A mine production scheduling auxiliary method and system based on multi-agent

Abstract

The application relates to the technical field of mine production scheduling optimization, and provides a mine production scheduling auxiliary method and system based on multiple agents, which comprises the following steps: obtaining multiple-source data related to mine production scheduling, processing the data and constructing a data base; constructing a candidate training sample set based on the data base, screening target training samples according to mine constraint conditions and production indexes, and updating model parameters of a scheduling agent base; obtaining a mine production scheduling request, generating a mine production scheduling task, calling a distributed collaboration framework constructed based on the data base and the optimized scheduling agent base model, splitting and executing the scheduling task, obtaining equipment state information, scheduling constraint conditions and a candidate scheduling suggestion set; arbitrating based on constraint priorities, obtaining a target scheduling suggestion, and converting the target scheduling suggestion into an executable scheduling instruction set. The method improves the adaptability of scheduling to on-site changes and the executability of scheduling instructions through hierarchical multiple-agent collaborative processing.

Description

Technical Field

[0001] This invention relates to the field of mine production scheduling optimization technology, specifically to a mine production scheduling auxiliary method and system based on multi-agent systems. Background Technology

[0002] Production equipment scheduling is a crucial aspect of underground mine production organization, involving multiple aspects such as mining, transportation, loading and unloading, access, and equipment maintenance. Its scheduling effectiveness directly impacts production continuity, equipment utilization, and operational safety.

[0003] In traditional underground mining operations, equipment scheduling typically relies on dispatchers' experience to make arrangements. Dispatchers need to simultaneously monitor factors such as equipment operating status, roadway accessibility, face production progress, loading / unloading capacity, and ore transportation demands. Due to the complex underground mining environment and frequent changes in on-site conditions, relying solely on manual methods makes it difficult to promptly grasp the overall production status, easily leading to problems such as vehicle waiting, equipment idleness, poor loading / unloading coordination, and transportation route conflicts. When unexpected situations arise, such as equipment failure, roadway congestion, or changes in face plans, manual dispatching often requires information verification, on-site communication, and plan adjustments, resulting in a slow response time and potentially impacting the continuity of ore flow and production organization efficiency.

[0004] Currently, intelligent scheduling technologies in mines can collect, store, and display data on personnel, equipment, environment, location, and production tasks, providing a data foundation for scheduling decisions. However, in practical applications, these technologies largely remain at the level of data aggregation and status monitoring, lacking the ability to perform real-time analysis and collaborative processing in complex production scenarios. Some scheduling models rely on historical data and fixed rules, making it difficult to adapt to the constantly changing production needs and operating conditions in mines, leading to discrepancies between scheduling results and actual operations. Furthermore, underground mines have a wide variety of equipment and closely interconnected operational processes; adjustments to a single piece of equipment or work point can affect subsequent transportation, loading and unloading, and production arrangements. Existing intelligent scheduling methods, when dealing with equipment failures, task changes, and restricted roadway access, do not adequately consider the interrelationships between different operational processes, easily resulting in better local processing outcomes but low overall efficiency. Summary of the Invention

[0005] This invention aims to solve the problems in the prior art and provide a multi-agent-based auxiliary method and system for mine production scheduling with high overall collaborative adjustment efficiency.

[0006] To achieve the above objectives, the first aspect of this application provides a multi-agent-based auxiliary method for mine production scheduling, comprising:

[0007] S1. Obtain multi-source data related to mine production scheduling, perform data governance on the multi-source data to obtain standard scheduling data, and construct a data foundation based on the standard scheduling data. S2. Construct a candidate training sample set for training the scheduling agent base model based on the data base; screen the candidate scheduling suggestions in the candidate training sample set according to the mine constraints and production indicators to obtain the target training sample; optimize the scheduling agent base model based on the target training sample to obtain the optimized scheduling agent base model. S3. Obtain a mine production scheduling request and generate a mine production scheduling task based on the mine production scheduling request; invoke a distributed collaboration framework built on the data base and the optimized scheduling agent base model, the distributed collaboration framework including a hierarchical multi-agent structure, and decompose the mine production scheduling task through the distributed collaboration framework to obtain a set of scheduling subtasks. S4. Execute the set of scheduling subtasks through the distributed collaboration framework to obtain device status information, scheduling constraints, and a set of candidate scheduling suggestions; based on the device status information and the scheduling constraints, arbitrate the set of candidate scheduling suggestions based on constraint priority to obtain the target scheduling suggestion; S5. Convert the target scheduling suggestion into a set of executable scheduling instructions and output the set of scheduling instructions.

[0008] A second aspect of this application provides a computer system including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned multi-agent-based mine production scheduling auxiliary method.

[0009] Compared with the prior art, this application has the following beneficial effects: This invention, based on the construction of a data foundation and an optimized scheduling agent foundation model, introduces a distributed collaborative framework including a hierarchical multi-agent structure. This allows mine production scheduling tasks to be planned and broken down by the overall agent, and then collaboratively processed by different agents focusing on equipment status identification, scheduling knowledge retrieval, and scheduling suggestion generation. Target scheduling suggestions are then formed through constraint priority arbitration. Since each agent undertakes different scheduling processing functions on a unified data basis and can aggregate their outputs into subsequent decision-making processes, it reduces the problems of biased judgment, untimely response, and difficulty in implementing scheduling suggestions that easily occur in single-model or manual experience-based scheduling. Through this approach, the adaptability of mine production scheduling to on-site changes can be improved, making the output scheduling instructions more suitable for current production conditions, thereby enhancing the executability of scheduling instructions and the efficiency of mine production organization. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating a method provided in an embodiment of this application; Figure 2 This is a framework diagram of a multi-agent-based mine production scheduling auxiliary system provided in a specific embodiment of this application. Detailed Implementation

[0012] To facilitate understanding of this application, the following description will be more comprehensive and detailed in conjunction with the accompanying drawings and preferred embodiments, but the scope of protection of this application is not limited to the following specific embodiments.

[0013] Unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the scope of this application.

[0014] Please see Figure 1 One embodiment of the present invention provides a mine production scheduling assistance method based on multi-agent systems, comprising the following steps.

