A Dynamic Control Method for Mines Based on System Instability Potential Energy and Collaborative Autonomy

Abstract

This application provides a dynamic control method for mines based on system instability potential energy and collaborative autonomy, relating to the field of intelligent control and automation technology for mine production. The method includes: collecting multi-source heterogeneous data from the mine production system, including equipment status data, environmental parameters, and production task flow data, and constructing a dynamic relationship graph, where nodes represent mine equipment, production areas, or environmental monitoring points, and edges represent the interaction relationships between nodes; based on the dynamic relationship graph, using a graph neural network to calculate the state and local instability potential energy of each node; based on the local instability potential energy of all nodes, generating a system-level instability potential energy field and predicting the future potential energy field to identify high-risk areas; based on the system-level instability potential energy field and gradient information, controlling an intelligent agent to perform autonomous control through a distributed decision-making approach to reduce the risk of system instability; after executing the control, collecting new data, updating the dynamic relationship graph and potential energy field, and fine-tuning the model by comparing predictions with actual results.

Description

Technical Field

[0001] This application belongs to the field of intelligent control and automation technology of mine production, specifically involving a dynamic control method for mines based on system instability potential energy and collaborative autonomy. Background Technology

[0002] In the field of mine production management, existing technologies generally employ production management systems based on data acquisition, simulation modeling, and centralized scheduling. For example, Chinese invention patent (CN 117371925 A) discloses a production management method that monitors and optimizes production line equipment, quality, energy consumption, and other factors by establishing a database, creating a simulation system, building a shared platform, and conducting operational analysis. However, this type of method is essentially passive and reactive; its optimization decisions are triggered only when the probability of a fault exceeds a threshold or data anomalies are detected, resulting in the following fundamental flaws:

[0003] The system only intervenes when the probability of equipment failure increases significantly or anomalies occur. By this time, the equipment may have already entered an irreversible stage of degradation, leading to high adjustment costs and increased risks of unplanned downtime. The optimization objectives of subsystems such as fault prediction, production scheduling, and safety monitoring are independent of each other, lacking a unified global optimization index, which easily leads to suboptimal decisions, such as sacrificing equipment health or safety to improve efficiency. Existing methods cannot perceive the "tight" or "loose" state of the entire mine production system (including equipment, environment, and task flow) at a macro level, and therefore cannot provide proactive and flexible intervention before the system becomes unstable, resulting in insufficient resilience of the system when faced with internal disturbances or external shocks.

[0004] Therefore, the aforementioned problems severely limit the level of intelligence and economic efficiency of mining production systems. Summary of the Invention

[0005] This application provides a dynamic control method for mines based on system instability potential energy and collaborative autonomy to solve one of the aforementioned technical problems.

[0006] The technical solution adopted in this application is as follows:

[0007] This application provides a mine dynamic control method based on system instability potential energy and collaborative autonomy, including:

[0008] Collect multi-source heterogeneous data from the mining production system, including equipment status data, environmental parameters, and production task flow data, and construct a dynamic relationship graph, where nodes represent mining equipment, production areas, or environmental monitoring points, and edges represent the interaction relationships between nodes.

[0009] Based on the dynamic relationship graph, a graph neural network is used to calculate the state and local instability potential energy of each node.

[0010] Based on the local instability potential energy of all nodes, a system-level instability potential energy field is generated, and the future potential energy field is predicted to identify high-risk areas.

[0011] Based on the system-level instability potential energy field and gradient information, the intelligent agent is controlled to perform autonomous regulation through a distributed decision-making approach to reduce the risk of system instability.

[0012] After implementing regulation, new data is collected to update the dynamic relationship diagram and potential energy field, and the model is fine-tuned by comparing the prediction with the actual situation.

[0013] According to one embodiment of this application, constructing the dynamic relationship graph includes:

[0014] The weights of edges are dynamically updated by analyzing the correlations of node state changes in historical data.

[0015] According to one embodiment of this application, the calculation of node states and local instability potential energy using a graph neural network includes:

[0016] Use graph convolutional networks to aggregate neighbor node information to update node state;

[0017] The local instability potential energy is calculated based on the node's own state anomaly degree, the risk transmission of neighboring nodes, and local environmental stress.

[0018] According to one embodiment of this application, the generation of the system-level instability potential energy field includes:

[0019] Local potential energy is mapped into a continuous potential energy field using spatial interpolation techniques;

[0020] The physical information neural network is used to predict the future potential energy field and identify potential well regions where potential energy rapidly accumulates above a critical threshold.

[0021] According to one embodiment of this application, controlling the intelligent agent for autonomous regulation through distributed decision-making includes:

[0022] Publish the current potential energy field and gradient information to the intelligent agent;

[0023] Intelligent agents autonomously adjust their behavior based on the potential energy field, including path planning, task reallocation, or maintenance of pre-set resources.

[0024] According to one embodiment of this application, in the path planning, the cost function includes a potential energy field integral, which causes the vehicle to bypass the potential energy peak region.

[0025] According to one embodiment of this application, the fine-tuning model includes:

[0026] The difference between the predicted potential field and the actual observed potential field is used as the loss function, and backpropagated to the graph neural network and the physical information neural network for online fine-tuning.

[0027] A second aspect of this application provides a dynamic intelligent control system for mines based on system instability potential energy and collaborative autonomy, comprising:

[0028] The dynamic relationship graph construction module is used to collect multi-source heterogeneous data and construct dynamic relationship graphs;

[0029] The node state and potential energy calculation module is used to calculate the node state and local instability potential energy based on the dynamic relationship graph using a graph neural network.

[0030] The potential energy field generation and prediction module is used to generate system-level unstable potential energy fields and predict future potential energy fields to identify high-risk areas.

[0031] The distributed control module is used to enable autonomous control of intelligent agents through distributed decision-making based on potential energy field and gradient information.

