Mine production plan intelligent optimization system

Abstract

The application discloses a mine production plan intelligent optimization system, which extracts features of collected data by using a DeepSeek-V3 multi-modal model to obtain mine state data; obtains a mine production plan optimization target and mine state data, models production plan making as a multi-target MDP Markov decision process by using an improved TD3 algorithm combined with a self-adaptive noise exploration strategy, and outputs an initial plan set; generates an initial production plan based on a production Agent, evaluates a risk level by a safety Agent and triggers an emergency plan, dynamically allocates equipment, manpower and energy by a resource Agent, and constructs a multi-collaborative decision model; and optimizes the initial plan set by the multi-collaborative decision model to obtain a target mine production plan. The plan can adapt to special working conditions, on-demand adjustment and dynamic balance can be realized, and production efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural irrigation, in particular to a mine production plan intelligent optimization system. BACKGROUND

[0002] The intelligent construction of mines has begun to enter the fast lane. In recent years, major coal mines have made substantial progress in the application of video analysis, machine learning and deep learning in hardware infrastructure construction and intelligentization. However, there are still deficiencies in safety production analysis. Data analysis is not thorough, and the relationship between specific data needs to be clarified. It is difficult to cope with dynamic changes in the production environment, market demand fluctuations, equipment failures, and changes in geological conditions, which seriously affect production efficiency. SUMMARY

[0003] The purpose of the present application is to solve the above problems, and a mine production plan intelligent optimization system is designed.

[0004] To achieve the above purpose, the technical scheme of the present application is that, further, in the above-mentioned mine production plan intelligent optimization system, the mine production plan intelligent optimization system method comprises the following modules:

[0005] The mine data processing module is used for collecting mine equipment state data, inventory data, environmental parameters and geological data in real time through sensors, using a DeepSeek-V3 multi-modal model to extract features from the collected data, and obtaining mine state data;

[0006] The production plan output module is used for obtaining mine production plan multi-objective optimization targets and mine state data, using an improved TD3 algorithm combined with an adaptive noise exploration strategy to model the production plan as a multi-objective MDP Markov decision process. The state space includes device state, inventory level, market demand, environmental parameter and safety risk level. The action space is a set of refined production scheduling instructions. A plan set is generated through a multi-Agent collaborative decision-making framework, and is optimized through a debate and consensus mechanism to output an initial plan set.

[0007] The decision model establishment module is used for generating an initial production plan based on a production Agent, evaluating a risk level and triggering an emergency plan based on a safety Agent, and dynamically allocating devices, manpower and energy based on a resource Agent, and constructing an Agent multi-collaborative decision-making model.

[0008] The production plan decision module is used for decision optimization of the initial plan set through the Agent multi-collaborative decision-making model to obtain a target mine production plan.

[0009] Further, in the above-mentioned mine production plan intelligent optimization system, the mine data processing module comprises the following units:

[0010] a data collection unit, configured to collect mine equipment state data, inventory data, environmental parameters and geological data in real time through sensors to obtain real-time collection data;

[0011] a data cleaning unit, configured to clean the real-time collection data, remove outliers in equipment vibration and gas concentration data through a 3σ criterion, and fill in missing values caused by sensor communication interruption by using a linear interpolation method to obtain first processing data;

[0012] a standardization processing unit, configured to convert temperature and vibration physical quantity parameters in the first processing data into dimensionless 0-1 interval values to unify data magnitudes, and perform time alignment on time series data to obtain second processing data;

[0013] a feature extraction unit, configured to input the second processing data into a DeepSeek-V3 multi-modal model, extract frequency domain features of equipment vibration signals through a convolution layer of the model, capture trend features of inventory data through a recurrent neural network, focus on environmental parameter features of high-risk areas through an attention mechanism, and analyze ore body distribution features of geological data through a three-dimensional convolution module to output mine state data.

[0014] Further, in the mine production plan intelligent optimization system, the production plan output module comprises the following units:

[0015] a state space construction unit, configured to integrate equipment health indexes, refine the equipment health indexes into three types of state labels, record current working conditions of equipment, divide raw material inventory states according to inventory health data, and divide demand intensity according to market demand on a weekly basis;

[0016] an action space definition unit, configured to divide production scheduling instructions into four types of actions, including equipment start-stop instructions, material allocation instructions, manpower scheduling instructions and capacity adjustment instructions, by using an improved TD3 algorithm combined with an adaptive noise exploration strategy; and action constraint conditions include that all instructions need to comply with equipment operation regulations, safety regulations and resource limitations.

[0017] Further, in the mine production plan intelligent optimization system, the production plan output module further comprises the following units:

[0018] a Markov decision unit, configured to update the state space according to fixed logic after executing a certain scheduling instruction, give positive rewards to actions that meet optimization objectives, and impose penalties when the objectives are violated;

[0019] a scheduling scheme generation unit, configured to randomly generate 50 groups of basic scheduling schemes, including equipment combinations, capacity distribution ratios and time arrangements;

[0020] The initial plan generation unit is used to retain the top 30% high-quality schemes by evaluating the cumulative reward values of each scheme and making local adjustments, and generate an initial plan set after 5 iterations.

[0021] Further, in the above-mentioned intelligent optimization system for mine production plan, the decision model establishment module comprises the following units:

[0022] The agent architecture establishment unit is used to establish the overall framework of the production agent for overall production plan, and the safety agent and the resource agent are used as auxiliary decision makers to make adjustment suggestions for the plan from the safety and resource angles respectively, and the three realize data synchronization through an information interaction channel.

[0023] Further, in the above-mentioned intelligent optimization system for mine production plan, the decision model establishment module further comprises the following units:

[0024] The safety level judgment unit is used to judge the fault risk by analyzing the equipment state data, and evaluate the safety level of the working face in combination with the dust concentration, gas content and surrounding rock stability parameters;

[0025] The risk evaluation judgment unit is used to suspend the execution of the plan by the production agent when the safety agent evaluates that the plan has medium and high risks, and restart after the risk is eliminated by the resource agent adjusting the equipment, and the resource agent allocates resources to meet the production task demand proposed by the production agent first when allocating resources.

