A metallurgical hidden danger identification method and device based on multi-agent cooperation

Abstract

This invention relates to the field of metallurgical safety management technology, and provides a method and device for identifying metallurgical hazards based on multi-agent collaboration. The invention triggers the identification task with multi-source data input. A sensing agent preprocesses the multi-source data input to obtain the object to be analyzed. A central scheduling agent performs deep semantic understanding and metallurgical safety rule matching on the object to be analyzed. Based on the output of the deep semantic understanding and metallurgical safety rule matching, it determines the scenario understanding agent and / or regulatory matching agent to be invoked. A comprehensive decision-making agent generates a final report and instructions based on the risk analysis results of the scenario understanding agent and / or the regulatory matching agent, thus completing the hazard identification. This invention solves the problems of existing metallurgical safety hazard identification technologies, such as high reliance on human experience, low identification efficiency, vague judgment criteria, and fragmented information at each stage.

Description

Technical Field

[0001] This invention relates to the field of metallurgical safety management technology, and in particular to a method and device for identifying metallurgical hazards based on multi-agent collaboration. Background Technology

[0002] Metallurgical production is characterized by complex processes, dense equipment, and hazardous media (e.g., high-temperature molten metal, high-pressure gases, toxic chemicals). Production is highly continuous, and safety risks are ever-present. Throughout the entire process, including sintering, ironmaking, steelmaking, rolling, and power generation, key locations such as the blast furnace body, converter oxygen lance area, gas holder perimeter, molten steel hoisting channels, and confined spaces (e.g., spray painting rooms, cable tunnels) inherently pose high risks, demanding extremely high accuracy and timely identification of potential hazards. The core objective is to promptly and accurately identify on-site hazards, clarify the criteria for hazard assessment, and provide scientific rectification recommendations.

[0003] However, the current mainstream safety hazard identification and management in metallurgical enterprises relies heavily on manual experience, resulting in low identification efficiency, vague judgment criteria, and fragmented information across different stages. This leads to limited coverage, accuracy, response speed, and standardization of safety hazard investigation in metallurgical enterprises.

[0004] Therefore, overcoming the shortcomings of the existing technology is an urgent problem to be solved in this technical field. Summary of the Invention

[0005] The technical problem to be solved by this invention is that the existing technology for identifying metallurgical safety hazards relies heavily on human experience, has low identification efficiency, vague judgment criteria, and fragmented information at each stage.

[0006] Firstly, a method for identifying potential metallurgical hazards based on multi-agent collaboration is provided, including: Multi-source data input triggers a recognition task, and the perceptual intelligent agent preprocesses the multi-source data input to obtain the object to be analyzed; The central scheduling agent performs deep semantic understanding and metallurgical safety rule matching on the object to be analyzed. Based on the output of deep semantic understanding and metallurgical safety rule matching, it selectively determines the scenario understanding agent and / or the regulatory matching agent to be invoked. The scenario understanding agent performs risk analysis on the object to be analyzed; the regulatory matching agent matches the object to be analyzed with relevant regulations. The integrated decision-making agent generates a final report and instructions based on risk analysis and / or matching results to complete the hazard identification.

[0007] Furthermore, the central scheduling agent performs deep semantic understanding and metallurgical safety rule matching on the object to be analyzed. Based on the output of the deep semantic understanding and metallurgical safety rule matching, it selectively determines the scenario understanding agent and / or regulatory matching agent to be invoked, including: The large language model is pre-trained and fine-tuned using a massive amount of metallurgical safety text to obtain a domain-fine-tuned large model; the domain-fine-tuned large model processes the object to be analyzed and outputs a first intent entity set. The object to be analyzed is subjected to keyword matching using a lightweight model and a security rule base, and a second intent entity set is output. If the consistency between the first intent entity set and the second intent entity set is greater than a preset value, then the first intent entity set or the second intent entity set is used to generate a structured instruction. If the consistency is less than or equal to a preset value, the arbitration big model analyzes the object to be analyzed, the first intention entity set, the second intention entity set, and the conflict context to generate an arbitration instruction; The urgency level is determined using the structured instructions or the arbitration instructions, and the invocation scenario understanding agent and / or regulatory matching agent are determined according to the urgency level.

[0008] Furthermore, the scenario understanding agent performs risk analysis on the object to be analyzed, including: The scenario understanding agent aligns and deeply fuses the multimodal data in the object to be analyzed to obtain fused features, thereby generating a description of potential hazards. Risk levels and hazard types are generated based on the fused features; The regulatory matching agent performs the following matching operations on the object to be analyzed: Constructing a metallurgical safety regulations map based on a graph database; The description of the potential hazards is converted into a graph query statement; The query statement described above is used to retrieve matching materials from the metallurgical safety regulations map.

[0009] Furthermore, the comprehensive decision-making agent generates a final report and instructions based on risk analysis and / or matching results, including: By integrating the matching materials, risk levels, and hazard types, and combining them with historical best practice templates, an actionable rectification decision is generated to obtain a final report and instructions.

[0010] Furthermore, the method also includes: The inspection task is initiated after the inspection personnel complete identity verification. Based on the identity authentication information of the inspection personnel, obtain the inspection monitoring dataset for the inspection task; Based on the inspection and monitoring dataset, the location and data of the objects to be investigated in the inspection task are determined to facilitate subsequent identification of metallurgical hazards.

[0011] Furthermore, before obtaining the inspection monitoring dataset for the inspection task based on the inspection personnel's identity authentication information, the method further includes: Each monitoring point records the start and end times of the patrol personnel entering its monitoring range; Once the inspection task is initiated, the monitoring points involved in the inspection task are determined.

[0012] Furthermore, obtaining the inspection monitoring dataset for the inspection task based on the identity authentication information of the inspection personnel includes: When the inspection personnel upload the inspection report of the inspection task, they use the identity authentication information to obtain the inspection task and the monitoring points involved in the inspection task. Based on the inspection report, the start time, and the end time, an inspection monitoring dataset is obtained from the monitoring points.

[0013] Further, the step of obtaining the inspection monitoring dataset based on the monitoring points according to the inspection report, the start time, and the end time includes: Obtain the multiple monitoring points involved in the inspection task; Extract the objects to be investigated from the inspection report; search for the objects to be investigated in the set of objects that can be covered by the multiple monitoring points, and obtain the monitoring points corresponding to the objects to be investigated, which are then used as the points to be investigated. The video stream from the corresponding start time to the end time in the monitoring video stream of the location to be investigated is taken as the video to be investigated; Identify the video frames and / or set of video frames in the video to be investigated that effectively reflect the object to be investigated, in order to obtain the inspection and monitoring dataset.

[0014] Secondly, a metallurgical hazard identification device based on multi-agent collaboration is provided, the metallurgical hazard identification device based on multi-agent collaboration includes: a processor and a memory for storing processor-executable instructions; The processor is configured to execute the metallurgical hazard identification method based on multi-agent collaboration.

[0015] Thirdly, a non-volatile computer storage medium is provided, which stores computer-executable instructions that are executed by one or more processors to perform the metallurgical hazard identification method based on multi-agent collaboration described in the first aspect.