[0032] The model update module is used to collect new data after the regulation is performed, update the dynamic relationship diagram and potential energy field, and fine-tune the model.

[0033] According to one embodiment of this application, the dynamic relationship graph construction module is configured to dynamically update the weights of edges by analyzing the correlation of node state changes in historical data;

[0034] The node state and potential energy calculation module is configured to use a graph convolutional network to aggregate neighbor node information to update the node state, and calculate the local instability potential energy based on the node's own state anomaly degree, the risk transmission of neighbor nodes, and local environmental stress.

[0035] According to one embodiment of this application, the potential energy field generation and prediction module is configured to generate a continuous potential energy field through spatial interpolation technology and use a physical information neural network to predict the future potential energy field to identify potential well regions.

[0036] The distributed control module is configured to publish potential energy field and gradient information to the intelligent agent, and the intelligent agent autonomously adjusts its behavior based on the potential energy field, including path planning, task reallocation, or maintenance resource pre-setting.

[0037] In the path planning, the cost function includes the potential energy field integral, which enables the vehicle to bypass the potential energy peak region.

[0038] The model update module is configured to use the difference between the predicted potential energy field and the actual observed potential energy field as a loss function, and backpropagate it to the graph neural network and the physical information neural network for online fine-tuning.

[0039] Due to the adoption of the above technical solution, the beneficial effects achieved by this application are as follows:

[0040] This application, by collecting multi-source heterogeneous data and constructing a dynamic relationship graph, and using graph neural networks to calculate node states and local instability potential energy, enables the system to explicitly model the complex relationships between devices, the environment, and task flows. Based on this, a system-level instability potential energy field is generated and its future evolution is predicted. This allows the system to identify systemic instability trends (such as potential well regions) before the probability of device failure increases significantly, thus enabling proactive and flexible control. This effectively avoids the problems of unplanned downtime and high intervention costs caused by decision-making lags in existing technologies.

[0041] Based on system-level instability potential energy field and gradient information, this approach uses a distributed decision-making method to control intelligent agents (such as unmanned transport vehicles and plant controllers) for autonomous regulation. With the unified goal of "flattening the potential energy field," it naturally achieves synergistic optimization of multiple objectives, including safety, efficiency, and equipment health. Compared to existing technologies that rely on centralized scheduling and manually set target weights, this solution guides autonomous behavior through the potential energy field gradient, automatically tending towards reducing overall system risk and avoiding suboptimal decisions that create a "whack-a-mole" effect.

[0042] The distributed control mechanism enables intelligent agents to autonomously adjust their behavior (such as path planning and task reallocation) based on local potential field information. When local disturbances occur, the potential field gradient guides other agents to automatically avoid or compensate, forming a response mechanism similar to an "immune system." This collaborative autonomy allows the system to maintain overall functional stability when faced with internal equipment failures or external environmental shocks, significantly improving the system's resilience.

[0043] By collecting new data after implementing interventions, updating the dynamic relationship diagram and potential energy field, and using the difference between the predicted and observed potential energy fields as a loss function to fine-tune the model (as shown in graph neural networks and physical information neural networks), the system can learn from each intervention and continuously improve its understanding and prediction capabilities regarding system dynamics. This deep feedback loop enables the core model of the system to continuously evolve, resulting in increasingly accurate and intelligent decision-making, achieving a qualitative leap in capabilities. Attached Figure Description

[0044] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0045] Figure 1 A flowchart illustrating a dynamic control method for mines based on system instability potential energy and collaborative autonomy, provided for an embodiment of this application;

[0046] Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0047] Figure label:

[0048] 810, Processor; 820, Communication interface; 830, Memory; 840, Communication bus. Detailed Implementation

[0049] To more clearly illustrate the overall concept of this application, a detailed explanation is provided below with reference to the accompanying drawings.

[0050] Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of this application is not limited to the specific embodiments disclosed below. It should be noted that, unless otherwise specified, the embodiments of this application and the features thereof can be combined with each other.

[0051] In this application, unless otherwise expressly specified and limited, the "above" or "below" of the second feature can mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. In the description of this specification, references to terms such as "an embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples.

[0052] Example 1

[0053] like Figure 1 As shown, a dynamic control method for mines based on system instability potential energy and collaborative autonomy includes:

[0054] S100. Collect multi-source heterogeneous data from the mining production system, including equipment status data, environmental parameters, and production task flow data, and construct a dynamic relationship diagram, where nodes represent mining equipment, production areas, or environmental monitoring points, and edges represent the interaction relationships between nodes.

[0055] As mentioned above, this step is a fundamental component of the dynamic intelligent control method for mines. Its core lies in explicitly modeling the complex internal relationships within the mine production system by integrating multi-source heterogeneous data and constructing a dynamic relationship graph. Specifically, multi-source heterogeneous data refers to real-time or historical data collected from sensors or systems of different physical sources, formats, and semantics. This includes equipment status data (such as physical parameters during equipment operation), environmental parameters (such as external natural or operational environmental indicators), and production task flow data (such as production plans and execution processes). After collection, this data requires spatiotemporal alignment and preprocessing to ensure consistency and usability. Constructing the dynamic relationship graph abstracts the mine system into a graph structure, where nodes represent physical or logical entities (such as mine equipment, production areas, or environmental monitoring points), and edges represent the interactions between these entities (such as physical connections, logical dependencies, or environmental influences). This process not only achieves structured data integration but also provides input for subsequent graph neural network-based analysis, thereby overcoming the shortcomings of data isolation and lack of correlation in existing technologies.

