An underground mine production equipment intelligent scheduling method and system

Abstract

The present application relates to the technical field of mine production scheduling optimization, and discloses an underground mine production equipment intelligent scheduling method and system, which comprises the following steps: collecting core data of underground mine production; constructing an underground mine scheduling model for underground mine production equipment scheduling; extracting process deep features, equipment deep features and global features of the underground mine scheduling model; integrating the extracted process deep features, equipment deep features and global features; constructing a reinforcement learning algorithm combined with contrast learning to train the underground mine scheduling model; inputting underground mine production data into the trained underground mine scheduling model, and inputting abnormal data into the trained underground mine scheduling model to obtain a real-time adjusted stope production connection plan and equipment scheduling scheme. Through the method, the problem that manual work cannot quantify complex correlations is solved, and the problems of low efficiency, high safety risk and weak dynamic adaptation of traditional manual scheduling are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of mine production scheduling optimization technology, specifically to an intelligent scheduling method and system for underground mine production equipment. Background Technology

[0002] The scheduling of production equipment in underground mines is a core aspect of mine production, directly impacting both safety and efficiency. It requires a balance between safe workflow coordination and high overall production efficiency, given the constraints of irreversible mining operations (such as support before ore extraction) and limited equipment resources (a single piece of equipment cannot simultaneously serve multiple mining areas). Traditional underground mine equipment scheduling relies entirely on manual methods. Schedulers must rely on experience to manually plan the sequence of mining operations and allocate equipment resources. This approach has significant limitations: firstly, manual methods struggle to accurately handle the complex relationships between mining areas, processes, and equipment, easily leading to errors in workflow coordination (such as delayed support operations increasing the risk of mining area exposure) or imbalances in equipment allocation (some equipment idle, others overloaded), affecting both production efficiency and creating safety hazards. Secondly, in the face of dynamic changes in mine production (such as temporary equipment failures or adjustments to mining operations), manually re-formulating scheduling plans is time-consuming and slow, often resulting in process stagnation or breaches of safety thresholds, making it difficult to meet the demands of modern mines for timely and reliable scheduling. As the construction of smart mines progresses, the problems of the extensiveness and lag of manual scheduling become increasingly prominent, and there is an urgent need for an intelligent method to break through the bottleneck.

[0003] Therefore, it is necessary to propose an intelligent scheduling method for underground mine production equipment to solve the problem that the complex relationships in underground mine production cannot be quantified manually, and to improve the generation efficiency and real-time adaptability of scheduling schemes. Summary of the Invention

[0004] The technical problem to be solved by this invention is to overcome the shortcomings of the existing technology and provide a highly efficient and adaptable intelligent scheduling method and system for underground mine production equipment. The specific technical solution is as follows:

[0005] Firstly, a method for intelligent scheduling of underground mining production equipment is provided, the method comprising:

[0006] S1: Collect core data on underground mine production, including mining parameters, process parameters, and equipment parameters;

[0007] S2: Based on the core data, the underground mine production equipment scheduling problem is abstracted into a flexible workshop scheduling problem, and an underground mine scheduling model for underground mine production equipment scheduling is constructed.

[0008] S3: Construct a dual-attention feature extraction network to extract deep process features, deep equipment features, and global features of the underground mine scheduling model;

[0009] S4: Construct a policy network to integrate the extracted deep features of the process, deep features of the equipment, and global features to obtain the action probability distribution of process-equipment pairing;

[0010] S5: Based on the action probability distribution of the process-equipment pairing, a reinforcement learning algorithm combining contrastive learning is constructed to train the underground mine scheduling model, thereby obtaining the trained underground mine scheduling model.

[0011] S6: Input the collected underground mine production data into the trained underground mine scheduling model to generate a mining area production succession plan and equipment scheduling scheme.

[0012] As a further improvement to the above technical solution:

[0013] The stope parameters include the number of stopes N and the reserves of the i-th stope. ;

[0014] The process parameters include the number of processes M, process sequence constraints, and the baseline operation time of the j-th process in the i-th mining area. ;

[0015] The equipment parameters include the equipment model and quantity for the j-th process. Equipment efficiency and equipment travel time loss coefficient ;

[0016] The sequence of processes in the process sequence constraint is as follows: rock drilling, charging, blasting, ventilation, support, ore extraction and filling.

[0017] As a further improvement to the above technical solution:

[0018] The abstraction of the underground mine production equipment scheduling problem into a flexible job shop scheduling problem includes:

[0019] The stopes are mapped to workpieces in the flexible workshop scheduling, and each stope needs to complete 7 fixed processes, thus defining a stope set. for:

[0020] ;

[0021] in, Represents the i-th mining area;

[0022] The single-process operation in the stope is mapped to the operation of the flexible work workshop scheduling. The processes in the same stope satisfy the process sequence constraints. For any stope... Its process sequence is as follows These correspond to the processes in the process sequence constraints, representing the complete set of processes in all mining areas. for:

[0023] ;

[0024] This refers to the machines in a flexible work workshop that correspond to mine-specific equipment and similar equipment, and where the same process can be handled by multiple machines of the same type. for:

[0025] ;

[0026] in, This includes the mine-specific equipment and similar equipment corresponding to each process in the process sequence; it also defines a process-equipment compatibility set. ,satisfy , This indicates that the process is compatible with the specific equipment type;

[0027] The total time spent on the equipment for each process is obtained based on the baseline time, equipment efficiency, and movement losses, including the following steps:

[0028] Process In the equipment Basic homework time Determined by the baseline operation time and equipment efficiency, it satisfies the following expression:

[0029] ;

[0030] in, For process The baseline operation time, For equipment The stage effect;

[0031] If the device It needs to be moved from the current mining area to the process area. For the mining area to which it belongs, an additional travel time is added, satisfying the following expression:

[0032] ;

[0033] Where d is the distance between mining areas. For equipment The travel time loss coefficient; at this time, the process In the equipment The total time spent on it is ;

[0034] The constraints of the underground mine scheduling model include the inherent constraints of the flexible job shop scheduling problem, while also incorporating safety and resource limitations inherent in the mining scenario, specifically including:

[0035] Process priority constraints must satisfy the following conditions:

[0036] For any mining area Adjacent processes ;

[0037] Explosion process With ventilation process Additional satisfaction ;

[0038] in, for The start time, for Completion time;

[0039] Equipment exclusivity constraint: The same equipment can process only one process at a time, satisfying the following conditions:

[0040] For any device and two processing steps , ,have or ;

[0041] The total cycle time for all processes to be completed in all mining areas corresponds to the minimum total completion time of flexible operation workshop scheduling, while taking into account the utilization rate of mining equipment, and satisfies the following expression:

[0042] ;

[0043] in, The total cycle time for completing all processes in all mining areas. For process The completion time.

