Mine safety interlocking control method and device based on reinforcement learning, equipment and medium

Abstract

The application discloses a mine safety interlocking control method and device based on reinforcement learning, equipment and medium, relates to the technical field of mine safety monitoring, and comprises the following steps: a mine safety situation awareness system is constructed by adopting a perception-transmission-processing three-level architecture, and multi-dimensional safety data is collected and preprocessed; a decision model is constructed based on reinforcement learning, safety data is used as a state space, interlocking actions are used as an action space, safety risks are used as a reward function, and the reinforcement learning decision model is obtained through offline training; and then the optimal strategy is output through a three-layer network to drive the safety interlocking equipment action, so that automatic control and closed-loop optimization are realized. The integrated mine safety interlocking control architecture is constructed, emergency interlocking closed-loop control is realized based on multi-dimensional safety data, the intelligent level and response speed of mine safety interlocking control are improved, manual dependence is reduced, and mine production safety is ensured.

Description

Technical Field

[0001] This invention relates to the field of mine safety monitoring technology, and in particular to a mine safety interlocking control method, device, equipment and medium based on reinforcement learning. Background Technology

[0002] Currently, most mine safety control systems use independent monitoring equipment to collect environmental, equipment, and personnel data. Simple alarms or manual interventions are triggered by preset fixed thresholds. Some systems use basic logic control to achieve simple interlocking actions of single equipment. A few intelligent solutions use conventional machine learning or shallow networks for data analysis, rely on manual rule configuration to achieve safety control, depend on historical experience to set control parameters, and put the model directly into use after offline training.

[0003] Existing technologies lack a unified and hierarchical situational awareness architecture, making it difficult to achieve collaborative collection and efficient preprocessing of multi-dimensional data such as environmental parameters, equipment operating parameters, and personnel location parameters. They do not employ reinforcement learning to build adaptive decision-making models, nor do they clearly define state and action spaces or construct reward functions adapted to mine safety scenarios, thus failing to achieve the autonomous generation of optimal control strategies. Furthermore, the models do not support an iterative optimization mechanism combining offline training and online incremental learning, resulting in rigid control strategies and poor adaptability. Simultaneously, they lack an emergency interlocking mechanism that automatically judges threshold exceedances, pauses optimization, and executes the highest-priority protective actions in emergency situations, leading to delayed control response, untimely risk handling, and difficulty in forming an adaptive, high-real-time safety interlocking control closed loop.

[0004] Therefore, how to construct an integrated mine safety interlocking control architecture and realize emergency interlocking closed-loop control based on multi-dimensional safety data, thereby improving the intelligence level and response speed of mine safety interlocking control, has become an urgent problem to be solved. Summary of the Invention

[0005] The main purpose of this application is to provide a method, device, equipment and medium for mine safety interlock control based on reinforcement learning, which aims to solve the technical problem of how to improve the intelligence level and response speed of mine safety interlock control.

[0006] To achieve the above objectives, this application proposes a mine safety interlocking control method based on reinforcement learning, comprising: A mine safety situation awareness system is constructed to collect and process multi-dimensional safety status data to obtain pre-processed data. The multi-dimensional safety status data includes environmental parameters, equipment operating parameters, and personnel location parameters. The mine safety situation awareness system adopts a three-level architecture of perception layer, transmission layer, and processing layer. The preprocessed data is input into a reinforcement learning decision model to obtain the optimal control strategy, wherein the reinforcement learning decision model includes a feature extraction layer, a fully connected layer and an output layer. Generate control commands based on the optimal control strategy; The control command is sent to the corresponding safety interlock device so that the safety interlock device performs the interlock action; Before the step of inputting the preprocessed data into the reinforcement learning decision model to obtain the optimal control strategy, the following steps are included: An initial decision model is constructed based on a reinforcement learning algorithm. The preprocessed data is used as the safety state space, and at least one interlocking control action among ventilation equipment adjustment, gas extraction equipment start-up and shutdown, mining equipment emergency stop, personnel evacuation early warning, power supply system power outage and waterproof gate closure is used as the action space. A reward function is constructed based on the safety risk level, the effect of control action execution and the emergency control result. Acquire historical safe operation data, accident case data, and simulated working condition data of the mine; The initial decision model is trained offline based on the mine's historical safe operation data, accident case data, and simulated working condition data to obtain a reinforcement learning decision model.