[0016] Fourthly, a computer program product containing instructions is provided, which, when executed on a computer or processor, causes the computer or processor to perform the metallurgical hazard identification method based on multi-agent collaboration as described in the first aspect.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a multi-agent architecture that simulates the collaborative work of an expert team. A perception agent performs preprocessing, a central scheduling agent performs deep semantic understanding, metallurgical safety rule matching, and decision-making, a scenario understanding agent performs risk analysis, a regulatory matching agent matches the relevant criteria, and a comprehensive decision-making agent generates the final report and instructions. This achieves hazard identification by realizing a fully automated and intelligent closed-loop process, from multimodal perception, intelligent analysis, cross-domain analysis, and accurate attribution of hazard information to the generation of executable decisions. This significantly reduces reliance on human experience and improves identification efficiency. With clear judgment criteria, close collaboration among all stages, and information integration, it fundamentally enhances the coverage, accuracy, response speed, and standardization of safety hazard investigation in metallurgical enterprises. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a metallurgical hazard identification method based on multi-agent collaboration provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall structure of a metallurgical hidden danger identification system based on multi-agent collaboration provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating step 20 provided in an embodiment of the present invention; Figure 4 This is a flowchart illustrating step 30 provided in an embodiment of the present invention; Figure 5 This is a flowchart illustrating another step 30 provided in an embodiment of the present invention; Figure 6 This is a flowchart illustrating a specific example of a metallurgical hazard identification system based on multi-agent collaboration provided in an embodiment of the present invention. Figure 7 This is a flowchart illustrating a method for quickly locating an object to be investigated, provided by an embodiment of the present invention. Figure 8 This is a flowchart illustrating step 502 provided in an embodiment of the present invention; Figure 9 This is a schematic diagram illustrating a specific example of obtaining an inspection and monitoring dataset provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of a metallurgical hazard identification device based on multi-agent collaboration provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] Unless the context otherwise requires, throughout the specification and claims, the term "comprising" is interpreted as openly inclusive, meaning "including, but not limited to." In the description of the specification, terms such as "one embodiment," "some embodiments," "exemplary embodiment," "example," "specific example," or "some examples" are intended to indicate that a particular feature, structure, material, or characteristic associated with that embodiment or example is included in at least one embodiment or example of this disclosure. The illustrative representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics mentioned may be included in any suitable manner in any one or more embodiments or examples; that is, although they may be incorporated into embodiments or examples using the above terms for reasons such as order and position, it does not limit them to be incorporated in combination by a single embodiment or example.

[0022] In the description of this invention, 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 indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more. Furthermore, for example, the description may use the prefix "A" or "B" to describe the same type of nouns as two independent entities. In this case, the corresponding features defined with "A" and "B" are used only to distinguish between similar entities and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features.

[0023] In describing some embodiments, the terms "coupled," "coupled," and "connected," and their derivative expressions, may be used. For example, the term "connected" may be used in describing some embodiments to indicate that two or more components have direct physical or electrical contact with each other. Similarly, the term "coupled" may be used in describing some embodiments to indicate that two or more components have direct physical or electrical contact. However, the terms "connected" or "coupled" may also refer to two or more components that do not have direct contact with each other but still cooperate or interact with each other, such as "optical coupling," "wireless connection," etc. The embodiments disclosed herein are not necessarily limited to the scope of this invention.

[0024] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0025] Currently, the mainstream safety hazard identification and management in metallurgical enterprises mainly relies on the following combination of technologies, forming a traditional model centered on "human-led and technology-assisted" approaches, as follows: Regular manual inspections and specialized checks: Companies assign dedicated safety officers or equipment inspectors to conduct on-site inspections according to fixed routes and checklists. Inspectors rely on their personal experience and senses (such as sight, hearing, and touch) to identify abnormalities, such as loose bolts, pipeline corrosion, and abnormal liquid levels, and record the problems on paper or in simple electronic spreadsheets. For more specialized areas (such as electrical and pressure vessel safety), specialized safety inspections are conducted regularly by experts.

[0026] Single-point sensors and fixed-threshold alarms: Independent sensors, such as temperature, pressure, and gas concentration detectors, are installed at key hazard locations and connected to local audible and visual alarms or the central control room's monitoring system. The system is preset with fixed safety thresholds; once the monitored data exceeds the threshold, an alarm is triggered. For example, if the carbon monoxide concentration in a gas-filled area exceeds a certain value, the alarm light in the control room will flash.

[0027] Video surveillance system and post-incident investigation: Network cameras are deployed in key areas throughout the plant to achieve real-time monitoring and video recording. This is primarily used for post-incident investigation and tracing, such as reviewing recordings to analyze the cause of an accident. Some advanced systems attempt to integrate simple video analytics algorithms, such as area intrusion detection or smoke detection, but these suffer from high false alarm rates and difficulty in understanding complex security semantics.

[0028] Information-based hazard reporting and process management: A safety hazard management system is used to achieve online reporting, transfer, rectification, and statistics of hazards. After inspectors discover a hazard, they take a photo with their mobile phone, fill in a description, and submit it. The system then generates a work order and assigns it to the responsible person. The core function of this type of system is process digitization and tracking, but the characterization, classification, regulatory matching, and rectification plan formulation of hazards still need to be completed manually by safety experts. The system itself does not have intelligent analysis and decision-making capabilities.

[0029] Static database queries of standards and regulations: Companies may purchase or maintain an electronic database of safety regulations and standards (such as the "Compilation of Metallurgical Safety Production Standards"). When a compliance assessment is required for a particular issue, safety personnel must manually search the database using keywords, read and filter relevant clauses themselves. This process is tedious and highly dependent on individual professional knowledge and search skills.

[0030] The essence of existing technical solutions is to partially "digitize" and "network" the traditional manual inspection process. However, the core decision-making and analysis aspects (i.e., "What problem was found?", "What regulations were violated?", and "How should it be scientifically rectified?") still heavily rely on subjective human judgment and experience. The systems are independent of each other, forming information silos, lacking the ability to intelligently integrate and deeply analyze cross-modal information. This makes it impossible to achieve dynamic, quantitative risk assessment and proactive early warning, resulting in a passive, sluggish, and inadequate response to the complex and ever-changing safety situation in metallurgical production. Steel enterprises are characterized by long process flows, large plant areas, and numerous hazardous sources, with thousands of employees, leading to frequent personnel accidents and injuries. Areas such as coke oven basements, blast furnace tapping areas, tuyeres, cold storage rooms, molten steel hoisting areas, electrostatic precipitator ash discharge ports, and crude benzene tar tanks present significant risks and difficulties in personnel management.