[0026] Further, in the above-mentioned intelligent optimization system for mine production plan, the production plan decision module comprises the following units:

[0027] The safety and resource evaluation unit is used for the production agent to select 20 representative plans from the initial plan set and submit them to the safety agent and the resource agent for evaluation;

[0028] The production plan optimization unit is used for the safety agent to feed back the risk evaluation result to the production agent, and the production agent modifies the plan in combination with the resource adjustment suggestion provided by the resource agent, and iteratively obtains the target mine production plan.

[0029] Further, in the method for realizing the intelligent optimization system for mine production plan, the method comprises the following steps:

[0030] The mine state data is obtained by collecting the mine equipment state data, inventory data, environmental parameters and geological data in real time through sensors, and using the DeepSeek-V3 multi-modal model to extract features from the collected data;

[0031] The multi-objective optimization target of the mine production plan and mine state data are acquired, the improved TD3 algorithm is combined with an adaptive noise exploration strategy, the production plan formulation is modeled as a multi-objective MDP Markov decision process, the state space includes device state, inventory level, market demand, environmental parameter and safety risk level, the action space is a set of fine production scheduling instructions, a plan set is generated through a multi-Agent collaborative decision framework, and the initial plan set is optimized through debate and consensus mechanism, and output;

[0032] The initial production plan is generated based on the production Agent, the safety Agent evaluates the risk level and triggers the emergency plan, the resource Agent dynamically allocates devices, manpower and energy, and an Agent multi-collaborative decision model is constructed.

[0033] The initial plan set is optimized through the Agent multi-collaborative decision model, and the target mine production plan is obtained.

[0034] Further, in the method for realizing the mine production plan intelligent optimization system, the method comprises the following steps:

[0035] The fault risk is judged by analyzing the device state data, the safety level of the working face is evaluated in combination with dust concentration, gas content and surrounding rock stability parameters;

[0036] When the safety Agent evaluates that the plan has medium and high risks, the production Agent suspends the execution of the plan, and restarts after the resource Agent adjusts the device to eliminate the risk, and the resource Agent allocates resources to preferentially meet the production task demand proposed by the production Agent.

[0037] Further, in the method for realizing the mine production plan intelligent optimization system, the method comprises the following steps:

[0038] The production Agent selects 20 representative plans from the initial plan set and submits them to the safety Agent and the resource Agent for evaluation;

[0039] The safety Agent feeds back the risk evaluation result to the production Agent, the production Agent modifies the plan in combination with the resource adjustment suggestion provided by the resource Agent, and iteratively obtains the target mine production plan.

[0040] The beneficial effects are that 1. AI is used to process queries and control, information acquisition speed is improved, and fault response time is shortened by more than 50%. Intelligent dialogue and knowledge retrieval shorten decision-making time and improve work quality. AI analyzes data and discovers hidden dangers in a timely manner, and the accuracy of production planning is high. 2. The off-peak energy supply mode reduces the energy consumption per unit of ore, and the labor is matched according to skills and the work intensity is distributed, reducing the cost of temporary labor, achieving the dual goals of cost reduction and efficiency improvement. 3. Ensure that the plan adapts to special working conditions; during the market demand peak, the efficiency target is prioritized to achieve on-demand adjustment and dynamic balance. The final output target plan includes a detailed execution list that can directly guide on-site operations, significantly reducing the cost and error rate of manual decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0041] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings. The drawings are for purposes of illustration only and are not intended to be limiting of the application.

[0042] Figure 1 A first embodiment schematic diagram of a mine production plan intelligent optimization system in an embodiment of the present application;

[0043] Figure 2 A second embodiment schematic diagram of a mine production plan intelligent optimization system in an embodiment of the present application;

[0044] Figure 3 A third embodiment schematic diagram of a mine production plan intelligent optimization system in an embodiment of the present application. DETAILED DESCRIPTION

[0045] In order to make the objects, technical solutions and advantages of the present application clearer, the following further describes the present application in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0046] Those skilled in the art can understand that, unless specifically stated otherwise, the singular form "one", "said" and "the" used herein can also include the plural form. It should be further understood that the wording used in the specification of the present application includes the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups.

[0047] The present application will be described in detail below with reference to the accompanying drawings, as shown in Figure 1 A mine production plan intelligent optimization system, which includes the following modules:

[0048] The mine data processing module is used for collecting mine equipment state data, inventory data, environmental parameters and geological data in real time through sensors, performing feature extraction on the collected data by using a DeepSeek-V3 multi-modal model, and obtaining mine state data.

[0049] Specifically, the embodiment further includes a data acquisition unit configured to collect mine equipment state data, inventory data, environmental parameters and geological data in real time through sensors, and obtain real-time acquisition data.

[0050] The data cleaning unit is configured to clean the real-time acquisition data, remove outliers in the device vibration and gas concentration data by using a 3σ criterion, and fill in missing values caused by sensor communication interruption by using a linear interpolation method, to obtain first processing data.

[0051] The standardization processing unit is configured to convert temperature and vibration physical quantity parameters in the first processing data into dimensionless 0-1 interval values, and unify the data magnitude; and perform time alignment on the time series data to obtain second processing data.

[0052] The feature extraction unit is configured to input the second processing data into the DeepSeek-V3 multi-modal model, extract frequency domain features of the device vibration signal through the convolution layer of the model, capture trend features of the inventory data through the recurrent neural network, focus on the environmental parameter features of the high-risk area through the attention mechanism, and analyze the ore body distribution features of the geological data through the three-dimensional convolution module, to output mine state data.