[0044] As a further improvement to the above technical solution:

[0045] The underground mine scheduling model also includes a Markov decision model, which includes:

[0046] The state space includes process characteristics, equipment characteristics, and process-equipment feasible pairing characteristics. Irrelevant information is excluded, and the state space gradually decreases as scheduling progresses.

[0047] Action space, the action space ,in, For the set of processes to be scheduled, A set of idle devices;

[0048] Reward function, the reward function , For a moment The lower bound of the maximum time to complete a process;

[0049] State transition is used to perform actions. Afterwards, update For already scheduled and For busy, regenerate , obtain a new state .

[0050] As a further improvement to the above technical solution:

[0051] S2 further includes constructing a disjunctive graph representation of the underground mine scheduling model, and adjusting the underground mine scheduling model based on the disjunctive graph representation. The disjunctive graph representation of the underground mine scheduling model includes:

[0052] Define disjunction graph ,in:

[0053] Node set Node clustering and These are respectively process nodes, equipment nodes, and mining site nodes;

[0054] Conjunction edge set This corresponds to the sequential constraints of the processes within the same mining area;

[0055] disjunctive edge set For equipment The set of process nodes that can be processed, when the process Assigned to device back, Deletion of parsing edges from other devices The disjunctive edges between the remaining process nodes are converted into directed edges to represent the process processing order on the equipment.

[0056] As a further improvement to the above technical solution:

[0057] The dual-attention feature extraction network includes a state representation module, an operation message attention module, and a machine message attention module;

[0058] The state representation module is used to construct state st at decision time t, and state st includes decision-related operational features, machine features, and feasible pairing features.

[0059] The dimensions of the decision-related operational features, machine features, and feasible pairing features satisfy the minimum sufficiency principle of the flexible job shop scheduling state, specifically:

[0060] Decision-related operational characteristics include the proportion of reserves in the mining area to which the process belongs, the baseline operation time of the process, and the waiting time attribute of the process;

[0061] Machine characteristics include the current status of the equipment, the estimated time for the equipment to move to the target mining area, and the equipment utilization rate attribute;

[0062] Feasible pairing characteristics include the actual operation time of process-equipment pairing and whether the safety constraint attributes are met after pairing;

[0063] And the state It contains only decision-related information, which excludes equipment with completed processes and no available processes. The state space gradually decreases as scheduling progresses, in order to meet the design requirements of a tight state in flexible workshop scheduling.

[0064] The operation message attention module is used to focus on the process associations within the same mining area and obtain deep process features, including:

[0065] The attention coefficient of the operation message attention module Calculate the priority constraints of operations in the flexible job shop scheduling problem, satisfying the following expression:

[0066] ;

[0067] ;

[0068] in, For process The original characteristics, It is a linear transformation matrix. For attention weight vectors, These correspond to the preceding process, the current process itself, and the following process, respectively. The normalized attention coefficient. For ELU activation function; For process The deep characteristics of the process;

[0069] The machine message attention module is used to introduce device contention intensity and obtain deep device features, including:

[0070] The attention coefficient of the machine message attention module The competition relationship between machines in the flexible job shop scheduling problem is calculated, satisfying the following expression:

[0071] ;

[0072] in, and respectively equipment and equipment The original characteristics, It is a linear transformation matrix. For attention weight vectors, For equipment The competition intensity represents the sum of the priority scores of unscheduled processes that the two devices can jointly handle. After softmax normalization, the features are used for weighted aggregation of competing device features, ultimately yielding deep device features. .

[0073] As a further improvement to the above technical solution:

[0074] The dual-attention feature extraction network also includes a global feature aggregation module, which satisfies the following expression:

[0075] ;

[0076] in, These are the process depth features and equipment depth features after passing through an L-layer dual-attention feature extraction network, respectively. Let L be the set of decision-related processes and the set of equipment, respectively, where L=2; These are global features used as decision inputs for the policy network. The aggregation method of global features corresponds to capturing the overall production status through global features in the flexible job shop scheduling problem.

[0077] As a further improvement to the above technical solution:

[0078] The policy network satisfies the following expression:

[0079] ;

[0080] ;

[0081] in, It is a 2-layer fully connected network with 64 hidden layers and the activation function is tanh. These are the process deep features and equipment deep features output by the dual-attention feature extraction network, respectively. As a global feature, For feasible pairing features, Rate the action. This represents the probability distribution of actions.

[0082] The policy network employs an action sampling strategy during the training phase to ensure exploratory behavior, and a greedy strategy during the inference phase to select the action with the highest probability.

[0083] As a further improvement to the above technical solution:

[0084] The reinforcement learning algorithm that combines contrastive learning includes:

[0085] The PPO-Clip module avoids excessive policy update amplitude by pruning the advantage function, thus stabilizing policy network updates and pruning the loss. Satisfy the following formula:

[0086] ;

[0087] in, , which is the ratio of the action probability of the current policy to that of the old policy, used to ensure the continuity of policy updates; , is the dominant function, For the value of the action, State value is used to measure the relative merits of actions; This is a pruning threshold used to avoid... Excessive size leads to strategy oscillation; Indicates the distribution of sampling trajectories Expectations;

[0088] The contrastive learning module is used to enhance the feature learning capabilities of the underground mine scheduling model, including:

[0089] Constraint Satisfaction and Total Cycle of Mine Scheduling Define positive and negative sample trajectories, specifically:

[0090] The positive sample trajectory satisfies all mine constraints, and the total period of the positive sample trajectory is... To find the near-optimal solution for the current schedule, the set of positive samples is defined as follows: K is the number of positive samples;