[0056] For example, in practical mining applications, equipment status data may include the vibration amplitude of the hoist, the current load of the crusher, or the operating speed of the conveyor belt; environmental parameters may cover the temperature, humidity, dust concentration, or gas content of the mining face; and production task flow data may involve the execution progress of the production plan, material flow rate, or task priority queue. When constructing a dynamic relationship graph, nodes can be specified as entities such as hoists, crushers, mining faces, or ventilation monitoring points, while edges can represent the conveyor belt connection (physical connection) between the hoist and the crusher, the process dependency (logical dependency) between the crusher and the concentrator, or the impact of environmental monitoring points on the operating status of adjacent equipment (environmental impact). For example, if historical data shows a strong correlation between temperature increases in a certain area and the failure rate of adjacent equipment, the weight of the corresponding edge can be dynamically adjusted through mutual information analysis to reflect this risk transmission relationship.

[0057] It should be noted that, in specific implementation scenarios, the acquisition of multi-source heterogeneous data can be expanded to include real-time image or audio data, and data quality can be improved through data cleaning and normalization. The construction of dynamic relationship graphs can incorporate edge weight update mechanisms based on time series analysis, such as calculating the correlation coefficient or information entropy of node state changes using a sliding window, to adaptively reflect the dynamic evolution of the system. Furthermore, the interaction relationships of edges can be extended to dimensions including energy flow, information transmission, or risk diffusion. For example, by introducing virtual nodes to represent logical gateways in the production task flow, the graph model's ability to characterize complex production processes can be enhanced.

[0058] S200. Based on the dynamic relationship graph, a graph neural network is used to calculate the state and local instability potential energy of each node.

[0059] As described above, this step is the core computational component of the method. Its purpose is to extract and fuse the node's own state information and the influence information of its neighboring nodes from the dynamic relationship graph using graph neural networks, an advanced graph structure data processing model. This enables a deep perception of the node's state and a quantitative assessment of its instability risk. Specifically, the graph neural network operates through a mechanism called "message passing" or "neighborhood aggregation": each node receives state information from its directly connected neighboring nodes and fuses this external information with its own state information to generate an updated node state representation that more comprehensively reflects its role and status in the system. This process is carried out in a hierarchical and iterative manner, allowing nodes to indirectly perceive the influence of nodes beyond multiple hops, thereby capturing long-range dependencies within the system. Based on the updated node state representation, the system further calculates the "local instability potential energy" of each node. This potential energy is a comprehensive quantitative indicator. It is not a single physical parameter, but a function output value that integrates the abnormality of the node's own health state, the risk transmission effect from neighboring nodes, and the degree of stress in its local environment. It is used to characterize the trend and potential possibility of the node evolving from its current stable state to an unstable state (such as failure or performance degradation).

[0060] For example, consider the "crusher" node in a specific mining production scenario. In the dynamic relationship graph, this node is connected to nodes such as the "upstream feeder," the "downstream conveyor belt," and the "temperature sensor of the surrounding area" via edges. When the graph neural network begins calculation, the crusher node first receives state information from these neighboring nodes, such as the current load data of the feeder, the operating speed of the conveyor belt, and the ambient temperature. Through multi-layer aggregation, the new state representation of the crusher node not only includes its own vibration and current data but also embeds related logic such as "excessive upstream feed may cause blockage," "slower downstream conveyor belt speed means poor material discharge," and "high ambient temperature may affect heat dissipation." Based on this fused, context-rich state representation, the system calculates the local instability potential energy of the crusher: its own excessive vibration contributes to the basic outlier value; the high load of the upstream feeder is amplified by the edge weights and superimposed as a risk transmission; at the same time, the high temperature environment is also included in the calculation as a stress factor. Ultimately, a high local instability potential energy value indicates that the crusher is in a high-risk state caused by both internal and external factors.

[0061] It should be noted that, in specific implementation scenarios, based on the above scheme, the aggregation function for node state updates can be, but is not limited to, weighted summation, maximum maximization, or attention mechanisms. The attention mechanism can dynamically assign different importance weights to different neighboring nodes, thereby more accurately capturing key risk sources. Secondly, when calculating local instability potential energy, the function can be concretized as a learnable neural network model or a multi-factor weighted scoring model. Its input parameters can be extended to include trend information of historical node state data (such as the upward slope of vibration indices), the criticality of the node in the current production task flow (such as the impact of equipment failure on the entire production line), and predefined equipment vulnerability coefficients. Furthermore, the graph neural network model can be extended to a temporal graph neural network containing gating mechanisms or memory units, enabling it to not only consider the graph structure at the current moment but also learn dynamic patterns from the time-series changes in node states, thereby further improving the accuracy and foresight of state perception and potential energy calculation.

[0062] S300 generates a system-level instability potential energy field based on the local instability potential energy of all nodes, and predicts the future potential energy field to identify high-risk areas.

[0063] As described above, this step represents a leap from discrete node risk assessment to continuous spatial risk field perception, which is crucial for forward-looking decision-making. Its core lies in transforming the spatially discrete local instability potential energy calculated at each node into a continuously distributed system-level instability potential energy field covering the entire physical space of the mine through spatial interpolation techniques. Mathematically, this potential energy field can be understood as a scalar field function defined on the mine's spatial coordinates, and its field strength directly reflects the comprehensive instability risk level at different spatial locations, thus forming a global "system pressure distribution map." Building upon this, to achieve truly forward-looking regulation, this scheme further introduces a potential field evolution model based on a physical information neural network. This model not only learns the temporal variation patterns in historical potential energy field data, but more importantly, it embeds the partial differential equations describing the physical processes of potential energy diffusion and conduction as constraints into the neural network's learning process, thereby enabling it to predict the potential energy field distribution in a more physically consistent manner. By analyzing the predicted future potential energy field, the system can identify regions where the field strength exceeds a preset critical threshold and energy is rapidly accumulating. These regions are defined as "potential wells" or high-risk regions, which are the most vulnerable and most likely to fail or collapse in efficiency in the system.

