Multi-agent collaborative basic model system and method based on agent chain normal form

Abstract

The invention relates to a multi-agent collaborative basic model system and method based on an agent chain normal form. The system is used for the coal industry, the system comprises a coal industry exclusive multi-agent collaborative basic model, the model comprises an input layer, a core layer and an output layer, the core layer receives a demand input by the input layer in a coal industry scene and outputs a processing result aiming at the demand to the output layer, and the output layer outputs the processing result aiming at the demand. The core layer comprises a role module part, the role module part comprises a plurality of agent roles, and each agent role comprises an industry decision agent and a field execution agent; and the dynamic arrangement part is used for dynamically activating the agent roles so as to realize the dynamic cooperation of the plurality of agent roles. According to the invention, multi-agent collaboration can be realized, so that multi-link collaboration in the coal industry is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent technology in the coal industry; specifically, it relates to a multi-agent collaborative basic model system and method based on the agent chain paradigm. Background Technology

[0002] In the coal industry, the demand for intelligentization is concentrated in scenarios such as mining, washing and beneficiation, safety monitoring, and equipment operation and maintenance. However, existing artificial intelligence applications are mostly single-function models (such as equipment fault detection only) and have not fully integrated the collaborative paradigm of the general intelligent agent domain, resulting in the industry's intelligentization still being in a fragmented stage. Summary of the Invention

[0003] In view of this, the present invention provides a multi-agent collaborative basic model system and method based on the agent chain paradigm, thereby solving or at least alleviating one or more of the above-mentioned problems and other problems existing in the prior art.

[0004] To achieve the aforementioned objectives, a first aspect of the present invention provides a multi-agent collaborative basic model system based on the agent chain paradigm, wherein the system is used in the coal industry, the system includes a coal industry-specific multi-agent collaborative basic model, the model includes an input layer, a core layer, and an output layer, the core layer receives the requirements of the coal industry scenario input by the input layer, and outputs the processing results for the requirements to the output layer, the core layer including: The role module includes multiple intelligent agent roles, including industry decision-making intelligent agents and on-site execution intelligent agents. The dynamic orchestration section dynamically activates the intelligent agent roles, enabling dynamic collaboration among the multiple intelligent agent roles.

[0005] In the system described above, optionally, the industry decision-making agent includes: The fully mechanized mining scheduling intelligent agent is used for production coordination in underground fully mechanized mining faces. The fully mechanized mining scheduling intelligent agent dynamically adjusts the task flow in the production of underground fully mechanized mining faces based on data including coal mining machine operation data, support pressure and coal seam thickness. A safety assessment intelligence agent is used to determine the risk level and trigger the risk response process. The basis for determining the risk level and triggering the risk response process includes coal mine safety regulations and historical hazard data. A coal washing optimization intelligent agent is used to optimize the parameters of the coal washing plant's production process, which includes raw coal crushing, raw coal sorting, and product dewatering; and An intelligent operation and maintenance planning agent is used to formulate equipment maintenance plans. The basis for formulating the equipment maintenance plans includes equipment runtime, fault history and production plan. The equipment includes fully mechanized mining equipment and washing and beneficiation equipment. The field execution intelligent agent includes: The device detection agent is used to collect the device's operating data, convert the operating data into a format suitable for the model, and filter out the operating data that contains anomalies. A hazard identification intelligent agent is used to identify hazards, including abnormalities in fully mechanized mining faces, mechanical equipment, and electrical equipment. The basis for identifying hazards includes data collected by underground cameras and infrared sensors. A device control agent is used to control the device to execute operation commands through the device's control system. A ventilation adjustment agent is used to adjust the parameters of the underground ventilation system, including adjusting the airflow and damper positions of the underground ventilation system based on gas and dust concentrations; and The spare parts management intelligent agent is used to query spare parts inventory and apply for spare parts procurement through the coal mine spare parts warehouse system.

[0006] In the system described above, optionally, the dynamic activation of the agent role is based on a state transition equation, which is: in, In time step The persistent inference state within the model, In time step Activated intelligent agent roles, For parameters State transition function, In time step Activated agent role The execution result, Based on the state The role sampling probability distribution Furthermore, the process of dynamically activating the agent role includes: Through the state transition function According to the time step The persistent inference state inside the model The intelligent agent role and the execution result Update the state ; Based on the state From the probability distribution Sampling is performed to determine the role of the intelligent agent. ; Repeat the update of the state The process and the determination of the agent role The process continues until the requirements are met.

[0007] In the system described above, optionally, the dynamic orchestration component optimizes the state transition equation by incorporating industry information, including security risk levels, device response latency, and an industry rule base.