[0091] Total period of negative sample trajectory Define the negative sample set as L is the number of negative samples;

[0092] Construct a set of contrast sample pairs For each positive sample trajectory, two different negative sample trajectories are randomly selected to form two comparison sample pairs, resulting in a set of all such comparison sample pairs. For any comparison sample pair The cosine similarity metric based on global features is used to compare the similarity between sample pairs of trajectories. Specifically:

[0093] Extract global features of each decision time t in the trajectory of the comparison sample pair. , Calculate the similarity at a single time step. It satisfies the following expression:

[0094] ;

[0095] in, For vector dot product, It is an L2 norm;

[0096] The trajectory-level similarity is calculated by averaging the similarity of the samples across all time periods. It satisfies the following expression:

[0097] ;

[0098] Using InfoNCE loss as the contrastive loss, we promote the clustering of positive sample features and the distancing of negative sample features, specifically:

[0099] For each comparison sample pair Comparison of losses Satisfy the following expression:

[0100] ;

[0101] in, The similarity between a positive sample and itself; Temperature is a parameter used to adjust the weight of similarity.

[0102] The global contrastive loss is obtained by averaging the losses of all contrastive sample pairs. It satisfies the following expression:

[0103] ;

[0104] Total loss function The weighted sum of the clipping loss and the contrast loss satisfies the following expression:

[0105] ;

[0106] in, These are the weighting coefficients;

[0107] Parameter updates are performed using the Adam optimizer, sampling multiple trajectories per training round to construct multiple comparison sample pairs. Regarding the total loss Gradient descent is performed iteratively multiple times to update the parameters of the underground mine scheduling model;

[0108] After multiple rounds of training, the model is validated on a mine scheduling test set. If the constraint satisfaction rate is ≥99% for 5 consecutive rounds of test set and the total cycle optimization rate is ≥8%, then training is stopped to ensure that the underground mine scheduling model has generalization ability in actual mine scenarios.

[0109] In a second aspect, a computer system is provided, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the intelligent scheduling method for underground mine production equipment as described above.

[0110] Compared with the prior art, the advantages of the present invention are as follows:

[0111] This invention discloses an intelligent scheduling method for underground mine production equipment. It systematically collects three core production data categories: mining area, processes, and equipment, providing a precise data foundation for scheduling. A flexible workshop scheduling problem model is then constructed, incorporating a "minimize total production cycle" objective function and constraints related to process sequence, equipment capacity, and mining area space. This model is transformed into a graph structure linking mining area, processes, and equipment, ensuring the scheduling model aligns with actual mine production logic and addressing the challenge of quantifying complex relationships manually. A dual-attention feature extraction network is then designed, using operation and machine message attention modules to focus on process relationships and equipment competition, accurately capturing key scheduling elements easily overlooked by traditional methods. A policy network is then used, integrating features through a fully connected layer and focusing on key elements via an attention mechanism. This outputs process sequence and equipment allocation schemes. A reinforcement learning algorithm combining contrastive learning dynamically optimizes the scheduling model parameters, improving model adaptability and training efficiency, and rapidly generating scheduling schemes adapted to dynamic production changes, solving the problems of slow manual response and poor scheme adaptability. Finally, a mining area succession plan and equipment scheduling scheme are generated based on real-time production data. Simultaneously, a monitoring system tracks equipment status and process progress in real time, triggering an intelligent agent to adjust the scheme online and issue warnings in case of anomalies. This provides mine managers with precise equipment allocation and process coordination, effectively solving the problems of low efficiency, high safety risks, and weak dynamic adaptability of traditional manual scheduling, and providing intelligent and highly reliable technical support for underground mine production scheduling. Attached Figure Description

[0112] Figure 1 This is a flowchart of an intelligent scheduling method for underground mine production equipment according to an embodiment of the present invention. Detailed Implementation

[0113] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0114] This implementation provides an intelligent scheduling method for underground mine production equipment, including the following steps:

[0115] S1: Collect core data on underground mine production, including mining parameters, process parameters, and equipment parameters;

[0116] The stope parameters include the number of stopes N and the reserves of the i-th stope. The process parameters include the number of processes M, process sequence constraints, and the baseline operation time for the j-th process in the i-th mining area. The equipment parameters include the equipment model and quantity for the j-th process. Equipment efficiency and equipment travel time loss coefficient The sequence of operations in the aforementioned process sequence constraint is as follows: rock drilling, charging, blasting, ventilation, support, ore extraction, and backfilling.

[0117] For stope parameters, the filing information of each stope is retrieved through the mine production execution system to determine the total number of stops and the oreable reserves (i.e., stope reserves) of each stope. For process parameters, the total number of processes to be performed in the mining process is determined by referring to the current production process of the mine. The fixed sequential constraint relationship of "rock drilling → charging → blasting → ventilation → support → ore extraction → backfilling" is determined according to the mining process logic specified in the regulations. At the same time, the actual execution time of the corresponding processes in each stope in the historical operation records is combined with the abnormal data to determine the benchmark operation time of each process in each stope. For equipment parameters, the ledger information of the special equipment for each process is retrieved through the mine equipment management system to determine the total number of equipment equipped for each process. The actual operation efficiency (i.e., equipment efficiency) of each piece of equipment is determined based on the rated operating efficiency in the equipment's technical manual and the actual on-site operation load records. Then, the time loss ratio during the equipment movement process is calculated based on the movement path planning of the equipment between different stops and the transportation road condition records. Finally, the equipment movement time loss coefficient is determined.

[0118] S2: Based on the aforementioned core data, the underground mine production equipment scheduling problem is abstracted into a Flexible Job Shop (FJSP) scheduling problem. First, the optimization direction is "minimizing the total production cycle." Then, three types of constraints are set: process sequence constraints (the logical order of "drilling → charging → blasting → ventilation → support → ore extraction → backfilling" must be satisfied; the next process cannot start until the previous process is completed), equipment capacity constraints (a single piece of equipment can only execute one process at a time, and the equipment's operating time cannot exceed its rated working time), and stope space constraints (only one process can be carried out in the same stope at the same time to avoid operational space conflicts). This yields an underground mine scheduling model for underground mine production equipment scheduling.