[0064] For example, in a digital model of an open-pit mine, the system first obtains the local instability potential energy values ​​of hundreds of nodes distributed across the mining face, transport lines, crushing stations, and other locations. Through spatial interpolation, a potential energy field cloud map covering the entire mine pit is generated, with color depth representing potential energy levels. The map might show that the area around the "A-section crusher" and the connected "B-section conveyor belt" is dark red, indicating the formation of a high potential energy zone. Subsequently, based on the current potential energy field, equipment operating inertia, and environmental trends (such as predicted high temperatures), the physical information neural network predicts that in the next 30 minutes, this high potential energy zone will not only continue to increase in intensity but its influence will also spread to the surrounding "C-section transfer station," thus forming a larger and deeper "potential well" in the prediction map. Based on this, the system determines that this "potential well" area is a high-risk area about to become unstable and requires immediate intervention.

[0065] It should be noted that, in specific implementation scenarios, the spatial interpolation method can be, but is not limited to, Kriging interpolation, inverse distance weighted interpolation, or radial basis function interpolation, based on the above scheme. Choosing different interpolation algorithms can adapt to the different characteristics and anisotropy of the mine's spatial structure. Secondly, the physical constraints introduced by the physical information neural network can be extended to include custom physical laws that conform to the characteristics of mine production, such as potential energy conduction equations derived from the laws of material flow and energy conservation, or empirical evolution rules based on the collaborative working logic of equipment groups. This makes the prediction model more industry-specific. Furthermore, the identification criteria for "high-risk areas" can be further refined. For example, it can be based not only on the exceeding of the absolute value of the potential energy field, but also on a comprehensive judgment combining the gradient change rate of the potential energy field (i.e., the speed of potential energy accumulation), the criticality index of the area in the overall production task, and the size of the connected area of ​​the risk area, thereby improving the accuracy of the early warning. The prediction process can also be multi-timescale, such as simultaneously predicting the potential energy field for the next 5 minutes, 15 minutes, and 1 hour, to support decision-making needs with different response speeds, such as real-time obstacle avoidance, production adjustment, and maintenance planning.

[0066] S400. Based on the system-level instability potential energy field and gradient information, the intelligent agent is controlled to perform autonomous regulation through a distributed decision-making method to reduce the risk of system instability.

[0067] As described above, this step is the execution stage of the dynamic intelligent control scheme. Its core lies in abandoning the traditional centralized scheduling command model and instead adopting a distributed collaborative autonomous mechanism based on global potential energy field information. The system publishes the generated system-level instability potential energy field and its spatial gradient vector field (i.e., the direction and rate of the fastest potential energy change) to various intelligent agents within the mining area (such as unmanned transport vehicles, intelligent crusher controllers, inspection robots, etc.). Each intelligent agent, as an independent decision-making unit, autonomously calculates and makes behavioral decisions based on its own location, task objectives, and the received local potential energy field and gradient information. Their common, implicit global optimization goal is to "flatten the potential energy field," that is, to facilitate the system's migration from a high-potential-energy (high-risk) state to a low-potential-energy (low-risk) state through their respective actions. Gradient information provides guidance for the intelligent agents' actions, for example, guiding them to move from high-potential-energy areas to low-potential-energy areas or transfer loads. This distributed decision-making mechanism decomposes the global optimization problem into multiple local autonomous problems, thereby achieving efficient, agile, and resilient system control, effectively reducing the risk of overall system instability caused by local disturbances.

[0068] For example, in a mining transportation scenario, several unmanned transport vehicles receive the current system instability potential energy field map. The map shows that the road area leading to crushing station No. 2 exhibits high potential energy (dark red) due to equipment overload. Simultaneously, the system calculates that the gradient direction of this area points towards the backup crushing station No. 3 (which has a lower load, and its potential energy is displayed in green). Based on this, an unmanned transport vehicle heading towards crushing station No. 2 will autonomously replan its route using its onboard decision-making system. The cost function for its route planning includes not only travel distance and time, but more importantly, the integral value of the potential energy field along the route. Therefore, it will automatically choose a route that, although slightly longer in logistical distance, has a lower overall potential energy risk, bypassing the high-risk area and heading towards crushing station No. 3. At the same time, the production scheduling system, as a virtual intelligent entity, will also autonomously decide to suspend the allocation of new crushing tasks to crushing station No. 2 and instead allocate them to crushing station No. 3, thereby "relieving pressure" on the high-potential energy area from the production source.

[0069] It should be noted that, in specific implementation scenarios, the types of intelligent agents can be expanded beyond the above scheme to include mobile devices (such as unmanned drilling rigs and excavators), fixed equipment controllers (such as ventilators and pumping stations), and robot swarms performing specific tasks (such as collaborative welding robots). Secondly, the autonomous decision-making rules of each intelligent agent can be customized according to its functional role. For example, for an inspection robot, its decision-making rule can be defined as actively moving towards the direction of increasing potential energy gradient (i.e., high-risk sources) to perform precise inspections; for an energy management system, its rule can be defined as dynamically allocating more energy to high-potential-energy production nodes to improve their stability and processing capacity. Furthermore, the distributed decision-making process can introduce a negotiation mechanism based on game theory or collaborative contracts, enabling multiple intelligent agents to communicate and coordinate lightweightly during decision-making, avoiding conflicts between autonomous decisions and forming a better group synergy effect. The decision-making rules themselves can also be an online-updable strategy library, allowing the system to continuously optimize the decision-making logic of various agents based on historical intervention effects and using methods such as reinforcement learning.

[0070] S500 collects new data after implementing regulation, updates the dynamic relationship diagram and potential energy field, and fine-tunes the model by comparing predictions with actual results.