[0008] In the system described above, optionally, the core layer further includes a training framework portion, the process of which the training framework portion trains the model includes: The industry oversight and fine-tuning phase is used to allow the model to learn from collaborative experiences within the coal industry through oversight and fine-tuning. The security reinforcement learning phase is used to optimize the model through reinforcement learning, and the optimization objectives of optimizing the model through reinforcement learning include security and efficiency.

[0009] In the system described above, optionally, the industry oversight fine-tuning phase includes: Based on historical operation data, the collaborative trajectory of scenarios in the mine is collected. The scenarios include fully mechanized mining face production, equipment failure and safety hazards. The collaborative trajectory includes the process of handling the needs generated in the scenario, as well as the relevant data of each link in the process. The relevant data includes equipment data, operation instructions and time nodes. The collaborative trajectory is converted into an intelligent agent chain trajectory, and the requirements of the coal mine safety regulations are embedded in the intelligent agent chain trajectory. Filter the intelligent agent's chain trajectory; The model is trained using the filtered agent chain trajectory; The process of filtering the intelligent agent's chain trajectory includes: Safety compliance filtering, which includes removing the intelligent agent chain trajectory that does not comply with coal mine safety regulations; Working condition adaptation filtering, wherein the working condition adaptation filtering filters the agent chain trajectory according to the working condition complexity, including removing the agent chain trajectory under normal working conditions. Expert verification filtering includes manually reviewing the agent chain trajectory and correcting deviations in the agent chain trajectory.

[0010] In the system described above, optionally, the security reinforcement learning phase includes: Select sample data from coal mining industry scenarios that pose risks; Design safety rewards and efficiency rewards, wherein the safety rewards are calculated based on operational compliance and the efficiency rewards are calculated based on production efficiency indicators; The total reward is designed based on the security reward and the efficiency reward, wherein the weight of the security reward is higher than the weight of the efficiency reward in the total reward; Using the sample data, the model is trained using reinforcement learning based on the total reward.

[0011] In the system described above, optionally, the calculation rules for the security reward include: One point will be awarded if the coal mine safety regulations are met. Five points will be deducted for any violations. 2 points will be added if the hazard handling procedure is triggered. The calculation rules for the efficiency bonus include: One point will be added if the daily output of the fully mechanized mining face meets the target. 2 points will be added if the equipment failure handling time is at least 30% less than the average equipment failure handling time. One point is awarded if the clean coal yield of the coal washing plant increases by at least 1%. And, the total reward ,in For the aforementioned security reward, The efficiency reward is given.

[0012] To achieve the aforementioned objective, a second aspect of the present invention provides a method for using a multi-agent cooperative basic model system based on the agent chain paradigm as described in any one of the first aspects above.

[0013] Optionally, in the method described above, the method includes the following steps: The model is pre-trained in a two-stage manner suitable for the coal industry, the two-stage training including supervised fine-tuning and reinforcement learning; Input the demands of the coal industry into a multi-agent collaborative basic model specific to the coal industry; By using a state transition equation applicable to the coal industry, intelligent agent roles are dynamically activated based on the aforementioned needs. The activated intelligent agent roles make decisions and coordinate or perform on-site tasks in response to the aforementioned needs, thereby achieving dynamic collaboration among multiple intelligent agent roles. Output the processing results of the model for the stated requirements.

[0014] The multi-agent collaborative basic model system based on the agent chain paradigm of the present invention can realize multi-agent collaboration, thereby realizing multi-stage collaboration in the coal industry.

[0015] The present invention further provides a multi-agent collaborative basic model method based on the agent chain paradigm, and therefore the method also has the above-mentioned advantages. Attached Figure Description

[0016] The disclosure of this invention will become more apparent from the accompanying drawings. It should be understood that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings: Figure 1 This is a schematic block diagram of an embodiment of the multi-agent cooperative basic model system based on the agent chain paradigm of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the multi-agent collaborative basic model method based on the agent chain paradigm of the present invention. Detailed Implementation

[0017] Referring to the accompanying drawings and specific embodiments, the structure, composition, characteristics, and advantages of the multi-agent collaborative basic model system and method based on the agent chain paradigm of the present invention will be described below by way of example. However, all descriptions should not be construed as limiting the present invention in any way.

[0018] Furthermore, for any single technical feature described or implied in the embodiments mentioned herein, or any single technical feature shown or implied in the various figures, the present invention still allows for any combination or deletion of these technical features (or their equivalents) without any technical obstacle, and thus these further embodiments according to the present invention should also be considered within the scope of this description.

[0019] Figure 1 This is a schematic block diagram of an embodiment of the multi-agent cooperative basic model system based on the agent chain paradigm of the present invention.