[0031] Currently, the safety management technology solutions widely used in the metallurgical industry, such as traditional video surveillance, sensor threshold alarms, electronic inspection systems, and isolated information management platforms, have improved management efficiency to a certain extent in some areas, but in essence, they have not broken away from the traditional paradigm of "human-led, passive response, and information silos". When dealing with the complex, dynamic, and professional safety risks in metallurgical production, these technical means have exposed the following fundamental defects: (1) Single modal processing and lack of information fusion capabilities: Existing solutions are mostly designed for single types of data. For example, video systems are only used for behavior monitoring, sensors are only used for over-limit alarms, and inspection systems are only used for recording and circulation. Data is not shared between systems and there are semantic barriers. It is impossible to conduct cross-modal correlation analysis and deep fusion of text descriptions, on-site images, real-time working condition data, and historical records, resulting in a superficial understanding of hidden dangers and difficulty in discerning the systemic risks caused by the coupling of multiple factors. (2) Reliance on human experience and static rules, and low level of decision-making intelligence: The final judgment of hidden dangers, the search for legal basis, and the formulation of rectification measures almost entirely rely on the personal experience and manual retrieval of safety personnel. Preset alarm rules (e.g., fixed thresholds) are rigid and cannot be dynamically adjusted according to the operating status of the equipment and changes in the environment. This method is inefficient, inconsistent in standards, and is prone to misjudgment or omission due to differences in the professional ability of personnel or fatigue and negligence, and cannot achieve stable and objective standardized decision-making. (3) "Separation of criteria" and knowledge gap: Existing technology separates "discovery of hidden dangers" and "matching of legal standards" into two independent and highly manual links. The system lacks a structured and computable legal knowledge system, and cannot automatically and accurately associate specific on-site problems with specific clauses in the vast legal provisions, resulting in vague basis for the judgment of hidden dangers and insufficient persuasiveness. The rectification suggestions often lack authoritative and accurate legal and regulatory support. (4) The system is closed and rigid and lacks continuous evolution capability: Once the functions of the traditional system are deployed, they are basically solidified. Its knowledge base and rule base are updated in a lagging manner and cannot learn autonomously from the experience and feedback accumulated in daily operation. Historical hidden danger data is simply archived and becomes a "data graveyard", which cannot effectively feed back and optimize the risk identification model, so that the system cannot adapt to the new risks brought by new processes and new equipment, and has weak sustainable development capability.

[0032] The core problem with existing technological solutions lies in their fragmentation, static nature, and superficiality, failing to construct a holistic intelligent closed loop capable of autonomous perception, intelligent analysis, precise decision-making, and continuous learning. This has led to a long-standing predicament in the safety management of metallurgical enterprises, characterized by "invisibility, inaccurate judgment, and insufficient management," hindering the qualitative leap from "passive response" to "proactive prevention and intelligent control." Therefore, a systematic technological innovation is urgently needed.

[0033] Furthermore, most existing personnel management measures in steel enterprises rely on simple video surveillance, facial recognition, personnel positioning, and access control, which improve the efficiency of personnel risk management to varying degrees but do not fully utilize dynamic data from production and monitoring systems. This fails to provide direct, rapid, accurate, and objective safety assistance for decisions regarding unauthorized entry, equipment malfunctions, and toxic gas leaks, thus failing to effectively protect the lives of enterprise personnel.

[0034] To address the problems of existing technologies, this embodiment proposes a metallurgical hazard identification method based on multi-agent collaboration. In one embodiment, such as... Figure 1 As shown, it includes: Step 10: Multi-source data input triggers the recognition task. The perceptual agent preprocesses the multi-source data input to obtain the object to be analyzed.

[0035] In one embodiment, the present invention provides a system, the overall system architecture of which is as follows: Figure 2 As shown, the system adopts a layered, decoupled, and co-evolutionary design philosophy, mainly composed of the following six core modules forming an organic whole: Multi-source data access and perception agent, serving as the system's senses and data portal; hereinafter referred to as the "perception agent"; Central scheduling and task analysis agent, serving as the system's command center and decision-making brain; hereinafter referred to as the "central scheduling agent"; Scene understanding and initial risk assessment agent, serving as the system's technical analysis expert; hereinafter referred to as the "scene understanding agent"; Regulatory knowledge graph and matching agent, serving as the system's regulatory compliance expert; hereinafter referred to as the "regulatory matching agent"; Comprehensive decision generation agent, serving as the system's report and strategy synthesizer; Decision execution and feedback module, serving as the system's execution arm and learning engine, connected to a structured knowledge base; hereinafter referred to as the "decision execution module". These modules are connected through standardized data interfaces and message communication protocols, separating data flow from control flow, forming a complete closed loop from perception to learning evolution; among them, message communication protocols include gRPC and Message Queuing Telemetry Transport (MQTT).

[0036] The sensing agent is responsible for the unified access, cleaning, and preprocessing of complex, multi-source, heterogeneous data from the metallurgical site. The sensing agent has access capabilities: it supports accessing real-time process data (such as temperature, pressure, and flow rate) from a Distributed Control System (DCS) or Programmable Logic Controller (PLC) via the Open Platform Communications Unified Architecture (OPC UA) protocol; accessing IoT sensor data (such as gas concentration and vibration) via the MQTT protocol; accessing enterprise management system data (such as electronic inspection records and work orders) via Application Programming Interface (API) or file parsing; and directly processing text descriptions, images, and voice information reported from the site, thus obtaining multi-source data input. It also has processing capabilities for preprocessing the multi-source data input: it has built-in data cleaning, format standardization, and spatiotemporal alignment algorithms. For image / video streams, a unified multimodal large model (e.g., Qwen3-Omni) is integrated to extract safety-related targets in real time (such as personnel not wearing protective equipment, the status of safety facilities, open flames and smoke); for text / speech, automatic transcription, word segmentation, and noise reduction are performed. In one embodiment, the perceptual agent outputs: generating data objects (i.e., the objects to be analyzed) with a unified spatiotemporal stamp, data source identifier, and standardized feature vector, providing high-quality "data fuel" for the upper layer. The perceptual agent serves as the system's unified data entry point, comprehensively collecting information from various channels. The types of data received include, but are not limited to, sensor data generated by field equipment (such as temperature and pressure), monitoring images / videos, text reports filled out by operators, and system operation log files. The perceptual agent enables the system to possess powerful multimodal data processing capabilities.

[0037] Step 20: The central scheduling agent performs deep semantic understanding and metallurgical safety rule matching on the object to be analyzed. Based on the output of deep semantic understanding and metallurgical safety rule matching, it selectively determines the scenario understanding agent and / or the regulatory matching agent to be invoked.

[0038] The central scheduling agent is responsible for accurate task understanding and intelligent scheduling. Internally, it employs a "dual-path voting-arbitration mechanism," meaning it performs deep semantic understanding and metallurgical safety rule matching on the object to be analyzed separately. Then, based on the output of this dual-path parallel analysis, it arbitrates and makes decisions, ultimately determining whether to invoke the scenario understanding agent, the rule matching agent, or both. A specific example will be provided below, and will not be elaborated further here.

[0039] Step 30: The scenario understanding agent performs risk analysis on the object to be analyzed; the regulatory matching agent performs legal matching on the object to be analyzed.

[0040] The scenario understanding agent is responsible for conducting multi-dimensional and quantitative analysis of specific potential hazard scenarios. It is invoked to perform risk analysis on the objects under analysis. The regulatory matching agent provides authoritative and accurate judgment criteria, ensuring system standardization and compliance. It is invoked to match the objects under analysis with relevant metallurgical industry safety regulations. Specific examples of risk analysis and regulatory matching will be provided below, and will not be elaborated further here.

[0041] Step 40: The intelligent agent generates a final report and instructions based on the risk analysis and / or matching results to complete the hazard identification.

