A method, system, equipment, and medium for collaborative scheduling of computing tasks in mines.

Abstract

This application relates to a method, system, equipment, and medium for collaborative scheduling of computing tasks in mines. The method includes: acquiring multimodal data from the mine and performing spatiotemporal alignment processing to obtain a unified spatiotemporal event stream; based on the unified spatiotemporal event stream, acquiring currently executing or planned tasks, and constructing a task state vector for each task; based on the task state vectors, constructing a production-dimensional task hypergraph, inputting the production-dimensional task hypergraph into a dynamic graph neural network, outputting an enhanced feature vector for each task node and a sequence of predicted hyperedge weights for multiple future time steps, and combining collaborative scheduling constraints and coordination objectives to generate a collaborative scheduling scheme. This method effectively addresses the fundamental deficiency of traditional methods in handling dynamic coupling relationships between tasks, significantly reducing unplanned equipment downtime and production delays caused by production-dimensional conflicts, and improving overall mine operational efficiency and resource allocation rationality.

Description

Technical Field

[0001] This invention belongs to the field of intelligent mine scheduling technology, and in particular relates to a method, system, equipment and medium for collaborative scheduling of mine computing tasks. Background Technology

[0002] With the rapid development of intelligent mining, the widespread deployment of various sensors and smart terminals has provided massive amounts of real-time data for refined production management. In the mining process, coordinating the orderly execution of production tasks such as coal mining with maintenance tasks such as equipment repair has become crucial for improving operational efficiency.

[0003] Currently, mine dispatching systems generally adopt an architecture that separates planning and scheduling. Production plans and maintenance plans are formulated by different departments. Production tasks are usually scheduled according to output targets to ensure the continuity of mining operations. Maintenance tasks are arranged according to fixed equipment maintenance cycles or inserted into production gaps.

[0004] However, the traditional scheduling method that separates production and maintenance has fundamental flaws: when production is under pressure, maintenance tasks are often postponed, leading to equipment operating with defects; conversely, when equipment malfunctions, production must be interrupted for emergency repairs, resulting in output losses. More critically, existing technology cannot accurately perceive and quantify the complex dynamic coupling between the two types of tasks. The intensity of production task execution affects equipment status in real time, dynamically altering the urgency and time window of maintenance tasks; conversely, sudden maintenance tasks consume critical resources, directly impacting the continuous execution of multiple production tasks. Existing technology treats tasks as independent scheduling objects, relying solely on simple priority rules for coordination. This results in scheduling schemes with poor adaptability to real-time changes in the field, failing to dynamically optimize and insert maintenance windows while maintaining production-sales balance constraints. Summary of the Invention

[0005] Therefore, it is necessary to provide a mine task scheduling method that can accurately describe the multidimensional coupling relationship between production and maintenance tasks and perform dynamic collaborative scheduling to address the above-mentioned technical problems.

[0006] Firstly, this application provides a method for collaborative scheduling of mining computing tasks, including:

[0007] Acquire multimodal data from underground mines and perform spatiotemporal alignment processing on the multimodal data to obtain a unified spatiotemporal event stream;

[0008] Based on the unified spatiotemporal event flow, obtain the tasks that are currently being executed or planned to be executed, and construct a task state vector for each task; among them, tasks include production tasks and operation and maintenance tasks;

[0009] Construct a production-dimensional task hypergraph based on the task state vector;

[0010] The hypergraph of the production and dimension task is input into a dynamic graph neural network, which outputs the enhanced feature vector of each task node and the sequence of predicted hyperedge weights for multiple future time steps.

[0011] Based on the enhanced feature vector and the predicted value sequence of hyperedge weights, a collaborative scheduling scheme is generated by combining the collaborative scheduling constraints and the collaborative scheduling objectives.

[0012] Furthermore, the method also includes:

[0013] During the execution of the collaborative scheduling scheme, a unified spatiotemporal event stream is continuously received, and based on the continuously received unified spatiotemporal event stream, the deviation between the actual execution status and the collaborative scheduling scheme is detected, and the detection results are obtained.

[0014] If the detection result shows a deviation, based on the real-time task execution status, the steps of constructing the task state vector to generate the enhanced feature vector and the hyperedge weight prediction value sequence are re-executed to obtain the actual production dimension task hypergraph and the actual hyperedge weight prediction value.

[0015] Based on the actual predicted values ​​of the superedge weights and the planned predicted values ​​in the collaborative scheduling scheme, the degree of entanglement disturbance caused by the deviation is calculated.

[0016] When the degree of entanglement disturbance exceeds the preset threshold, a decoupling operation is performed on the actual production and maintenance task hypergraph to generate a production and maintenance task subgraph.

[0017] Each production and maintenance task subgraph is restructured and scheduled separately to obtain a restructured subgraph scheme. The restructured subgraph schemes are then merged to generate an adjusted collaborative scheduling scheme.

[0018] Furthermore, based on the task state vector, a production-dimensional task hypergraph is constructed, including:

[0019] Based on the resource consumption, spatial range, and time window in the task state vector of each task, the relationships between tasks are identified, and a hyperedge is constructed for each relationship; among them, the relationships include resource competition, state dependency, and spatial mutual exclusion.

[0020] The weight of each hyperedge is calculated based on the dynamic priority and task progress in the task state vector.

[0021] Generate a production-dimensional task hypergraph based on tasks, hyperedges, and weights.

[0022] Furthermore, the hypergraph of the task is input into a dynamic graph neural network, which outputs the enhanced feature vector of each task node and a sequence of predicted hyperedge weights for multiple future time steps, including:

[0023] Based on the node feature matrix, hyperedge association matrix, and hyperedge weight vector of the task hypergraph, the hyperedge feature vector is generated through the first-stage aggregation operation of the spatial convolutional layer of the dynamic graph neural network; wherein, the first-stage aggregation operation is to aggregate the node features onto the hyperedge according to the first attention coefficient of each task node to its hyperedge.

[0024] Based on the hyperedge feature vector, the aggregated node feature vector is generated through the second-stage aggregation operation of the spatial convolutional layer; wherein, the second-stage aggregation operation is to aggregate the hyperedge feature vector back to the task node according to the second attention coefficient of each hyperedge to the task node.

[0025] The aggregated node feature vectors obtained after processing through multiple spatial convolutional layers are combined to generate the node feature vector sequence at the current time.

[0026] The current node feature vector sequence is input into the temporal convolutional layer of the dynamic graph neural network, and the output is the current node hidden state vector that integrates historical information. The current node hidden state vector is then determined as the enhanced feature vector.

[0027] The enhanced feature vectors of all task nodes are combined to generate the first feature matrix. The first feature matrix is ​​then input into the sequence prediction head of the dynamic graph neural network to generate a sequence of hyperedge weight prediction values ​​for multiple future time steps.

[0028] Furthermore, based on the enhanced feature vectors and the predicted sequence of hyperedge weights, combined with preset constraints and preset objectives, a collaborative scheduling scheme is generated, including:

[0029] Based on the expected duration, resource demand vector, spatial region, dynamic priority, earliest start time and latest end time of all tasks, and combined with the inter-task correlation information in the enhanced feature vector, a mathematical model of the scheduling problem is constructed. The constraints of the mathematical model of the scheduling problem include time window constraints, resource capacity constraints, spatial mutual exclusion constraints and production-sales balance constraints. The objective function of the mathematical model of the scheduling problem consists of task delay penalty, resource idle loss and production-dimensional conflict cost.

[0030] Based on the constraints in the mathematical model of the scheduling problem, all tasks are arranged in descending order of dynamic priority to obtain the sequence of tasks to be assigned;

[0031] For each task in the sequence of tasks to be assigned, based on the sequence of predicted superedge weights, task type, and expected output, generate the tentative start time and resource allocation details for all tasks.

[0032] Based on each provisional start time and each resource allocation detail, a complete scheduling scheme is obtained, and the objective function value of the complete scheduling scheme is calculated according to the objective function in the mathematical model of the scheduling problem.

[0033] Randomly select a subset of tasks from the complete scheduling scheme, and based on the subset of tasks, update the sequence of tasks to be assigned and the current remaining resource vector to obtain the updated sequence of tasks to be assigned and the updated current remaining resource vector.

[0034] Based on the updated sequence of tasks to be assigned and the updated current remaining resource vector, the steps of generating the provisional start time and resource allocation details and updating the sequence of tasks to be assigned and the current remaining resource vector are repeated until the preset number of iterations is reached, and all complete scheduling schemes and the objective function value of the complete scheduling schemes are obtained.

[0035] The complete scheduling scheme that minimizes the objective function value is determined as the cooperative scheduling scheme.