[0053] Specifically,

[0054] 1.1 Mine equipment state data acquisition

[0055] For core production equipment such as mining machines, conveyors, crushers and hoists, multiple types of sensors are deployed: a thermocouple temperature sensor is installed on the engine cylinder of the mining machine to monitor the operating temperature in real time; a piezoelectric vibration sensor is installed on the bearing of the rotating part to collect vibration acceleration and frequency; a Hall effect speed sensor is installed on the transmission system to record the main shaft speed; and a pressure sensor is installed on the hydraulic system pipeline to monitor the hydraulic oil pressure.

[0056] The sensor data is uploaded to the edge computing gateway at a frequency of 1 minute / time through an industrial Ethernet or 5G wireless transmission module, to ensure the timeliness of the data. For key equipment (main shaft hoist), a high-frequency acquisition mode of 5 seconds / time is adopted to ensure the accuracy of the high-risk equipment state monitoring.

[0057] 1.2 Inventory data acquisition

[0058] In raw material warehouse (explosive, detonator, steel) deployment of radio frequency identification (RFID) access control system, combined with shelf weight sensor to realize dynamic counting: each batch of materials into the warehouse attached RFID tag, through the warehouse entrance card reader to automatically record the material type and quantity; the strain gauge type weighing sensor at the bottom of the shelf real-time feedback of inventory weight, converted to the actual inventory.

[0059] Finished product ore bin uses laser level meter and ultrasonic flowmeter combination monitoring: level meter scans the material height in the warehouse every 30 minutes, combined with the geometric parameters of the warehouse body to calculate the inventory volume; the flowmeter installed on the discharge belt synchronously records the real-time output, and the inventory data is corrected by difference method.

[0060] 1.3 Environmental parameter acquisition:

[0061] Working face area: set up an environmental monitoring terminal every 50 meters in the driving face and the mining face, integrating non-spectroscopic infrared gas sensor (monitoring gas and carbon dioxide concentration), laser dust instrument (PM10 / PM2.5), temperature and humidity sensor, data sampling interval is 10 seconds / time.

[0062] Ventilation system: install wind speed sensor at the outlet of main fan and local fan, and install wind pressure sensor at the elbow of air duct to monitor the ventilation efficiency in real time, and the data is updated every 30 seconds.

[0063] Surface environment: set up meteorological station at the mine boundary and surrounding sensitive areas to collect wind speed, wind direction, precipitation and other data, updated once a day.

[0064] 1.4 Geological data acquisition:

[0065] Use three-dimensional laser scanning technology to scan the stope and roadway once a week to generate point cloud data to obtain the shape change of the ore body; set up optical fiber sensor in the drill hole to monitor the rock stress and displacement change in real time, with a sampling frequency of 1 hour / time.

[0066] Use geological radar to detect ahead of the driving face, collect geological structure data (fault, fracture distribution) every 5 meters, and combine with historical drilling data to build a dynamic geological model.

[0067] 1.5 Data preprocessing:

[0068] Clean the collected raw data: remove outliers in the data such as equipment vibration and gas concentration by 3σ criterion; use linear interpolation method to fill in missing values caused by sensor communication interruption (single missing not more than 5 minutes).

[0069] Data standardization: Convert physical parameters such as temperature and vibration into dimensionless 0-1 interval values to unify data magnitudes; align time series data (device speed curve) to ensure consistent sampling times.

[0070] 1.6 Feature extraction based on DeepSeek-V3 multi-modal model:

[0071] Model input layer receives multi-source heterogeneous data: Convert time series signals of device status, numerical changes of inventory, spatial distribution data of environment, and three-dimensional structure information of geology into tensor format recognizable by the model.

[0072] Feature extraction process: Extract frequency domain features of device vibration signals (abnormal vibration frequency peaks) through convolution layers, capture trend features of inventory data (raw material consumption rate) through recurrent neural networks, focus on environmental parameter features of high-risk areas (gas concentration gradient changes) through attention mechanisms, and analyze ore body distribution features of geological data (ore body strike and thickness changes) through three-dimensional convolution modules.

[0073] Output mine status data: Integrate the above features to form a comprehensive status vector containing device health index (based on temperature / vibration fault warning value), inventory health degree (safe inventory margin), environmental safety level (dust / gas exceeding probability), and geological stability coefficient (rock integrity score).

[0074] Production plan output module, used to obtain mine production plan multi-objective optimization target and mine status data, uses improved TD3 algorithm combined with adaptive noise exploration strategy to model production planning as a multi-objective MDP Markov decision process, state space includes device status, inventory level, market demand, environmental parameters and safety risk level, action space is a set of refined production scheduling instructions, generates a plan set through a multi-Agent collaborative decision-making framework, and optimizes it through debate and consensus mechanism, outputs an initial plan set;

[0075] Specifically, the embodiment also includes a state space construction unit for integrating device health index, refining it into three types of state labels, recording the current working condition of the device, and dividing the raw material inventory state according to the inventory health degree data; according to market demand, divide the demand intensity by week.

[0076] Action space definition unit, used to divide production scheduling instructions into four types of actions using improved TD3 algorithm combined with adaptive noise exploration strategy, including device start-stop instructions, material allocation instructions, manpower scheduling instructions, and capacity adjustment instructions; action constraints include that all instructions must comply with device operation procedures, safety procedures, and resource limitations.

[0077] Markov decision unit, for updating the state space according to fixed logic after executing a certain scheduling instruction, giving positive rewards to actions that meet the optimization goal, and imposing penalties when the goal is violated;

[0078] Scheduling scheme generation unit, for randomly generating 50 groups of basic scheduling schemes, including device combination, capacity allocation ratio and time arrangement;

[0079] Initial plan generation unit, for retaining the top 30% of high-quality schemes by evaluating the cumulative reward value of each scheme and performing local adjustment, and generating an initial plan set after 5 iterations.