[0071] As described above, this step constitutes the core of the closed-loop optimization and self-evolution of the dynamic intelligent control method, essentially establishing a deep feedback learning loop. After the system executes distributed control commands based on the potential energy field distribution, it collects a new round of equipment status data, environmental parameters, and production task flow data through a sensor network deployed throughout the mine. These data reflect the actual response state of the system after the implementation of control measures. Using this new real-time data, the system first re-executes the construction process of the dynamic relationship graph, updating the connection relationships and weights between nodes to capture the changes in the internal interaction relationships of the system caused by control. Subsequently, based on the updated dynamic relationship graph, it recalculates the node states and generates a new, actually observed potential energy field that reflects the latest state of the system. Most importantly, the system accurately compares the future potential energy field predicted based on historical data with the currently observed new potential energy field; the difference between the two constitutes the model prediction error. This error is systematically used as a supervisory signal. Through machine learning mechanisms such as backpropagation, the parameters in the graph neural network model and the physical information neural network potential field evolution model, which constitute the core of the system's cognition, are fine-tuned and optimized online. This enables the model to learn from the interaction between each decision and the actual feedback from the system, and gradually correct its cognitive biases regarding the dynamic behavior of complex mining systems.

[0072] For example, the system previously predicted that the potential energy in the "Transport Lane 5" area would rise sharply to a critical state within the next 15 minutes, and accordingly dispatched three unmanned transport vehicles to detour around the area to "depressurize." Fifteen minutes after the intervention, the system collected new data and found that the actual increase in potential energy in the area was far lower than the predicted value. Through in-depth analysis, the system identified that the prediction error mainly stemmed from not fully considering the improvement effect of the adjacent "Ventilation Shaft 3" on the roadway's environmental parameters. Therefore, when updating the dynamic relationship graph, the system strengthened the edge weights of the environmental impact between the "Ventilation Shaft 3" node and the "Transport Lane 5" node. Simultaneously, the difference between the prediction and the actual value, as a loss signal, was transmitted back to the physical information neural network to fine-tune its internal parameters. This allows the model to incorporate the intervention effect of the ventilation system into its calculations when predicting similar scenarios in the future, thus making more accurate predictions.

[0073] It should be noted that, in specific implementation scenarios, the acquisition of new data can be expanded to include novel sensor data (such as acoustic imaging data and infrared thermal imaging data), thereby providing the model with richer feature inputs. Secondly, the model update and fine-tuning mechanism can be flexibly configured. For example, different time-scale update strategies can be adopted—periodic batch updates of graph neural networks on a daily or weekly basis, while more frequent incremental online updates of physical information neural networks on a minute or hourly basis, to adapt to the different sensitivities of model parameters to dynamic changes in the system. Furthermore, the fine-tuning process can be extended into a multi-model collaborative adjustment framework, not only adjusting the internal parameters of individual models but also dynamically selecting or fusing the outputs of multiple alternative models (such as graph neural networks with different architectures) based on long-term learning effects to improve system robustness. The objective function of feedback learning can also be further enriched. In addition to minimizing the prediction error of the potential energy field, a multi-objective trade-off between control costs (such as additional energy consumption of equipment and path extension distance) and risk reduction effects can be added, thereby guiding the model to learn more economical decision-making strategies.

[0074] According to one embodiment of this application, constructing the dynamic relationship graph includes:

[0075] The weights of edges are dynamically updated by analyzing the correlations of node state changes in historical data.

[0076] As mentioned above, in the process of constructing a dynamic relationship graph, dynamically updating edge weights by analyzing the correlation of node state changes in historical data means that the system continuously monitors and records the state data of each node at different points in time, and uses data analysis methods to quantitatively evaluate the degree of mutual influence between the state changes of any two connected nodes in the graph. This correlation analysis aims to capture the dynamic and potentially time-evolving interaction strength between nodes, rather than relying on preset, fixed connection weights.

[0077] Specifically, the system calculates quantitative indicators characterizing the degree of correlation between node states based on historical time-series data, such as the mutual information value or correlation coefficient between the state sequences of two nodes. Mutual information can capture linear and nonlinear dependencies, while the correlation coefficient mainly reflects linear correlations. Through such calculations, the extent to which a change in the state of one node contains information about the change in the state of another node can be quantified. Subsequently, the system assigns or updates the weights of the edges connecting the two nodes based on the results of this quantitative indicator. For example, if historical data analysis shows that an abnormal increase in the vibration amplitude of node A (such as a critical pumping station) is always accompanied by a significant increase in the temperature of node B (such as a transformer on its power supply line), and this association has high statistical significance, then the weight of the edge connecting A and B will be set to a correspondingly higher value. Conversely, if the state changes of two nodes are proven to be independent or weakly correlated, the weight of the edge between them will be reduced, and in some implementations, the edge may even be removed.

[0078] This dynamic update mechanism enables the constructed relationship graph to transcend static physical connections or logical presuppositions, truly reflecting the dynamic coupling relationships and risk transmission paths within the system based on actual operational data. This lays the foundation for subsequent precise analysis based on the graph structure. The edge weights here become dynamically evolving, data-driven parameters, allowing the graph model to adaptively characterize the actual operational status of the mining production system.

[0079] According to one embodiment of this application, the calculation of node states and local instability potential energy using a graph neural network includes:

[0080] Use graph convolutional networks to aggregate neighbor node information to update node state;

[0081] The local instability potential energy is calculated based on the node's own state anomaly degree, the risk transmission of neighboring nodes, and local environmental stress.

[0082] As described above, the system utilizes the message-passing mechanism of graph convolutional networks when updating node states. Each node receives state feature information from its direct neighbors through the edges it connects to. The network integrates this neighbor feature information using a learnable aggregation function. This aggregation process typically involves a weighted average of the neighbor node features, where the weights are related to the weights of edges in the dynamic graph, thus giving greater influence to neighbors with closer interactions with the current node on its state update. The aggregated information is then combined with the node's own current state features and passed through a nonlinear transformation layer to ultimately generate a new state representation for the node that incorporates its local graph structure context information. This process ensures that each node's state is no longer isolated but includes the interconnected influences of its local network.