[0119] The problem of scheduling production equipment in underground mines is abstracted into a flexible job shop scheduling problem, which is achieved through element mapping and mathematical definition, specifically including:

[0120] The core entities of mine stope scheduling are mapped one-to-one with the three core elements of flexible workshop scheduling: "workpiece-operation-machine," ensuring that the abstracted result conforms to the flexible scheduling characteristics of flexible workshop scheduling.

[0121] The stopes are mapped to workpieces in the flexible workshop scheduling, and each stope needs to complete 7 fixed processes, thus defining a stope set. for:

[0122] ;

[0123] in, Represents the i-th mining area;

[0124] The single-stage operation in a stope is mapped to the operation of a flexible work workshop. The operations within the same stope satisfy the aforementioned sequence constraints. The operations within the same stope must be executed in the order of "rock drilling → charging → blasting → ventilation → support → ore extraction → backfilling," which conforms to the priority constraints of operations in the flexible work workshop scheduling problem. This applies to any stope. Its process sequence is as follows These correspond to the processes in the process sequence constraint. Complete set of all mining operations for:

[0125] ;

[0126] This refers to the machines in a flexible work workshop that correspond to mine-specific equipment and similar equipment, and where the same process can be handled by multiple machines of the same type. for:

[0127] ;

[0128] in, This includes the mine-specific equipment and similar equipment corresponding to each process in the process sequence; it also defines a process-equipment compatibility set. ,satisfy , This indicates that the process is compatible with the specific equipment type;

[0129] The total time spent on the equipment for a process is obtained based on the baseline time, equipment efficiency, and movement loss. This is consistent with the definition of processing time in flexible job shop scheduling and includes the following steps:

[0130] Process In the equipment Basic homework time Determined by the baseline operation time and equipment efficiency, it satisfies the following expression:

[0131] ;

[0132] in, For process The baseline operation time, For equipment The stage effect;

[0133] If the device It needs to be moved from the current mining area to the process area. For the mining area to which it belongs, an additional travel time is added, satisfying the following expression:

[0134] ;

[0135] Where d is the distance between mining areas. For equipment The travel time loss coefficient; at this time, the process In the equipment The total time spent on it is ;

[0136] The constraints of the underground mine scheduling model include the inherent constraints of the flexible job shop scheduling problem, while also incorporating safety and resource limitations inherent in the mining scenario, specifically including:

[0137] Process priority constraint means that processes within the same mining area must be executed sequentially, and blasting and ventilation must meet safe intervals, subject to the following specific conditions:

[0138] For any mining area Adjacent processes ;

[0139] Explosion process With ventilation process Additional satisfaction (Unit: minutes);

[0140] in, for The start time, for Completion time;

[0141] Equipment exclusivity constraint: The same equipment can process only one process at a time, satisfying the following conditions:

[0142] For any device and two processing steps , ,have or ;

[0143] The total cycle time for all processes to be completed in all mining areas corresponds to the minimum total completion time of flexible operation workshop scheduling, while taking into account the utilization rate of mining equipment, and satisfies the following expression:

[0144] ;

[0145] in, The total cycle time for completing all processes in all mining areas. For process The completion time.

[0146] In this embodiment, after obtaining the underground mine scheduling model, a disjunctive graph representation of the underground mine scheduling model is constructed. The graph structure of the disjunctive graph uses "stopping area-process-equipment" as the node set and "relationships" as the edge set, which is used to adjust and make the underground mine scheduling model conform to the actual production logic. Specifically, it includes:

[0147] The dissociative graph represents a standard structure of a dissociative graph that perfectly matches the flexible job shop scheduling problem;

[0148] Define disjunction graph ,in:

[0149] Node set Node clustering and These are respectively process nodes, equipment nodes, and mining site nodes;

[0150] Conjunction edge set This corresponds to the sequential constraints of the procedures within the same mining area (e.g., rock drilling → charging).

[0151] disjunctive edge set For equipment The set of process nodes that can be processed (e.g.) (For all drilling process nodes that the rock drilling rig can handle), when the process Assigned to device back, Deletion of parsing edges from other devices The disjunctive edges between the remaining process nodes are converted into directed edges to represent the process processing order on the equipment.

[0152] Furthermore, S2 of this embodiment also includes modeling the underground mine scheduling model as a Markov Decision Process (MDP), wherein the modeling of the Markov Decision Process satisfies:

[0153] The state space includes process characteristics, equipment characteristics, and process-equipment feasible pairing characteristics. Irrelevant information is excluded, and the state space gradually decreases as scheduling progresses.

[0154] Action space, the action space ,in, For the set of processes to be scheduled, A set of idle devices;

[0155] Reward function, the reward function , For a moment The lower bound of the maximum time to complete a process;

[0156] State transition is used to perform actions. Afterwards, update For already scheduled and For busy, regenerate , obtain a new state .

[0157] S3: After completing the scheduling model construction, a dual-attention feature extraction network is built to extract the deep process features, deep equipment features, and global features of the underground mine scheduling model. The state module of the dual-attention feature extraction network constructs state st at decision time t, including process features such as the percentage of reserves in the corresponding stope, waiting time, and baseline operation time; machine features such as the current state, estimated time to move to the target stope, and cumulative utilization rate; and machine-process pair features such as the actual time consumed by the equipment in processing the process and whether the safety interval requirements are met. Furthermore, as scheduling progresses, equipment with completed processes and no suitable tasks is automatically removed, compressing the state space; operation... The message attention module focuses on the process relationships within the same mining area. By calculating the attention coefficients between the current process and its preceding and succeeding processes, it aggregates relevant features to strengthen the expression of process dependencies. The machine message attention module introduces equipment competition intensity, capturing resource contention relationships between equipment through an attention mechanism and optimizing equipment feature output. The policy network integrates dual-attention feature extraction from the network output—deep process features, deep equipment features, and global features—through a fully connected layer. It leverages the attention mechanism to highlight key elements and ultimately outputs the action probability distribution of "process-equipment" pairings. The pairing with the highest probability is selected as the current scheduling decision. Specifically:

[0158] The dual-attention feature extraction network includes a state representation module, an operation message attention module, and a machine message attention module;

[0159] The state representation module is used to construct state st at decision time t, and state st includes decision-related operational features, machine features, and feasible pairing features.