[0036] Furthermore, for each task in the sequence of tasks to be assigned, based on the sequence of predicted superedge weights, task type, and expected output, a provisional start time and resource allocation details for all tasks are generated, including:

[0037] The current task is retrieved sequentially from the sequence of tasks to be assigned, and multiple candidate start times are generated for the current task within the allowed time window;

[0038] For each candidate start time, the fitness impact factor is calculated based on the current task's resource requirement vector, the current task's task type, the current task's expected output, the current remaining resource vector, and the hyperedge weight prediction value sequence.

[0039] The fitness factor is weighted and combined to obtain the overall fitness of each candidate start time;

[0040] The candidate start time with the highest overall fit is determined as the provisional start time of the current task, and a resource allocation detail for the current task is generated based on the resource requirement vector and the current remaining resource vector of the current task.

[0041] Once all tasks in the task sequence have been processed, the tentative start time and resource allocation details for all tasks are obtained.

[0042] Furthermore, the fitness influencing factors include resource availability factors, entanglement risk factors, and production-sales balance influencing factors. For each candidate start time, the fitness influencing factors are calculated based on the current task's resource demand vector, the current task's task type, the current task's projected output, the current remaining resource vector, and the hyperedge weight prediction value sequence. These factors include:

[0043] For each candidate start time, the resource availability factor is calculated based on the resource demand vector of the current task and the current remaining resource vector.

[0044] For each candidate start time, the entanglement risk factor between the task executed at the candidate start time and other tasks is calculated based on the superedge weight prediction value sequence.

[0045] For each candidate start time, the production and sales balance impact factor is calculated based on the task type and the expected output of the current task.

[0046] Secondly, this application also provides a collaborative scheduling system for mine computing tasks, comprising:

[0047] The data acquisition module is used to acquire multimodal data from underground mines and perform spatiotemporal alignment processing on the multimodal data to obtain a unified spatiotemporal event stream.

[0048] The vector construction module is used to obtain the currently executing or planned tasks based on the unified spatiotemporal event stream, and to construct a task state vector for each task; among which, tasks include production tasks and operation and maintenance tasks;

[0049] The graph structure construction module is used to construct a multidimensional task hypergraph based on the task state vector;

[0050] The sequence generation module is used to input the production task hypergraph into the dynamic graph neural network and output the enhanced feature vector of each task node and the sequence of predicted hyperedge weights for multiple future time steps.

[0051] The scheme generation module is used to generate a collaborative scheduling scheme based on the enhanced feature vector and the predicted value sequence of the hyperedge weight, combined with preset constraints and preset objectives.

[0052] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, the at least one program, the code set or instruction set is loaded and executed by the processor to implement a mine computing task collaborative scheduling method as described in any of the embodiments of this application.

[0053] Fourthly, this application also provides a computer-readable storage medium storing at least one piece of program code, which is loaded and executed by a processor to implement a mine computing task collaborative scheduling method as described in any of the embodiments of this application.

[0054] The aforementioned method, system, equipment, and medium for collaborative scheduling of mining computing tasks obtain a unified spatiotemporal event stream by collecting multimodal data from the mine and performing spatiotemporal alignment. Based on this unified spatiotemporal event stream, a production-dimensional task hypergraph is constructed and input into a dynamic graph neural network to extract deep features of task nodes and predict future entanglement evolution trends. Under multiple constraints, an optimized scheduling scheme that balances resource utilization and task collaboration is generated. This effectively addresses the fundamental deficiency of traditional methods in handling dynamic coupling relationships between tasks, significantly reduces unplanned equipment downtime and production delays caused by production-dimensional conflicts, and improves overall mine operational efficiency and resource allocation rationality. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating a collaborative scheduling method for mining computing tasks in one embodiment;

[0057] Figure 2 This is a flowchart illustrating the steps of constructing a production-dimensional task hypergraph based on task state vectors in one embodiment.

[0058] Figure 3 This is a schematic diagram of the structure of a collaborative scheduling system for mining computing tasks in one embodiment. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0060] In one embodiment, a collaborative scheduling method for mining computing tasks is provided. This embodiment illustrates the method by applying it to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through the interaction between the terminal and the server. Figure 1 As shown, in this embodiment, the method includes the following steps:

[0061] Step S101: Obtain multimodal data from the mine and perform spatiotemporal alignment processing on the multimodal data to obtain a unified spatiotemporal event stream.

[0062] Spatiotemporal alignment refers to the process of accurately matching and coordinating data from different sources, modalities, or perspectives in both time and space dimensions.

[0063] For example, multimodal data from underground mines is acquired, including: equipment operating status data such as the start-up and shutdown times of coal mining machines; working condition data of roadway support structures, such as the cutting current of the coal mining machine; environmental monitoring data, such as gas concentration; and personnel location data, including the specific roadway coordinates and movement trajectories of miners. The multimodal data is then cleaned: for outliers with instantaneous jumps, sliding window mean filtering can be used; for missing data, time-series-based linear interpolation can be used for completion. The cleaned data is then synchronized using a unified clock source, employing an event window-based alignment strategy to package all events occurring per second into a spatiotemporal event slice. Finally, each event record becomes a standardized unified spatiotemporal event stream, with each event containing five core fields: a UNIX timestamp accurate to milliseconds, spatial coordinates in a three-dimensional roadway coordinate system, entity identifiers distinguishing different equipment or personnel, an event type describing the nature of the event, and the corresponding event value.

[0064] Among them, a coal mining machine is a large-scale mechanical device used for coal mining, capable of cutting, crushing, and loading coal in the mine; data cleaning refers to a series of operations to check, correct, and standardize raw data; sliding window mean filtering refers to defining a fixed-size window on the input data sequence, which slides along the data point by point, calculating the arithmetic mean of all data points within the window at each window position, and using this mean as the output value at the current window center position; linear interpolation based on time series refers to assuming that the change between two known time points is linear, estimating the value at missing time points by connecting adjacent observation points with a straight line; time synchronization refers to aligning with the same standard time to ensure that the timestamps of all data are consistent and comparable; the alignment strategy based on event windows uses the event's own timestamp as a benchmark, defining time windows to map multi-source data, event streams, or time series segments into the same window to complete matching and association; packaging refers to organizing and encapsulating scattered files, data, etc., into a whole.

[0065] Step S102: Based on the unified spatiotemporal event stream, obtain the tasks that are currently being executed or planned to be executed, and construct a task state vector for each task; wherein, tasks include production tasks and operation and maintenance tasks.

[0066] For example, currently executing or soon-to-be-started production and maintenance task instances are parsed from a unified spatiotemporal event stream. A multi-dimensional task state vector is constructed for each task instance, with the following dimensions: Task progress dimension (for production tasks, this refers to the proportion of completed work to total work; for maintenance tasks, it refers to the proportion of completed work steps to total steps, quantified by accumulating equipment running time or completed operation steps); Resource occupancy dimension (updated in real-time by extracting equipment lock status and personnel location information from the event stream, recording the current equipment list, personnel list, and material consumption status); Spatial range dimension (describing the geographical area involved in the task, which can be a coal mining face, tunnel, or a specific coordinate range of a particular piece of equipment); Time window dimension (including the planned start time, planned end time, actual start time, and estimated remaining time based on the current progress); and Dynamic priority dimension (obtained by weighted summation based on factors such as task urgency, potential production loss due to delays, and equipment failure risk level, adjusted in real-time according to changes in external conditions, with each task's state vector updated synchronously during event stream updates). Weighted summation refers to assigning weights to different data according to their importance, multiplying them, and then summing them. Production tasks are primarily identified based on work instructions issued by the mine production planning system. These instructions specify the amount of coal mined or the tunneling footage to be completed at a designated working face within a specific time period. When equipment operation events such as the start of the coal mining machine occur in the event stream, the actual commencement of the production task can be confirmed by combining the planned instructions. Maintenance tasks are identified through three triggering mechanisms: first, preventative maintenance plans triggered based on equipment runtime, such as those automatically generated when the cumulative running time of the coal mining machine reaches a set value. The system includes three main types of maintenance tasks: 1) maintenance tasks; 2) condition-based maintenance requirements triggered when equipment status monitoring data exceeds set standard values; 3) emergency repair tasks triggered by sudden fault alarms, such as fault troubleshooting tasks automatically generated after equipment shutdown protection actions; 4) a digital planning hub for mines (open-pit / underground), which automatically generates and dynamically optimizes long-term, medium-term, and short-term production plans based on geological, resource, and other data, achieving full-process collaboration and closed-loop management of mining, transportation, etc.; and 5) a preventive maintenance plan, which proactively performs maintenance and repairs according to a plan before equipment breaks down, avoiding sudden faults, downtime, and safety accidents.

[0067] Step S103: Construct a production-dimensional task hypergraph based on the task state vector.

[0068] For example, after obtaining the task state vector of each task, a production and maintenance task hypergraph is constructed based on these state vectors. This hypergraph is used to characterize the complex coupling relationship between production tasks and maintenance tasks.

[0069] Step S104: Input the production dimension task hypergraph into the dynamic graph neural network, and output the enhanced feature vector of each task node and the sequence of predicted hyperedge weights for multiple future time steps.