[0042] This invention constructs a multi-agent architecture that simulates the collaborative work of an expert team. A perception agent performs preprocessing, a central scheduling agent performs deep semantic understanding, metallurgical safety rule matching, and decision-making, a scenario understanding agent performs risk analysis, a regulatory matching agent matches the relevant criteria, and a comprehensive decision-making agent generates the final report and instructions. This achieves hazard identification by realizing a fully automated and intelligent closed-loop process, from multimodal perception, intelligent analysis, cross-domain analysis, and accurate attribution of hazard information to the generation of executable decisions. This significantly reduces reliance on human experience and improves identification efficiency. With clear judgment criteria, close collaboration among all stages, and information integration, it fundamentally enhances the coverage, accuracy, response speed, and standardization of safety hazard investigation in metallurgical enterprises.

[0043] The following is a further description of the metallurgical hazard identification method based on multi-agent cooperation according to an embodiment of the present invention: Figure 2 This paper presents a three-layer architecture and core process of a metallurgical hidden danger identification system based on multi-agent collaboration according to an embodiment of the present invention: The data perception layer, namely, module (1), is responsible for multi-source data access and preprocessing, and is the system's sensor. The agent collaboration layer modules (2) to (5) act as the "brain". Through central scheduling, namely module (2), tasks are distributed and the agents of the risk preliminary judgment module (3) and the regulation matching module (4) are called in parallel. Finally, the comprehensive decision-making module (5) generates the results. The resource and execution layer, namely module (6) and the knowledge base, executes decisions and collects feedback. The feedback data is deposited into the structured knowledge base. The knowledge base optimizes the agents of the collaboration layer in reverse, forming a closed loop of "perception-analysis-decision-feedback".

[0044] The central dispatching agent acts as a "central brain," performing preliminary understanding, intent recognition, and task definition on the input raw data. For example, it can deduce the core task of "suspected electrical circuit overheating" from a vague text description and a scene image. To further illustrate the central dispatching agent, in one embodiment, such as... Figure 3 As shown, step 20 includes: Step 201: Use massive amounts of metallurgical safety text to pre-train and fine-tune the large language model to obtain a domain-fine-tuned large model; the domain-fine-tuned large model processes the object to be analyzed and outputs a first intent entity set.

[0045] Task triggering of the centrally dispatching intelligent agent: It can passively receive tasks from sensing intelligent agents or artificial systems, or actively generate periodic inspection tasks according to plan.

[0046] The following explains the dual-path parallel parsing of the central scheduling agent: The first approach (deep semantic understanding approach): employs a large-scale language model (i.e., a domain-fine-tuned large model) pre-trained and fine-tuned on massive amounts of metallurgical safety texts (e.g., accident reports, procedures, case studies). This approach excels at understanding complex contextual semantics, ambiguous expressions, and implicit intentions, ultimately outputting the first set of intent entities from the domain-fine-tuned large model. Examples of large-scale language models (i.e., large language models) include Qwen3 and the Chat Generative Language Model (ChatGLM4).

[0047] Step 202: Use a lightweight model and a security rule base to perform keyword matching on the object to be analyzed, and output a second intent entity set.

[0048] The second approach (fast rule matching approach): This approach employs a lightweight, transformer-based Bidirectional Encoder Representations from Transformers (BERT) model, combined with a robust rule base for the metallurgical safety domain. This rule base includes standard equipment KKS codes, hazard classification trees, and high-frequency violation verb phrases. This approach prioritizes fast keyword matching and standardized mapping.

[0049] The dual-path parsing stage is the entry point for system perception and understanding, employing a strategy combining "deep understanding" and "rapid verification." In step 10, raw task input and preprocessing occur: the system receives raw task instructions from text, speech, or multimodal channels and performs preprocessing operations such as feature extraction, cleaning, and standardization to prepare for subsequent parsing. In step 20, parallel parsing occurs. First, a domain-fine-tuned large model serves as the main parser, leveraging its powerful semantic understanding and reasoning capabilities to perform a deep and comprehensive analysis of the task, outputting preliminary intent and entity sets. Then, a "lightweight model + rule knowledge base" serves as an auxiliary verifier, utilizing its high efficiency and determinism to quickly parse the task and match it with the rule base, providing a lightweight parsing result.

[0050] Step 203: If the consistency between the first intent entity set and the second intent entity set is greater than a preset value, then use the first intent entity set or the second intent entity set to generate a structured instruction.

[0051] The preset values ​​are determined by those skilled in the art based on the specific usage scenario, and are not limited here.

[0052] Arbitration and Decision-Making: A consistency comparison arbitrator is established to compare the core intent and key entity set of the two outputs. If they match, the result is directly adopted as the high-confidence resolution result. Arbitration Discriminant Model Activation: The system invokes a specially trained arbitration model with stronger logical reasoning and contextual understanding capabilities. This model comprehensively analyzes the original task, the results of the two conflicts, and the context of the conflicts to make a final ruling, provide an explanation, and then generate an arbitration instruction. This path ensures the reliability of the system's decision-making in complex and ambiguous scenarios.

[0053] Consistency arbitration is a core element in ensuring system accuracy, handling uncertainties in the parsing results through intelligent comparison and arbitration. The system automatically compares the core intent and key entities output by two parsing models. When the results of the two models are highly consistent, the system considers the parsing result highly credible, directly adopts it, and generates high-confidence structured instructions. This path ensures the system's efficient operation under normal circumstances.

[0054] Step 204: If the consistency is less than or equal to a preset value, the arbitration big model analyzes the object to be analyzed, the first intention entity set, the second intention entity set, and the context of the conflict to generate an arbitration instruction.

[0055] When the results of the two models differ significantly, the system determines it as a "conflict" and automatically triggers a more advanced arbitration mechanism. That is, if a conflict occurs, an arbitration agreement is immediately triggered, submitting the original input (i.e., with regional variations) and the results from both parties (i.e., the first set of intent entities, the second set of intent entities, and the context of the conflict) to a dedicated, more reasoning-capable arbitration model (based on a large-scale deep-thinking model such as DeepseekR1). The arbitration model analyzes the root causes of the conflict and outputs the final ruling and a brief explanation of the reasoning.

[0056] The system employs a quality control mechanism by verifying the confidence level of the parsed results to ensure the reliability of task parsing. The system self-evaluates the initial parsing results. For high-confidence paths, when the system is highly confident in the parsing results, it directly proceeds to the next step, generating task instructions to ensure efficient workflow. For low-confidence paths, when the system deems the parsing results ambiguous or unclear, it activates a more advanced discrimination mechanism. An arbitration model acts as an "expert referee" to address uncertainties under low-confidence conditions. This could be a more complex deep learning model or a rule-based expert system, designed to provide the most accurate ruling and prevent the assignment of erroneous tasks.

[0057] Step 205: Determine the urgency level using the structured instructions or the arbitration instructions, and determine the invocation scenario understanding agent and / or regulatory matching agent according to the urgency level.

[0058] After parsing, the agent breaks down the task into atomic subtasks (such as "risk assessment" and "regulation matching"), and dynamically schedules subsequent agents to execute them based on task type and current system load, acting as a global resource coordinator. The validated parsing results are then transformed into machine-readable, standardized task instructions, providing a clear work list for subsequent agent collaboration.