[0020] In this embodiment, the system adopts a coal industry-specific multi-agent collaborative foundation model (Coal-AFM, where AFM stands for Agent Foundation Models) built on the Chain-of-Agents (CoA) paradigm.

[0021] Figure 1 The hierarchical structure of the model (Coal-AFM) is shown. For example... Figure 1 As shown, the model receives demand input from the coal industry and processes this input through a core layer, which consists of three parts: a role module, dynamic orchestration, and a training framework. The role module includes nine agent roles: four industry decision-making agents and five field execution agents. The dynamic orchestration mechanism employs a dynamic orchestration mechanism based on state transition equations and an industry rule base. The training framework uses a two-stage training framework that includes industry-supervised fine-tuning (SFT) and secure reinforcement learning (RL). The processing results from each part of the core layer are transmitted to the output layer to form the coal scenario output, which may include results related to production, safety, or operation and maintenance.

[0022] This model, based on the agent chain paradigm, internalizes the collaborative capabilities of multiple agents into a single model. It transforms industry collaboration logic, such as integrated mining scheduling and safety handling, into a role-switching task chain within the model. This eliminates the need for manual coordination of multiple systems and allows for the dynamic activation of multiple internally defined agent roles within a single inference iteration. Furthermore, agent roles can exchange information, simulating a highly efficient team working seamlessly together. Existing agent system technologies primarily follow two main routes: Multi-Agent Systems (MAS) and Tool-Integrated Reasoning (TIR). In MAS, each agent maintains an independent internal state and interacts via messaging mechanisms, resulting in significant communication overhead. Moreover, its collaboration process heavily relies on manually designed complex workflows. Additionally, this distributed, non-end-to-end architecture prevents data-driven unified learning. Tool-Integrated Reasoning, on the other hand, is essentially a single-threaded model and cannot handle complex scenarios requiring dynamic collaboration among multiple roles. In contrast, Coal-AFM can handle coal industry scenarios that require dynamic collaboration among multiple agents, internalizing multi-agent collaboration capabilities into a single model and fundamentally eliminating the overhead of external communication. In this embodiment, the model's token consumption is reduced by 84.6% compared to general multi-agent systems, making it more suitable for downhole computing power.

[0023] Furthermore, the model is adapted to the coal industry, internalizing the multi-role collaboration process of the coal industry into a single model. By designing exclusive role modules for the coal industry and optimizing the dynamic orchestration mechanism according to the characteristics and needs of the coal industry, the model can complete tasks by dynamically activating industry-specific intelligent agent roles. Industry-specific training is designed in a two-stage training framework to ensure that the model is adapted to the coal scenario and can achieve intelligent collaboration throughout the entire coal scenario process.

[0024] The aforementioned role module is designed for four core scenarios in the coal industry: mining, washing, safety, and operation and maintenance. Based on the intelligent agent chain paradigm, it optimizes and designs two types of nine intelligent agent roles.

[0025] The aforementioned industry decision-making intelligence agents are responsible for coordination and judgment in various scenarios such as coal mining, washing, safety, and operation and maintenance. They undertake high-level tasks such as production scheduling, safety assessment, and process optimization, integrating coal industry standards and expert experience, such as the "Coal Mine Safety Regulations" (Order No. 17 of the Ministry of Emergency Management of the People's Republic of China). The aforementioned four industry decision-making intelligence agents are the fully mechanized mining scheduling intelligence agent, the safety assessment intelligence agent, the washing and beneficiation optimization intelligence agent, and the operation and maintenance planning intelligence agent.

[0026] The fully mechanized mining dispatching intelligent agent acts as the overall production commander, focusing on the coordinated production of the fully mechanized mining face underground. It dynamically adjusts the task flow based on data such as coal mining machine operation data, support pressure, and coal seam thickness. For example, when it detects an increase in the cutting resistance of the coal mining machine, the fully mechanized mining dispatching intelligent agent automatically activates the equipment control intelligent agent in the field execution intelligent agent to adjust the support and increase the initial support force. At the same time, it notifies the hazard identification intelligent agent in the field execution intelligent agent to check the coal wall condition and avoid roof and floor accidents.

[0027] The safety assessment agent acts as a safety director, judging the on-site risk level and triggering response procedures based on coal industry standards such as the "Coal Mine Safety Regulations" and historical hazard data. Furthermore, in some embodiments, the agent role also includes a personnel evacuation agent, responsible for generating evacuation routes based on the location information of underground personnel. For example, when the underground gas concentration reaches a critical value such as 0.8%, the safety assessment agent immediately activates the ventilation adjustment agent within the on-site execution agent to control the underground ventilation system to increase airflow, while simultaneously activating the personnel evacuation agent to generate evacuation routes and notifying the dispatch center and the underground personnel positioning system.