[0015] S1. Obtain multi-source data related to mine production scheduling, perform data governance on the multi-source data to obtain standard scheduling data, and build a data foundation based on the standard scheduling data. In this step, multi-source data comes from the mine's production execution system, IoT platform, and scheduling interaction process. Data from different sources differs in field names, collection frequency, data format, and time recording methods, making it unsuitable for direct use in scheduling inference and model training. Therefore, the multi-source data is first uniformly accessed, and then the information is processed to form standard scheduling data that can be correlated by equipment, work area, shift, and time. After the standard scheduling data is written into the data base, the data base is used to store information such as production plans, equipment status, on-site sensing results, scheduling procedures, shift handover records, and manual feedback. The data base can also store mine production status information and structured scheduling knowledge extracted from the standard scheduling data. Mine production status information reflects the current equipment location, equipment operating status, work point load, transportation channel status, and on-site anomalies. Structured scheduling knowledge reflects equipment capacity boundaries, roadway access restrictions, safety procedures, loading and unloading point constraints, and historical scheduling experience.

[0016] S2. Construct a candidate training sample set based on the data foundation for training the scheduling agent foundation model; screen the candidate scheduling suggestions in the candidate training sample set according to the mine constraints and production indicators to obtain the target training samples; optimize the scheduling agent foundation model based on the target training samples to obtain the optimized scheduling agent foundation model. In this step, the candidate training sample set consists of both real scheduling cases and synthetic scheduling samples. Real scheduling cases mainly come from historical shift plans, mining operation orders, scheduling logs, equipment execution results, scheduling personnel adoption records, and manual modification records. Synthetic scheduling samples are generated by the scheduling agent base model under a given scheduling input, and are used to supplement the problem of insufficient number of real cases or inadequate coverage of working conditions.

[0017] The scheduling inputs in the candidate training samples can include the current shift plan, available equipment, equipment location, equipment status, work point requirements, road traffic status, and scheduling constraints. Candidate scheduling suggestions can include equipment allocation, transportation routes, loading and unloading sequences, work schedules, and anomaly handling methods. The system performs quality screening on the candidate scheduling suggestions based on mine constraints and production indicators. Mine constraints include safety constraints, equipment capacity constraints, roadway access constraints, work time window constraints, and material level constraints. Production indicators include ore output per unit time, equipment waiting time, equipment idle time rate, ore level fluctuations in the ore pass, and the number of violations.

[0018] After screening, candidate scheduling suggestions that clearly violate safety regulations, equipment capacity boundaries, or on-site access restrictions were eliminated, while those that met the constraints and had better production indicators were retained as target training samples. The scheduling agent base model was optimized based on the target training samples and its parameters were updated. Through continuous optimization and iteration of the model parameters, the scheduling suggestions output by the model gradually aligned with the scheduling habits and production constraints of the mine site.

[0019] S3. Obtain the mine production scheduling request and generate the mine production scheduling task based on the mine production scheduling request; call the distributed collaboration framework built on the data foundation and the optimized scheduling agent foundation model. The distributed collaboration framework includes a hierarchical multi-agent structure. The mine production scheduling task is split through the distributed collaboration framework to obtain a set of scheduling sub-tasks. In this step, mine production scheduling requests can originate from shift production plans, equipment malfunction alarms, changes in transport road conditions, adjustments to work point requirements, or temporary instructions from dispatchers. After receiving a scheduling request, the system converts it into a unified task representation. This unified task representation records the work object, loading point, unloading point, demand, time requirements, available equipment, constraint tags, and current production status.

[0020] The distributed collaboration framework adopts a hierarchical multi-agent structure. The central agent is responsible for task planning and decomposition, dividing the mine production scheduling task into scheduling sub-tasks such as status monitoring, knowledge retrieval, scheme generation, and instruction execution. Lower-level agents execute sub-tasks according to their respective responsibilities. The early warning agent is responsible for analyzing equipment operating status, the knowledge base agent is responsible for querying equipment parameters and scheduling procedures, the scheduling agent is responsible for generating a set of candidate scheduling suggestions and arbitrating based on constraint priorities to obtain the target scheduling suggestion, and the decision agent is responsible for converting the target scheduling suggestion into a set of executable scheduling instructions and outputting the set of scheduling instructions. Through this decomposition, complex scheduling tasks are no longer completed by a single model all at once, but are handled by multiple agents under unified constraints.

[0021] S4. Execute a set of scheduling subtasks through a distributed collaboration framework to obtain device status information, scheduling constraints, and a set of candidate scheduling suggestions; through the scheduling agent in the distributed collaboration framework, arbitrate the set of candidate scheduling suggestions based on constraint priority according to the device status information and scheduling constraints to obtain the target scheduling suggestion. In this step, the early warning agent uses device location data, device operating status data, and sensor sampling data to determine anomalies in the device's operating status, and incorporates the anomaly determination result as part of the device status information. Specifically, let the device... At any moment The state vector is ,equipment The historical average is The historical standard deviation is And it is a scalar, with an anomaly threshold coefficient of . Then the anomaly detection result satisfies:

[0022] Among them, when When, it indicates the equipment At any moment An anomaly is detected; the alert agent outputs the anomaly event, the abnormal device, and the confidence level. When, it indicates the equipment At any moment It was not identified as abnormal.

[0023] The knowledge base agent retrieves scheduling procedures, equipment capacity limits, roadway access restrictions, maintenance records, and historical similar cases from the data base based on the current equipment and work area to obtain scheduling constraints. These constraints include material level constraints and access constraints. For ore passes... To determine the upper limit of the material level safety, the following constraint verification function is constructed:

[0024] in, Indicates candidate scheduling suggestions chute The results of the constraint verification of the upper limit of material level safety. Indicates candidate scheduling suggestions chute The expected feed volume generated Indicates well chute At the upper limit of the safe material level The allowed capacity.

[0025] For a closed roadway, the following access constraint verification function is constructed:

[0026] in, Indicates the side of a closed alleyway. Indicates candidate scheduling suggestions The corresponding transportation route This is an indicator function. This constraint is not satisfied when the route contains closed lane edges.