[0070] Among them, dynamic graph neural networks are deep learning models that integrate graph structure modeling and temporal modeling. They are specifically designed to handle dynamic graph data in which nodes, edges, and attributes change over time, thus overcoming the limitation of static graph neural networks in capturing temporal evolution. Static graph neural networks, on the other hand, are deep learning models specifically designed to handle graph structure data. By performing message passing and aggregation operations on the nodes, edges, and global attributes of the graph, each node can integrate information from its neighboring nodes and learn the representation of nodes, edges, or the entire graph.

[0071] For example, the production task hypergraph is input into a dynamic graph neural network, which outputs an enhanced feature vector for each task node and a sequence of predicted hyperedge weights for multiple future time steps.

[0072] Step S105: Based on the enhanced feature vector and the predicted value sequence of hyperedge weights, and combined with the cooperative scheduling constraints and cooperative scheduling objectives, a cooperative scheduling scheme is generated.

[0073] The constraints of collaborative scheduling include time window constraints, resource capacity constraints, spatial mutual exclusion constraints, and production-sales balance constraints; the objectives of collaborative scheduling include minimizing task delay penalties, minimizing resource idle losses, and minimizing additional costs caused by production-maintenance conflicts.

[0074] For example, based on the enhanced feature vector and the predicted sequence of hyperedge weights, combined with the cooperative scheduling constraints and objectives, a heuristic search algorithm can be used to generate a cooperative scheduling scheme. Here, the heuristic search algorithm refers to using a heuristic function (evaluation function) to estimate the cost / distance from the current point to the target, prioritizing paths that are more likely to lead to the answer.

[0075] In this embodiment, a unified spatiotemporal event stream is obtained by collecting multimodal data from the mine and performing spatiotemporal alignment. Based on this unified spatiotemporal event stream, a production and maintenance task hypergraph is constructed and input into a dynamic graph neural network to extract deep features of task nodes and predict future entanglement evolution trends. Under multiple constraints, an optimized scheduling scheme that balances resource utilization and task coordination is generated. This effectively solves the fundamental deficiency of traditional methods in dealing with dynamic coupling relationships between tasks, significantly reduces unplanned equipment downtime and production delays caused by production and maintenance conflicts, and improves the overall operational efficiency and resource allocation rationality of the mine.

[0076] In one exemplary embodiment, the method further includes:

[0077] Step S201: During the execution of the collaborative scheduling scheme, a unified spatiotemporal event stream is continuously received, and based on the continuously received unified spatiotemporal event stream, the deviation between the actual execution state and the collaborative scheduling scheme is detected, and the detection result is obtained.

[0078] For example, after the collaborative scheduling scheme is issued to each execution terminal underground and begins implementation, a unified spatiotemporal event stream is continuously received. This event stream pushes real-time data, including timestamps, spatial coordinates, entity identifiers, event types, and event values, once per second. Each event in the event stream is compared item by item with the currently executing collaborative scheduling scheme: for production tasks, it compares whether the actual coal mining machine start time matches the planned start time and whether the current cumulative output meets the planned progress requirements; for maintenance tasks, it compares whether the equipment shutdown time matches the planned maintenance time and whether maintenance personnel are in place as planned. When a task start time is detected to be earlier or later than allowed, the task execution speed deviates from the expected progress, the actual resource occupancy does not match the planned allocation, or a sudden task such as a newly added equipment fault alarm in the event stream that is not included in the scheme is detected, a deviation is determined to exist. The detection results are stored in the form of a Boolean flag; if a deviation exists, the flag is set to true, and the deviation type and deviation amount are recorded. Among them, the execution terminal refers to all the positions, teams, equipment and monitoring points responsible for specific execution underground; the Boolean flag is a switch or mark used as a Boolean type variable, which has only two states: true or false.

[0079] Step S202: If the detection result indicates a deviation, based on the real-time task execution status, re-execute the steps of constructing the task state vector to generate the enhanced feature vector and the hyperedge weight prediction value sequence to obtain the actual production dimension task hypergraph and the actual hyperedge weight prediction value.

[0080] For example, when the detection result indicates a deviation, the real-time task execution status at the current moment is obtained. This status is determined by the latest received unified spatiotemporal event stream, including the actual progress (ratio of completed workload to total workload) of each currently executing or planned task, the current list of occupied equipment and human resources, the precise spatial coordinates of the task, the actual time consumed by the task, and information on newly triggered emergencies. Based on this real-time status, a new real-time task state vector is constructed for each task. Based on these state vectors, a real-time production-dimensional task hypergraph is constructed. Nodes in this hypergraph represent real-time tasks, and hyperedges represent real-time resource competition, state dependencies, or spatial mutual exclusion relationships. The hyperedge weights are calculated based on the real-time state vectors. This real-time hypergraph is input into a dynamic graph neural network, which outputs the actual enhanced feature vector of each task node at the current moment and a sequence of predicted actual hyperedge weights for multiple future time steps.

[0081] Step S203: Based on the actual predicted value of the superedge weight and the planned predicted value in the collaborative scheduling scheme, the degree of entanglement disturbance caused by the deviation is calculated.

[0082] For example, the sequence of planned superedge weight predictions corresponding to the current time is extracted from the cooperative scheduling scheme generated in step S105. These planned predictions are superedge weight predictions for multiple future time steps saved during the initial scheduling generation. The actual superedge weight predictions and the planned predictions are compared edge-by-edge and time-step-by-time to calculate the degree of entanglement disturbance caused by the deviation. A greater degree of entanglement disturbance indicates that the deviation has spread to a wider range, and the original scheduling scheme is no longer applicable in the entangled network, requiring significant adjustments. The formula for calculating the degree of entanglement disturbance is as follows: .in, This indicates the degree of entanglement perturbation at the current time t. Let be the set of all hyperedges in the current hypergraph. The actual weight prediction value of hyperedge e at time t obtained in step S202. This represents the predicted weight of the superedge e at time t, stored in the initial collaborative scheduling scheme.

[0083] Step S204: When the degree of entanglement disturbance exceeds the preset threshold, perform a decoupling operation on the actual production and maintenance task hypergraph to generate a production and maintenance task subgraph.

[0084] Among them, the preset threshold is an empirical value determined by offline training based on historical mine operation data, used to balance the frequency of scheduling adjustments and system stability; determined by offline training means that it is not calculated online in real time, but rather obtained by post-event, centralized, and batch analysis using algorithms / models; the decoupling operation means breaking down the originally strongly bound, many-to-many, and intertwined production and maintenance tasks, resources, dependencies, and processes into independent, low-dependency, and individually schedulable units, so that the complex hypergraph becomes a clear and controllable modular structure.

[0085] For example, the degree of entanglement perturbation is compared with a preset threshold. When the degree of entanglement perturbation exceeds the threshold, the current scheduling scheme is determined to need to be restructured. A multi-level hypergraph segmentation strategy can be used to perform decoupling operations on the actual production and maintenance task hypergraph: the hypergraph is coarsened by merging tightly coupled node pairs to reduce the hypergraph size; an initial segmentation algorithm is performed on the coarsened hypergraph to divide the hypergraph into several preliminary subgraphs with the goal of minimizing the sum of hyperedge weights between subgraphs; the segmentation results are mapped back to the original hypergraph and fine-tuned, and the final set of production and maintenance task subgraphs is obtained through boundary optimization. There are strong entanglement relationships between the task nodes contained within each subgraph, while the entanglement relationships between subgraphs are weak. Among them, multi-level hypergraph segmentation strategy is the mainstream strategy for solving the efficient and high-quality partitioning of large-scale hypergraphs. It first coarseens and reduces dimensionality, then performs initial subgraph segmentation, and finally refines and restores the structure layer by layer. Coarsening is the process of generating a smaller but coarsened hypergraph by merging strongly related nodes and aggregating hyperedges in a large-scale hypergraph. Initial segmentation algorithm refers to the basic segmentation method used in data processing and other fields to divide the original input into several preliminary sub-regions or sub-units, such as multi-level segmentation: a method of decomposing and classifying research objects, problems, etc., layer by layer according to different dimensions, standards, or granularities. Mapping refers to establishing the correspondence between elements of two sets, that is, for each element in one set, a unique element in another set is assigned to it according to a certain rule. Fine-tuning refers to improving the accuracy, efficiency, quality, or adaptability of a thing or process by fine-tuning, calibrating, and improving details on the existing basis. Boundary optimization refers to the process of adjusting, refining, or reconstructing the constraint boundary of a problem to improve the system performance, accuracy, stability, or efficiency.

[0086] Step S205: Reconstruct and schedule each production and maintenance task subgraph separately to obtain the reconstructed subgraph scheme, and merge the reconstructed subgraph schemes to generate the adjusted collaborative scheduling scheme.