[0028] The coal washing optimization agent acts as a coal washing engineer, optimizing parameters for processes such as raw coal crushing, raw coal sorting, and product dewatering in coal washing plants. For example, based on parameters such as ash content, moisture, and particle size distribution of the raw coal, the optimization agent automatically adjusts the jig amplitude and flotation reagent dosage, and activates the equipment monitoring agent in the field execution agent to monitor the sorting process and verify the sorting effect in real time, ensuring a 2%-3% increase in clean coal yield.

[0029] The operation and maintenance planning intelligent agent acts as an equipment steward, formulating equipment maintenance plans based on equipment runtime, fault history, and production plans in the coal industry. Furthermore, in some embodiments, the intelligent agent role also includes a shutdown and maintenance intelligent agent, responsible for scheduling maintenance windows based on production plans, equipment status, and other information, assisting the operation and maintenance planning intelligent agent in equipment operation and maintenance management. For example, when it is predicted that the scraper conveyor reducer has only a preset number of days remaining in its service life, such as 7 days, the shutdown coordination intelligent agent is activated to schedule maintenance windows without affecting the fully mechanized mining progress. Simultaneously, the spare parts management intelligent agent in the field execution intelligent agent is invoked to confirm spare parts inventory, avoiding unplanned downtime.

[0030] The aforementioned on-site execution intelligent agents are responsible for implementing tasks in various scenarios within the coal industry, including mining, washing, safety, and operation and maintenance. They connect with specialized equipment and systems in the coal industry to perform on-site tasks such as data acquisition, equipment control, and hazard identification. The aforementioned five on-site execution intelligent agents are: equipment monitoring intelligent agent, hazard identification intelligent agent, equipment control intelligent agent, ventilation adjustment intelligent agent, and spare parts management intelligent agent.

[0031] The equipment monitoring intelligent agent acts as a data collector, connecting to the coal mine industrial internet platform to collect real-time operating data of fully mechanized mining equipment such as coal mining machines and scraper conveyors, as well as washing and beneficiation equipment such as jigs and filter presses, including current, temperature, and vibration frequency. It converts the collected operating data into a format that the model can understand, while filtering out abnormal data, such as sensor false alarms.

[0032] The hazard identification intelligent agent acts as an underground inspector, identifying hazards such as coal wall spalling, hydraulic support leakage, and cable damage based on data collected by devices such as underground cameras and infrared sensors. For example, the hazard identification intelligent agent can locate the leakage point of a hydraulic support column through image recognition and output information such as the leakage location and leakage amount. Furthermore, the intelligent agent role also includes a maintenance intelligent agent. When the hazard identification intelligent agent identifies a hazard, in addition to outputting hazard information, it simultaneously triggers the maintenance intelligent agent to generate a maintenance task.

[0033] The equipment control agent acts as a field operator, directly interfacing with the mine equipment control system to execute specific operational instructions. For example, based on instructions from the fully mechanized mining scheduling agent, the equipment control agent remotely adjusts the coal cutting speed of the coal mining machine, such as reducing it from 8 m / min to 6 m / min, or controls the screen inclination angle of the vibrating screen in the coal washing plant. The equipment control agent can automatically and remotely control designated equipment without requiring manual on-site operation.

[0034] The ventilation adjustment agent acts as a ventilation operator, coordinating with the underground ventilation system to adjust parameters such as airflow and damper positions based on factors like methane and dust concentrations. For example, when the safety assessment agent detects excessive methane concentration, the ventilation adjustment agent, once activated, automatically opens a backup damper, increasing the airflow at the working face, for instance, from 800 m³ / min to 1200 m³ / min, to quickly reduce the methane concentration.

[0035] The spare parts management intelligent agent acts as a warehouse manager, connecting to the coal mine's spare parts warehouse system to perform operations such as querying spare parts inventory and requesting purchases. For example, when the maintenance planning intelligent agent formulates a plan to overhaul the scraper conveyor reducer, the spare parts management intelligent agent, after being activated, automatically checks the inventory of scraper conveyor reducer gears. If the inventory is insufficient, it immediately generates a purchase request and pushes it to the supplier system to ensure that spare parts arrive in a timely manner.

[0036] The aforementioned intelligent agent role can reduce the number of on-site personnel required through intelligent decision-making and remote operation, and improve the efficiency of dispatch command response through intelligentization, thereby saving costs. In this embodiment, the number of on-site personnel is reduced by 30%, and the efficiency of dispatch command response is improved by 60%. Simultaneously, the improved response efficiency also includes shorter response times to safety hazards, thus enhancing system safety. Furthermore, the intelligent agent role incorporates industry standards such as the "Coal Mine Safety Regulations" in both decision-making and execution, further improving safety.