[0027] The scheduling agent combines equipment status information and scheduling constraints to generate a set of candidate scheduling suggestions. These suggestions include equipment allocation, transportation routes, loading / unloading sequences, operation schedules, and anomaly handling methods. Specifically, for transportation routes in the candidate scheduling suggestions, the mine roadway network is modeled as a roadway diagram. ,in, Represents the set of roadway nodes. Represents the set of lane edges; for the starting node and target node In the alleyway map China has determined from arrive Path set The transportation path is determined according to the following path cost function:

[0028] in, In the tunnel diagram, from node To the node The minimum path cost, Indicates from node To the node One of the candidate paths, It is the side of the alleyway in the path. For the side of the alley Length, This represents the congestion coefficient along the alleyway. Congestion weighting.

[0029] When multiple candidate scheduling suggestions compete for resources—for example, multiple pieces of equipment simultaneously occupying the same roadway, loading point, or unloading point—or when a scheduling suggestion, while highly efficient, poses a safety risk, the system arbitrates according to constraint priority. In this priority structure, safety constraints take precedence over production efficiency constraints, and equipment capacity and access constraints take precedence over simple capacity improvement goals. The system first retains candidate scheduling suggestions that meet high-priority constraints, then performs a comprehensive evaluation of the retained suggestions to arrive at the target scheduling suggestion. This approach prevents scheduling suggestions from solely pursuing efficiency while neglecting safety and on-site feasibility.

[0030] S5. Convert the target scheduling proposal into a set of executable scheduling instructions and output the set of scheduling instructions.

[0031] In this step, the target scheduling suggestion is a solution-level result and needs to be converted into scheduling instructions that can be recognized by the production execution system, scheduling terminal, or vehicle terminal. The set of scheduling instructions can include equipment number, job type, starting location, target location, transportation route, loading and unloading requirements, execution time, obstacle avoidance requirements, and anomaly handling requirements. When outputting the set of scheduling instructions, the instruction version, effective time, associated scheduling request, generation basis, and execution feedback identifier are recorded simultaneously for easy traceability later.

[0032] In this embodiment, a data foundation capable of unified access to the production execution system, IoT platform, and scheduling interaction feedback is established. Then, the scheduling agent foundation model is continuously optimized using real-world cases and synthetic samples. Subsequently, a hierarchical multi-agent structure is used to complete task decomposition, state analysis, knowledge retrieval, solution generation, and conflict arbitration, finally outputting a set of executable scheduling instructions. This allows the mine scheduling process to simultaneously consider production plans, equipment status, on-site constraints, and human feedback, reducing issues such as equipment allocation imbalances, excessively long transportation waiting times, poor process coordination, and delayed responses to anomalies.

[0033] like Figure 2As shown, the data source, end-to-end data hub, dynamic training system, distributed collaboration framework, and output layer correspond to the data access, data foundation construction, model optimization and parameter update, multi-agent collaborative scheduling, and scheduling instruction output processes in this embodiment, respectively. The data source includes the production execution system, the IoT platform, and user interaction streams, providing production execution system data, IoT platform data, and scheduling interaction feedback data, respectively. The end-to-end data hub includes a data gateway, a data processing module, and a knowledge base. The data gateway accesses multi-source data, the data processing module performs data governance on the multi-source data to obtain standard scheduling data, and the knowledge base stores scheduling procedures, equipment parameters, roadway access restrictions, and historical scheduling experience. The data gateway, data processing module, and knowledge base together form the aforementioned data foundation. The dynamic training system includes a sample construction module, a quality screening module, and a strategy optimization module, used to construct a candidate training sample set based on the data foundation, screen target training samples according to mine constraints and production indicators, and optimize the scheduling agent foundation model based on the target training samples, respectively. The distributed collaboration framework comprises a central intelligent agent, an early warning agent, a knowledge base agent, a scheduling agent, and a decision agent. It is used to break down and execute mine production scheduling tasks, arbitrate candidate scheduling suggestions, and obtain the target scheduling suggestion. The output layer includes scheduling instructions and execution logs. The scheduling instructions correspond to the aforementioned set of executable scheduling instructions, and the execution logs record the execution results and human feedback of the scheduling instructions, serving as a data source for subsequent optimization of the scheduling agent's underlying model.

[0034] In one embodiment, in S1, the multi-source data includes: production execution system data, IoT platform data, and scheduling interaction feedback data; The production execution system data includes shift plans, mining operation orders, equipment ledgers, and maintenance records. IoT platform data includes device location data, device operating status data, sensor sampling data, and on-site image acquisition data; The dispatch interaction feedback data includes dispatch logs, dispatch procedures, shift handover records, dispatcher adoption records, and manual modification records; Production execution system data and scheduling interaction feedback data mainly form text data and structured form data. Device positioning data, device operating status data and sensor sampling data in the Internet of Things platform mainly form time series data, and on-site image acquisition data form image data.

[0035] Data governance is performed on multi-source data to obtain standard scheduling data, specifically including: Text data, image data, and time-series data are extracted from multi-source data according to their data formats. The text data, image data, and time-series data are then cleaned, formatted, and time-stamped respectively to obtain standard text data, standard image data, and standard time-series data. Specifically, time-series data undergoes denoising, missing data completion, resampling, and time alignment; text data undergoes segmentation, deduplication, field normalization, and entity normalization; and image data undergoes key region cropping, compression, and time-stamp association.

[0036] Feature extraction was performed on standard text data, standard image data, and standard time-series data respectively to obtain text features, image features, and time-series features; Text features, image features, and temporal features are used as feature data of different modalities for modal alignment and fusion to obtain a multimodal fusion vector. Standard text data, standard image data, standard temporal data, and multimodal fusion vector are used as standard scheduling data and written into the data base. Modality alignment and fusion are implemented using a multi-head attention mechanism, specifically including: Text features Image features With time series characteristics Each modal vector of uniform dimension d is obtained through linear mapping:

[0037] in, , , All are learnable projection matrices. , , These are the original feature dimensions for each modality; Input a mode sequence composed of mode vectors:

[0038] Let the number of attention heads be , No. The query dimension of the size is Value dimension is Calculate the query matrix, key matrix, and value matrix:

[0039] in For the first The projection matrix of each attention head; No. The output of each attention head is:

[0040] Attention output is obtained after splicing multiple heads:

[0041] in, To output the projection matrix; The attention output is average-pooled to obtain the multimodal fusion vector:

[0042] in, Representation matrix The Row vectors.