[0087] Reconfiguration scheduling refers to the process of redesigning and optimizing the original task scheduling mechanism.

[0088] For example, for each production and dimension task subgraph, the set of task nodes contained in the subgraph, the real-time state vector of each task, and the list of available equipment and human resources in the subgraph are extracted. Simultaneously, the enhancement features of the corresponding task nodes in the subgraph are extracted from the actual enhancement feature vector, and the predicted weights of the hyperedges involving the tasks in the subgraph are extracted from the actual hyperedge weight prediction values. The heuristic search algorithm in step S105 is executed independently on a subgraph basis to generate a local reconstruction scheduling scheme for each subgraph. This scheme specifies the start time and resource allocation details for each task in the subgraph. After all subgraph reconstruction schemes are generated, they are merged to form a global adjusted scheduling scheme. During the merging process, the global resource capacity constraints and spatial mutual exclusion constraints are checked. If cross-subgraph resource conflicts exist, the conflicts are resolved by fine-tuning the start time of tasks at the boundaries of adjacent subgraphs, ultimately generating an adjusted collaborative scheduling scheme.

[0089] In this embodiment, during the execution of the scheduling scheme, the execution status is continuously monitored based on the continuously received unified spatiotemporal event stream. Once a deviation is detected, the production and maintenance task hypergraph is immediately reconstructed based on real-time operating conditions, and the degree of entanglement disturbance is calculated. When the degree of entanglement disturbance exceeds a preset threshold, the actual production and maintenance task hypergraph is decoupled, and a production and maintenance task subgraph is generated. Each production and maintenance task subgraph is reconstructed and scheduled separately to obtain a reconstructed subgraph scheme. The reconstructed subgraph schemes are then merged to generate an adjusted collaborative scheduling scheme. This approach can adapt to the dynamic changes in downhole operating conditions in a timely manner, ensuring that production and maintenance tasks always operate in an optimal collaborative state, significantly enhancing the robustness and adaptability of the scheduling system.

[0090] In one embodiment, such as Figure 2 As shown, a production-dimensional task hypergraph is constructed based on the task state vector, including:

[0091] Step S301: Based on the resource consumption, spatial range, and time window in the task state vector of each task, identify the relationships between tasks and construct a hyperedge for each relationship; wherein, the relationships include resource competition relationship, state dependency relationship, and spatial mutual exclusion relationship.

[0092] For example, resource occupancy dimension, spatial range dimension, and time window dimension are extracted from the task state vector of each task. The resource occupancy dimension records the equipment type, quantity, personnel type, and number of personnel required by the task, stored in the form of a resource demand vector. The spatial range dimension describes the geographical area involved in the task, represented by a coordinate interval in a three-dimensional tunnel coordinate system. The three-dimensional tunnel coordinate system is a local right-handed three-dimensional rectangular coordinate system established in mining / tunnel engineering based on the tunnel's own direction and cross-section. The time window dimension includes the task's planned start time, planned end time, and expected duration. Resource competition is identified when the resource occupancy dimensions of multiple tasks intersect, meaning they require the same type of equipment or personnel, and their time windows overlap, preventing simultaneous resource fulfillment. State dependency is identified when there is equipment association between production and maintenance tasks; for example, the equipment used by a production task is the service object of another maintenance task, and the intensity parameter of the production task affects the urgency of the maintenance task. Spatial exclusivity is identified when the spatial ranges of multiple tasks overlap or the safety distance is insufficient, and safety regulations prohibit such operations from being performed simultaneously. For example, based on these dimensional data, pairwise and multi-task joint analyses are performed on any two or more tasks to identify the types of relationships between them. For each identified relationship, a corresponding hyperedge is constructed. Here, pairwise analysis refers to examining the relationship between only two tasks at a time; multi-task joint analysis refers to modeling and analyzing all tasks together as a whole; the hypergraph model is a higher-order extension of the probabilistic model that uses a graph (nodes + edges) to represent the relationship between variables, replacing ordinary edges with hyperedges, which can connect any number of vertices simultaneously; a hyperedge is an edge structure that can connect any number of nodes, used to connect all task nodes with specific relationships together to form the basic connection skeleton of the hypergraph.

[0093] Step S302: Calculate the weight of each hyperedge based on the dynamic priority and task progress in the task state vector.

[0094] For example, different weighting methods are used for different types of hyperedges based on the dynamic priority dimension and task progress dimension in the task state vector of each task. For the weight of resource competition hyperedges: obtain all task nodes connected to the hyperedge, determine whether each task is actually occupying the competing resource at the current moment through real-time event stream, take its demand intensity value if it occupies it, otherwise take zero, and sum the demand intensity values ​​of all tasks to obtain the current total resource occupation; divide the total occupation by the total available resource to obtain the basic resource competition intensity; finally, multiply by the dynamic conflict coefficient used to reflect the urgency when multiple high-priority tasks compete for the same resource at the same time. The coefficient is calculated by adding a preset conflict sensitivity coefficient and the product of the sum of the dynamic priorities of all tasks in the current hyperedge and the ratio of the highest priority among them to the base of 1; for state dependency hyperedges, the weight is obtained by multiplying the sensitivity of the impact of the state change of the preceding task on the urgency of the subsequent task by the state change rate of the preceding task. The impact sensitivity can be determined by the equipment wear model fitted by historical data, and the state change rate is calculated from the real-time changes in the task progress dimension; for spatially mutually exclusive hyperedges, the weight is determined by the degree of overlap of the task spatial regions, and the degree of overlap is calculated by the intersection of the coordinate intervals of the spatial range dimension. The weights of all hyperedges are functions of time variables, changing in real time as the task state vector is updated, so that the hypergraph can dynamically reflect the evolution of entanglement relationships between tasks. The preset conflict sensitivity coefficient is a pre-defined quantization threshold. The equipment wear model fitted with historical data refers to a formula / curve that describes the wear pattern of equipment using historical data such as past operation, failures, maintenance, and working conditions, generated through mathematical / statistical methods.

[0095] Step S303: Generate the production dimension task hypergraph based on tasks, hyperedges, and weights.

[0096] For example, all identified task nodes, all constructed hyperedges, and the current weight of each hyperedge are integrated to generate a complete production-dimensional task hypergraph. This hypergraph can be stored in the form of an adjacency list, where the node table records the task identifier, task type, and corresponding task state vector for each task; the hyperedge table records the identifier, hyperedge type, a list of all connected task node identifiers, and the current weight value for each hyperedge. An adjacency list is a common method for representing graph data structures, describing the connections between vertices by maintaining a list of adjacent vertices for each vertex in the graph.

[0097] In this embodiment, the originally abstract and scattered task state vectors are transformed into a production-dimensional task hypergraph with a clear mathematical structure. This fundamentally solves the problem that traditional scheduling methods cannot model the dynamic coupling relationships between tasks, enabling production tasks and operation and maintenance tasks to be collaboratively optimized within a unified mathematical framework.

[0098] In one embodiment, the hypergraph of the task is input into a dynamic graph neural network, which outputs an enhanced feature vector for each task node and a sequence of predicted hyperedge weights for multiple future time steps, including:

[0099] Step S401: Based on the node feature matrix, hyperedge association matrix, and hyperedge weight vector of the task hypergraph, the hyperedge feature vector is generated through the first-stage aggregation operation of the spatial convolutional layer of the dynamic graph neural network; wherein, the first-stage aggregation operation is to aggregate the node features onto the hyperedge according to the first attention coefficient of each task node to its hyperedge.

[0100] In this system, each row of the node feature matrix corresponds to the initial feature vector of a task node, which is obtained by numerical encoding of the task state vector; the hyperedge association matrix is ​​a Boolean matrix of size equal to the number of hyperedges multiplied by the number of nodes, used to record which nodes each hyperedge is connected to; the hyperedge weight vector records the weight value of each hyperedge at the current time; the spatial convolutional layer is one of the components of the convolutional neural network in deep learning, specifically used to process data with two-dimensional or three-dimensional spatial structures; the convolutional neural network is a neural network specifically used to process grid-like data; numerical encoding refers to converting non-numerical information such as text and categories into numbers that can be processed by computers.

[0101] For example, the task hypergraph is transformed into structured input data, including node feature matrices, hyperedge association matrices, and hyperedge weight vectors. This input data is fed into the spatial convolutional layer of a dynamic graph neural network. For each hyperedge, the first attention coefficient of all task nodes connected to that hyperedge is calculated. This coefficient is calculated based on the similarity between node features and hyperedge type embedding vectors, and is normalized using a learnable attention parameter vector and a softmax function. The feature vector of each node is multiplied by its corresponding first attention coefficient and then weighted and summed to obtain the aggregated feature vector of that hyperedge. The aggregated feature vectors of all hyperedges are combined to form the hyperedge feature matrix. Among them, similarity can be calculated using cosine similarity; cosine similarity is to represent two things as vectors and calculate the cosine of the angle between them, which is used to measure the degree of similarity in a direction; the learnable attention parameter vector is a trainable parameter in the attention mechanism that is learned by the model and used to generate queries / keys / values; the softmax function is an activation function that transforms a set of real numbers into a probability distribution, with the output summing to 1 and each value between 0 and 1; normalization refers to scaling data of different magnitudes and units to the same uniform range.