[0160] The dimensions of the decision-related operational features, machine features, and feasible pairing features satisfy the minimum sufficiency principle of the flexible job shop scheduling state, specifically:

[0161] Decision-related operational characteristics include static and dynamic attributes such as the proportion of reserves in the mining area to which the process belongs, the baseline operation time of the process, and the waiting time of the process.

[0162] Machine characteristics include the current status of the equipment (0 = idle, 1 = busy), the estimated time for the equipment to move to the target mining area, and the equipment utilization rate, among other attributes.

[0163] Feasible pairing features include specific attributes such as the actual operation time of process-equipment pairing and whether safety constraints are met after pairing;

[0164] And the state It contains only decision-related information, which excludes equipment with completed processes and no available processes, and the state space gradually decreases as scheduling progresses;

[0165] In this embodiment, after obtaining the state st, the dual attention feature extraction network is input to extract deep features. The dual attention feature extraction network also includes an operation message attention module and a machine message attention module. The operation-machine pairing action probability distribution π(at|st) is generated through the policy network. The action is selected and multi-dimensional constraint verification is performed. If the verification passes, the scheduling decision is executed; otherwise, constraint conflict processing is triggered.

[0166] The operation message attention module is used to focus on the process associations within the same mining area and obtain deep process features, including:

[0167] The attention coefficient of the operation message attention module Calculate the priority constraints of operations in the flexible job shop scheduling problem, satisfying the following expression:

[0168] ;

[0169] ;

[0170] in, For process The original characteristics, It is a linear transformation matrix. For attention weight vectors, These correspond to the preceding process, the current process itself, and the following process, respectively. The normalized attention coefficient. For ELU activation function; For process The deep characteristics of the process;

[0171] The machine message attention module is used to introduce device contention intensity and obtain deep device features, including:

[0172] The attention coefficient of the machine message attention module The competition relationship between machines in the flexible job shop scheduling problem is calculated, satisfying the following expression:

[0173] ;

[0174] in, and respectively equipment and equipment The original characteristics, It is a linear transformation matrix. For attention weight vectors, For equipment The competition intensity (i.e., the sum of priority scores of unscheduled processes that the two devices can jointly process) represents the sum of priority scores of unscheduled processes that the two devices can jointly process. After softmax normalization, the features are used for weighted aggregation of competing device features, ultimately yielding deep device features. .

[0175] The dual-attention feature extraction network also includes a global feature aggregation module, which satisfies the following expression:

[0176] ;

[0177] in, These are the process depth features and equipment depth features after passing through an L-layer dual-attention feature extraction network, respectively. Let L be the set of decision-related processes and the set of equipment, respectively, where L=2; These are global features used as decision inputs for the policy network. The aggregation method of global features corresponds to capturing the overall production status through global features in the flexible job shop scheduling problem.

[0178] Repeat S3 until all operation scheduling is complete.

[0179] S4: Construct a policy network to integrate the extracted deep features of the process, deep features of the equipment, and global features to obtain the action probability distribution of process-equipment pairing;

[0180] The policy network is constructed based on the actor-critic framework of the flexible job shop scheduling problem, and the policy network satisfies the following expression:

[0181] ;

[0182] ;

[0183] in, It is a 2-layer fully connected network with 64 hidden layers and the activation function is tanh. These are the process deep features and equipment deep features output by the dual-attention feature extraction network, respectively. As a global feature, For feasible pairing features, Rate the action. This represents the probability distribution of actions.

[0184] The policy network employs an action sampling strategy during the training phase to ensure exploratory behavior, and a greedy strategy during the inference phase to select the action with the highest probability.

[0185] S5: Based on the action probability distribution of the process-equipment pairing, a reinforcement learning algorithm combining contrastive learning is constructed to train the underground mine scheduling model, resulting in a trained underground mine scheduling model. The reinforcement learning algorithm combining contrastive learning is based on the PPO-Clip framework, and improves the model's ability to learn high-quality scheduling trajectories by incorporating a contrastive learning module. Specifically, it includes three parts: the PPO-Clip module, the contrastive learning module, and total loss optimization, as detailed below:

[0186] After the policy network is constructed, the scheduling model is trained using the PPO deep reinforcement learning training algorithm based on contrastive learning. The model training phase employs a reinforcement learning algorithm combining PPO-Clip and contrastive learning: first, the initial parameters of the policy network, value network, and dual-attention feature extraction network are configured, and a reasonable number of training iterations and learning rate are set to lay the foundation for model training. The core function of the PPO-Clip module is to ensure the stability of policy updates. By calculating the ratio of the action probabilities of the current policy and the old policy under the same state, combined with the advantage function used to measure the relative merits of actions, a pruning loss mechanism is used to limit the magnitude of policy updates, avoiding policy oscillations caused by excessively large single updates, allowing the policy to be gradually optimized through stable iterations.

[0187] The contrastive learning module enhances feature learning capabilities through three steps: First, it constructs positive and negative sample trajectories. Positive sample trajectories represent high-quality scheduling schemes that satisfy all production constraints (such as process sequence, safety intervals, and resource capacity) and have a total production cycle close to the optimal solution (pre-calculated using traditional algorithms). Negative sample trajectories represent inferior schemes with constraint conflicts (such as simultaneous equipment occupancy or insufficient safety intervals) or a total production cycle significantly longer than the near-optimal solution. Second, it calculates the similarity between different trajectories by extracting global features at each decision point in the trajectory and using the average cosine similarity to measure the overall similarity between different trajectories. Third, it drives feature optimization through the InfoNCE loss function, causing features of high-quality trajectories to cluster together and features of high-quality and low-quality trajectories to diverge, enhancing the model's ability to identify high-quality scheduling patterns. The total loss function is a weighted sum of the pruning loss of the PPO-Clip module and the InfoNCE loss of the contrastive learning module. The weight coefficients are determined through repeated optimization on the validation set to balance strategy stability and feature discriminative power. During training, a certain number of scheduling trajectories are sampled in each round to construct corresponding comparison sample pairs. The Adam optimizer is used for mini-batch gradient descent, and the network parameters are updated through multiple rounds of iteration. When the constraint satisfaction rate and total cycle optimization rate of the test set reach the preset standard for multiple consecutive rounds, training is stopped to ensure that the model has stable optimization effect and generalization ability in real-world scenarios.