[0080] Specifically;

[0081] 2.1 Define the optimization goal of mine production plan:

[0082] The core goals include three categories: efficiency goals, such as improving equipment utilization rate to more than 90% and shortening material transportation cycle to less than 4 hours; cost goals, such as controlling unit ore energy consumption below 8 kilowatt-hours per ton and reducing labor scheduling cost by 15%; and safety goals, ensuring that the monthly safety accident rate is 0 and the downtime of major equipment failure is not more than 24 hours.

[0083] Dynamic adjustment of goal priority: when market demand surges, the efficiency goal weight increases to 60%; during the rainy season or high-geological-disaster period, the safety goal weight increases to 70%; when the raw material price rises by more than 30%, the cost goal becomes the primary consideration.

[0084] 2.2 Construct the state space:

[0085] Device state dimension: integrate the device health index output in step 1 into three state labels—normal, warning, and failure. At the same time, record the current working condition of the device, such as the real-time working position of the mining machine, the load rate of the transport machine, and the processing capacity of the crusher.

[0086] Inventory level dimension: based on the inventory health data, divide the raw material inventory into three states—sufficient, critical, and short; and the finished product inventory is marked as full, regular, and low.

[0087] Market demand dimension: through the interface of the enterprise resource planning (ERP) system, divide the demand intensity into peak (weekly order volume ≥50,000 tons), stable (2-4.9 tons), and valley (≤1.9 tons) by week, and record the demand product type (iron concentrate, copper concentrate) and delivery deadline.

[0088] 2.3 Define the action space:

[0089] Production scheduling instructions are divided into four core actions: equipment start-up and shutdown instructions, such as starting mining machine A shift at 8:00 and stopping at 18:00, and transport machine C line running continuously for 24 hours; material allocation instructions, including prioritizing the transport of iron ore from warehouse No. 3 to the sintering plant and allocating 500 detonators from the spare warehouse to tunneling face No. 2; manpower scheduling instructions, such as arranging maintenance team A to support crusher maintenance and dispatching 3 blasters to the new mining area; and capacity adjustment instructions, such as increasing the daily mining volume from 8,000 tons to 10,000 tons and reducing the ball mill speed to 80% of the rated value.

[0090] Action constraints: All instructions must comply with equipment operating procedures (the crusher must be preheated for 3 minutes before starting), safety procedures (blasting operations must avoid thunderstorms) and resource restrictions (the total number of equipment dispatched at the same time shall not exceed 80% of the total number of equipment in use).

[0091] 2.4 Markov Decision Process Modeling:

[0092] State transition rule: After a certain scheduling instruction is executed, the state space is updated according to a fixed logic. For example, after executing the action of increasing the working time of mining machine B by 2 hours, the health index of the machine in the equipment state decreases by 5 points, the amount of finished ore in the inventory level increases by 300 tons, and the remaining order quantity in market demand decreases by 300 tons.

[0093] Reward mechanism design: Positive rewards are given for actions that meet the optimization goals (10 points are awarded when equipment utilization reaches the target), and penalties are imposed for actions that violate the goals (20 points are deducted for safety violations). The reward value changes dynamically with the status; for example, in a state of inventory shortage, the reward for successfully replenishing raw materials is doubled.

[0094] 2.5 Generate the initial plan set:

[0095] The improved TD3 algorithm, combined with an adaptive noise exploration strategy, generates plans through multiple rounds of simulation decision-making: the first round randomly generates 50 basic scheduling schemes, covering different equipment combinations, capacity allocation ratios, and time arrangements; subsequently, by evaluating the cumulative reward value of each scheme, the top 30% of high-quality schemes are retained and locally adjusted (replacing 1-2 scheduling instructions), and after 5 rounds of iteration, 100 initial plans are formed.

[0096] The plan must include all necessary elements: clearly define the daily mining volume for each mining area (accurate to 50 tons), equipment handover time (accurate to 15 minutes), material transportation routes (marking specific roadway numbers), and a list of emergency backup equipment (the standby locations of 2 backup transport machines).

[0097] The improved TD3 algorithm combined with an adaptive noise exploration strategy for multi-objective MDP modeling in mine production planning is a complex system integrating deep reinforcement learning and dynamic decision optimization. This model transforms mine production planning into a multi-objective Markov decision process by constructing a state space, action space, reward function, and state transition rules, achieving dynamic resource scheduling and multi-objective optimization.

[0098] In terms of model architecture design, the improved TD3 algorithm adopts a dual-Critic network architecture and an Actor network working collaboratively. The dual-Critic network reduces the risk of overfitting in value estimation and enhances the stability of policy learning by independently evaluating the value of actions. The Actor network uses a three-layer LSTM structure to process time-dependent production data, capturing the time-series features of state variables such as equipment status, inventory levels, and market demand. The Critic network introduces a convolutional neural network to extract multi-dimensional features of the equipment state space. Through a combination of convolutional layers, pooling layers, and fully connected layers, it achieves deep feature extraction of state parameters such as equipment health index, load rate, and processing capacity. In the action output layer, a Gaussian noise module is integrated, with its noise amplitude controlled by a dynamic adjustment factor η. This factor is adjusted in real time according to the current health index, inventory level, and safety risk level to achieve a balance between exploration and utilization. For example, when the equipment health index is below the warning threshold, the system automatically increases the noise amplitude to 150% of the original value to promote the exploration of alternative equipment combinations; when the inventory is at a critical state, the noise direction is biased towards material allocation instructions to enhance supply chain resilience.