[0102] Step S402: Based on the hyperedge feature vector, the aggregated node feature vector is generated through the second-stage aggregation operation of the spatial convolutional layer; wherein, the second-stage aggregation operation is to aggregate the hyperedge feature vector back to the task node according to the second attention coefficient of each hyperedge to the task node.

[0103] For example, for each task node, the second attention coefficient of all hyperedges containing that node is calculated. This coefficient is calculated based on the similarity between the hyperedge feature vector and the original feature vector of the node. It is also normalized using a learnable attention parameter vector and the Softmax function. The feature vector of each hyperedge is multiplied by its corresponding second attention coefficient and then weighted and summed to obtain the aggregated feature vector of that node. The aggregated feature vectors of all nodes are combined to form a new node feature matrix.

[0104] Step S403: Combine the aggregated node feature vectors obtained after processing through multiple spatial convolutional layers to generate the node feature vector sequence at the current time.

[0105] For example, the aggregated node feature vectors are used as the result of the first spatial convolution layer and input into the second spatial convolution layer, repeating steps S401 and S402. After stacking multiple spatial convolution layers, the feature vectors of each task node gradually incorporate information from multi-hop neighbors. After completing the spatial convolution processing of the preset number of layers, the node feature vectors output by each layer are combined in chronological order. Since the task hypergraph evolves dynamically over time, a new set of node feature vectors is generated at each time step. The node feature vectors of the current time step and multiple past time steps are stacked along the time dimension to form the node feature vector sequence of the current time step. This sequence is a three-dimensional tensor, whose dimensions correspond to the time step, task node, and feature dimension, respectively. Among them, multi-layer spatial convolution refers to stacking spatial convolution layers to extract features from the data from simple to complex; the preset number of layers refers to the number of spatial convolution layers to be stacked in advance.

[0106] Step S404: Input the current node feature vector sequence into the temporal convolutional layer of the dynamic graph neural network, output the current node hidden state vector that integrates historical information, and determine the current node hidden state vector as the enhanced feature vector.

[0107] The temporal convolutional layer is a convolutional neural network layer specifically designed for processing temporal data. It can employ an extended causal convolutional structure, characterized by the convolutional kernel sampling at exponentially increasing intervals in the time dimension, thereby expanding the receptive field without increasing the number of parameters. This temporal convolutional layer contains multiple parallel convolutional filters, each performing convolution operations on the input sequence with a different expansion rate. The expansion rate increases exponentially with the depth of the layer, allowing deeper layers to cover longer historical time windows.

[0108] For example, the current node feature vector sequence is input into the temporal convolutional layer of a dynamic graph neural network for convolution operations that follow causal constraints. That is, the output at the current time depends only on the inputs at the current and past times, and not on information from future times, to ensure the correctness of the causal relationship in the temporal processing. After processing through multiple temporal convolutional layers, each task node obtains a hidden state vector at the current time that incorporates historical evolution information. This hidden state vector is then used as the enhanced feature vector for that task node. The convolutional filter is a core learnable parameter in the convolutional neural network, typically represented as a small-sized multidimensional array (convolutional kernel), used to slide across the input data and perform convolution operations. Multiple parallel convolutional filters refer to using multiple different sets of convolutional kernels simultaneously in the same layer, each extracting different features, and then concatenating the results for output. The convolution operation involves using a small filter to slide across the data, performing a dot product and summation at each position to extract local features.

[0109] Step S405: Combine the enhanced feature vectors of all task nodes to generate the first feature matrix, and input the first feature matrix into the sequence prediction head of the dynamic graph neural network to generate a sequence of hyperedge weight prediction values ​​for multiple future time steps.

[0110] The sequence-to-sequence prediction head can be implemented using an encoder-decoder architecture. During the training phase, the prediction head is optimized using a mean squared error loss function to ensure the predicted values ​​are as close as possible to the actual observations. The encoder is a device or program that transforms information from one form into a simpler, more standard form. It can use a recurrent neural network to encode the first feature matrix of the input, converting it into a fixed-dimensional context vector. This vector integrates the global information of all task nodes at the current time step. The decoder is a program that restores the encoded data into understandable and usable original information. The decoder uses this context vector as its initial state and gradually generates the output sequence for multiple future time steps. A recurrent neural network is a neural network specifically designed for processing sequence data. Encoding refers to transforming information into another form according to a set of rules.

[0111] For example, the enhanced feature vectors of all task nodes are combined in node order to obtain a feature matrix of size equal to the number of nodes multiplied by the feature dimension. This feature matrix is ​​then input into the sequence prediction head of the dynamic graph neural network. The decoder uses this context vector as its initial state and progressively generates the output sequence for multiple future time steps. At each decoding time step, the decoder predicts the hyperedge weight vector for the next time step based on the output of the previous time step and the current hidden state. The output sequence of predicted hyperedge weights for multiple future time steps is a two-dimensional matrix, where rows correspond to future time steps and columns correspond to hyperedge identifiers.

[0112] In this embodiment, the task hypergraph is input into a dynamic graph neural network. A two-stage spatial convolution is used to generate the current node feature vector sequence. An extended causal convolutional structure with temporal convolutional layers is employed to capture the evolution of node features over time. A sequence-to-sequence prediction head generates a sequence of predicted hyperedge weights for multiple future time steps. This approach progressively abstracts the original task state information into enhanced features rich in structural and temporal information, enabling accurate prediction of the future evolution of entangled relationships between tasks.

[0113] In one embodiment, a collaborative scheduling scheme is generated based on the enhanced feature vector and the predicted sequence of hyperedge weights, combined with preset constraints and preset objectives, including:

[0114] Step S501: Based on the expected duration, resource demand vector, spatial region, dynamic priority, earliest start time and latest end time of all tasks, and combined with the inter-task correlation information in the enhanced feature vector, a mathematical model of the scheduling problem is constructed. The constraints of the mathematical model of the scheduling problem include time window constraints, resource capacity constraints, spatial mutual exclusion constraints and production-sales balance constraints. The objective function of the mathematical model of the scheduling problem consists of task delay penalty, resource idle loss and production-dimensional conflict cost.

[0115] Among these constraints, the time window constraint requires that the start time of each task must fall within a closed interval formed by subtracting the expected duration from its earliest start time and latest end time; the resource capacity constraint requires that at any given time, the total demand for a certain type of resource by all executing tasks must not exceed the total available resources of that type, including equipment such as coal mining machines and various types of personnel; the spatial mutual exclusion constraint requires that tasks with mutually exclusive relationships in a spatial region cannot be executed simultaneously, and the mutual exclusion relationships are predefined by safety procedures and stored in a spatial mutual exclusion table; the spatial mutual exclusion table is a reference table based on the theory of valence layer electron pair mutual exclusion; the production and sales balance constraint requires that the total output of production tasks match the sales plan; the task delay penalty is used to quantify the loss caused by the actual completion time of a task exceeding the planned time window; the resource idle loss is used to measure the waste caused by resources being idle while waiting for tasks; and the production and maintenance conflict cost refers to the additional cost incurred due to mutual interference between production and maintenance tasks, calculated based on the correlation information in the enhanced feature vector.

[0116] For example, the expected duration, resource demand vector, spatial region, dynamic priority, earliest start time, and latest end time of each task are extracted from the task state vector. Inter-task correlation information is extracted from the enhanced feature vector, reflecting the coupling strength between tasks in terms of resource competition, state dependency, and spatial mutual exclusion. Based on the above data, a mathematical model for the scheduling problem is constructed. The constraints of this model include time window constraints, resource capacity constraints, spatial mutual exclusion constraints, and production-demand balance constraints. The objective function of this model consists of three parts: task delay penalty, resource idle loss, and production-dimensional conflict cost.

[0117] Step S502: Based on the constraints in the mathematical model of the scheduling problem, arrange all tasks in descending order of dynamic priority to obtain the sequence of tasks to be assigned.

[0118] For example, the dynamic priority of all currently scheduled tasks is obtained. This priority is calculated based on the task's urgency, delay cost, and risk level. A higher dynamic priority value indicates a more urgent or important task. All tasks are sorted from highest to lowest dynamic priority to generate an ordered task list, i.e., a sequence of tasks to be assigned. The generation of this sequence follows the time window constraint in the constraints, ensuring that higher-priority tasks receive scheduling opportunities first, but subsequent assignments still need to meet other constraints.

[0119] Step S503: For each task in the task sequence to be assigned, generate the tentative start time and resource allocation details of all tasks based on the superedge weight prediction value sequence, task type and expected output.