[0188] The PPO-Clip module (a basic reinforcement learning framework) avoids excessive policy updates by pruning the advantage function, thus stabilizing policy network updates and pruning the loss function. Satisfy the following formula:

[0189] ;

[0190] in, , which is the ratio of the action probability of the current policy to that of the old policy, used to ensure the continuity of policy updates; , is the dominant function, For the value of the action, State value is used to measure the relative merits of actions; The pruning threshold (an empirical value, determined through optimization using a mine scheduling validation set) is used to avoid... Excessive size leads to strategy oscillation; Indicates the distribution of sampling trajectories The expected sampling trajectory comes from the mine scheduling instance library;

[0191] The contrastive learning module enhances the feature learning capabilities of the underground mine scheduling model. Addressing the issue of "confusion between high-quality and low-quality trajectory features" in mine scheduling, this module constructs contrastive signals based on three principles to guide the model in learning more discriminative scheduling state features, including:

[0192] Constraint Satisfaction and Total Cycle of Mine Scheduling Define positive and negative sample trajectories to ensure that the samples reflect the difference between "efficient" and "inefficient" results.

[0193] The positive sample trajectory satisfies all mining constraints (process priority, safety interval, resource carrying capacity constraints), and the total period of the positive sample trajectory is... To find the near-optimal solution for the current scheduling instance, the set of positive samples is defined as follows: K is the number of positive samples;

[0194] Total period of negative sample trajectory Define the negative sample set as L is the number of negative samples;

[0195] Construct a set of contrast sample pairs For each positive sample trajectory, two different negative sample trajectories are randomly selected to form two comparison sample pairs, resulting in a set of all such comparison sample pairs. To ensure that each positive sample has sufficient negative sample references, a global feature cosine similarity metric is used to compare the similarity between sample pairs' trajectories. The global features are derived from the scheduling state global features output by the dual attention feature extraction network. To ensure that similarity reflects the "overall consistency of the scheduling strategy", specifically:

[0196] Extract global features of each decision time t in the trajectory of the comparison sample pair. , Calculate the similarity at a single time step. It satisfies the following expression:

[0197] ;

[0198] in, For vector dot product, It is an L2 norm;

[0199] The trajectory-level similarity is calculated by averaging the similarity of the samples across all time periods. It satisfies the following expression:

[0200] ;

[0201] The purpose of this metric is to enable high-quality trajectories. The SIM between them approaches 1, indicating the difference between high-quality and low-quality trajectories. The SIM between the two approaches 0, providing a clear signal for subsequent loss optimization.

[0202] The InfoNCE (Information Noise Contrastive Estimation) loss is used as the contrastive loss to promote the clustering of positive sample features and the dispersal of negative sample features, ensuring that the model can accurately identify high-quality scheduling strategies. Specifically:

[0203] For each comparison sample pair Comparison of losses Satisfy the following expression:

[0204] ;

[0205] in, This represents the similarity between a positive sample and itself (theoretically, the value is 1, but in practice, it is the average similarity between a positive sample and other positive samples of the same type). The temperature parameter (determined through optimization using the mine scheduling training set) is used to adjust the weight of similarity and avoid the influence of extreme values.

[0206] The global contrastive loss is obtained by averaging the losses of all contrastive sample pairs. It satisfies the following expression:

[0207] ;

[0208] The role of this loss: when the similarity between high-quality and low-quality trajectories is low. When it decreases, Reduce and guide the model to learn deeper features that make the characteristics of high-quality scheduling states more unique and the characteristics of low-quality states easier to distinguish.

[0209] Total loss function The weighted sum of the pruning loss and the contrast loss, balancing "stable policy updates" and "improved feature discriminative power," satisfies the following expression:

[0210] ;

[0211] in, The weighting coefficients are used to achieve an optimal balance between the "constraint satisfaction rate" and the "total cycle optimization rate" of the model.

[0212] Parameter updates, using the Adam optimizer, learning rate Weight decay coefficient To avoid overfitting; 200 trajectories are sampled per training round ( ), construct 500 comparison sample pairs ; Regarding the total loss Perform mini-batch (batchsize=32) gradient descent, iterating 4 times to update the parameters of the underground mine scheduling model;

[0213] After every 10 rounds of training, the model is validated on a mine scheduling test set (including 50 unseen mining scenarios). If the "constraint satisfaction rate ≥ 99%" and "total cycle optimization rate ≥ 8%" (compared to traditional flexible workshop scheduling problem methods) are achieved for 5 consecutive rounds of test set, training is stopped to ensure that the underground mine scheduling model has generalization ability in actual mine scenarios.

[0214] S6: The collected underground mine production data is used to construct a real-time graph structure according to S2, and then input into the trained underground mine scheduling model to generate a stop production succession plan (including the start and stop times of each stop process) and an equipment scheduling scheme (including the operation allocation sequence of each piece of equipment). Specifically:

[0215] Once the model is trained, it can be used to generate real-time mine scheduling schemes. First, the real-time production data collected from the mine (such as the remaining reserves of each stope, the real-time operating status of equipment, and the progress of completed processes) is used to generate a real-time graph structure according to the aforementioned method for constructing the scheduling model graph structure. The remaining reserves information of the stope nodes, the real-time efficiency of the equipment nodes (such as whether the equipment is working normally), and the progress of the completed processes are updated. At the same time, the edge set state (such as the connection arc state between completed processes and subsequent processes) is adjusted. Then, this real-time graph structure is input into the trained scheduling model. After the model is calculated by the policy network module, it outputs two types of results: one is the stope production succession plan, which clarifies the planned start time and planned completion time of each unfinished process in each stope to ensure that the process connection meets the sequential constraints; the other is the equipment scheduling scheme, which clarifies the work allocation order of each piece of equipment (i.e., the processes that the equipment needs to execute in sequence and the corresponding stopes).

