, key risks, risk priorities, handling suggestions, and human review suggestions. The expert feedback and final output objects are represented as follows: JR={task_summary,professional_results,review_results,hallucination_check_result,multi_domain_result,priority_list,final_judgement,suggestions,manual_review_items,memory_update}.

[0090] Among them, `task_summary` represents the task summary; `professional_results` represents the output results of each professional assessment agent; `review_results` represents the output results of the adversarial review agent; `hallucination_check_result` represents the hallucination risk check results; `multi_domain_result` represents the multi-domain consultation results; `priority_list` represents the risk priority list; `final_judgement` represents the final auxiliary assessment conclusion; `suggestions` represents the treatment recommendations; `manual_review_items` represents the items recommended for manual review; and `memory_update` represents the updated rule memory, case memory, risk pattern memory, or failure case memory based on expert opinions. Through this step, the system transforms a one-time assessment task into a closed-loop process of continuous learning and evolution.

[0091] In summary, this embodiment achieves a closed-loop process of "professional assessment, adversarial review, expert memory, and multi-domain consultation" in the auxiliary assessment of coal mine safety risks through a hallucination suppression mechanism of paired collaboration between professional assessment agents and adversarial review agents, an expert memory enhancement and continuous update mechanism, and a comprehensive consultation mechanism for risks in multiple areas of coal mines. This effectively reduces the risk of a single agent directly generating erroneous or unfounded conclusions, enabling the system to gradually improve its assessment capabilities in subsequent similar tasks. It can also discover the inherent connections between risks such as gas, water hazards, fires, rock bursts, ventilation, roof collapse, and electromechanical transportation, forming comprehensive risk judgments and priority handling suggestions, thereby improving the accuracy, reliability, professional collaboration, and continuous optimization capabilities of auxiliary assessment of coal mine safety risks.

[0092] To achieve the above embodiments, this application also proposes a coal mine safety risk assessment system based on multi-agent collaboration and adversarial review, such as... Figure 3 As shown, the system includes: The receiving module 100 is used to receive coal mine safety risk auxiliary assessment tasks and safety production information related to the assessment tasks, and generate a risk task profile, wherein the risk task profile is used to provide multi-dimensional basic data of the assessment tasks. The calling module 200 is used to select a corresponding professional judgment agent and an adversarial review agent for each type of risk in the task based on the risk task profile, and to call multiple expert memories; The generation module 300 is used to generate preliminary risk conclusions by combining the expert memory with the various professional judgment agents, and to conduct reverse review of the preliminary risk conclusions by the adversarial review agent to identify hallucination risks and make feedback corrections, thereby generating a closed-loop corrected professional judgment result. The output module 400 is used to perform multi-domain correlation analysis and priority ranking on the professional judgment results and review opinions through the comprehensive consultation intelligent agent, output the comprehensive judgment result for the task, and update the expert memory according to expert feedback.

[0093] It should be noted that the aforementioned descriptions of the embodiments of the coal mine safety risk assessment method based on multi-agent collaboration and adversarial review also apply to the system of this application, and will not be repeated here.

[0094] In summary, the coal mine safety risk assessment system based on multi-agent collaboration and adversarial review proposed in this application can effectively suppress the illusion of assessment by a single agent, achieve continuous accumulation of experience, and improve the accuracy and reliability of comprehensive assessment of coal mine safety risks through multi-domain consultation.

[0095] To implement the above embodiments, this application also proposes an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the coal mine safety risk assessment method based on multi-agent collaborative and adversarial review as described in any of the first aspect embodiments above.

[0096] To implement the above embodiments, this application also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the coal mine safety risk assessment method based on multi-agent collaborative and adversarial review as described in any one of the first aspects of the embodiments above.

[0097] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0098] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0099] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0100] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0101] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0102] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0103] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0104] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A coal mine safety risk research and judgment method based on multi-agent cooperation and confrontation review, characterized in that, Includes the following steps: Receive coal mine safety risk auxiliary assessment tasks and safety production information related to the assessment tasks, and generate risk task profiles, wherein the risk task profiles are used to provide multi-dimensional basic data for the assessment tasks; Based on the risk task profile, select a corresponding professional assessment agent and adversarial review agent for each type of risk in the task, and call upon multiple expert memories; The preliminary risk conclusion is generated by combining the expert memory with the various professional judgment agents. The preliminary risk conclusion is then reviewed by the adversarial review agent to identify hallucination risks and provide feedback correction, thus generating a closed-loop corrected professional judgment result. The comprehensive consultation intelligent agent performs multi-domain correlation analysis and priority ranking on the professional judgment results and review opinions, outputs a comprehensive judgment result for the task, and updates the expert memory based on expert feedback.

2. The method of claim 1, wherein, The generation of the risk task profile includes: The assessment task is identified, and the risk objects, involved fields, abnormal signals, analysis targets, credibility requirements, and output requirements of the assessment task are determined. Based on the identification and determination results of the assessment task, a risk task profile is generated, which includes the task number, source of input information, risk professional field, assessment target object, assessment trigger signal, task intention, assessment conclusion requirements, and output requirements.