[0043] Specifically, in the feature fusion stage, text features, image features, and temporal features are first mapped to the same dimension and then input into a multi-head attention structure. The multi-head attention structure is used to calculate the degree of correlation between different modalities. For example, equipment anomalies recorded in scheduling logs, the status of the work area shown in on-site images, and operational fluctuations reflected in sensor data can be correlated within the same vector space. The multi-modal fusion vector obtained after mean pooling is written into a vector database for similar working condition retrieval, and can be represented as:

[0044] in, This is the current operating condition vector. For the first in the vector database Historical working condition vectors, The similarity between the current and historical scheduling conditions is used. Based on the similarity, similar historical scheduling conditions are retrieved from the vector database, and the retrieved historical scheduling cases, execution results, and human feedback are used as reference information for the scheduling agent to generate candidate scheduling suggestions. Through this process, the data foundation can not only preserve the original scheduling data, but also provide the vectorized information and structured knowledge required for model inference and task decomposition.

[0045] In one embodiment, S2, the candidate training sample set includes real scheduling cases and synthetic scheduling samples. The candidate scheduling proposals in the synthetic scheduling samples are obtained by a rejection sampling algorithm, specifically including: For a given scheduling input From the proposed distribution Sampling yields candidate scheduling suggestions Its acceptance probability is:

[0046] in, It is a constant, and satisfies the condition for any have:

[0047] Target distribution Constructed from the reward function and constraint verification, specifically as follows:

[0048] The acceptance probability then simplifies to:

[0049] in, It is the scaling factor, and , The reward function is obtained by weighting the capacity gain, waiting time penalty, violation penalty, and safety risk penalty.

[0050] Specifically, the number of real-world scheduling cases is usually limited and may not cover all production disruption scenarios such as equipment failures, road closures, and temporary adjustments to work sites. To improve the coverage of the training samples, synthetic scheduling samples are introduced in addition to real-world scheduling cases. For a given scheduling input... The scheduling agent base model or basic large model can generate multiple candidate scheduling suggestions. After candidate scheduling suggestions are generated, the system does not directly use them as training samples. Instead, it first filters them based on the reward function and constraint verification. The reward function is specifically expressed as follows:

[0051] in, Indicates capacity gain. This indicates a waiting time penalty. This indicates a penalty for violation. Indicates penalties for security risks. , , , These are the corresponding weighting coefficients.

[0052] reward function The calculation is a weighted average of capacity gain, waiting time penalty, violation penalty, and safety risk penalty. Capacity gain is used to evaluate the contribution of scheduling recommendations to ore extraction efficiency or work completion volume. Waiting time penalty is used to evaluate the impact of equipment queuing, idle time, and poor process coordination. Violation penalty is used to evaluate whether scheduling recommendations violate roadway access rules, work time windows, and equipment capacity boundaries. Safety risk penalty is used to evaluate whether scheduling recommendations involve high-risk paths, continued operation of abnormal equipment, or excessive material levels in the ore pass.

[0053] The rejection sampling algorithm decides whether to retain candidate scheduling suggestions based on the acceptance probability. Candidate scheduling suggestions with higher reward values ​​and those that pass constraint checks are more likely to be accepted. Candidate scheduling suggestions with lower reward values ​​or those with violation risks are more likely to be rejected. In this way, synthetic samples are not simply piled up into the training set, but are screened by production rules and business metrics before participating in training, thereby improving the quality of training samples.

[0054] In one embodiment, S2, the scheduling agent base model is optimized using the GRPO algorithm based on the target training samples, specifically including: For each input The sampling yielded a set of candidate scheduling suggestions:

[0055] make For the first The reward value for each candidate, within the group baseline. with standard deviation They are respectively:

[0056]

[0057] in, The number of candidate scheduling suggestions, ,and It is a numerically stable term; No. The relative advantages of each candidate scheduling suggestion for:

[0058] make To optimize the generation strategy for the scheduling agent, The sampling strategy has the following importance ratio:

[0059] The objective function is:

[0060] in The cropping threshold, The regularization coefficient is . For reference strategy, Let KL divergence be denoted as KL divergence.

[0061] Specifically, the GRPO algorithm is used to compare the merits of multiple candidate scheduling suggestions under the same scheduling input. For the same input... The system samples obtained Each candidate scheduling suggestion is given, and the reward value is calculated according to the reward function. Intragroup baseline This represents the average reward level of this set of candidate scheduling suggestions. Standard deviation Used to measure the differences between the candidate scheduling recommendations in this group. Relative advantage Indicates the first The relative merits of each candidate scheduling suggestion compared to other suggestions in the same group.

[0062] When a candidate scheduling suggestion reduces equipment waiting time, increases capacity, and does not violate safety and access constraints, its reward value is typically higher than the group average, and its relative advantage is positive. During model optimization, the generation probability of this type of output is increased. When a candidate scheduling suggestion involves excessively long detours, equipment conflicts, route restrictions, or safety risks, its reward value is lower than the group average, and its relative advantage is negative. During model optimization, the generation probability of this type of output is decreased.

[0063] Importance ratio Used to constrain the magnitude of change between the old and new strategies. (Clipping item) To prevent excessive policy changes during a single training session, the KL divergence term is used to ensure that the scheduling agent base model does not deviate excessively from the reference policy while learning from field feedback, avoiding significant fluctuations in output format, scheduling logic, or safety boundaries. Through this optimization method, the scheduling agent base model can learn the scheduling method preferred by the mine through comparison of multiple candidate suggestions.

[0064] In one embodiment, in S2, the scheduling agent base model uses LoRA fine-tuning for parameter updates, specifically including: The linear layer weights in the scheduling agent base model Its update format is as follows:

[0065]

[0066] in, Let be the lower projection matrix. Let be the upward projection matrix. ,and It is a low-rank dimension. This is the scaling factor; For the linear layer input vector The forward computation is as follows:

[0067] Freeze during training and only update and .

[0068] Specifically, LoRA fine-tuning is used to reduce the training cost during continuous optimization of the scheduling agent base model. The original weights in the scheduling agent base model. Keep the process frozen during training, using only low-rank matrices. and Forming incremental weights Due to low-rank dimensions With dimensions smaller than the input and output dimensions, fewer parameters need to be updated during training, resulting in faster model training convergence.