[0083] When calculating the local instability potential energy, the system comprehensively evaluates information from three aspects. First, the node's own state anomaly degree is determined by comparing the node's updated state characteristics with a preset normal operating baseline state, such as assessing the degree to which key parameters like vibration and temperature deviate from normal ranges. Second, the risk transmission from neighboring nodes is quantified by analyzing the potential energy or state anomaly information of its neighboring nodes and weighting them according to the weights of the connecting edges. This reflects the cascading risks caused by failures or performance degradation of related equipment. Third, local environmental stress directly originates from environmental sensor data deployed near the node, such as the stress effect of temperature, humidity, or dust concentration on equipment stability. Finally, the system uses a comprehensive function, such as a trainable fully connected neural network or a multi-factor weighted fusion model, to map the inputs of these three dimensions into a scalar value, namely the node's local instability potential energy. The higher this value, the greater the risk of the node tending towards instability.

[0084] According to one embodiment of this application, the generation of the system-level instability potential energy field includes:

[0085] Local potential energy is mapped into a continuous potential energy field using spatial interpolation techniques;

[0086] The physical information neural network is used to predict the future potential energy field and identify potential well regions where potential energy rapidly accumulates above a critical threshold.

[0087] As described above, at the first level, local potential energy is mapped to a continuous potential energy field using spatial interpolation techniques. The system uses the locally distributed, spatially discrete instability potential energy values ​​calculated at each node as known data points. Based on these data points, a spatial interpolation algorithm is used to fit a continuous function across the entire physical spatial coordinate range of the mine, thereby generating a continuous scalar field covering the entire region, i.e., a system-level instability potential energy field. This potential energy field has a definite potential energy value at any spatial coordinate point, and its spatial distribution directly characterizes the macroscopic geographical distribution of instability risk in the entire mine system.

[0088] At the second level, a physical information neural network is used to predict the future potential energy field and identify potential well regions. This physical information neural network is a machine learning model that embeds physical constraints into the learning process. It uses the potential energy field from the previous or current time step as initial conditions, and combines physical equations (such as partial differential equations) describing the intrinsic evolution of potential energy conduction and diffusion as constraints. Through its network structure, it calculates and outputs the predicted potential energy field for a specific future time step. The system then analyzes the predicted future potential energy field and identifies regions that simultaneously meet the following two conditions as potential well regions: first, the potential energy field value in this region exceeds a preset system stability critical threshold; second, the potential energy field in this region exhibits a rapid accumulation trend in space, meaning its field strength has a significant growth gradient and trend within the prediction time. These identified potential well regions are potential high-risk outbreak points for the system.

[0089] According to one embodiment of this application, controlling the intelligent agent for autonomous regulation through distributed decision-making includes:

[0090] Publish the current potential energy field and gradient information to the intelligent agent;

[0091] Intelligent agents autonomously adjust their behavior based on the potential energy field, including path planning, task reallocation, or maintenance of pre-set resources.

[0092] As described above, the system publishes current system-level unstable potential energy field data and its spatial gradient information to each intelligent agent. The potential energy field data is provided in the form of a data layer covering the geographic space of the mine, and the intelligent agents receive the potential energy field distribution of their location and surrounding area through a communication network. The gradient information characterizes the direction and rate of change of the potential energy field at each spatial location, providing directional guidance for the agents' decision-making.

[0093] Based on the received potential energy field and gradient information, each intelligent agent autonomously adjusts its operational behavior according to its own task objectives. In path planning, intelligent transportation equipment incorporates the potential energy field integral into the path cost function, proactively avoiding high-potential-energy regions, even if the path is geometrically shorter. Regarding task redistribution, the production scheduling system dynamically migrates production tasks from high-potential-energy nodes to low-potential-energy nodes based on the real-time potential energy values ​​of each production node (such as crushers and mineral processing equipment), achieving load balancing and risk diversification. In terms of pre-positioning maintenance resources, the system autonomously instructs inspection robots or maintenance resources to move and deploy in advance to areas where potential energy is rapidly accumulating and about to exceed a critical threshold, based on the potential energy field prediction results, achieving proactive maintenance intervention. These distributed decisions by each agent work together to achieve the goal of reducing the overall instability risk.

[0094] According to one embodiment of this application, in the path planning, the cost function includes a potential energy field integral, which causes the vehicle to bypass the potential energy peak region.

[0095] As described above, in the path planning process, the cost function includes integrating the potential energy field of the space traversed by the candidate path. Specifically, for each feasible path to be evaluated from the starting point to the ending point, the system calculates the integral of the potential energy field value of all points on the path along the path, and uses this integral value as one of the key cost factors for path evaluation.

[0096] This means that the quality of a route is no longer determined solely by traditional indicators such as route length or travel time, but rather by the overall risk exposure along the route as the core decision-making criterion. By incorporating this integral term into the cost function and optimizing it, the system automatically tends to select routes that, even if physically longer, have lower overall potential energy accumulation. The direct technical effect is to guide the vehicle to actively avoid areas of high potential energy throughout the entire potential energy field, achieving autonomous risk avoidance decisions and effectively reducing the overall instability risk faced by the vehicle during operation at the route planning level.

[0097] According to one embodiment of this application, the fine-tuning model includes:

[0098] The difference between the predicted potential field and the actual observed potential field is used as the loss function, and backpropagated to the graph neural network and the physical information neural network for online fine-tuning.