[0059] The instruction distribution phase serves as the interface between the system and the external environment. Whether it's a high-confidence structured instruction or an arbitration instruction, it will ultimately be unified into a standard, machine-readable structured format. This structured instruction is then sent to downstream intelligent agent clusters (such as scene understanding and regulatory matching agents) to initiate subsequent business processing flows.

[0060] In one embodiment, such as Figure 4 As shown, in step 30, the scenario understanding agent performs risk analysis on the object to be analyzed, including: Step 301a: The scene understanding agent aligns and deeply fuses the multimodal data in the object to be analyzed to obtain fused features, thereby generating a hazard feature description.

[0061] Multimodal fusion analysis: This process receives standardized multi-source data from perceptual agents (such as a device image, a historical sensor curve, or a text description). Using a Transformer-based cross-modal attention fusion network, visual features, temporal data features, and text features are mapped to a unified semantic space for alignment and deep fusion. This yields fused features, generating a comprehensive and structured description of the potential hazard scenario (i.e., a hazard feature description).

[0062] Step 302a: Generate risk level and hazard type based on the fused features.

[0063] Preliminary risk quantification assessment: Based on the fused features, a built-in risk assessment model (such as an integrable multimodal model for dynamic risk prediction) is invoked to output a preliminary risk level (e.g., "low", "medium", "high", "critical") and hazard type code. This assessment provides preliminary data-driven conclusions for subsequent work.

[0064] In one embodiment, such as Figure 5 As shown, in step 30, the regulatory matching agent performs a matching process on the object to be analyzed, including: Step 301b: Construct a metallurgical safety regulations map based on a graph database.

[0065] In one embodiment, the knowledge graph is constructed as follows: a metallurgical safety regulations knowledge graph is built based on a graph database (such as Neo4j). Graph nodes include: specific legal and regulatory clauses (such as Article X of the Production Safety Law, Section Y of GB 4053.3), hazard instances, equipment and facility entities, and rectification measures; edge relationship definitions include "violation," "applicable to," "cause of," and "reference cases." The graph is semi-automatically constructed and updated from a structured regulatory database and historical cases using ETL tools.

[0066] Step 302b: Convert the description of the hidden danger characteristics into a graph query statement.

[0067] Intelligent matching and reasoning: Receive the description of potential hazards output by the scene understanding agent and convert it into a graph query (such as a Cypher statement).

[0068] Step 303b: Use the above-mentioned map query statement to query the metallurgical safety regulations map to obtain the matching materials.

[0069] By using graph traversal and multi-hop reasoning, we can not only accurately match the most directly relevant legal provisions, but also discover related superior legal principles and special standard requirements, and retrieve similar historical cases and their handling results, forming a complete chain of evidence.

[0070] In one embodiment, step 40 includes: integrating the matching materials, the risk level, and the hazard type, combined with historical best practice templates, to generate an actionable rectification decision, so as to obtain a final report and instructions.

[0071] The integrated decision-making agent serves as the endpoint for information aggregation and output. It is used for decision synthesis: integrating quantitative risk analysis provided by the scenario understanding agent, precise legal basis and historical cases provided by the regulatory matching agent, and best practice templates from a structured knowledge base, to generate specific and actionable rectification decisions. These decisions include: hazard location and description, multimodal evidence (such as images), judgment conclusions, full text of cited regulatory clauses, index of related historical cases, specific rectification measures recommendations (including technical standards, material requirements, responsible departments, and suggested timeframes), and dynamic risk levels.

[0072] This invention employs multiple specialized intelligent agents working in parallel, simulating the collaborative model of a human expert team, to conduct in-depth and professional analysis. Through a parallel processing framework of multi-agent collaborative analysis, multiple agents with different professional capabilities are simultaneously invoked, achieving an analytical effect of "1+1>2". The scene understanding agent utilizes multimodal fusion technology to analyze image, video, and sensor data to understand what is happening on-site (e.g., identifying flames, smoke, abnormal equipment behavior), and provides a preliminary risk level assessment based on historical data. The regulatory matching agent performs high-speed and accurate matching queries between the identified scene and problem and a built-in, vast regulatory knowledge graph to determine whether the current situation violates specific safety regulations, operating procedures, or industry standards. The comprehensive decision generation agent combines the analysis results of the first two agents (i.e., "actual on-site risk" and "regulatory compliance") for comprehensive weighing and decision-making, generating a final, comprehensive risk assessment report that clarifies the risk level, possible consequences, and regulatory basis. Based on the risk assessment results, the corresponding handling process is automatically triggered. The handling path is determined based on preset risk thresholds (e.g., "high risk," "critical"). When the risk level exceeds the threshold, the system automatically executes a series of emergency operations, including but not limited to: sending emergency notifications to relevant personnel, activating audible and visual alarms, linking fire protection / security systems, and initiating evacuation plans, achieving a response time within seconds. When the risk level is within a controllable range, the system automatically generates a formatted hazard report and rectification notice, clearly defining the problem description, risk level, rectification requirements, and deadline, and distributes them to relevant personnel, forming a closed-loop management system.

[0073] In one embodiment, the system is finally implemented and self-evolved through a decision execution and feedback module. The generated decision reports are executed through multiple channels: pushed to the safety management system to form electronic work orders; sent to relevant personnel's mobile terminals; and for high-risk emergencies, sent warnings or interlocking instructions to the DCS or broadcast system via standard interfaces. The system then automatically collects data throughout the entire decision-making lifecycle, including final manual review and confirmation opinions, rectification execution results, and effect verification. This feedback data is structured and stored in a structured knowledge base: as new samples for incremental training of agent models such as task parsing and scenario understanding, optimizing their parameters; and as new cases for enriching and revising the regulatory knowledge graph, enhancing the weight of related edges, or supplementing new relationships. This closed loop enables the system to continuously and autonomously optimize; the longer it operates, the more accurate its identification and the richer its knowledge. To ensure the system can learn from each action and continuously improve its capabilities, whether it's emergency response or work order dispatch, the system records all execution actions and results. On-site personnel or managers then provide final confirmation of the system's handling results and risk assessments; the system is iteratively upgraded using the execution results and manually confirmed data. By leveraging new case data, the AI ​​model is continuously trained, making future judgments more accurate and intelligent. Resolved hazard cases and new regulatory provisions are automatically or semi-automatically updated into the knowledge graph and case library.

[0074] It should be noted that the "model" in the embodiments of the present invention refers to a machine learning (e.g., deep learning) model, which is determined by those skilled in the art according to the specific use case. For example, a large model can be used.

[0075] like Figure 6 As shown, the system in this embodiment of the invention achieves self-evolution, forming a complete learning loop. The system actively collects the execution results of downstream agents, as well as key manual review and correction information, to obtain high-quality "real labels." The system determines whether the received feedback indicates an error in the front-end task parsing process. If the feedback confirms an error, the system automatically adds this "original task + correct result" sample to the reinforcement training set of the corresponding model. If the feedback confirms that the parsing is correct, the process ends normally, and the system enters a continuous monitoring state. The newly added training samples will be used to perform targeted incremental training or rule optimization on the three core models (domain parsing large model, lightweight model, and arbitration model) through the feedback path shown by the dotted line. This allows the system to learn from each error, continuously adapt to new task types and language patterns, and achieve continuous evolution of its capabilities.