[0037] The nine intelligent agent roles described in this embodiment are optimized for industry scenarios, embedded with the "Coal Mine Safety Regulations" and mine equipment control logic, and adapted to the use of underground tools. Furthermore, these intelligent agent roles, deeply adapted to mine equipment, environment, and processes, cover industry decision-making coordination and on-site task execution, making this embodiment widely applicable to scenarios such as underground fully mechanized mining, coal washing and processing, safety monitoring, equipment operation and maintenance, and emergency rescue in the coal industry.

[0038] In other alternative embodiments, the agent role design can be adjusted, for example, designing the aforementioned agent roles such as personnel evacuation agent, shutdown maintenance agent, or maintenance worker agent.

[0039] The aforementioned dynamic orchestration part is based on a state transition equation to automate model state updates and role sampling. This state transition equation is: in, In time step The persistent reasoning state within the model is the model's working memory, which includes historical thinking, action results, and observation information; In time step Activated intelligent agent roles; For parameters State transition function; In time step Activated agent role The execution result is the result of the previous activated agent role. Based on the current (time step t) state The role sampling probability distribution.

[0040] The state transition process includes the following steps: Step S11, the model uses a state transition function. According to the previous time step (time step) The persistent inference state within the model Activated intelligent agent role and its execution results Update the model's current state (at time step t). ; Step S12, based on the current state From the probability distribution Sampling is performed to determine the next activated agent role. ; Step S13: Repeat steps S11 and S12 above until the current task of the model is completed.

[0041] The state transition process described above is completed in a single model inference, achieving efficient internalized collaboration.

[0042] Based on the characteristics of the agent chain paradigm, this dynamic orchestration mechanism enables automated model state updates and role sampling without additional communication overhead, supporting efficient collaboration within inference.

[0043] Furthermore, in this embodiment, factors such as safety risk level, equipment response latency, and industry rule base are incorporated into the state transition equation to optimize it. For example, parameters or constraints are introduced into the state transition function based on the aforementioned factors to achieve state transitions adapted to coal mining scenarios, thereby enabling dynamic collaboration of intelligent agents suitable for the coal industry. Specifically, considering the significant differences in the urgency of tasks in the coal industry (e.g., the need for immediate handling of excessive gas concentration) and equipment response latency (limited underground network bandwidth), the dynamic orchestration mechanism is optimized in the state transition function. The system introduces a weighting factor based on the above characteristics and uses a priority-weighted state transition equation to achieve dynamic activation of roles, ensuring that safety tasks are executed first and equipment instructions are transmitted accurately.

[0044] The aforementioned training framework incorporates coal industry data and experience into the model. Based on the intelligent agent chain paradigm, the training framework focuses on integrating historical mine data, industry rules, and expert experience. Through two stages—industry-supervised fine-tuning and safety reinforcement learning—it ensures that the model is adapted to the coal industry scenario, enabling the model to master the collaborative logic of the coal industry.

[0045] The industry-monitored fine-tuning phase allows the model to learn from the collaborative experience of the coal industry. From data collection to model output, it highlights the coal industry's unique filtering rules (such as safety compliance and working condition adaptation) and trajectory transformation logic, ultimately resulting in a model with preliminary industry capabilities. For example, this phase includes the following steps: mine collaborative trajectory collection, trajectory industry-specific conversion, mine data quality filtering, and industry-monitored fine-tuning training.

[0046] In the mine collaboration trajectory collection process, based on historical operational data, expert collaboration trajectories for scenarios such as fully mechanized mining production, equipment failure, and safety hazards are collected. For example, the complete process for handling a scraper conveyor chain breakage failure is recorded as "inspector discovers chain breakage → reports to dispatch center → shutdown → maintenance team repairs → spare parts warehouse dispatches → production resumes," and equipment data, operating instructions, time nodes, and other information at each stage of the process are collected simultaneously.

[0047] In the process of converting tracks into industry-specific data, scattered tracks such as those from dispatch centers, maintenance teams, and inspection teams are transformed into intelligent agent chain tracks that conform to the coal mining process. For example, the track for handling scraper conveyor chain breakage faults can be integrated into "[Monitoring: Scraper conveyor current abnormality → Activate equipment monitoring intelligent agent → Determine chain breakage → Observation: Chain breakage confirmed → Consideration: Shutdown and maintenance required → Activate fully mechanized mining dispatch intelligent agent → Output shutdown command → Observation: Equipment shutdown → ...]", while embedding the requirements of the "Coal Mine Safety Regulations" into the track, such as the requirement to disconnect the power supply before shutdown.