[0099] As described above, the fine-tuning process of the model is as follows: The system compares the potential energy field predicted by the physical information neural network at a future moment with the actual potential energy field generated from actual monitoring data at the same moment, point by point, and calculates the difference between the two. This difference is quantified into a scalar loss function, typically using mean squared error or mean absolute error as its specific mathematical form. The value of this loss function characterizes the accuracy of the model's prediction; a larger value indicates a larger prediction error. The system then uses an error backpropagation algorithm to propagate this loss value backward along the computational path of the neural network, calculating the gradient of this loss with respect to each trainable parameter in both the graph neural network model and the physical information neural network model. Based on the calculated gradient direction, the weights and bias parameters in the two types of neural network models are slightly adjusted and updated using gradient descent or its variant optimization algorithms. This fine-tuning process is performed periodically while the system is running online. Its direct purpose is to continuously optimize and correct the model's internal parameters by constantly utilizing the differences between the latest field data and the prediction results, thereby gradually improving the model's ability to predict the future evolution of the system's potential energy field over time.

[0100] A second aspect of this application provides a dynamic intelligent control system for mines based on system instability potential energy and collaborative autonomy, comprising:

[0101] The dynamic relationship graph construction module is used to collect multi-source heterogeneous data and construct dynamic relationship graphs;

[0102] The node state and potential energy calculation module is used to calculate the node state and local instability potential energy based on the dynamic relationship graph using a graph neural network.

[0103] The potential energy field generation and prediction module is used to generate system-level unstable potential energy fields and predict future potential energy fields to identify high-risk areas.

[0104] The distributed control module is used to enable autonomous control of intelligent agents through distributed decision-making based on potential energy field and gradient information.

[0105] The model update module is used to collect new data after the regulation is performed, update the dynamic relationship diagram and potential energy field, and fine-tune the model.

[0106] According to one embodiment of this application, the dynamic relationship graph construction module is configured to dynamically update the weights of edges by analyzing the correlation of node state changes in historical data;

[0107] The node state and potential energy calculation module is configured to use a graph convolutional network to aggregate neighbor node information to update the node state, and calculate the local instability potential energy based on the node's own state anomaly degree, the risk transmission of neighbor nodes, and local environmental stress.

[0108] According to one embodiment of this application, the potential energy field generation and prediction module is configured to generate a continuous potential energy field through spatial interpolation technology and use a physical information neural network to predict the future potential energy field to identify potential well regions.

[0109] The distributed control module is configured to publish potential energy field and gradient information to the intelligent agent, and the intelligent agent autonomously adjusts its behavior based on the potential energy field, including path planning, task reallocation, or maintenance resource pre-setting.

[0110] In the path planning, the cost function includes the potential energy field integral, which enables the vehicle to bypass the potential energy peak region.

[0111] The model update module is configured to use the difference between the predicted potential energy field and the actual observed potential energy field as a loss function, and backpropagate it to the graph neural network and the physical information neural network for online fine-tuning.

[0112] Example 2

[0113] 1. System Initialization and Data Acquisition

[0114] The intelligent control system of this invention was deployed in a large open-pit iron mine. The system first collects multi-source heterogeneous data through a sensor network distributed throughout the mine:

[0115] Equipment status data: including real-time vibration, temperature, current, and operating speed data for 10 electric wheel mining trucks, 3 large crushers, and 5 belt conveyors.

[0116] Environmental parameters: These include temperature, humidity, and dust concentration data for each mining face, as well as environmental monitoring data around key equipment.

[0117] Production task flow data includes the daily ore mining plan, crushing and processing targets, and transportation task queues from the production execution system.

[0118] 2. Construction of Dynamic Relationship Diagrams

[0119] The system constructs a dynamic relationship graph G(V, U, t), where:

[0120] Node set V contains all monitored objects: {electric wheel trucks 1-10, crushers 1-3, conveyor belts 1-5, mining faces AC, environmental monitoring points 1-8};

[0121] Boundary set U represents interaction relationships, such as:

[0122] Physical connection: Crusher 1 → Conveyor Belt 1 (weight determined through historical data analysis);

[0123] Environmental impact: Environmental monitoring point 3 (high temperature) → Crusher 2 (weight reflects the correlation strength between temperature and equipment failure rate);

[0124] The edge weights W_ij are dynamically updated by calculating the mutual information values ​​of the node state sequences.

[0125] MI(i,j) = ΣΣ p(i,j) log[p(i,j) / (p(i)p(j))]

[0126] Where p(i) and p(j) are the node state distributions, and p(i,j) is the joint distribution.

[0127] 3. Calculation of nodal states and local instability potential energy

[0128] Use a Graph Convolutional Network (GCN) to update the node state. For crusher node 2:

[0129] Aggregate the status information of its neighboring nodes (upstream feeding equipment, connected conveyor belts, environmental monitoring points);

[0130] Node state update formula:

[0131]

[0132] in Let N(2) be the feature representation of crusher 2 at layer l, and let N(2) be its set of neighbor nodes.

[0133] Calculate the local instability potential energy Φ2 of crusher 2:

[0134] Φ2 = α·A2 + β·ΣW 2j Φ j + γ·E2

[0135] in:

[0136] A2 represents the degree of abnormality in its own state (the degree of vibration exceeding the standard).

[0137] ΣW 2j Φ j Weighted sum of risk propagation from neighboring nodes;

[0138] E2 represents local environmental stress (the difference between the current temperature and the rated operating temperature).

[0139] α, β, γ are weight coefficients determined through training.