[0120] For example, each task in the sequence of tasks to be assigned is traversed sequentially. For the current task, candidate start times are obtained. For each candidate start time, a comprehensive fit is calculated based on the sequence of predicted hyperedge weights, and the candidate start time with the highest comprehensive fit is selected as the provisional start time for that task. The corresponding resources are then locked from the remaining resource vector based on the resource demand vector, generating the resource allocation details for that task. After processing all tasks, the provisional start times and resource allocation details for all tasks are obtained.

[0121] Step S504: Based on each provisional start time and each resource allocation detail, obtain the complete scheduling scheme, and calculate the objective function value of the complete scheduling scheme according to the objective function in the mathematical model of the scheduling problem.

[0122] For example, the tentative start time and resource allocation details determined for each task are summarized to obtain a complete time-resource allocation table, which is a complete scheduling scheme. This scheme includes the start time, expected end time, and a list of specific equipment and personnel used for all tasks. The objective function is then called to evaluate the scheduling scheme. The calculation of the objective function requires traversing the entire scheduling cycle: the task delay penalty term is obtained by comparing the actual start time and planned time window of each task. If the actual start time is later than the latest end time minus the duration, a delay penalty is incurred; the resource idle loss term is obtained by statistically analyzing the cumulative idle time of various resources between two consecutive tasks; the production and dimension conflict cost term is based on the task correlation information in the enhanced feature vector, penalizing task pairs with strong coupling relationships within the same period. The weighted sum of the three terms yields the objective function value of the scheduling scheme. The smaller the value, the better the scheme.

[0123] Step S505: Randomly select a portion of tasks from the complete scheduling scheme, and based on the portion of tasks, update the sequence of tasks to be assigned and the current remaining resource vector to obtain the updated sequence of tasks to be assigned and the updated current remaining resource vector.

[0124] For example, from the complete scheduling scheme, a subset of tasks are randomly selected according to a preset ratio. All resources occupied by these tasks are released, and these tasks are removed from the allocated state. The number of released resources is added back to the current remaining resource vector, restoring its availability. Simultaneously, these selected tasks are returned to the task-to-be-assigned sequence, and their order in the sequence is randomly shuffled to change the priority order of subsequent allocations. Unselected tasks retain their allocated state, and their occupied resources remain locked. Finally, an updated task-to-be-assigned sequence and an updated current remaining resource vector are obtained. Here, the preset ratio refers to a pre-defined and fixed ratio.

[0125] Step S506: Based on the updated sequence of tasks to be assigned and the updated current remaining resource vector, repeatedly execute the steps of generating the provisional start time and resource allocation details to update the sequence of tasks to be assigned and the current remaining resource vector until the preset number of iterations is reached, and obtain all complete scheduling schemes and the objective function value of the complete scheduling scheme.

[0126] For example, starting with the updated sequence of tasks to be assigned and the updated remaining resource vector, steps S503 to S505 are re-executed to form an iterative loop. In each iteration, a new complete scheduling scheme is first generated and its objective function value is calculated. Then, some tasks are randomly selected again to release resources and update the sequence and resource vector. The iterative process continues until the preset maximum number of iterations is reached. The complete scheduling scheme and its objective function value generated in each iteration are recorded in the candidate scheme list, ultimately resulting in a set containing multiple different scheduling schemes, each scheme corresponding to an objective function value. The preset maximum number of iterations is a hyperparameter, set during system initialization based on the mine task size and computational resource limitations, typically ranging from several hundred to several thousand iterations to ensure sufficient search performance.

[0127] Step S507: The complete scheduling scheme with the minimum objective function value is determined as the cooperative scheduling scheme.

[0128] For example, after the iteration, the accumulated list of candidate solutions is traversed, the objective function value of each solution is compared, and the solution with the smallest value is found. This solution is determined as the final collaborative scheduling solution, and the start time and resource allocation details of each task are extracted from the solution and formatted into a standard instruction set. This instruction set is used to instruct each execution terminal to perform actual production and maintenance operations. At the same time, the core data of this collaborative scheduling solution, namely the start time of each task and the sequence of predicted hyperedge weights used in the generation process, are persistently stored. Persistent storage means storing the data in a location that will not be lost even after a power outage, preserving it long-term; formatting means organizing the messy, inconsistent, and unstructured content into a unified, standardized instruction format that can be directly executed by machines / humans.

[0129] In this embodiment, the complex collaborative scheduling problem of production and maintenance tasks is transformed into a mathematical model with explicit constraints and an objective function. Dynamic priority sorting is used to obtain a sequence of tasks to be assigned. A provisional start time and resource allocation details are generated for each task in the sequence, resulting in a complete scheduling scheme. The objective function value of this scheme is then calculated based on the objective function. Resources for some tasks are randomly released from the complete scheduling scheme. The steps of generating provisional start times and resource allocation details, updating the sequence of tasks to be assigned, and the current remaining resource vector are repeated until a preset number of iterations is reached, resulting in all complete scheduling schemes and their corresponding objective function values. The complete scheduling scheme with the smallest objective function value is determined as the collaborative scheduling scheme. This approach ensures the global optimality of the scheduling scheme and possesses computational feasibility for practical engineering applications.

[0130] In one embodiment, for each task in the sequence of tasks to be assigned, based on the sequence of predicted superedge weights, task type, and expected output, a provisional start time and resource allocation details for all tasks are generated, including:

[0131] Step S601: Obtain the current task sequentially from the task sequence to be assigned, and generate multiple candidate start times for the current task within the allowed time window.

[0132] For example, tasks to be processed are obtained sequentially according to the order in the task sequence. For each task, the earliest start time and latest end time are extracted from its task state vector. These two time parameters together constitute the time window during which the task is allowed to execute. The earliest start time is usually determined by material preparation, the completion time of the previous process, or the planned start time, while the latest end time is determined by the requirements of downstream processes or the delivery date constraints of sales contracts. Using this time window as the feasible region, a series of candidate start times are generated according to a preset sampling step size. The preset sampling step size can be flexibly set according to the scheduling accuracy requirements.

[0133] Step S602: For each candidate start time, calculate the fitness impact factor based on the current task's resource requirement vector, the current task's task type, the current task's expected output, the current remaining resource vector, and the hyperedge weight prediction value sequence.

[0134] For example, each candidate start time generated for the current task is traversed, and the fitness impact factor is calculated based on the resource requirement vector of the current task, the task type of the current task, the expected output of the current task, the current remaining resource vector, and the sequence of predicted hyperedge weights.

[0135] Step S603: The fitness influencing factors are weighted and combined to obtain the comprehensive fitness of each candidate start time.

[0136] For example, the fit influencing factors are weighted and summed according to preset weight coefficients to obtain the comprehensive fit of each candidate start time. The setting of the weight coefficients reflects the mine manager's preference for different scheduling objectives. For example, if more emphasis is placed on resource utilization, the weight of the resource availability factor is increased.

[0137] Step S604: The candidate start time with the highest overall fit is determined as the provisional start time of the current task, and the resource allocation details of the current task are generated based on the resource requirement vector and the current remaining resource vector of the current task.

[0138] For example, the overall fit of all candidate start times is compared, and the candidate time with the largest value is determined as the provisional start time of the current task. Based on the resource requirement vector of the current task, specific resource instances are allocated to the task from the current remaining resource vector according to the proximity principle: for equipment resources, the equipment closest to the task space area and currently idle is selected; for human resources, personnel with the corresponding skills and currently idle are selected. During the allocation process, the specific time period for each resource to be occupied needs to be recorded, i.e., the interval from the provisional start time to the provisional start time plus the expected duration. After allocation, the current remaining resource vector is updated, and the number of occupied resources is deducted from the remaining amount. Simultaneously, a resource allocation detail for the current task is generated, including the task identifier, provisional start time, expected end time, list and number of occupied equipment, and list and number of occupied personnel.

[0139] Step S605: After all tasks in the task sequence to be assigned have been processed, the provisional start time of all tasks and the resource allocation details of the tasks are obtained.

[0140] For example, steps S601 to S604 are repeated to process each task in the task sequence to be assigned in turn until the sequence is empty. In each round of processing, the current remaining resource vector is updated as new tasks are locked for resources, and the candidate start time evaluation of subsequent tasks is based on the updated resource availability status, thereby ensuring that resource capacity constraints are always satisfied throughout the scheduling process. After all tasks have been processed, the provisional start time of all tasks and the resource allocation details of the tasks are obtained.

[0141] In this embodiment, the complex multi-task scheduling problem is decomposed into a series of sequentially solved individual task decision-making processes. Each task's decision comprehensively considers resource availability, future entanglement risks, and production-sales balance requirements. Through operations such as candidate start time generation, multi-dimensional influencing factor calculation, weighted combination comparison, and resource locking, the provisional start times and resource allocation details for all tasks are obtained. This enables the current task's decision to proactively avoid potential future production-sales conflicts, while real-time updates to the resource allocation details ensure that resource competition between tasks is properly handled.