[0216] S7: Based on the mining area production succession plan and equipment scheduling scheme, monitor the equipment operation status and process progress in real time. If equipment failure or mining area status abnormality occurs, input abnormal data into the trained underground mine scheduling model to adjust the mining area production succession plan and equipment scheduling scheme in real time.

[0217] While generating the scheduling plan, a real-time monitoring system for mine scheduling is built to achieve dynamic management: the system collects the operating parameters of the equipment (such as speed, temperature, and load) in real time through equipment sensors, analyzes the collected data in real time, and judges whether the equipment is operating normally and whether the process progress meets the plan; if the equipment operating parameters are detected to exceed the normal threshold (determined as equipment failure) or the process progress is behind the plan by more than the set threshold (determined as abnormal mining status), the system immediately inputs the real-time graph structure updated with abnormal data, recalculates and outputs the adjusted scheduling plan (such as assigning the process of the faulty equipment to the standby equipment and adjusting the start and stop time of subsequent processes).

[0218] Through the above process, dynamic scheduling and intelligent control of underground mine production can be achieved, taking into account both the need for target optimization and the flexibility and safety of the production process.

[0219] This embodiment 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 intelligent scheduling method for underground mine production equipment as described above.

[0220] The computer system can be a computing device such as a mobile phone, desktop computer, laptop, handheld computer, or cloud server. The computer system may include, but is not limited to, a processor and memory. For example, the computer system may also include input / output devices, network access devices, and a bus.

[0221] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. For those skilled in the art, improvements and modifications obtained without departing from the inventive concept should also be considered within the scope of protection of the present invention.

Claims

1. An underground mine production facility intelligent scheduling method, characterized in that, The method comprises: S1: collecting core data of underground mine production, wherein the core data comprises stope parameters, process parameters and equipment parameters; The stope parameters include the number of stopes N and the reserves of the ith stope ; The process parameters include the number of processes M, process sequence constraints, and reference operation time of the jth process of the ith stope ; The device parameters include a device model of the jth process, a device quantity , a device table efficiency , and a device moving time loss coefficient ; The sequence of the process sequence constraint is drilling, charging, blasting, ventilation, supporting, ore extraction and filling in turn; S2: based on the core data, abstracting the underground mine production equipment scheduling problem as a flexible job shop scheduling problem, and constructing an underground mine scheduling model for underground mine production equipment scheduling; S3: constructing a double-attention feature extraction network for extracting process deep features, equipment deep features and global features of the underground mine scheduling model; The double-attention feature extraction network comprises a state representation module, an operation message attention module and a machine message attention module; The state representation module is used to construct a state st at a decision time t, wherein the state st comprises decision-related operation features, machine features and feasible pairing features; The dimensions of the decision-related operation features, machine features and feasible pairing features satisfy the minimum sufficient principle of the state of the flexible job shop scheduling, specifically: The decision-related operation features include the stope reserve proportion of the process, the process benchmark operation time and the process waiting time attributes; The machine features include the current state of the equipment, the estimated time for the equipment to move to the target stope and the equipment utilization rate attributes; The feasible pairing features include the actual operation time of the process-equipment pairing and whether the pairing satisfies the safety constraint attribute; and the state contains only decision relevant information, excluding finished processes and devices with no available processes, the state space gradually decreases as the schedule progresses, to meet the design requirement of compact states in flexible job-shop scheduling; The operation message attention module is used to focus on the process association of the same stope and obtain process deep features; The machine message attention module is used to introduce the equipment competition intensity and obtain equipment deep features; S4: constructing a strategy network to integrate the extracted process deep features, equipment deep features and global features to obtain the action probability distribution of the process-equipment pairing; S5: based on the action probability distribution of the process-equipment pairing, constructing a reinforcement learning algorithm combined with contrastive learning to train the underground mine scheduling model, and obtaining the trained underground mine scheduling model; S6: inputting the collected underground mine production data into the trained underground mine scheduling model to generate a stope production connection plan and an equipment scheduling scheme.

2. The underground mine production equipment intelligent scheduling method according to claim 1, characterized in that, The underground mine production equipment scheduling problem is abstracted as a flexible job shop scheduling problem, which comprises: The workpieces corresponding to the flexible job shop scheduling are divided into mining fields, each mining field needs to complete 7 fixed processes, and a mining field set is defined is: ； wherein, represents the ith stope; Corresponding to the operation of flexible job-shop scheduling, the single pass procedure of the mining field, the procedure of the same mining field satisfies the procedure sequence constraint, and for any mining field , the procedure sequence is , corresponding to the procedure in the procedure sequence constraint respectively, and the whole set of procedures of all mining fields is: ； The mine exclusive equipment and the same type equipment corresponding to the flexible job shop scheduling machine, and the same process can be handled by multiple same type equipment, define the equipment set Is: ； wherein, is the mine-specific equipment and the same type of equipment corresponding to each process in the process sequence; at the same time, define the process-equipment compatible set , meet , , indicating that the process and the specific equipment type are compatible one by one; Based on the benchmark time, equipment efficiency and movement loss, the total time of the process on the equipment is obtained, comprising the following steps: Process Basic operation time on the equipment The basic operation time is determined from the reference operation time and the equipment efficiency, and satisfies the following expression:​ ； wherein, is the reference operation time of the process , is the station efficiency of the equipment ; If the device needs to be moved from the current face to the process belonging to the face, then the additional movement time is added, which satisfies the following expression: ； where d is the stope spacing, is the equipment of the mobile time loss factor; the total time spent in the process on the equipment is ; The constraints of the underground mine scheduling model include the inherent constraints of the flexible job shop scheduling problem, as well as the safety and resource limitations of the mine scene, specifically including: Process priority constraint, satisfying the following conditions: for any stope adjacent process ; To the blasting procedure With the ventilation procedure , additionally satisfied ; wherein is the start time of is the completion time of Device exclusive constraint, the same device only processes one process at the same time, satisfying the following conditions: For any device and two processes to be processed , , there are or ; The total cycle of all stope completing all processes corresponds to the minimized total completion time of the flexible job shop scheduling, while considering the utilization rate of mine equipment, satisfying the following expression: ； wherein, is the total cycle time for completing all processes for all stope, is the completion time of the process .