[0069] In mine production scheduling scenarios, the model needs to continuously absorb new scheduling feedback, but the production site requires the inference service to remain stable. By adopting LoRA fine-tuning, the system can store newly learned scheduling preferences, changes in equipment constraints, and manual correction experience in the LoRA weights. During online deployment, only the approved LoRA weights need to be loaded into the inference nodes; the entire base model does not need to be replaced. This reduces training and deployment costs and minimizes the impact of model parameter updates on the online scheduling service.

[0070] In one embodiment, in S2, the continuous optimization of the scheduling agent base model adopts a mechanism that separates online inference from asynchronous offline retraining, specifically including: An experience pool is set up to collect scheduling inputs, candidate scheduling suggestions, target scheduling suggestions, execution results, and human feedback. Specifically, execution results include ore output per unit time, equipment idle rate, average waiting time, ore level fluctuation in the chute, and the number of violations. Human feedback includes the scheduler's adoption status of candidate scheduling suggestions and the modifications made. The system correlates scheduling inputs, candidate scheduling suggestions, target scheduling suggestions, execution results, and human feedback to form feedback samples, which are then written into the experience pool for subsequent rejection sampling and updates to the relative advantage within groups.

[0071] When the cumulative penalties for scheduling violations exceed the preset threshold, or when the preset business downturn period is reached, training data is extracted from the experience pool. Based on the extracted training data, the scheduling agent base model is subjected to batch iterative optimization and testing using the GRPO algorithm and LoRA fine-tuning to obtain the parameters of the model to be reviewed. After the parameters of the model to be reviewed pass the review, the approved LoRA weights will be loaded into the inference node to update the scheduling agent base model parameters.

[0072] Specifically, online inference continuously receives scheduling requests and outputs scheduling suggestions during the production process. Asynchronous offline retraining periodically updates the parameters of the scheduling agent's underlying model without affecting online scheduling. The experience pool stores key data from the scheduling process, including scheduling inputs, candidate scheduling suggestions generated by the model, the final adopted scheduling scheme, equipment execution results, violations, scheduling personnel adoption records, and manual modification records. This data reflects the differences between model suggestions and actual on-site execution.

[0073] When the early warning agent detects that the cumulative penalties for scheduling violations exceed a preset threshold, it indicates that the current model's output has deviated significantly under certain operating conditions, requiring offline retraining. During shift changes, maintenance breaks, or other low-business periods, training data can also be extracted from the experience pool for batch iterative optimization. During the offline retraining phase, the scheduling agent's underlying model is fine-tuned using the GRPO algorithm and LoRA, and its updated parameters are tested. Testing may include whether safety constraints are met, whether access constraints are met, whether equipment capabilities are exceeded, whether the output format is stable, and whether production indicators have improved.

[0074] After the model parameters pass testing and manual review, the corresponding LoRA weights are loaded into the inference nodes. The inference nodes then continue to provide services for online scheduling. This mechanism separates model parameter updates from on-site scheduling operations, preventing the training process from affecting scheduling instruction generation, while allowing the model to continuously absorb new scheduling experience and execution feedback.

[0075] In one embodiment, S3, the hierarchical multi-agent structure includes a general agent, an early warning agent, a knowledge base agent, a scheduling agent, and a decision agent; The overall intelligent agent is used to plan and decompose the mine production scheduling tasks to obtain a set of scheduling sub-tasks. The early warning agent is used to analyze the operating status of equipment and output equipment status information; The knowledge base agent is used to query equipment parameters, scheduling procedures, roadway access restrictions and historical scheduling experience in the data base, and output the scheduling constraints and historical scheduling experience. The scheduling agent is used to generate a set of candidate scheduling suggestions based on device status information, scheduling constraints and historical scheduling experience, and to arbitrate the set of candidate scheduling suggestions based on constraint priority to obtain the target scheduling suggestion; The decision agent is used to translate target scheduling recommendations into a set of executable scheduling instructions.

[0076] Specifically, the hierarchical multi-agent structure is used to assign different processing tasks in the mine scheduling process to different agents for execution. The overall agent is located at the task entry point and is responsible for understanding the mine production scheduling request and breaking down the scheduling task into multiple sub-tasks based on the current production status, task type, and resource conditions. When breaking down tasks, the overall agent can assign content related to equipment status to the early warning agent, content related to procedures, equipment parameters, and historical cases to the knowledge base agent, and content related to solution generation and solution comparison to the scheduling agent.

[0077] The early warning agent primarily handles equipment operating status. Based on equipment location data, equipment operating status data, and sensor sampling data, the early warning agent can determine whether there are anomalies in vehicles, loading equipment, transport channels, and work sites, and output the abnormal event, abnormal equipment, abnormal time, and confidence level. The equipment status information output by the early warning agent can be used as input for subsequent scheduling agents to generate plans, preventing abnormal equipment from being assigned to high-risk operations.

[0078] The knowledge base agent primarily handles scheduling rules and historical experience. Based on the current equipment, work area, and task type, the knowledge base agent queries the data base for equipment capacity boundaries, maintenance records, scheduling procedures, roadway access restrictions, and historically similar working conditions. For roadways that are not permitted to pass, equipment load capacity, work point loading and unloading capacity, and chute material level limits, the knowledge base agent can generate scheduling constraints that the scheduling agent can verify. In some embodiments, the knowledge base agent can query equipment parameters and scheduling procedures in the knowledge base, and the scheduling agent provides suggested processing methods for scheduling based on the scheduling procedures and historical experience.

[0079] The scheduling agent primarily handles candidate solution generation and conflict resolution. After receiving equipment status information from the early warning agent and scheduling constraints from the knowledge base agent, the scheduling agent generates a set of candidate scheduling suggestions based on historical scheduling experience. These suggestions can include equipment allocation, transportation routes, loading / unloading sequences, operation schedules, and anomaly handling methods. When there is competition between candidate scheduling suggestions for equipment, roadways, loading points, or unloading points, the scheduling agent arbitrates the suggestions according to constraint priority and outputs the target scheduling suggestion.