[0142] In one embodiment, the fitness impact factor includes a resource availability factor, an entanglement risk factor, and a production-sales balance impact factor. For each candidate start time, the fitness impact factor is calculated based on the current task's resource demand vector, the current task's task type, the current task's expected output, the current remaining resource vector, and the hyperedge weight prediction value sequence, including:

[0143] Step S701: For each candidate start time, calculate the resource availability factor based on the current task's resource demand vector and the current remaining resource vector.

[0144] For example, the resource requirement vector for the current task is extracted, which records the specific quantities of various types of equipment and personnel required by the task; the current remaining resource vector is extracted, which reflects the available quantities of various types of resources at the current moment. By comparing the resource requirement vector and the current remaining resource vector, a resource availability factor is calculated: if, at the candidate start time and throughout the entire task duration, the required resources are not occupied by other tasks and are in sufficient quantities, the resource availability factor is set to 1; otherwise, it is set to 0.

[0145] Step S702: For each candidate start time, calculate the entanglement risk factor that will be generated by executing the task at the candidate start time and other tasks based on the superedge weight prediction value sequence.

[0146] For example, from the sequence of predicted superedge weights generated in step S405, all predicted superedge weights related to the current task are extracted from the candidate start time to the end of the expected duration of the task. These predicted values ​​are then summed to obtain the entanglement risk factor.

[0147] Step S703: For each candidate start time, calculate the production and sales balance impact factor based on the task type and the expected output of the current task.

[0148] For example, based on the task type and expected output of the current task, combined with the cumulative output of the assigned tasks and the production-sales balance target, the production-sales balance impact factor is calculated: if the current task is a production task and the expected output helps to make up for the output gap, the factor takes a positive value; if it will lead to an overproduction, the factor takes a negative value.

[0149] In this embodiment, for each candidate start time, a fitness influence factor is calculated based on the current task's resource requirement vector, task type, expected output, remaining resource vector, and predicted hyperedge weights. This provides a comprehensive and multi-dimensional evaluation system for candidate start times, offering a scientific and quantitative basis for subsequent weighted combination and optimal selection.

[0150] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0151] Based on the same inventive concept, this application also provides a mine computing task collaborative scheduling system for implementing the aforementioned mine computing task collaborative scheduling method. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more mine computing task collaborative scheduling system embodiments provided below can be found in the limitations of the mine computing task collaborative scheduling method described above, and will not be repeated here.

[0152] In one exemplary embodiment, such as Figure 3 As shown, a mine computing task collaborative scheduling system 300 is provided, comprising:

[0153] The data acquisition module 301 is used to acquire multimodal data from the mine and perform spatiotemporal alignment processing on the multimodal data to obtain a unified spatiotemporal event stream.

[0154] The vector construction module 302 is used to obtain the currently executing or planned tasks based on the unified spatiotemporal event stream, and to construct a task state vector for each task; wherein, tasks include production tasks and operation and maintenance tasks;

[0155] Graph structure construction module 303 is used to construct a multidimensional task hypergraph based on the task state vector;

[0156] The sequence generation module 304 is used to input the production task hypergraph into the dynamic graph neural network and output the enhanced feature vector of each task node and the sequence of predicted hyperedge weights for multiple future time steps.

[0157] The scheme generation module 305 is used to generate a collaborative scheduling scheme based on the enhanced feature vector and the predicted value sequence of the hyperedge weight, combined with preset constraints and preset objectives.

[0158] In one exemplary embodiment, the system further includes:

[0159] The data detection module is used to continuously receive a unified spatiotemporal event stream during the execution of the collaborative scheduling scheme, and based on the continuously received unified spatiotemporal event stream, detect whether there is a deviation between the actual execution status and the collaborative scheduling scheme, and obtain the detection results.

[0160] The data update module is used to re-execute the steps of constructing the task state vector to generate the enhanced feature vector and the hyperedge weight prediction value sequence based on the real-time task execution status if the detection result is biased, so as to obtain the actual production dimension task hypergraph and the actual hyperedge weight prediction value.

[0161] The data calculation module is used to calculate the degree of entanglement disturbance caused by the deviation based on the actual predicted value of the hyperedge weight and the planned predicted value in the collaborative scheduling scheme.

[0162] The subgraph generation module is used to perform decoupling operations on the actual production and maintenance task hypergraph and generate a production and maintenance task subgraph when the degree of entanglement perturbation exceeds a preset threshold.

[0163] The scheme adjustment module is used to reconstruct and schedule each production and maintenance task subgraph separately, obtain the reconstructed subgraph scheme, and merge the reconstructed subgraph schemes to generate the adjusted collaborative scheduling scheme.

[0164] In one embodiment, the graph structure construction module 303 is further configured to:

[0165] Based on the resource consumption, spatial range, and time window in the task state vector of each task, the relationships between tasks are identified, and a hyperedge is constructed for each relationship; among them, the relationships include resource competition, state dependency, and spatial mutual exclusion.

[0166] The weight of each hyperedge is calculated based on the dynamic priority and task progress in the task state vector.

[0167] Generate a production-dimensional task hypergraph based on tasks, hyperedges, and weights.

[0168] In one embodiment, the sequence generation module 304 is further configured to:

[0169] Based on the node feature matrix, hyperedge association matrix, and hyperedge weight vector of the task hypergraph, the hyperedge feature vector is generated through the first-stage aggregation operation of the spatial convolutional layer of the dynamic graph neural network; wherein, the first-stage aggregation operation is to aggregate the node features onto the hyperedge according to the first attention coefficient of each task node to its hyperedge.

[0170] Based on the hyperedge feature vector, the aggregated node feature vector is generated through the second-stage aggregation operation of the spatial convolutional layer; wherein, the second-stage aggregation operation is to aggregate the hyperedge feature vector back to the task node according to the second attention coefficient of each hyperedge to the task node.

[0171] The aggregated node feature vectors obtained after processing through multiple spatial convolutional layers are combined to generate the node feature vector sequence at the current time.

[0172] The current node feature vector sequence is input into the temporal convolutional layer of the dynamic graph neural network, and the output is the current node hidden state vector that integrates historical information. The current node hidden state vector is then determined as the enhanced feature vector.

[0173] The enhanced feature vectors of all task nodes are combined to generate the first feature matrix. The first feature matrix is ​​then input into the sequence prediction head of the dynamic graph neural network to generate a sequence of hyperedge weight prediction values ​​for multiple future time steps.

[0174] In one embodiment, the scheme generation module 305 is further configured to:

[0175] Based on the expected duration, resource demand vector, spatial region, dynamic priority, earliest start time and latest end time of all tasks, and combined with the inter-task correlation information in the enhanced feature vector, a mathematical model of the scheduling problem is constructed. The constraints of the mathematical model of the scheduling problem include time window constraints, resource capacity constraints, spatial mutual exclusion constraints and production-sales balance constraints. The objective function of the mathematical model of the scheduling problem consists of task delay penalty, resource idle loss and production-dimensional conflict cost.

[0176] Based on the constraints in the mathematical model of the scheduling problem, all tasks are arranged in descending order of dynamic priority to obtain the sequence of tasks to be assigned;

[0177] For each task in the sequence of tasks to be assigned, based on the sequence of predicted superedge weights, task type, and expected output, generate the tentative start time and resource allocation details for all tasks.

[0178] Based on each provisional start time and each resource allocation detail, a complete scheduling scheme is obtained, and the objective function value of the complete scheduling scheme is calculated according to the objective function in the mathematical model of the scheduling problem.

[0179] Randomly select a subset of tasks from the complete scheduling scheme, and based on the subset of tasks, update the sequence of tasks to be assigned and the current remaining resource vector to obtain the updated sequence of tasks to be assigned and the updated current remaining resource vector.

[0180] Based on the updated sequence of tasks to be assigned and the updated current remaining resource vector, the steps of generating the provisional start time and resource allocation details and updating the sequence of tasks to be assigned and the current remaining resource vector are repeated until the preset number of iterations is reached, and all complete scheduling schemes and the objective function value of the complete scheduling schemes are obtained.

[0181] The complete scheduling scheme that minimizes the objective function value is determined as the cooperative scheduling scheme.

[0182] In one embodiment, the scheme generation module 305 is further configured to:

[0183] The current task is retrieved sequentially from the sequence of tasks to be assigned, and multiple candidate start times are generated for the current task within the allowed time window;

[0184] For each candidate start time, the fitness impact factor is calculated based on the current task's resource requirement vector, the current task's task type, the current task's expected output, the current remaining resource vector, and the hyperedge weight prediction value sequence.

[0185] The fitness factor is weighted and combined to obtain the overall fitness of each candidate start time;

[0186] The candidate start time with the highest overall fit is determined as the provisional start time of the current task, and a resource allocation detail for the current task is generated based on the resource requirement vector and the current remaining resource vector of the current task.