[0099] The adaptive noise exploration mechanism is one of the core innovations of the model. This mechanism achieves dynamic noise adjustment through state sensitivity analysis, ensuring that the algorithm maintains efficient exploration capabilities in complex mining production scenarios. The noise attenuation coefficient is updated using an exponential moving average to ensure convergence stability in the later stages of training. Specifically, the system calculates state sensitivity indicators based on parameters such as the current state of equipment health index, inventory level, and safety risk level, and then adjusts the noise amplitude and direction accordingly. For example, during periods of high geological disaster incidence, the system automatically increases the weight of safety targets to 70% and adjusts the noise strategy accordingly, prioritizing the exploration of safe and compliant scheduling instructions. During periods of surging market demand, the system increases the weight of efficiency targets to 60%, and the noise strategy favors equipment start-up / shutdown and capacity adjustment instructions to quickly respond to changes in market demand.

[0100] Multi-objective MDP modeling transforms mine production planning into a dynamic process of state transition and reward optimization. The state space comprises five dimensions: equipment status, inventory level, market demand, environmental parameters, and safety risk level. The equipment status dimension is further refined into three categories: normal, warning, and fault, recording real-time equipment operating parameters such as mining machine location, conveyor load rate, and crusher processing capacity. The inventory level dimension is categorized into three states: sufficient, critical, and shortage based on health data, while finished ore inventory is labeled as full, normal, and low. The market demand dimension connects to order data through the ERP system, classifying demand intensity into three levels on a weekly basis: peak, stable, and trough, while simultaneously recording product type and delivery deadline. Environmental parameters include external influencing factors such as geological conditions, weather conditions, and energy prices, while the safety risk level is dynamically assessed by comprehensively considering factors such as equipment failure probability, personnel operational risks, and environmental safety hazards.

[0101] The action space is defined as a set of four core production scheduling instructions: equipment start / stop instructions, material allocation instructions, manpower scheduling instructions, and capacity adjustment instructions. Each instruction is subject to strict constraints to ensure compliance with equipment operating procedures, safety regulations, and resource limitations. For example, equipment start / stop instructions must adhere to the requirement of 3 minutes of no-load preheating before crusher startup, blasting operations must avoid thunderstorms, and the total number of equipment scheduled at the same time must not exceed 80% of the total number of in-use equipment. State transition rules define the update logic of the state space after executing a scheduling instruction. For example, after executing an instruction to increase the operating time of mining machine B by 2 hours, the machine's health index in the equipment status decreases by 5 points, the finished ore quantity in the inventory level increases by 300 tons, and the remaining order quantity in market demand decreases by 300 tons. The reward mechanism adopts a hierarchical weighted structure combined with dynamic priority adjustments. Actions that meet the optimization goals are positively rewarded, such as a 10-point reward for achieving the equipment utilization rate target; penalties are imposed for violating the goals, such as a 20-point deduction for safety violations. The reward value changes dynamically with the state; for example, in a state of inventory shortage, the reward for successfully replenishing raw materials is doubled.

[0102] The training and optimization strategy employs a hybrid offline-online mode. In the offline phase, a basic policy network is pre-trained based on historical data, generating 50 basic scheduling schemes covering different equipment combinations, capacity allocation ratios, and time schedules. In the online phase, the state space is updated through real-time data streaming, and the schemes are iteratively optimized through multiple rounds of debate and consensus mechanisms. The debate mechanism, using a multi-agent collaborative decision-making framework, simulates the conflict and negotiation process of different scheduling strategies to form a consensus-based optimization scheme. The consensus mechanism aggregates the decision results of multiple agents to select the optimal set of scheduling instructions. During iteration, the cumulative reward value is statistically analyzed using a sliding window, retaining the top 30% of high-quality schemes for local adjustments, such as replacing 1-2 scheduling instructions. After 5 rounds of iteration, 100 initial plan sets are formed. Each iteration dynamically adjusts the noise amplitude to adapt to changes in the scenario, ensuring the algorithm maintains a high-efficiency balance between exploration and utilization in dynamic production environments.

[0103] In a real-world scenario, the algorithm was validated at a large copper mine. Through a dynamic noise adjustment mechanism, the algorithm maintained the stability of the scheduling scheme even when market demand fluctuated by 30%. During the rainy season, a period of high geological disaster risk, the system automatically triggered adjustments to safety priority targets, keeping downtime for major equipment failures to within 20 hours. Regarding cost control, energy consumption per unit of ore was controlled at 7.8 kWh / ton, and manual scheduling costs were reduced by 15%. Equipment utilization increased to 92%, and material transportation cycles were shortened to 3.8 hours. By sensing the mine's production status in real time and dynamically adjusting scheduling strategies, this model achieved efficient resource allocation and multi-objective collaborative optimization, providing an intelligent and dynamic solution for mine production planning.

[0104] The decision model building module is used to generate an initial production plan based on the production agent, the safety agent to assess the risk level and trigger emergency plans, and the resource agent to dynamically allocate equipment, manpower and energy, and build a multi-agent collaborative decision model.

[0105] Specifically, this embodiment also includes an Agent architecture establishment unit, which is used to coordinate the overall framework of the production Agent in production planning. The security Agent and resource Agent act as auxiliary decision-makers, respectively proposing adjustment suggestions for the plan from the perspectives of security and resources. The three agents achieve data synchronization through an information interaction channel.

[0106] The safety level assessment unit is used to determine the risk of failure by analyzing equipment status data and to evaluate the safety level of the working face by combining dust concentration, gas content and surrounding rock stability parameters.

[0107] The risk assessment unit is used to pause the execution of the production agent's plan when the security agent assesses that the plan has medium to high risks. The plan will be restarted after the resource agent adjusts the equipment to eliminate the risks. When allocating resources, the resource agent will prioritize meeting the production task requirements proposed by the production agent.

[0108] Specific;

[0109] 3.1 The production agent generates the initial production plan:

[0110] The production agent generates plans based on the initial plan set output in step 2, combined with historical mine production data (equipment operating efficiency and mining area capacity fluctuation patterns over the past 12 months) and real-time operating conditions (ore grade and roadway access status in each mining area).