[0076] By integrating multi-source heterogeneous data, artificial intelligence technology is used to achieve automated, intelligent, and closed-loop management of safety hazards. Its core objective is to shift from passive post-event response to proactive pre-event warning and in-event intervention, comprehensively improving safety management levels and emergency response efficiency. The entire process is divided into four core stages: data access and task analysis, multi-agent collaborative analysis, decision-making and handling, and closed-loop feedback and optimization, forming a complete intelligent closed loop from perception to action and then to self-evolution.

[0077] The workflow of this invention is described in detail below with reference to four specific embodiments. These embodiments demonstrate how the system initiates a multi-agent collaborative workflow to address different types and risk levels of metallurgical safety hazards, achieving the entire process from intelligent perception to closed-loop management.

[0078] (1) Identification and handling of the hidden danger of insufficient height of the guardrail at the blast furnace tapping area. Scenario description: The inspector found that the height of a section of guardrail on the main platform of blast furnace No. 2 was insufficient and reported it through a mobile terminal. The sensing agent received two pieces of information at the same time: one was the text reported by the inspector, "The height of the guardrail on the west side of the main platform of blast furnace No. 2 is insufficient, about 1 meter"; the other was the real-time monitoring video stream of the area. The dual arbitration mechanism of the central dispatch agent worked in parallel: the first semantic model parsed out "check the compliance of the guardrail height"; the second rule model matched the category of "safety protection facilities". The results of the two paths were consistent, and a structured task was generated: "{target: guardrail, attribute: height, location: west side of the main platform of blast furnace No. 2, action: compliance judgment}", and the scenario understanding and regulatory matching agents were dispatched simultaneously. The scenario understanding agent analyzed the video stream and confirmed that the guardrail height was 1.05 meters through a visual measurement algorithm, which was judged as a high-risk level. The regulatory matching agent performs parallel queries of the knowledge graph, precisely matching GB4053.3-2009, Clause 5.1.1 (requiring the railing height of the site to be ≥1200mm), and associates it with similar historical cases. The comprehensive decision-making agent automatically generates a standardized work order based on the above results, including: a judgment conclusion (non-compliance with national standards), precise evidence (with the original clause text), and rectification suggestions ("increase the height to ≥1200mm within 24 hours"). The decision execution and feedback module pushes the work order to the maintenance department and tracks the closed loop. After rectification, photos are uploaded, the system confirms the closed loop, and this data is added to the database as a positive sample to optimize the model.

[0079] (2) Emergency Response to Suspected Gas Pipeline Leakage. Scenario Description: Sensors in the coke oven gas main pipeline area issue an abnormal alarm. The sensing agent detects a sudden increase in CO concentration from 20ppm to 60ppm, while an infrared thermal imager shows an abnormally low temperature point at a flange. The central dispatch agent receives this multimodal abnormal signal stream, and both dual-path analysis determine it as a "suspected gas leak" emergency event. The highest priority dispatch command is immediately triggered, requiring the scenario understanding and regulatory matching agents and the system emergency interface to respond collaboratively. The scenario understanding agent integrates the concentration rise rate, spatial distribution, and thermal imaging characteristics, combined with pipe pressure data, to determine it as a "medium-scale combustible gas leak," and dynamically delineates the real-time danger zone. The regulatory matching agent quickly matches the GB 6222 emergency response clauses and the enterprise emergency plan number. The comprehensive decision generation agent generates an emergency response decision set. The decision execution and feedback module simultaneously executes multiple automated actions: sending audible and visual alarms and evacuation instructions to personnel positioning terminals in the danger zone; sending a signal to the production control system suggesting the cut-off of the gas source; automatically notifying the emergency team and generating an event report. Afterwards, the data from the handling process was fed back to optimize the early-stage leak identification model.

[0080] (3) Visual predictive maintenance of longitudinal tearing of conveyor belt in raw material yard. Scenario description: The system detects early damage to the belt through intelligent video inspection. The dedicated video analysis submodule of the sensing agent detects a longitudinal bright stripe about 0.8 meters long (a typical feature of the early stage of tearing) in the real-time video stream of the "B102 ore belt". The central dispatch agent analyzes this visual event as "structural damage risk to critical transportation equipment". Although the vibration sensor data is currently normal, the scenario understanding agent is still dispatched to perform in-depth analysis, and the regulatory matching agent is dispatched to find maintenance standards. The scenario understanding agent retrieves the recent current and vibration history data of the belt. Although there are no drastic fluctuations, combined with clear visual evidence, it is still assessed that it has a high risk of causing catastrophic belt breakage. The regulatory matching agent matches the relevant inspection standards in the "Safety Specifications for Belt Conveyors" and pushes historical similar accident cases. The comprehensive decision generation agent generates a predictive maintenance decision and recommends "immediate shutdown for inspection". The report includes a close-up image of the defect, a risk prediction description, and case references. The decision execution and feedback module directly transmits work orders to maintenance teams and the dispatch center, preventing unplanned downtime. This successful identification strengthens the system's logic for identifying latent risks such as "minor visual anomalies + no data alarms".

[0081] (4) Intelligent compliance supervision of confined space operation lack of monitoring. Scenario description: The system automatically supervises the entire process of the approved "crude benzene storage tank cleaning" operation. The sensing agent continuously monitors three data during the operation: gas sensor data inside the tank, facial recognition camera data at the tank opening, and process status of the electronic work ticket system. The central dispatching agent establishes a supervision task chain at the start of the operation. When it detects that "the monitoring personnel are away from their posts for more than 5 minutes" and "the monitoring data upload is interrupted", it is judged as "lack of monitoring violation", and the dispatching regulation matching agent performs a matching based on the data. The regulation matching agent accurately matches the mandatory clauses on continuous monitoring and surveillance in the "Regulations on Safety Management of Confined Space Operations in Industrial and Commercial Enterprises". The comprehensive decision generation agent generates a "three-level alarm" decision. The decision execution and feedback module executes synchronously: 1. Send a vibration alarm to the smart safety helmet of the personnel inside the tank; 2. Send a violation notification to the safety administrator's mobile phone; 3. Highlight the violation on the large screen of the factory safety cockpit. The entire violation process automatically generates an unalterable audit log, realizing mandatory technical supervision of safety procedures.

[0082] As can be seen from the above four embodiments, the system described in this invention can flexibly and accurately respond to various hidden dangers, from static facility defects, dynamic hazard source leakage, early equipment failures to violations of work behavior, fully demonstrating the adaptability, accuracy and reliability of the multi-agent collaborative architecture, and realizing intelligent closed-loop management of metallurgical safety control.

[0083] Building upon this, to facilitate the rapid and convenient acquisition, organization, and processing of multimodal data by the sensing agent, this embodiment of the invention also provides a rapid filtering method. This method organizes relevant data based on inspection tasks for monitoring videos received by the sensing agent and text reports (e.g., inspection logs) filled out by operators (e.g., metallurgical safety inspectors). Specifically, in one embodiment, as... Figure 7 As shown, the method further includes: Step 501: After the inspection personnel complete the identity verification, the inspection task will be started.