3. The underground mine production equipment intelligent scheduling method according to claim 2, characterized in that, The underground mine scheduling model also includes a Markov decision model, which comprises: State space, containing process features, equipment features and process-equipment feasible pairing features, excluding irrelevant information, and the state space gradually decreases as the scheduling progresses; action space, the action space wherein, is a set of to-be-scheduled processes, is a set of idle devices; Reward function, the reward function , For a moment The lower bound of the maximum time to complete a process; state transition, for performing an action after, update for scheduled and for busy, regenerate , resulting in a new state .

4. The underground mine production equipment intelligent scheduling method according to claim 3, characterized in that, The S2 further comprises constructing a disjunctive graph representation of the underground mine scheduling model, and adjusting the underground mine scheduling model based on the disjunctive graph representation, and the constructing the disjunctive graph representation of the underground mine scheduling model comprises: Defining a disjunctive graph wherein: set of nodes , the set of nodes and are respectively a process node, a plant node and a mine node; conjunct edge set corresponding to the same mining face disjunct edge set for the device a set of process nodes that can be handled, when a process assigned to the device after, disjunct edge deletion with other devices, disjunct edges between the remaining process nodes are converted into directed edges to represent the process handling order on the device.

5. The underground mine production equipment intelligent scheduling method according to claim 4, characterized in that, The attention coefficient of the operation message attention module The priority constraint based on the operation in the flexible job-shop scheduling problem is calculated, satisfying the following expression: ； ； wherein, is the original feature of the process , is a linear transformation matrix, is an attention weight vector, respectively correspond to the previous process, the current process and the subsequent process of the current process, is the normalized attention coefficient, is an ELU activation function; is the process deep feature of the process . attention coefficient of the machine message attention module The competition relationship of machines in the flexible job-shop scheduling problem is calculated to satisfy the following expression: ； wherein, and are original features of the device and the device , is a linear transformation matrix, is an attention weight vector, is the competitive strength of the device , representing the sum of the unscheduled process priority scores that the two devices can jointly handle, after softmax normalization, is used to weight and aggregate the competitive device features, and finally the device deep features are obtained.

6. The underground mine production equipment intelligent scheduling method according to claim 5, characterized in that, The double-attention feature extraction network further comprises a global feature aggregation module, and satisfies the following expression: ； wherein, respectively are the process deep feature and the equipment deep feature after passing through the L-layer double attention feature extraction network, respectively are the decision-related process set and the equipment set, L=2; is the global feature, which is used for the decision input of the strategy network, and the aggregation manner of the global feature aggregation corresponds to capturing the overall production state through the global feature in the flexible job shop scheduling problem.

7. The underground mine production equipment intelligent scheduling method according to claim 6, characterized in that, The policy network satisfies the following expression: ； ； wherein, is a 2-layer fully connected network with hidden layer dimension 64 and activation function tanh; are the procedure deep feature and the device deep feature output by the double attention feature extraction network respectively, is the global feature, is the feasible pairing feature, is the action score, is the action probability distribution; The policy network adopts an action sampling strategy in a training stage to ensure exploration, and adopts a greedy strategy in an inference stage to select an action with the maximum probability.

8. The underground mine production equipment intelligent scheduling method according to claim 7, characterized in that, The reinforcement learning algorithm combined with the contrastive learning comprises: The PPO-Clip module avoids a too large strategy update range by clipping the advantage function, stabilizes the strategy network update, and clips the loss satisfies the following formula: ； wherein, is the ratio of the action probability of the current policy to that of the old policy, used to ensure the continuity of policy update; is the advantage function, is the action value, is the state value, used to measure the relative merits of actions; is the clipping threshold, used to avoid too large to cause policy shock; denotes the expectation of the distribution of the sampled trajectory ; The contrastive learning module is configured to strengthen the feature learning capability of the underground mine scheduling model, and comprises: Constraint satisfaction degree and total period based on mine dispatching , define positive sample trajectories and negative sample trajectories, specifically: The positive sample trajectory satisfies all mine constraints, and the total period of the positive sample trajectory is a near-optimal solution for the current scheduling, and the positive sample set is defined as K is the number of positive samples; total cycle of negative sample track , define the negative sample set as , L is the number of negative samples; Constructing a set of contrast sample pairs For each positive sample trajectory, randomly select 2 different negative sample trajectories to form 2 contrast sample pairs, and obtain a set of all such contrast sample pairs For any contrast sample pair The similarity between the trajectories of the contrast sample pair is measured by the global feature cosine similarity, specifically: extracting global features of the contrast sample at each decision time t in the trajectory , , calculating single-time similarity , satisfying the following expression: ； wherein, is the vector inner product, is the L2 norm; The trajectory-level similarity is calculated by taking the average of the similarity of the contrast sample to the entire time period of the trajectory , satisfying the following expression: ； An InfoNCE loss is used as a contrastive loss to promote feature clustering of positive samples and feature divergence of negative samples, and specifically: for each pair of contrast samples a contrast loss satisfies the following expression: ； wherein, is the similarity of the positive sample to itself; is a temperature parameter for adjusting the weight of the similarity; The global contrast loss is obtained by averaging the loss for all pairs of contrast samples satisfies the following expression: ； Total loss function The total loss function L is a weighted sum of the cropping loss Lcrop and the contrast loss Lcontsatisfying the following expression: ； wherein are weight coefficients; Parameter update, using Adam optimizer, sampling multiple trajectories per round of training, constructing multiple contrast samples , gradient descent is performed on the total loss , and the parameters of the underground mine scheduling model are updated multiple times. After multiple rounds of training, if the constraint satisfaction rate of a test set of the mine scheduling is greater than or equal to 99% and the total cycle optimization rate is greater than or equal to 8% for 5 consecutive rounds, the training is stopped, and the underground mine scheduling model is ensured to have a generalization capability in an actual mine scene.

9. A computer system, characterized by The computer program comprises a processor, a memory, and a computer program stored in the memory and executable by the processor, and the processor implements the underground mine production equipment intelligent scheduling method according to any one of claims 1-8 when executing the computer program.

Images