[0080] The decision agent primarily handles the conversion of scheduling suggestions into scheduling instructions. Once the target scheduling suggestion is determined, the decision agent transforms it into a set of scheduling instructions that can be recognized by the production execution system, scheduling terminal, or vehicle-mounted terminal. This set of scheduling instructions may include equipment number, job type, starting location, target location, transportation route, execution time, loading / unloading requirements, and obstacle avoidance requirements. The decision agent can also write the instruction version, effective time, and execution feedback identifier into the production execution system for subsequent traceability and model optimization.

[0081] In one embodiment, in S3, the mine production scheduling task is broken down using a distributed collaboration framework to obtain a set of scheduling subtasks, specifically including: The mine production scheduling request is represented as a set of tasks:

[0082] Each task Includes load point Unload point Demand Time window With constraint label set ; The overall intelligent agent, based on the constraint label set, resource type, and current production state, assigns the task set... Divided into a set of early warning subtasks Knowledge retrieval subtask set And schedule to generate a set of subtasks The tasks are then assigned to the corresponding intelligent agents, and the set of warning subtasks is further divided. The knowledge retrieval subtask set and the scheduling generates a set of subtasks As the set of scheduled subtasks.

[0083] Specifically, mine production scheduling requests are typically not single operational instructions, but rather encompass multiple operational requirements. For example, a shift scheduling request may simultaneously involve loading tasks at several mining areas, transportation tasks at multiple unloading points, replacement tasks after equipment malfunctions, and route adjustment tasks due to roadway congestion. Therefore, the system first represents the scheduling request as a set of tasks. Each task in the task set This corresponds to a relatively independent scheduling item.

[0084] Each task Loading point in Used to indicate the starting location of material handling and unloading point. Used to indicate the target location and required quantity of materials. A time window is used to represent the amount of transportation or work that needs to be completed for the task. Used to define the start and completion time range of a task, constraining the set of tags. Used to mark the safety restrictions, equipment restrictions, road restrictions, material level restrictions, or priority requirements involved in the task.

[0085] When splitting tasks, the overall intelligent agent does not simply distribute them equally according to the number of tasks, but rather divides them based on the constraint label set, resource type, and current production status. Tasks involving equipment status, anomaly monitoring, and equipment fault diagnosis are categorized into the early warning sub-task set. The tasks are then assigned to the early warning agent. Tasks involving equipment parameters, scheduling procedures, roadway access restrictions, and historical scheduling experience queries are categorized into the knowledge retrieval subtask set. The tasks are then assigned to the knowledge base agent. Tasks involving equipment allocation, path selection, job sequence, and solution generation are categorized into a set of scheduling generation subtasks. And assign it to the scheduling agent.

[0086] Through the above-described decomposition method, the distributed collaboration framework enables different agents to handle states, rules, and solutions separately around the same scheduling request, avoiding the concentration of all information in a single processing node. The set of early warning subtasks, the set of knowledge retrieval subtasks, and the set of scheduling generation subtasks together constitute the scheduling subtask set. After each agent completes its execution, its output is then aggregated and processed by the scheduling agent and the decision agent to form a subsequent set of candidate scheduling suggestions and scheduling instructions.

[0087] In one embodiment, in S4, the candidate scheduling suggestion set is generated using hybrid sampling, specifically including: Based on the multimodal fusion vector constructed using S1, the top-ranked vectors with similarity to the current working condition are retrieved from the data base vector library. Historically feasible scheduling schemes; simultaneously, the scheduling agent calls the optimized scheduling agent base model, combining device status and scheduling constraints to generate... A new scheduling scheme is proposed. After removing duplicates from the two schemes, a set of candidate scheduling suggestions is obtained by integrating them. .

[0088] Specifically, the text features, image features, and temporal features corresponding to the current mining production conditions are first fused using a multi-head attention mechanism to obtain a multimodal fusion vector, which is then used as the current working condition vector. The data base vector library pre-stores historical operating condition vectors corresponding to each historical operating condition. Furthermore, each historical work condition vector is bound to a corresponding historical feasible scheduling scheme. The degree of matching between the current work condition and each historical work condition is calculated using a work condition similarity formula:

[0089] in, This is the current operating condition vector. For the first in the vector database Historical working condition vectors, The similarity between the two is used as the benchmark. All historical scheduling schemes are sorted from highest to lowest similarity score, and the top-ranked schemes are selected. Historically feasible scheduling schemes, It is a preset positive integer.

[0090] Meanwhile, the scheduling agent takes the device status information output by the early warning agent and the scheduling constraints output by the knowledge base agent as input, calls the optimized scheduling agent base model, and generates data in batches. A brand new scheduling scheme The preset positive integer; a single new scheduling scheme also includes equipment allocation, transportation route, loading and unloading sequence, operation time arrangement and abnormal handling method.

[0091] After acquiring both types of scheduling schemes, a duplicate determination is performed based on the core elements of the scheduling schemes: if two schemes have completely identical loading points, unloading points, main operating equipment, and core transportation routes, they are determined to be duplicate schemes. The retrieved historical scheduling schemes are compared pairwise with the newly generated scheduling schemes from the model, and all duplicate content is removed. Finally, the deduplicated historical scheduling schemes are merged with the new scheduling schemes to form a complete set of candidate scheduling suggestions. This is used in subsequent arbitration processes based on constraint priority.

[0092] In one embodiment, in S4, arbitration based on constraint priority specifically includes: Define the constraint priority sequence:

[0093] in, As the highest priority safety constraint, it belongs to the rigid red line constraint of mine safety production and covers four categories: equipment safety, roadway access safety, storage material level safety, and on-site working environment safety. Examples include operating equipment with defects, passing through closed roadways, material levels exceeding safety thresholds, and dispatching vehicles to operate in dangerous areas. Lowest priority efficiency constraint; The constraints listed above are merely typical implementation examples and do not constitute a limitation on the implementation of the above. Limitations on the scope of constraints; Establish index mapping relationship subscript i corresponds one-to-one with the constraint priority number, the first... Priority constraints All decision conditions are determined by the set All check functions within the scope are characterized together. For a single check function... Define a single constraint compliance identifier:

[0094] If the same priority any one of the next Then the scheduling scheme Not satisfied Level constraint. If the first level constraint. If all individual constraints under priority constraints have a compliance flag of 1, then =1; otherwise =0.