[0187] Once all tasks in the task sequence have been processed, the tentative start time and resource allocation details for all tasks are obtained.

[0188] In one embodiment, the scheme generation module 305 is further configured to:

[0189] For each candidate start time, the resource availability factor is calculated based on the resource demand vector of the current task and the current remaining resource vector.

[0190] For each candidate start time, the entanglement risk factor between the task executed at the candidate start time and other tasks is calculated based on the superedge weight prediction value sequence.

[0191] For each candidate start time, the production and sales balance impact factor is calculated based on the task type and the expected output of the current task.

[0192] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the previously described mine computing task collaborative scheduling method.

[0193] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0194] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0195] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A collaborative scheduling method for mining computing tasks, characterized in that, The method includes: Multimodal data from underground mines is acquired, and the multimodal data is spatiotemporally aligned to obtain a unified spatiotemporal event stream. Based on the unified spatiotemporal event stream, the tasks currently being executed or planned to be executed are obtained, and a task state vector is constructed for each task; wherein, the tasks include production tasks and operation and maintenance tasks; Based on the task state vector, construct a production-dimensional task hypergraph; The hypergraph of the production task is input into a dynamic graph neural network, which outputs the enhanced feature vector of each task node and the sequence of predicted hyperedge weights for multiple future time steps. Based on the enhanced feature vector and the predicted sequence of hyperedge weights, a collaborative scheduling scheme is generated by combining the collaborative scheduling constraints and the collaborative scheduling objective.

2. The method according to claim 1, characterized in that, The method further includes: During the execution of the collaborative scheduling scheme, the unified spatiotemporal event stream is continuously received, and based on the continuously received unified spatiotemporal event stream, the deviation between the actual execution state and the collaborative scheduling scheme is detected, and the detection result is obtained. If the detection result indicates the existence of the deviation, based on the real-time task execution status, the step of re-executing the task state vector to generate the enhanced feature vector and the hyperedge weight prediction value sequence is re-executed to obtain the actual production dimension task hypergraph and the actual hyperedge weight prediction value. Based on the actual predicted value of the hyperedge weight and the planned predicted value in the cooperative scheduling scheme, the degree of entanglement disturbance caused by the deviation is calculated. When the degree of entanglement disturbance exceeds a preset threshold, a decoupling operation is performed on the actual production and maintenance task hypergraph to generate a production and maintenance task subgraph. Each of the production and maintenance task subgraphs is reconstructed and scheduled separately to obtain a reconstructed subgraph scheme. The reconstructed subgraph schemes are then merged to generate an adjusted collaborative scheduling scheme.

3. The method according to claim 1, characterized in that, The construction of the production-dimensional task hypergraph based on the task state vector includes: Based on the resource consumption, spatial range, and time window in the task state vector of each task, the relationships between the tasks are identified, and a hyperedge is constructed for each relationship; wherein, the relationships include resource competition, state dependency, and spatial mutual exclusion. The weight of each hyperedge is calculated based on the dynamic priority and the task progress in the task state vector. Based on the task, the hyperedge, and the weight, the product dimension task hypergraph is generated.

4. The method according to claim 1, characterized in that, The step of inputting the hypergraph of the production task into a dynamic graph neural network and outputting the enhanced feature vector of each task node and the sequence of predicted hyperedge weights for multiple future time steps includes: Based on the node feature matrix, hyperedge association matrix, and hyperedge weight vector of the aforementioned task hypergraph, a hyperedge feature vector is generated through the first-stage aggregation operation of the spatial convolutional layer of the dynamic graph neural network; wherein, the first-stage aggregation operation is to aggregate the node features onto the hyperedge according to the first attention coefficient of each task node to its respective hyperedge. Based on the hyperedge feature vector, the aggregated node feature vector is generated through the second-stage aggregation operation of the spatial convolutional layer; wherein, the second-stage aggregation operation is to aggregate the hyperedge feature vector back onto the task node according to the second attention coefficient of each hyperedge pair on the task node; The aggregated node feature vectors obtained after processing through multiple spatial convolutional layers are combined to generate the node feature vector sequence at the current time. The current node feature vector sequence is input into the temporal convolutional layer of the dynamic graph neural network, and the current node hidden state vector with historical information is output. The current node hidden state vector is then determined as the enhanced feature vector. The enhanced feature vectors of all the task nodes are combined to generate a first feature matrix, and the first feature matrix is ​​input into the sequence prediction head of the dynamic graph neural network to generate the sequence of predicted superedge weights for multiple future time steps.

5. The method according to claim 1, characterized in that, The process of generating a collaborative scheduling scheme based on the enhanced feature vector and the predicted sequence of hyperedge weights, combined with preset constraints and preset objectives, includes: Based on the expected duration, resource demand vector, spatial region, dynamic priority, earliest start time, and latest end time of all the tasks, and combined with the inter-task correlation information in the enhanced feature vector, a mathematical model for the scheduling problem is constructed. The constraints of the mathematical model for the scheduling problem include time window constraints, resource capacity constraints, spatial mutual exclusion constraints, and production-sales balance constraints. The objective function of the mathematical model for the scheduling problem consists of task delay penalties, resource idle losses, and production-dimensional conflict costs. Based on the constraints in the mathematical model of the scheduling problem, all tasks are arranged in descending order of dynamic priority to obtain a sequence of tasks to be assigned. For each task in the sequence of tasks to be assigned, based on the sequence of predicted superedge weights, task type, and expected output, a provisional start time and resource allocation details for all tasks are generated. Based on the provisional start times and resource allocation details, a complete scheduling scheme is obtained, and the objective function value of the complete scheduling scheme is calculated according to the objective function in the mathematical model of the scheduling problem. Randomly select a portion of tasks from the complete scheduling scheme, and update the task sequence to be assigned and the current remaining resource vector based on the selected tasks to obtain the updated task sequence to be assigned and the updated current remaining resource vector. Based on the updated sequence of tasks to be assigned and the updated current remaining resource vector, the steps of generating the provisional start time and the resource allocation details and updating the sequence of tasks to be assigned and the current remaining resource vector are repeated until a preset number of iterations is reached, and all the complete scheduling schemes and the objective function value of the complete scheduling schemes are obtained. The complete scheduling scheme that minimizes the objective function value is determined as the cooperative scheduling scheme.

6. The method according to claim 5, characterized in that, For each task in the sequence of tasks to be assigned, based on the superedge weight prediction sequence, task type, and expected output, a provisional start time and resource allocation details for all tasks are generated, including: The current task is obtained sequentially from the sequence of tasks to be assigned, and multiple candidate start times are generated for the current task within the allowed time window; For each candidate start time, the fitness impact factor is calculated based on the resource requirement vector of the current task, the task type of the current task, the expected output of the current task, the current remaining resource vector, and the hyperedge weight prediction value sequence. The fitness influencing factors are weighted and combined to obtain the comprehensive fitness of each candidate start time; The candidate start time with the highest overall fit is determined as the provisional start time of the current task, and a resource allocation detail for the current task is generated based on the resource requirement vector and the current remaining resource vector of the current task. Once all tasks in the sequence of tasks to be assigned have been processed, the provisional start time of all tasks and the resource allocation details of the tasks are obtained.

7. The method according to claim 6, characterized in that, The fitness influencing factors include resource availability factors, entanglement risk factors, and production-sales balance influencing factors. For each candidate start time, the fitness influencing factors are calculated based on the resource demand vector of the current task, the task type of the current task, the expected output of the current task, the current remaining resource vector, and the predicted value sequence of the hyperedge weights, including: For each candidate start time, the resource availability factor is calculated based on the resource demand vector of the current task and the current remaining resource vector. For each candidate start time, the entanglement risk factor that will be generated by the task and other tasks when the task is executed at the candidate start time is calculated based on the superedge weight prediction value sequence. For each candidate start time, the production-sales balance impact factor is calculated based on the task type of the current task and the expected output of the current task.

8. A collaborative scheduling system for mining computing tasks, characterized in that, The system includes: The data acquisition module is used to acquire multimodal data from underground mines and perform spatiotemporal alignment processing on the multimodal data to obtain a unified spatiotemporal event stream. The vector construction module is used to obtain the currently executing or planned tasks based on the unified spatiotemporal event stream, and construct a task state vector for each task; wherein, the tasks include production tasks and operation and maintenance tasks; The graph structure construction module is used to construct a multidimensional task hypergraph based on the task state vector. The sequence generation module is used to input the production task hypergraph into a dynamic graph neural network and output the enhanced feature vector of each task node and the sequence of predicted hyperedge weights for multiple future time steps. The scheme generation module is used to generate a collaborative scheduling scheme based on the enhanced feature vector and the predicted value sequence of the hyperedge weights, combined with preset constraints and preset objectives.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images