[0111] During the generation process, the connection logic of each production link needs to be clearly defined: for example, the ore in mining area A is transported to crusher 2 by conveyor 1 for processing, and the processed ore is transported to finished product bin No. 3 by elevator 3. The start and end time of each link is accurate to 30 minutes, and a 15% buffer time is reserved to deal with sudden delays.

[0112] Adjust the plan details according to the characteristics of different mining areas: prioritize the allocation of efficient mining equipment to high-grade mining areas, and appropriately reduce the daily mining volume and extend the operation cycle in geologically complex mining areas to ensure that the plan matches the site conditions.

[0113] 3.2 The security agent assesses the risk level and triggers the emergency response plan:

[0114] The risk assessment covers three main categories: equipment, environment, and operation. For equipment, the risk of failure is determined by analyzing equipment status data (abnormal vibration frequency, continuous temperature rise). For the environment, the safety level of the working face is assessed by combining parameters such as dust concentration, gas content, and surrounding rock stability. For operation, the safety of dispatch instructions is checked to ensure compliance with safety procedures (qualification of blasting personnel and compliance of equipment maintenance procedures).

[0115] The risk level is divided into four levels: no risk (all indicators are within the safety threshold), low risk (a single indicator slightly exceeds the standard, such as dust concentration slightly higher than the standard value by 5%), medium risk (two or more indicators exceed the standard or one indicator significantly exceeds the standard, such as gas concentration reaching 1.2 times the safety threshold), and high risk (there is a direct safety hazard, such as obvious cracks in the surrounding rock accompanied by rockfall).

[0116] Emergency response plan triggering mechanism: In low-risk situations, a warning notice is sent to the on-site team leader, requiring them to increase the frequency of inspections (from once every 2 hours to once every 1 hour); in medium-risk situations, operations in high-risk areas are suspended, and control measures such as local ventilation and equipment load reduction are initiated; in high-risk situations, all operations in the affected areas are immediately stopped, personnel are organized to evacuate to safe shelters, and emergency rescue teams are put on standby.

[0117] 3.3 Resource Agent dynamically allocates equipment, manpower, and energy:

[0118] Equipment allocation follows the principles of efficiency priority and load balance: based on the production tasks of each mining area, the best-performing mining machines and transport machines are prioritized for allocation to high-yield mining areas, while ensuring that a single piece of equipment operates continuously for no more than 8 hours (to avoid excessive fatigue), and the number of spare equipment in the same mining area is no less than 20% of the equipment in use.

[0119] Human resource allocation is based on skill matching and work intensity: blasters with more than 10 years of experience are assigned to complex geological mining areas, and electricians and mechanical maintenance workers are stationed nearby according to the equipment distribution area (at least 1 electrician and 1 mechanical maintenance worker are assigned to each work face), and the daily working time of a single person does not exceed 10 hours to avoid fatigue work.

[0120] Energy allocation is dynamically adjusted and staggered: during peak mining periods (8:00-12:00, 14:00-18:00), priority is given to ensuring the power supply of core equipment such as mining machines and hoists, while the power of non-essential equipment (auxiliary lighting, some transport machines) is reduced by 20%; during off-peak periods at night (0:00-6:00), energy supply is appropriately reduced to maintain only the minimum operating requirements of necessary equipment, thereby reducing energy costs.

[0121] 3.4 Constructing a Multi-Agent Collaborative Decision-Making Model:

[0122] The model architecture adopts a layered collaborative model: the production agent acts as the core decision-maker, responsible for coordinating the overall framework of the production plan; the security agent and resource agent act as auxiliary decision-makers, respectively proposing adjustment suggestions for the plan from the perspectives of security and resources. The three parties achieve data synchronization through an internal information exchange channel (real-time data sharing platform), with information transmission delay not exceeding 10 seconds.

[0123] The collaborative rules clearly state that when a security agent assesses that a plan has medium to high risks, the production agent must suspend the execution of the plan and restart it after the resource agent adjusts equipment or manpower to eliminate the risks. When allocating resources, the resource agent must prioritize meeting the critical production task requirements proposed by the production agent, while also taking into account the risk control requirements of the security agent (prioritizing the allocation of emergency power generation equipment to high-risk areas).

[0124] The production planning decision module is used to optimize the initial set of plans through an agent-based multi-collaborative decision-making model to obtain the target mine production plan.

[0125] Specifically, this embodiment also includes a security and resource assessment unit, which is used by the production agent to select 20 representative plans from the initial plan set and submit them to the security agent and resource agent for assessment;

[0126] The production planning optimization unit is used by the safety agent to feed back the risk assessment results to the production agent. The production agent then modifies the plan by combining the resource adjustment suggestions provided by the resource agent, and iterates to obtain the target mine production plan.

[0127] Specifically,

[0128] 4.1 Optimization process of agent multi-collaborative decision-making model:

[0129] The first round of optimization: The production agent selects 20 representative plans (covering different capacities, energy consumption, and safety redundancy types) from the initial plan set and submits them to the safety agent and resource agent for evaluation. The safety agent focuses on checking whether the safety measures for high-risk operations (deep-hole blasting, high-gas mining area operations) in the plan are in place, and the resource agent calculates whether the equipment, manpower, and energy required for the plan are within the available resource range.

[0130] Feedback and Adjustment: The safety agent feeds back the risk assessment results (the gas warning frequency in mining area B in a certain plan is too high) to the production agent. The production agent, together with the resource adjustment suggestions provided by the resource agent (adding 1 gas extraction device), modifies the plan. The adjusted plan is then resubmitted for evaluation until the safety risk level is reduced to low risk or below and the supply and demand of resources are balanced.

[0131] Multiple iterations: Repeat the above evaluation-adjustment process, optimizing 5-8 plans in each iteration, focusing on plans with prominent problems in the previous evaluation (plans with serious resource conflicts and high risk levels). After 3-4 iterations, select 30 better plans to enter the final optimization stage.