[0048] In the data quality filtering process for mines, considering the characteristics of coal data such as high noise levels and high safety priority, the filtering strategy is optimized. For example, the following three-layer filtering is designed: Safety and compliance filtering eliminates tracks that do not comply with the "Coal Mine Safety Regulations", such as violations such as failure to shut down operations when gas concentration exceeds the standard, ensuring that the model learns compliant processes; Working condition adaptation filtering retains the trajectory of complex working conditions, such as the trajectory of unconventional working conditions such as thin coal seam mining and steeply inclined fully mechanized mining, while removing simple working condition data, such as the working condition of stable production in normal coal seams, thereby improving the model's ability to cope with complex scenarios. Experts validated the filtering process, and senior engineers from coal mines (such as longwall mining team leaders and safety managers) were invited to review the trajectory and correct empirical deviations, such as differences in support adjustment parameters for different coal seam thicknesses.

[0049] In the industry supervision fine-tuning training step, a high-quality mine trajectory training model is used, which includes data on fully mechanized mining, washing and beneficiation, safety, operation and maintenance, etc. The model predicts the activation and operation commands of intelligent agent roles that conform to industry rules in the trajectory, ensuring that the output meets the actual needs of the mine.

[0050] The safety reinforcement learning phase enhances model safety and efficiency, with high-risk sample sampling and a dual-objective reward function design at its core. This phase's sample selection and reward design emphasizes the safety-first reward weighting in the coal industry, ensuring the model optimization direction aligns with mine needs and ultimately outputting a high-performance industry model. For example, since safety takes precedence over efficiency in the coal industry, this phase focuses on optimizing the balance between safety compliance and production efficiency. This is achieved through high-risk sample sampling and the design of a coal-specific reward function, enabling reinforcement learning training for mining scenarios.

[0051] When sampling high-risk samples, challenging samples that are likely to cause safety accidents or affect production efficiency are screened out, such as fluctuations in gas concentration, abnormal noises from the coal cutting section of the coal mining machine, and decreased sorting accuracy in the coal washing plant, to ensure that the model's capabilities are enhanced in high-risk scenarios. In this embodiment, high-risk event data from the coal mine over the past three years were collected, including 200 gas anomalies and 150 equipment failures.

[0052] When designing a reward function specific to coal mining, both safety and efficiency objectives are considered, with safety taking priority. For example, the reward function is as follows: in, For the total reward, As a safety reward, Rewards for efficiency. Rewards for security. Based on operational compliance calculations, the rules can be set as follows: 1 point for compliance with the "Coal Mine Safety Regulations," 5 points deducted for violations (such as continuing production despite excessive gas levels), and 2 points awarded for triggering hazard handling procedures. Efficiency bonuses are also available. Based on the production efficiency index, the calculation rules can be set as follows: 1 point is awarded for meeting the daily output target of the fully mechanized mining face, 2 points are awarded for shortening the equipment failure handling time, and 1 point is awarded for increasing the clean coal yield of the coal washing plant by 1%. The standard for shortening the equipment failure handling time can be set as 30% less than the average time.

[0053] When conducting reinforcement learning training in mining scenarios, the model is trained using high-risk sample sampling and reward functions, and high-risk event data from coal mines. This allows the model to optimize operations while maintaining safety and compliance. In this embodiment, through training, the model can adjust ventilation in advance when the gas concentration reaches 0.7%, rather than waiting for the set critical value of 0.8%, thus avoiding risks without affecting the production schedule.

[0054] In this embodiment, after training, the model's response time from hazard detection to problem resolution was reduced to 45 seconds, a 70% speedup compared to traditional manual methods, with a 92% reduction in violations and improved safety. Furthermore, daily output at the fully mechanized mining face increased by 8%, the clean coal yield at the coal washing plant increased by 2.5%, and equipment idling time decreased by 15%, achieving efficiency optimization.

[0055] This embodiment can be widely applied to scenarios such as underground fully mechanized mining, coal washing and processing, safety monitoring, equipment operation and maintenance, and emergency rescue in the coal industry, such as the intelligent mine construction of large coal enterprises like China Energy Investment Corporation and China Coal Group. It can also respond to new working conditions with zero samples, such as changes in coal seam thickness.

[0056] Furthermore, in optional embodiments, the system can also be extended to open-pit coal mine scenarios. Even further, in other optional embodiments, the system can also be extended to similar scenarios in non-coal mines, such as metal mines, with the system design tailored to the extended scenarios.

[0057] Figure 2 This is a flowchart illustrating an embodiment of the multi-agent collaborative basic model method based on the agent chain paradigm of the present invention.