[0084] In one embodiment, when an inspection personnel need to conduct an inspection, they first log in to the identity authentication system to start the inspection task, indicating to the system that they have started the inspection. After the identity authentication is completed, they then set off to begin the inspection.

[0085] In one optional embodiment, for each inspection task, the inspector can pre-fill the area involved in the inspection task and the specific inspection task items after identity authentication.

[0086] Step 502: Obtain the inspection monitoring dataset for the inspection task based on the identity authentication information of the inspection personnel.

[0087] Since each area has corresponding monitoring points and is equipped with monitoring equipment such as surveillance cameras, in one optional embodiment, after the inspection personnel complete their identity authentication, the monitoring equipment involved in this inspection task is determined based on their authentication information, and the corresponding inspection monitoring dataset is obtained accordingly. A specific example will be given below, and will not be elaborated further here.

[0088] Step 503: Based on the inspection monitoring dataset, determine the location and data of the objects to be investigated in the inspection task, so as to facilitate subsequent identification of metallurgical hazards.

[0089] In one embodiment, inspection personnel, by performing inspection tasks, identify objects requiring further investigation through on-site manual inspection (e.g., a piece of equipment with loose bolts). Using the inspection monitoring dataset obtained in step 502, the location and data of the objects to be investigated can be obtained or extracted (e.g., displaying the location of equipment with loose bolts through images). Subsequently, those skilled in the art determine, based on the specific application scenario, how to further identify metallurgical hazards according to the location and data of the investigated objects; for example, based on the inspection monitoring dataset, the safety risk level of the investigated objects can be initially classified, and according to the urgency of handling corresponding to the safety risk level, the identification of metallurgical hazards can be achieved either by manually going to the site for confirmation or by continuing to use image recognition methods.

[0090] The following is a further explanation: In one embodiment, before step 502, the method further includes: each monitoring point recording the start and end times of the inspection personnel entering its monitoring range. Each monitoring point may be equipped with one or more cameras. Each camera records the start time of the inspection personnel entering its monitoring range and the end time of their departure from its monitoring range. Furthermore, for security, the system continuously collects monitoring video streams from each monitoring point to be inspected, enabling real-time video retrieval. When the inspection task is initiated, the monitoring points involved in the inspection task are determined. Since inspection tasks (e.g., routine security checks) are often planned in advance, specifying the locations and equipment to be inspected, to expedite the subsequent retrieval, after the inspection personnel initiate the inspection task, the system can pre-determine which monitoring points and their cameras involve the corresponding monitoring angles based on the planned inspection locations and equipment.

[0091] In an optional embodiment, when the inspection personnel start the inspection task, the inspection log can be generated according to the planned items so that the inspection personnel can report the inspection situation through text information to fill in the inspection log and form an inspection report.

[0092] After the inspection task is started, such as Figure 9 As shown, during the on-site inspection, the inspectors fill out an inspection report based on the actual inspection situation; at the same time, the monitoring cameras at each monitoring point continuously record the start and end times of the inspector entering their monitoring range.

[0093] In one embodiment, step 502 includes: when the inspection personnel upload the inspection report of the inspection task, using the identity authentication information to obtain the inspection task and the monitoring points involved in the inspection task. When the inspection personnel upload the inspection report of the inspection task (i.e., the reported text information), and the system receives the inspection report upload event, it obtains the identity authentication information of the inspection personnel; after the inspection personnel upload the inspection report through identity authentication, the inspection task corresponding to the inspection task is obtained through the identity authentication information index, and the monitoring points involved are obtained. Then, based on the inspection report, the start time, and the end time, the inspection monitoring dataset is obtained based on the monitoring points. In one embodiment, specifically, as shown... Figure 8 As shown, it includes: Step 5021: Obtain the multiple monitoring points involved in the inspection task.

[0094] like Figure 9 As shown, for example, an inspection task involves multiple monitoring points, and the corresponding monitoring camera for each monitoring point is determined.

[0095] Step 5022: Extract the object to be investigated from the inspection report; search for the object to be investigated in the set of coverable objects of the multiple monitoring points, and obtain the monitoring point corresponding to the object to be investigated as the point to be investigated.

[0096] After manual inspections by patrol personnel, there are still objects requiring further investigation and subsequent identification of metallurgical hazards. The inspection reports uploaded by the patrol personnel will indicate these objects. The specific method for extracting relevant information about these objects from the inspection reports will be determined by those skilled in the art based on the specific application scenario. In one embodiment, the relevant information about the objects to be investigated is presented in text form in the inspection report, using a large language model to understand the semantics; in another embodiment, patrol personnel are required to explicitly specify the objects to be investigated in text form in the inspection report beforehand, obtaining their unique identifier (e.g., name or number).

[0097] In one embodiment, the monitored objects that can be covered by each surveillance camera at each monitoring point are pre-identified and recorded, forming a view coverage list in document form. The corresponding objects to be investigated are searched in the view coverage lists of the surveillance cameras at the multiple monitoring points involved in the inspection task, thereby obtaining the monitoring points corresponding to these objects to be investigated, which are then designated as the points to be investigated.

[0098] Step 5023: Select the video stream from the corresponding start time to the end time in the monitoring video stream of the location to be investigated as the video to be investigated.

[0099] The corresponding surveillance video streams are obtained from the surveillance cameras at each location to be investigated. For each surveillance camera's video stream, the video stream from the start time to the end time of the patrol personnel is extracted as the video to be investigated. In a specific example, after extracting the objects to be investigated from the patrol reports submitted by the patrol personnel, the objects to be investigated are searched in the set of coverable objects (e.g., cameras 1 to 30) of the surveillance points involved in the patrol task, and the surveillance point corresponding to one of the objects to be investigated is obtained as the location to be investigated. Camera 6 belongs to this location to be investigated, and the surveillance video stream of camera 6 can be located immediately; and the surveillance video stream of camera 6 is extracted according to the start time and end time of the patrol personnel corresponding to camera 6 to obtain the video to be investigated.

[0100] Step 5024: Determine the video frames and / or set of video frames in the video to be investigated that effectively reflect the object to be investigated, so as to obtain the inspection monitoring dataset.

[0101] The specific method for determining the video frames and / or set of video frames that effectively reflect the objects to be investigated in the video to be investigated shall be determined by those skilled in the art based on the specific application scenario. In one embodiment, a safety status dataset of each object to be investigated under both a safety hazard and a normal state can be constructed. The safety status dataset is used to train a visual model, enabling the visual model to learn to effectively reflect the image features of each object to be investigated. The trained visual model is used to identify the video to be investigated, thereby determining the video frames and / or set of video frames that effectively reflect the objects to be investigated through the output of the visual model. In a specific instance, when the object to be investigated is the boiler opening of boiler platform No. 3, after determining at least one viewpoint of the boiler opening to be investigated according to the above steps, based on the operating cycle status of the boiler platform, for the boiler opening of the boiler platform, according to the output of the visual model, determine which / which viewpoints are most effective in reflecting the metallurgical safety hazards corresponding to the boiler opening, and obtain the inspection and monitoring dataset. This dataset can then be used for further manual judgment or intelligent identification, combined with other data (e.g., relevant data on the operation of boiler platform No. 3) for comprehensive analysis, or to continue to go to the site for further on-site investigation of potential hazards.