[0140] 4. Generation and prediction of system-level instability potential energy field

[0141] By interpolating the Φ values ​​of all nodes using the Kriging space interpolation method, a continuous potential energy field Ψ(X,t) covering the entire mining area is generated. This potential energy field is displayed spatially as follows:

[0142] A potential energy peak (Ψ = 0.85) is formed in the crushing station area;

[0143] The main transportation routes form potential energy ridges (Ψ = 0.60-0.75);

[0144] Predicting future potential energy fields using a Physical Information Neural Network (PINN). The loss function of PINN includes:

[0145] L = L_data + λL_physics

[0146] in:

[0147] L_data is the mean squared error between the predicted value and the training data;

[0148] L_physics ensures that the predictions satisfy the potential energy diffusion equation: ;

[0149] λ is the physical constraint weight coefficient.

[0150] The prediction results show that after 30 minutes, the potential energy in the crushing station area will increase to Ψ = 0.92, exceeding the critical threshold Ψ_critical = 0.90, and the system will identify it as a "potential well" area.

[0151] 5. Distributed collaborative autonomous regulation

[0152] The system publishes the current potential energy field Ψ(X,t) and gradient information to relevant intelligent agents. Ψ:

[0153] Path planning: After receiving potential energy field data, mining truck 7, which is heading towards the crushing station, recalculates the path cost:

[0154] Cost = ∫_path[Ψ(X,t) + κ·distance]ds

[0155] Where κ is the distance weighting coefficient. The truck autonomously chooses a detour route to avoid high-potential-energy areas.

[0156] Task reallocation: The production scheduling system dynamically adjusts the 200 tons / hour task originally planned for crusher 2 to crushers 1 and 3, which have lower potential energy.

[0157] Pre-positioning of maintenance resources: The system instructs the inspection robot to move to the vicinity of the crushing station in advance to stand by and prepare for troubleshooting.

[0158] 6. Feedback Learning and Model Evolution

[0159] After the control measures are implemented, the system collects new monitoring data:

[0160] The actual potential energy field Ψ(X,t+Δt) shows that the potential energy in the crushing station area rose to 0.88, which is lower than the predicted value of 0.92;

[0161] The prediction error ΔΨ = 0.04 is used as the loss function and backpropagated to the GCN and PINN models;

[0162] The model parameters are fine-tuned online, with a particular emphasis on increasing the weight of the influence of ambient temperature on the equipment's potential energy.

[0163] For any parts not mentioned in this application, existing technologies may be used or referenced.

[0164] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0165] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A dynamic control method for mines based on system instability potential energy and collaborative autonomy, characterized in that, include: Collect multi-source heterogeneous data from the mining production system, including equipment status data, environmental parameters, and production task flow data, and construct a dynamic relationship graph, where nodes represent mining equipment, production areas, or environmental monitoring points, and edges represent the interaction relationships between nodes. The construction of the dynamic relationship graph includes: The weights of edges are dynamically updated by analyzing the correlation of node state changes in historical data. Based on the dynamic relationship graph, a graph neural network is used to calculate the state and local instability potential energy of each node, specifically as follows: Use graph convolutional networks to aggregate neighbor node information to update node state; The local instability potential energy is calculated based on the node's own state anomaly, the risk transmission of neighboring nodes, and the local environmental stress. Based on the local instability potential energy of all nodes, a system-level instability potential energy field is generated, and the future potential energy field is predicted to identify high-risk areas. The generated system-level instability potential energy field includes: Local potential energy is mapped into a continuous potential energy field using spatial interpolation techniques; The physical information neural network is used to predict the future potential energy field and identify potential well regions where potential energy rapidly accumulates beyond a critical threshold. Based on the aforementioned system-level instability potential energy field and gradient information, a distributed decision-making approach is used to control the intelligent agent for autonomous regulation, thereby reducing the risk of system instability. Specifically: Publish the current potential energy field and gradient information to the intelligent agent; Intelligent agents autonomously adjust their behavior based on the potential energy field, including path planning, task reallocation, or maintenance of pre-set resources; In the path planning, the cost function includes the potential energy field integral, which enables the vehicle to bypass the potential energy peak region. After implementing regulation, new data is collected to update the dynamic relationship diagram and potential energy field, and the model is fine-tuned by comparing the prediction with the actual situation. The fine-tuning model includes: The difference between the predicted potential field and the actual observed potential field is used as the loss function, and backpropagated to the graph neural network and the physical information neural network for online fine-tuning.

2. A dynamic control system for mines based on system instability potential energy and collaborative autonomy, characterized in that, include: The dynamic relationship graph construction module is used to collect multi-source heterogeneous data and construct dynamic relationship graphs; The dynamic relationship graph construction module is configured to dynamically update the edge weights by analyzing the correlation of node state changes in historical data. The node state and potential energy calculation module is configured to use a graph convolutional network to aggregate neighbor node information to update the node state, and calculate the local instability potential energy based on the node's own state anomaly degree, neighbor node risk transmission, and local environmental stress. The node state and potential energy calculation module is used to calculate the node state and local instability potential energy based on the dynamic relationship graph using a graph neural network. The potential energy field generation and prediction module is used to generate system-level unstable potential energy fields and predict future potential energy fields to identify high-risk areas. The potential energy field generation and prediction module is configured to generate a continuous potential energy field through spatial interpolation technology and use a physical information neural network to predict the future potential energy field in order to identify potential well regions. The distributed control module is configured to publish potential energy field and gradient information to the intelligent agent, and the intelligent agent autonomously adjusts its behavior based on the potential energy field, including path planning, task reallocation, or maintenance resource pre-setting. In the path planning, the cost function includes the potential energy field integral, which enables the vehicle to bypass the potential energy peak region. The model update module is configured to use the difference between the predicted potential energy field and the actual observed potential energy field as a loss function, and backpropagate it to the graph neural network and the physical information neural network for online fine-tuning. The distributed control module is used to enable autonomous control of intelligent agents through distributed decision-making based on potential energy field and gradient information. The model update module is used to collect new data after the regulation is performed, update the dynamic relationship diagram and potential energy field, and fine-tune the model.

Images