3. The method of claim 2, wherein, The aforementioned risk areas include: gas, water hazards, fire, rock bursts, ventilation, roof support, and electromechanical transportation. The judgment trigger signals include abnormal signals and risk clues that trigger the judgment. The stated mission objectives include: risk identification, anomaly assessment, hazard analysis, handling recommendations, and comprehensive evaluation.

4. The method of claim 1, wherein, The process involves selecting a corresponding expert assessment agent and an adversarial review agent for each type of risk in the task, and invoking multiple expert memories, including: Based on the risk task profile, identify the risk professional fields involved in the current assessment task, and based on the risk professional fields, select the corresponding professional assessment agent for each type of risk and set up an adversarial review agent in pairs with the professional assessment agent, and call up expert memory. Generate a collaborative orchestration and memory retrieval object, wherein the collaborative orchestration and memory retrieval object includes a risk professional domain, a selected professional judgment agent, a corresponding adversarial review agent, a set of expert memories to be retrieved, review focus, and collaborative mode.

5. The method of claim 1, wherein, The process of generating preliminary risk conclusions by combining the expertise of each professional assessment agent with the expert's memory includes: Each of the aforementioned professional judgment agents is controlled to independently generate a corresponding initial judgment result by combining the assigned sub-tasks, current task information, and expert memory. The initial judgment result includes the professional judgment agent identifier, professional field, identified risk items, preliminary risk level, risk formation cause or judgment logic, accessed expert memory content, judgment basis, preliminary handling suggestions, and professional judgment agent confidence level.

6. The method of claim 1, wherein, The process of reverse-examining the preliminary risk conclusion through the adversarial review agent, identifying hallucination risks, providing feedback corrections, and generating a closed-loop corrected professional assessment result includes: The adversarial review agent checks whether the preliminary risk conclusions are generated without basis, have insufficient basis, skip causal relationships, ignore key conditions, have biased risk levels, misjudge relevance as causation, or fail to exclude alternative explanations. Based on the problem inspection results, adversarial review and closed-loop correction objects are generated. The adversarial review and closed-loop correction objects include the adversarial review agent identifier, the professional judgment agent being reviewed, the judgment conclusion being reviewed, possible hallucination risk items, key points of questioning, missing information or conditions that need to be supplemented, possible alternative interpretations, risk level adjustment suggestions, and the judgment result and review conclusion after feedback correction. If the doubts to be reviewed are identified based on the adversarial review and closed-loop correction, the adversarial review opinions will be fed back to the original professional judgment agent or related professional agents for re-analysis.

7. The method of claim 1, wherein, The process involves using a comprehensive consultation intelligent agent to perform multi-domain correlation analysis and priority ranking on the professional assessment results and review opinions, outputting a comprehensive assessment result for the task, including: The comprehensive consultation intelligent agent performs a unified analysis of all professional conclusions, review opinions, disputed items and correction results to identify the inherent connections between different professional risks. These inherent connections include causal relationships, transmission relationships, superposition relationships, conflict relationships and treatment dependence relationships. Based on the inherent connections, risks are merged, classified, and prioritized to generate comprehensive consultation and priority review objects. These comprehensive consultation and priority review objects include the output results of each professional judgment agent, the output results of each adversarial review agent, the correlation between different risks, risk items with coupled or superimposed effects, conflict items that are not yet fully consistent among multiple agents, risk priority ranking results, high-priority risks, suggested handling order, and suggested items for manual review.

8. The method of claim 1, wherein, The step of updating the expert memory based on expert feedback includes: The intelligent agent that understands and updates expert opinions analyzes expert feedback and extracts risk correction conclusions, judgment basis, handling experience, professional rules, risk patterns and error cases contained in the expert feedback. The expert memory is updated based on different extraction results, including writing rules that are confirmed to be valid by experts into rule memory, writing typical risk scenarios into case memory, writing recurring abnormal combinations into risk pattern memory, and writing erroneous conclusions that are rejected by experts into failure case memory.

9. A coal mine safety risk research and judgment system based on multi-agent cooperation and confrontation review, characterized in that, Includes the following modules: The receiving module is used to receive coal mine safety risk auxiliary assessment tasks and safety production information related to the assessment tasks, and generate risk task profiles, wherein the risk task profiles are used to provide multi-dimensional basic data for the assessment tasks; The calling module is used to select the corresponding professional judgment intelligent agent and adversarial review intelligent agent for each type of risk in the task based on the risk task profile, and to call multiple expert memories; The generation module is used to generate preliminary risk conclusions by combining the expert memory with the various professional judgment agents, and to conduct reverse review of the preliminary risk conclusions by the adversarial review agent to identify hallucination risks and make feedback corrections, thereby generating a closed-loop corrected professional judgment result. The output module is used to perform multi-domain correlation analysis and priority ranking on the professional judgment results and review opinions through the comprehensive consultation intelligent agent, output the comprehensive judgment result for the task, and update the expert memory based on expert feedback.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the coal mine safety risk assessment method based on multi-agent collaborative and adversarial review as described in any one of claims 1-8.