[0102] In this embodiment of the invention, the monitoring video received by the sensing agent and the text report filled out by the inspection personnel are stored in a simple text format. In addition to the actual application scenario, which already requires storing the monitoring video stream, processing and extracting the data from the inspection report, and storing the view coverage list, the monitoring agent only needs to store the start and end times of the inspection personnel entering the monitoring range of each camera after identity authentication. By recording these key elements, the monitoring video streams of a large number of monitoring cameras can be quickly sorted and filtered for the object to be investigated while saving storage resources as much as possible. This enables the organization of relevant data based on the inspection task, completes the rapid data location, and greatly reduces time and human and material resource costs.

[0103] The foregoing embodiments provided a method for identifying metallurgical hazards based on multi-agent collaboration. In this embodiment, another device for identifying metallurgical hazards based on multi-agent collaboration will be proposed. The device for identifying metallurgical hazards based on multi-agent collaboration includes: a processor and a memory for storing processor-executable instructions; wherein the processor is configured to execute the method for identifying metallurgical hazards based on multi-agent collaboration described in the foregoing embodiments.

[0104] like Figure 10 As shown, the metallurgical hazard identification device based on multi-agent collaboration includes a processor 21 and a memory 22, wherein the processor 21 and the memory 22 can be connected by a bus or other means.

[0105] Processor 21 can be a CPU. Processor 21 can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0106] The memory 22, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the multi-agent cooperative metallurgical hazard identification method in the aforementioned embodiments. The processor executes various functional applications and training processes by running the non-transitory software programs, instructions, and modules stored in the memory.

[0107] The memory 22 may include a program storage area and a training storage area. The program storage area may store the operating system and applications required for at least one function; the training storage area may store training data created by the processor. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory, or other non-transitory solid-state storage device. In some embodiments, the memory 22 may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. The one or more modules stored in the memory 22, when executed by the processor 21, perform the metallurgical hazard identification method based on multi-agent collaboration as shown in the embodiments of the present invention. Specific details of the above-described metallurgical hazard identification method based on multi-agent collaboration can be understood by referring to the corresponding descriptions and effects in the embodiments of the present invention, and will not be repeated here.

[0108] This embodiment also provides a computer storage medium storing a computer program that can be executed by a processor to complete the metallurgical hazard identification method based on multi-agent collaboration described in the foregoing embodiments.

[0109] The computer storage medium stores computer-executable instructions, which can execute the metallurgical hazard identification method based on multi-agent collaboration in any of the above method embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium may also include combinations of the above types of memory.

[0110] The specific steps of the metallurgical hazard identification method based on multi-agent collaboration are described in the foregoing embodiments and will not be repeated in this embodiment.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying potential metallurgical hazards based on multi-agent collaboration, characterized in that, include: Multi-source data input triggers a recognition task, and the perceptual intelligent agent preprocesses the multi-source data input to obtain the object to be analyzed; The central scheduling agent performs deep semantic understanding and metallurgical safety rule matching on the object to be analyzed. Based on the output of deep semantic understanding and metallurgical safety rule matching, it selectively determines the scenario understanding agent and / or the regulatory matching agent to be invoked. The scenario understanding agent performs risk analysis on the object to be analyzed; the regulatory matching agent matches the object to be analyzed with relevant regulations. The integrated decision-making agent generates a final report and instructions based on risk analysis and / or matching results to complete the hazard identification.

2. The metallurgical hazard identification method based on multi-agent collaboration according to claim 1, characterized in that, The method includes: The large language model is pre-trained and fine-tuned using a massive amount of metallurgical safety text to obtain a domain-fine-tuned large model; the domain-fine-tuned large model processes the object to be analyzed and outputs a first intent entity set. The object to be analyzed is subjected to keyword matching using a lightweight model and a security rule base, and a second intent entity set is output. If the consistency between the first intent entity set and the second intent entity set is greater than a preset value, then the first intent entity set or the second intent entity set is used to generate a structured instruction. If the consistency is less than or equal to a preset value, the arbitration big model analyzes the object to be analyzed, the first set of intent entities, the second set of intent entities, and the context of the conflict to generate an arbitration instruction; The urgency level is determined using the structured instructions or the arbitration instructions, and the invocation scenario understanding agent and / or regulatory matching agent are determined according to the urgency level.

3. The metallurgical hazard identification method based on multi-agent collaboration according to claim 2, characterized in that, The method includes: The scenario understanding agent aligns and deeply fuses the multimodal data in the object to be analyzed to obtain fused features, thereby generating a description of potential hazards. Risk levels and hazard types are generated based on the fused features; The method includes: Constructing a metallurgical safety regulations map based on a graph database; The description of the potential hazards is converted into a graph query statement; The query statement described above is used to retrieve matching materials from the metallurgical safety regulations map.

4. The metallurgical hazard identification method based on multi-agent collaboration according to claim 3, characterized in that, The method includes: By integrating the matching materials, risk levels, and hazard types, and combining them with historical best practice templates, an actionable rectification decision is generated to obtain a final report and instructions.

5. The metallurgical hazard identification method based on multi-agent collaboration according to claim 1, characterized in that, The method further includes: The inspection task is initiated after the inspection personnel complete identity verification. Based on the identity authentication information of the inspection personnel, obtain the inspection monitoring dataset for the inspection task; Based on the inspection and monitoring dataset, the location and data of the objects to be investigated in the inspection task are determined to facilitate subsequent identification of metallurgical hazards.

6. The metallurgical hazard identification method based on multi-agent collaboration according to claim 5, characterized in that, The method further includes: Each monitoring point records the start and end times of the patrol personnel entering its monitoring range; Once the inspection task is initiated, the monitoring points involved in the inspection task are determined.

7. The metallurgical hazard identification method based on multi-agent collaboration according to claim 6, characterized in that, The method includes: When the inspection personnel upload the inspection report of the inspection task, they use the identity authentication information to obtain the inspection task and the monitoring points involved in the inspection task. Based on the inspection report, the start time, and the end time, an inspection monitoring dataset is obtained from the monitoring points.

8. The metallurgical hazard identification method based on multi-agent collaboration according to claim 7, characterized in that, The method includes: Obtain the multiple monitoring points involved in the inspection task; Extract the objects to be investigated from the inspection report; search for the objects to be investigated in the set of objects that can be covered by the multiple monitoring points, and obtain the monitoring points corresponding to the objects to be investigated, which are then used as the points to be investigated. The video stream from the corresponding start time to the end time in the monitoring video stream of the location to be investigated is taken as the video to be investigated; Identify the video frames and / or set of video frames in the video to be investigated that effectively reflect the object to be investigated, in order to obtain the inspection and monitoring dataset.

9. A metallurgical hazard identification device based on multi-agent collaboration, characterized in that, The metallurgical hazard identification device based on multi-agent collaboration includes: a processor and a memory for storing processor-executable instructions; The processor is configured to execute the metallurgical hazard identification method based on multi-agent collaboration as described in any one of claims 1 to 8.

10. A non-volatile computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, which are executed by one or more processors to perform the metallurgical hazard identification method based on multi-agent collaboration as described in any one of claims 1 to 8.

Images