[0095] initialization ,according to Filter by priority from highest to lowest:

[0096] After all levels of screening, the final set of compliant candidates ; any level If the set is empty after filtering, the filtering process terminates, indicating that there are no candidate scheduling suggestions that meet the constraints.

[0097] In the candidate set The candidate scheduling suggestion with the highest comprehensive score is selected as the target scheduling suggestion. (Comprehensive scoring function) for:

[0098] in, For the first Weight coefficients of class constraints For scheduling suggestions In the Scores under class constraints The number of constraint types. Sub-item scores corresponding to class constraints The calculation method is as follows:

[0099] In the formula: , The score is a preset constant. For the first Overall compliance indicators for priority constraints; For scheduling suggestions The corresponding average transportation path cost. Specifically, in mine production scheduling, multiple candidate scheduling suggestions often occupy the same resource simultaneously. For example, multiple vehicles may be assigned to the same narrow roadway at the same time, multiple mining areas may request the same loading equipment simultaneously, or multiple transportation tasks may simultaneously point to the same unloading point. If only the solution is selected based on capacity or distance, problems such as roadway congestion, equipment conflicts, queuing at loading and unloading points, or increased safety risks may occur. Therefore, arbitration based on constraint priority is required among candidate scheduling suggestions.

[0100] In addition, if the compliance candidate set is empty, the preset security scheduling suggestion used to control the relevant equipment to maintain the current security status and request manual handling will be used as the target scheduling suggestion; if the compliance candidate set is not empty, the candidate scheduling suggestion with the highest comprehensive score will be used as the target scheduling suggestion.

[0101] Constraint Priority Sequence These are used to indicate the relative importance of different constraints. Safety constraints are typically of the highest priority; for example, high-risk tasks must not continue if equipment malfunctions, closed roadways must not be traversed, and chute material levels must not exceed safety limits. Equipment capacity constraints, roadway access constraints, and operation time window constraints can be of medium priority, used to ensure that scheduling recommendations can be implemented on-site. Efficiency constraints can be of lower priority, used to further increase capacity, reduce waiting time, and lower empty-running rates after safety and feasibility conditions are met.

[0102] When multiple candidate scheduling suggestions compete for resources, the scheduling agent first filters according to the highest priority constraint, eliminating candidate scheduling suggestions that do not meet the safety constraints. Then, it continues filtering according to equipment capacity constraints, access constraints, time window constraints, and efficiency constraints. After this step-by-step filtering, a set of candidate suggestions that meet the high-priority constraints is obtained. .

[0103] In the candidate set In the process, the scheduling agent then calculates a comprehensive score for each candidate scheduling suggestion. .in, For the first Weight coefficients of class constraints For scheduling suggestions In the Scoring under class constraints. For rigid constraints such as safety, accessibility, and equipment capacity (high priority)... ): Directly adopt the overall compliance identifier of this level of constraints. Scoring: A plan that fully complies with the rules will receive a base score. Violations will result in 0 points; for efficiency indicators such as waiting time, transportation distance, and equipment utilization (lowest priority) The scoring uses the average path cost in reverse; the lower the path cost and the higher the scheduling efficiency, the higher the score. (The constant is not specified in the original text.) , Pre-configured according to the mine's on-site business rules. The scheduling agent selects the candidate scheduling suggestion with the highest comprehensive score as the target scheduling suggestion.

[0104] In this embodiment, a hierarchical multi-agent structure is used to divide the scheduling tasks, and resource competition among candidate scheduling suggestions is resolved through constraint priority arbitration. This allows the scheduling results to simultaneously meet safety requirements, equipment capacity requirements, roadway access requirements, and production efficiency requirements. Compared to a single model directly outputting a scheduling scheme, this process makes it easier to explain the sources of various information and to trace the scheduling suggestions.

[0105] This application also provides a computer system, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the multi-agent-based mine production scheduling assistance method described above. When the program or instructions are executed by the processor, they implement the various steps of the above method embodiments and achieve the same technical effects. To avoid repetition, they will not be described again here.

[0106] This application also provides a readable storage medium storing a program or instructions that, when executed by a processor, implement the various processes of the above-described method embodiments and achieve the same technical effects. To avoid repetition, these will not be described again here.

[0107] The processor can be the processor in the aforementioned system. Readable storage media include computer-readable storage media, such as computer read-only memory, random access memory, magnetic disks, or optical disks.

[0108] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0109] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as a read-only memory). The device includes a number of instructions in a ROM (random access memory), RAM (magnetic disk), or optical disk to cause a terminal to execute the methods described in the various embodiments of this application.

[0110] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A multi-agent-based auxiliary method for mine production scheduling, characterized in that, include: S1. Obtain multi-source data related to mine production scheduling, perform data governance on the multi-source data to obtain standard scheduling data, and construct a data foundation based on the standard scheduling data. S2. Construct a candidate training sample set for training the scheduling agent base model based on the data base; Based on the mine constraints and production indicators, the candidate scheduling suggestions in the candidate training sample set are screened to obtain the target training sample; The scheduling agent base model is optimized based on the target training samples to obtain the optimized scheduling agent base model. S3. Obtain a mine production scheduling request and generate a mine production scheduling task based on the mine production scheduling request; A distributed collaboration framework based on the data base and the optimized scheduling agent base model is invoked. The distributed collaboration framework includes a hierarchical multi-agent structure. The mine production scheduling task is split into a set of scheduling subtasks through the distributed collaboration framework. S4. Execute the set of scheduling subtasks through the distributed collaboration framework to obtain device status information, scheduling constraints, and a set of candidate scheduling suggestions; Based on the device status information and the scheduling constraints, the candidate scheduling suggestion set is arbitrated based on constraint priority to obtain the target scheduling suggestion; S5. Convert the target scheduling suggestion into a set of executable scheduling instructions, and output the set of scheduling instructions; In S2, the continuous optimization of the scheduling agent base model adopts a mechanism that separates online inference from asynchronous offline retraining, specifically including: An experience pool is set up to collect scheduling inputs, candidate scheduling suggestions, target scheduling suggestions, execution results, and human feedback. When the cumulative penalties for scheduling violations exceed a preset threshold, or when a preset low-business period is reached, training data is extracted from the experience pool. Based on the extracted training data, perform batch iterative optimization and testing using the GRPO algorithm and LoRA fine-tuning to obtain the parameters of the model to be reviewed; After the parameters of the model to be reviewed pass the review, the approved LoRA weights are loaded into the inference node to update the scheduling agent base model; In S3, the hierarchical multi-agent structure includes a general agent, an early warning agent, a knowledge base agent, a scheduling agent, and a decision agent; The overall intelligent agent is used to plan and decompose the mine production scheduling task to obtain the set of scheduling sub-tasks. The early warning agent is used to analyze the operating status of the equipment and output the equipment status information; The knowledge base agent is used to query equipment parameters, scheduling procedures, roadway access restrictions and historical scheduling experience in the data base, and output the scheduling constraints and the historical scheduling experience. The scheduling agent is used to generate the candidate scheduling suggestion set based on the device status information, the scheduling constraints and the historical scheduling experience, and to arbitrate the candidate scheduling suggestion set based on constraint priority to obtain the target scheduling suggestion; The decision agent is used to convert the target scheduling suggestion into the set of scheduling instructions.