[0132] 4.2 Optimization Direction and Decision Basis:

[0133] Efficiency optimization: The production agent analyzes and plans inefficient processes such as equipment idle time and material transportation waiting time. Combined with the equipment allocation plan provided by the resource agent (adjusting the transport aircraft route to reduce idle time), the production cycle is shortened, with the goal of increasing the overall equipment utilization rate to over 92%.

[0134] Cost optimization: By comparing the manpower and energy consumption data of different plans, the resource agent proposes cost reduction suggestions (reducing the number of temporary workers during low-load periods and adjusting equipment operating parameters to reduce energy consumption). The production agent adjusts the plan accordingly to strive to keep the unit ore production cost within the budget.

[0135] Safety Optimization: Based on the latest environmental parameters (sudden heavy rain causing water accumulation in the mining area) and equipment status (decreased sensitivity of a transport aircraft's braking system), the safety agent requests the production agent to modify the plan (suspend open-pit mining operations and replace the transport aircraft) to ensure that the plan complies with the latest safety standards.

[0136] Collaborative decision-making basis: The decision-making is based on multi-dimensional evaluation indicators, including efficiency indicators (equipment utilization rate, production cycle), cost indicators (energy consumption per unit of ore, labor cost), and safety indicators (risk level, completeness of emergency plan). Each indicator is comprehensively scored according to dynamic weight (the weight of safety indicators in the rainy season is increased to 40%), and the top 10% of the scores are selected as candidate target plans.

[0137] 4.3 Determination of the target mine's production plan:

[0138] Final selection: Select the best performing plan from the candidate target plans, which must simultaneously meet the following requirements: equipment utilization rate ≥ 90%, unit ore energy consumption ≤ 8 kWh / ton, safety risk level of no risk or low risk, and resource supply and demand gap of 0.

[0139] Plan Confirmation and Output: The production agent, safety agent, and resource agent jointly confirm the final plan. Confirmation includes ensuring consistency across all timelines, resource allocation schemes, and safety measures. The output plan must include a detailed execution checklist: daily work team rosters for each mining area, equipment maintenance schedules, material transport dispatch sheets, safety checkpoint locations and frequencies, emergency contact persons and response procedures, etc., ensuring on-site personnel can directly execute the plan.

[0140] Its beneficial effects are as follows: 1. Utilizing AI for query and control improves information acquisition speed and reduces fault response time by more than 50%. Intelligent dialogue and knowledge retrieval shorten decision-making time and improve work quality. AI analysis of data promptly identifies potential problems, resulting in high accuracy in production planning. 2. The off-peak energy supply model reduces energy consumption per unit of ore, and manpower is allocated according to skill matching and work intensity, reducing temporary labor costs and achieving the dual goals of cost reduction and efficiency improvement. 3. It ensures that the plan is adaptable to special working conditions; during peak market demand, efficiency targets are prioritized, achieving on-demand adjustments and dynamic balance, thereby improving production efficiency. The final output target plan includes a detailed execution checklist, which can directly guide on-site operations, significantly reducing manual decision-making costs and error rates.

[0141] Please see Figure 2 In a mine production planning intelligent optimization system, the mine data processing module includes the following units:

[0142] The data cleaning unit is used to clean the real-time acquired data, remove outliers from equipment vibration and gas concentration data using the 3σ criterion, and fill in missing values ​​caused by sensor communication interruptions using linear interpolation to obtain the first processed data.

[0143] The standardization processing unit is used to convert the temperature and vibration physical quantity parameters in the first processed data into dimensionless values ​​in the 0-1 range, unifying the data magnitude; and to perform time alignment on the time-series data to obtain the second processed data.

[0144] The feature extraction unit is used to input the second-processed data into the DeepSeek-V3 multimodal model. It extracts the frequency domain features of the equipment vibration signal through the model's convolutional layer, captures the trend features of the inventory data through the recurrent neural network, focuses the environmental parameter features of high-risk areas through the attention mechanism, and analyzes the ore body distribution features of the geological data through the three-dimensional convolutional module to output mine status data.

[0145] Please see Figure 3 In a mine production planning intelligent optimization system, the production planning output module includes the following units:

[0146] A Markov decision unit is used to update the state space according to a fixed logic after executing a certain scheduling instruction, giving positive rewards to actions that meet the optimization objective and imposing penalties when the objective is violated.

[0147] The scheduling scheme generation unit is used to randomly generate 50 basic scheduling schemes, including equipment combinations, capacity allocation ratios, and time arrangements.

[0148] The initial plan generation unit is used to evaluate the cumulative reward value of each plan, retain the top 30% of high-quality plans and make local adjustments, and generate an initial plan set after 5 iterations.

[0149] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A mine production planning intelligent optimization system, characterized in that, The intelligent optimization system for mine production planning includes the following modules: The mine data processing module is used to collect real-time data on mine equipment status, inventory, environmental parameters, and geological data through sensors. This results in real-time data collection, which is then cleaned by removing outliers in equipment vibration and gas concentration data using the 3σ criterion. Linear interpolation is used to fill in missing values ​​caused by sensor communication interruptions, resulting in the first processed data. Temperature and vibration physical quantities in the first processed data are converted to dimensionless 0-1 range values ​​to unify the data magnitude. Time-series data is then time-aligned to obtain the second processed data. This second processed data is input into the DeepSeek-V3 multimodal model. The model's convolutional layers extract the frequency domain features of equipment vibration signals, a recurrent neural network captures trend features of inventory data, an attention mechanism focuses on environmental parameter features in high-risk areas, and a 3D convolutional module analyzes the orebody distribution features of the geological data, outputting mine status data. The production planning output module is used to acquire multi-objective optimization objectives and mine status data for mine production planning. It uses an improved TD3 algorithm combined with an adaptive noise exploration strategy to model the production planning as a multi-objective MDP Markov decision process. The state space includes equipment status, inventory level, market demand, environmental parameters and safety risk level, and the action space is a set of refined production scheduling instructions. The plan set is generated through a multi-agent collaborative decision-making framework and optimized through debate and consensus mechanisms to output the initial plan set. The decision model building module is used to generate an initial production plan based on the production agent, the safety agent to assess the risk level and trigger emergency plans, and the resource agent to dynamically allocate equipment, manpower and energy, and build a multi-agent collaborative decision model. The production planning decision module is used to optimize the initial plan set through an Agent multi-collaborative decision-making model to obtain the target mine production plan.