[0058] like Figure 2 As shown, in this embodiment, the method includes the following steps: Step S20: The model is pre-trained in two stages: supervised fine-tuning and reinforcement learning, which are suitable for the coal industry. Step S21: Input the demand from the coal industry into the model; Step S22: By using the state transition equation applicable to the coal industry, the intelligent agent roles are dynamically activated to make decisions and coordinate or perform on-site tasks, thereby realizing dynamic collaboration among multiple intelligent agents. Step S23: Output the model processing results.

[0059] In step S20, the two-stage training injects coal industry data and experience into the model, including historical mine data, industry rules, and expert experience, ensuring that the model is adapted to the coal industry scenario and enabling the model to master the collaborative logic of the coal industry. After training in step S20, the model is put into application, and in actual application, steps S21-S23 are executed for each input.

[0060] Steps S21-S23 constitute the entire process of the model processing and outputting a single input requirement. Based on a dynamic orchestration mechanism, automated model state updates and role sampling are achieved without additional communication overhead, supporting efficient collaboration within inference. Step S22 is completed within a single model inference iteration, realizing efficient internalized collaboration. The activated agent roles cover industry decision-making coordination and on-site task execution, and can be widely applied to various scenarios in the coal industry, including underground fully mechanized mining, coal washing and processing, safety monitoring, equipment operation and maintenance, and emergency rescue.

[0061] Furthermore, in other alternative embodiments, the method can also be extended to similar scenarios, such as non-coal mines in metal mines, and the model training framework, state transition equations, and multi-agent role design can be adjusted for the extended scenarios.

[0062] Some embodiments of the present invention, based on the agent chain paradigm, can overcome the shortcomings of general multi-agent systems in the coal industry, such as poor process adaptation, slow response, and high resource consumption, and realize the native internalization of industry collaboration processes.

[0063] Some embodiments of the present invention, based on the intelligent agent chain paradigm, can overcome the single-threaded limitation of tool integration reasoning, simulate the parallel collaboration mode of multiple positions in the coal industry, and adapt to mining-specific tools and equipment.

[0064] Some embodiments of the present invention construct an industry-specific intelligent agent covering the entire process, ensuring that the operation complies with the "Coal Mine Safety Regulations", which can solve or at least alleviate one or more of the problems in existing coal mine artificial intelligence solutions, such as fragmentation, strong reliance on human intervention, and weak generalization ability.

[0065] As mentioned above, in one embodiment of the present invention, the response time of the trained model from hazard detection to problem resolution is reduced to 45 seconds, which is 70% faster than traditional manual methods, and violations are reduced by 92%, thus improving safety. Furthermore, in one embodiment of the present invention, the trained model achieves an 8% increase in daily output at the fully mechanized mining face, a 2.5% increase in the yield of clean coal at the coal washing plant, and a 15% reduction in equipment idling time, achieving efficiency optimization. In addition, in one embodiment of the present invention, after the model is applied to coal mines, the number of on-site personnel is reduced by 30%, the response efficiency of dispatch instructions is improved by 60%, and the model token consumption is reduced by 84.6% compared to general multi-agent systems, saving costs. Moreover, in one embodiment of the present invention, the model can be adapted to multiple scenarios such as open-pit coal mines, underground fully mechanized mining, and coal washing and processing, and can handle new working conditions with zero samples, such as changes in coal seam thickness, demonstrating strong generalization ability.

[0066] The technical scope of this invention is not limited to the contents of the above specification. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the scope of this invention.

Claims

1. A multi-agent collaborative base model system based on the agent chain paradigm, characterized in that, The system is used in the coal industry, and the system comprises a multi-agent collaborative basic model specific to the coal industry, the model comprising an input layer, a core layer and an output layer, the core layer receiving demands in a coal industry scene input by the input layer and outputting processing results for the demands to the output layer, the core layer comprising: a role module part, the role module part comprising a plurality of agent roles, the agent roles comprising industry decision-making agents and on-site execution agents; a dynamic arrangement part, the dynamic arrangement part dynamically activating the agent roles to realize dynamic collaboration of the plurality of agent roles.