2. The mine production scheduling assistance method based on multi-agent technology according to claim 1, characterized in that, In S1, the multi-source data includes: production execution system data, Internet of Things platform data, and scheduling interaction feedback data; The production execution system data includes shift plans, mining operation orders, equipment ledgers, and maintenance records. The IoT platform data includes device location data, device operating status data, sensor sampling data, and on-site image acquisition data; The dispatching interaction feedback data includes dispatching logs, dispatching procedures, shift handover records, dispatching personnel adoption records, and manual modification records; Data governance is performed on the multi-source data to obtain standard scheduling data, specifically including: Text data, image data, and time-series data are extracted from the multi-source data according to the data format, and the text data, image data, and time-series data are cleaned, formatted, and time-stamped respectively to obtain standard text data, standard image data, and standard time-series data. Feature extraction is performed on the standard text data, the standard image data, and the standard time-series data respectively to obtain text features, image features, and time-series features; The text features, image features, and temporal features are used as feature data of different modalities for modal alignment and fusion to obtain a multimodal fusion vector. The standard text data, standard image data, standard temporal data, and multimodal fusion vector are then written into the data base as the standard scheduling data. The modality alignment and fusion are implemented using a multi-head attention mechanism.

3. The mine production scheduling assistance method based on multi-agent technology according to claim 1, characterized in that, In step S2, the candidate training sample set includes real scheduling cases and synthetic scheduling samples. The candidate scheduling suggestions in the synthetic scheduling samples are obtained through a rejection sampling algorithm, specifically including: For a given scheduling input From the proposed distribution Sampling yields candidate scheduling suggestions Its acceptance probability is: in, It is a constant, and satisfies the condition for any have: Target distribution Constructed from the reward function and constraint verification, specifically as follows: The acceptance probability then simplifies to: in, It is the scaling factor, and , The reward function is obtained by weighting production capacity gain, waiting time penalty, violation penalty and safety risk penalty. The violation penalty and the safety risk penalty are determined based on the constraint verification results of the candidate scheduling suggestions on the mine constraints.

4. The mine production scheduling assistance method based on multi-agent technology according to claim 3, characterized in that, In step S2, the scheduling agent base model is optimized using the GRPO algorithm based on the target training samples, specifically including: For each input The sampling yielded a set of candidate scheduling suggestions: make For the first The reward value for each candidate, within the group baseline. with standard deviation They are respectively: in, The number of candidate scheduling suggestions, ,and It is a numerically stable term; No. The relative advantages of each candidate scheduling suggestion are: make To optimize the generation strategy for the scheduling agent, The sampling strategy, importance ratio for: The objective function is: in The cropping threshold, The regularization coefficient is . For reference strategy, Let KL divergence be the KL divergence. For the model parameters to be optimized for the scheduling agent, Indicates the scheduling input The expectation.

5. The mine production scheduling assistance method based on multi-agent technology according to claim 4, characterized in that, In step S2, the scheduling agent base model uses LoRA fine-tuning for parameter updates, specifically including: The linear layer weights in the scheduling agent base model Its update format is as follows: in, For the updated linear layer weights, For low-rank incremental weights, Let be the lower projection matrix. Let be the upward projection matrix. ,and It is a low-rank dimension. This is the scaling factor; For the linear layer input vector The forward computation is as follows: Freeze during training and only update and ,in, This is the output vector of the linear layer.

6. The mine production scheduling assistance method based on multi-agent technology according to claim 1, characterized in that, In S3, the mine production scheduling task is broken down using the distributed collaboration framework to obtain a set of scheduling subtasks, specifically including: The mine production scheduling request is represented as a set of tasks: Each task Includes load point Unload point Demand Time window With constraint label set ; The overall intelligent agent, based on the constraint label set, resource type, and current production state, assigns the task set... Divided into a set of early warning subtasks Knowledge retrieval subtask set And schedule to generate a set of subtasks The tasks are then assigned to the corresponding intelligent agents, and the set of warning subtasks is further divided. The knowledge retrieval subtask set and the scheduling generates a set of subtasks As the set of scheduled subtasks.

7. The mine production scheduling assistance method based on multi-agent technology according to claim 1, characterized in that, In S4, the arbitration based on constraint priority specifically includes: Each candidate scheduling suggestion in the candidate scheduling suggestion set is verified step by step according to a preset constraint priority sequence; wherein, the constraint priority sequence includes safety constraints and rigid rule constraints set in descending order; For any candidate scheduling suggestion, if it does not meet the constraints corresponding to the current priority, the candidate scheduling suggestion is removed from the candidate scheduling suggestion set; if it meets the constraints corresponding to the current priority, the constraint verification of the next priority continues until the constraint verification of all priorities is completed, and a compliant candidate set is obtained. In the set of compliance candidates, the comprehensive score of each candidate scheduling suggestion is determined based on the degree to which each candidate scheduling suggestion meets the safety constraints, rigid rule constraints, and efficiency constraints, as well as the production efficiency evaluation results corresponding to each candidate scheduling suggestion. The candidate scheduling suggestion with the highest comprehensive score is selected as the target scheduling suggestion.

8. A computer system, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the multi-agent-based mine production scheduling assistance method according to any one of claims 1 to 7.

Images