2. The intelligent optimization system for mine production planning as described in claim 1, characterized in that, The production planning output module includes the following units: The state space construction unit is used to integrate equipment health indices, which are refined into three types of state labels to record the current operating condition of the equipment, classify the raw material inventory status based on inventory health data, and classify the demand intensity according to market demand on a weekly basis. The action space definition unit is used to divide production scheduling instructions into four types of actions using the improved TD3 algorithm combined with an adaptive noise exploration strategy: equipment start / stop instructions, material allocation instructions, manpower scheduling instructions, and capacity adjustment instructions. The action constraints include that all instructions must comply with equipment operating procedures, safety procedures, and resource limitations.

3. The intelligent optimization system for mine production planning as described in claim 1, characterized in that, The production planning output module also includes the following units: A Markov decision unit is used to update the state space according to a fixed logic after executing a certain scheduling instruction, giving positive rewards to actions that meet the optimization objective and imposing penalties when the objective is violated. The scheduling scheme generation unit is used to randomly generate 50 basic scheduling schemes, including equipment combinations, capacity allocation ratios, and time arrangements. The initial plan generation unit is used to evaluate the cumulative reward value of each plan, retain the top 30% of high-quality plans and make local adjustments, and generate an initial plan set after 5 iterations.

4. The intelligent optimization system for mine production planning as described in claim 1, characterized in that, The decision model building module includes the following units: The Agent architecture establishes a framework for the overall production plan, with the production agent coordinating production planning. The security agent and resource agent act as auxiliary decision-makers, providing adjustment suggestions for the plan from the perspectives of security and resources, respectively. The three agents synchronize data through an information exchange channel.

5. The intelligent optimization system for mine production planning as described in claim 1, characterized in that, The decision model building module also includes the following units: The safety level assessment unit is used to determine the risk of failure by analyzing equipment status data and to evaluate the safety level of the working face by combining dust concentration, gas content and surrounding rock stability parameters. The risk assessment unit is used to pause the execution of the production agent's plan when the security agent assesses that the plan has medium to high risks. The plan will be restarted after the resource agent adjusts the equipment to eliminate the risks. When allocating resources, the resource agent will prioritize meeting the production task requirements proposed by the production agent.

6. The intelligent optimization system for mine production planning as described in claim 1, characterized in that, The production planning decision module includes the following units: The security and resource assessment unit is used by the production agent to select 20 representative plans from the initial plan set and submit them to the security agent and resource agent for assessment. The production planning optimization unit is used by the safety agent to feed back the risk assessment results to the production agent. The production agent then modifies the plan by combining the resource adjustment suggestions provided by the resource agent, and iterates to obtain the target mine production plan.

7. A method for implementing a mine production planning intelligent optimization system as described in claim 1, characterized in that, The method includes the following steps: Real-time data acquisition is achieved by collecting mine equipment status data, inventory data, environmental parameters, and geological data in real time using sensors. This real-time data is then cleaned, and outliers in equipment vibration and gas concentration data are removed using the 3σ criterion. Linear interpolation is used to fill in missing values ​​caused by sensor communication interruptions, resulting in the first processed data. Temperature and vibration physical quantities in the first processed data are converted to dimensionless 0-1 range values ​​to unify the data magnitude. Time-series data is then time-aligned to obtain the second processed data. This second processed data is input into the DeepSeek-V3 multimodal model. The model's convolutional layers extract the frequency domain features of equipment vibration signals, recurrent neural networks capture trend features of inventory data, an attention mechanism focuses on environmental parameter features in high-risk areas, and a 3D convolutional module analyzes the ore body distribution features of the geological data, outputting mine status data. The system acquires multi-objective optimization objectives and mine status data for mine production planning. It then uses an improved TD3 algorithm combined with an adaptive noise exploration strategy to model production planning as a multi-objective MDP Markov decision process. The state space includes equipment status, inventory level, market demand, environmental parameters, and safety risk level, while the action space is a set of refined production scheduling instructions. The system generates a plan set through a multi-agent collaborative decision-making framework and optimizes it through a debate and consensus mechanism, outputting an initial plan set. The production agent generates an initial production plan, the safety agent assesses the risk level and triggers emergency plans, and the resource agent dynamically allocates equipment, manpower and energy to build a multi-agent collaborative decision-making model. The initial set of plans is optimized using an agent-based multi-collaborative decision-making model to obtain the target mine production plan.

8. The method for implementing an intelligent optimization system for mine production planning as described in claim 7, characterized in that, The method includes the following steps: By analyzing equipment status data, the risk of failure is determined, and the safety level of the working face is assessed by combining dust concentration, gas content, and surrounding rock stability parameters. When the security agent assesses a plan with medium to high risk, the production agent suspends the execution of the plan and restarts it after the resource agent adjusts the equipment to eliminate the risk. When allocating resources, the resource agent prioritizes meeting the production task requirements proposed by the production agent.

9. The method for implementing an intelligent optimization system for mine production planning as described in claim 7, characterized in that, The method includes the following steps: The production agent selects 20 representative plans from the initial plan set and submits them to the security agent and resource agent for evaluation; The safety agent feeds back the risk assessment results to the production agent, which then modifies the plan based on the resource adjustment suggestions provided by the resource agent, and iterates to obtain the target mine production plan.

Images