2. The system of claim 1, wherein, The industry decision-making agents comprise: fully mechanized mining scheduling agents for underground fully mechanized mining face production collaboration, the fully mechanized mining scheduling agents dynamically adjusting task processes in underground fully mechanized mining face production according to data including shearer operation data, support pressure and coal seam thickness; safety research and judgment agents for judging risk levels and triggering risk response processes, the basis for judging risk levels and triggering risk response processes including coal mine safety regulations and historical hidden danger data; washing and sorting optimization agents for optimizing parameters of a coal washing plant production process, the process including raw coal crushing, raw coal sorting and product dewatering; and operation and maintenance planning agents for formulating equipment maintenance plans, the basis for formulating equipment maintenance plans including equipment operation time, fault history and production plans, the equipment including fully mechanized mining equipment and washing and sorting equipment; The on-site execution agents comprise: equipment detection agents for collecting operation data of the equipment, converting the operation data into a format suitable for the model, and filtering the operation data with abnormalities; hidden danger identification agents for identifying hidden dangers, the hidden dangers including hidden dangers of abnormalities of fully mechanized mining faces, mechanical equipment and electrical equipment, the basis for identifying hidden dangers including data collected by underground cameras and infrared sensors; equipment control agents for controlling the equipment to execute operation instructions through control systems of the equipment; ventilation adjustment agents for adjusting parameters of an underground ventilation system, the adjusting parameters of the underground ventilation system including adjusting air volume and damper position of the underground ventilation system according to gas concentration and dust concentration; and spare parts management agents for querying spare parts inventory and applying for procurement of spare parts through a coal mine spare parts warehouse system.

3. The system of claim 1, wherein, The dynamic activation of the agent roles is based on a state transition equation, the state transition equation being wherein, is a persistent inference state inside the model at time step is an agent role activated at time step is a state transition function parameterized by is an execution result of agent role is a role sampling probability distribution based on the state ​​​​​​ Furthermore, the process of the dynamic activation of the agent roles comprises: by the state transition function , according to the persistent inference state inside the model at the time step , the agent role and the execution result , update the state ; based on the state from the probability distribution sampling, determining the agent role ; repeating the updating the state and the determining the agent role processes until the processing of the demand is completed.

4. The system of claim 3, wherein, The dynamic arrangement part optimizes the state transition equation by incorporating industry information, the industry information including safety risk levels, equipment response delays and an industry rule library.

5. The system of claim 1, wherein, The core layer further comprises a training framework part, the process of the training framework part training the model comprising: an industry supervision fine-tuning stage for enabling the model to learn coal industry collaboration experience through supervised fine-tuning; a safety reinforcement learning stage for optimizing the model through reinforcement learning, the optimization target of the optimizing the model through reinforcement learning including safety and efficiency.

6. The system of claim 5, wherein, The industry supervision fine-tuning stage comprises: Based on historical operation data, collect collaborative trajectories of scenes in the mine, the scenes including fully mechanized mining face production, equipment failure and safety hazards, the collaborative trajectories including a flow for processing demands generated under the scenes, and related data of each link in the flow, the related data including equipment data, operation instructions and time nodes; Convert the collaborative trajectories into agent chain trajectories, and embed requirements of coal mine safety regulations in the agent chain trajectories; Filter the agent chain trajectories; Train the model using the filtered agent chain trajectories; The filtering of the agent chain trajectories includes: Safety compliance filtering, which includes eliminating the agent chain trajectories that do not comply with coal mine safety regulations; Working condition adaptation filtering, which filters the agent chain trajectories according to working condition complexity, including eliminating the agent chain trajectories under regular working conditions; Expert verification filtering, which includes manually auditing the agent chain trajectories and correcting deviations in the agent chain trajectories.

7. The system of claim 5, wherein, The safety reinforcement learning stage includes: Screening sample data of coal mine industry scenes that exist risks; Designing safety rewards and efficiency rewards, the safety rewards being calculated according to operation compliance, and the efficiency rewards being calculated according to production efficiency indicators; Designing total rewards according to the safety rewards and the efficiency rewards, the weight of the safety rewards being higher than the weight of the efficiency rewards in the total rewards; Using the sample data to perform reinforcement learning training on the model based on the total rewards.

8. The system of claim 7, wherein, The calculation rules of the safety rewards include: Adding 1 point in the case of compliance with coal mine safety regulations, Deducting 5 points in the case of existence of illegal operation, Adding 2 points in the case of triggering hidden danger processing flow; The calculation rules of the efficiency rewards include: Adding 1 point in the case of daily output of fully mechanized mining face meeting the standard, Adding 2 points in the case of equipment failure handling time being at least 30% less than average equipment failure handling time, Adding 1 point in the case of coal preparation plant clean coal yield increasing by at least 1%; and, the total reward wherein is the safety reward, is the efficiency reward.

9. A method applied to the system of any one of claims 1-8.

10. The method of claim 9, wherein, The method includes the following steps: Pre-training the model for two stages suitable for the coal industry, the two-stage training including supervised fine-tuning and reinforcement learning; Inputting demands of the coal industry into a coal industry-specific multi-agent collaborative basic model; Dynamically activating agent roles based on the demands through a state transition equation suitable for the coal industry, the activated agent roles making decision coordination or on-site task execution for the demands, realizing dynamic collaboration of multiple agent roles; Outputting processing results of the model for the demands.

Images