[0007] In one embodiment, the steps of constructing a mine safety situation awareness system, collecting and processing multi-dimensional safety status data, and obtaining preprocessed data include: Acquire mine geological exploration data, historical accident distribution data, and operational area planning data, and determine the mine safety risk assessment results through risk area identification algorithms; A three-tier architecture of perception layer, transmission layer, and processing layer is adopted to construct a mine safety situation awareness system. The perception layer is used to deploy environmental perception devices, equipment perception devices, and personnel perception devices in a differentiated manner according to the mine safety risk assessment results. The transmission layer is used for data transmission. The processing layer is used to deploy edge computing nodes with data caching function to complete data processing nearby. Based on the risk assessment results, environmental sensing devices are deployed in environmental risk areas to monitor environmental safety parameters, equipment sensing devices are deployed in key equipment parts to monitor equipment operating parameters, and personnel sensing devices are deployed in personnel work areas to monitor personnel location parameters. The deployment density of the environmental sensing devices, equipment sensing devices, and personnel sensing devices is positively correlated with the risk level in the risk assessment results. The environmental sensing device, equipment sensing device, and personnel sensing device collect multi-dimensional safety status data in real time. The multi-dimensional security status data is processed by edge computing nodes to unify the format, remove anomalies, and fill in missing data, resulting in preprocessed data.

[0008] In one embodiment, the step of constructing an initial decision model based on a reinforcement learning algorithm includes: The initial decision model is constructed by selecting a policy optimization-based reinforcement learning algorithm as the core algorithm. The preprocessed data is standardized using the min-max normalization method to obtain standardized data, and the standardized data is used as the safe state space of the initial decision model. The requirements for mine safety interlocking control are defined as the discrete action space of the initial decision model. The action space includes at least one interlocking control action among ventilation equipment adjustment, gas extraction equipment start-up and shutdown, mining equipment emergency stop, personnel evacuation early warning, power supply system power failure and waterproof gate closure. Each action corresponds to a unique action identifier. A base reward value is set based on the safety risk level, a change reward value is set based on the change in safety risk level before and after the control action is executed, and an emergency reward value is set based on the execution result of the emergency interlock control process. The safety risk level is divided into three levels: high risk, medium risk, and low risk based on a preset risk threshold. The reward function of the initial decision model is constructed by weighting and summing the base reward value, the variable reward value, and the emergency reward value.

[0009] In one embodiment, the step of offline training the initial decision model based on the mine's historical safe operation data, accident case data, and simulated working condition data to obtain a reinforcement learning decision model includes: The historical safe operation data, accident case data, and simulated working condition data of the mine are standardized and preprocessed to obtain a training dataset. Initialize the model parameters of the initial decision model, and set the maximum number of iterations and the first learning rate; The training dataset is input into the initial decision model, and random perturbations are added to simulate random changes in mining conditions to obtain the cumulative reward value. The model parameters are updated based on the first learning rate to obtain the optimized model; When the cumulative reward value tends to stabilize or the number of iterations reaches the maximum number of iterations, the optimized model at this time is used as the reinforcement learning decision model.

[0010] In one embodiment, the step of inputting the preprocessed data into a reinforcement learning decision model to obtain the optimal control strategy includes: The preprocessed data is input into the feature extraction layer of the reinforcement learning decision model to obtain the current security state. The current security state is input into the fully connected layer to calculate the expected reward value of each interlocking control action, thus obtaining the expected reward value of each interlocking control action; The expected reward value is input into the output layer, the action with the highest expected reward value is selected as the optimal control action, and the corresponding optimal control strategy is output. The optimal control strategy includes the action identifier and action execution parameters of the optimal control action. The action execution parameters include at least one of the following: ventilation equipment air volume adjustment value, warning duration, and power supply cut-off range.

[0011] In one embodiment, after the step of sending the control command to the corresponding safety interlock device to cause the safety interlock device to perform an interlocking action, the method includes: Collect safety status feedback data after the safety interlocking device performs its actions; When the safety status feedback data does not exceed the safety status emergency threshold, the reinforcement learning decision model is iteratively optimized online using the safety status feedback data through incremental learning to obtain an updated reinforcement learning decision model. When the safety status feedback data exceeds the safety status emergency threshold, the emergency interlock control process is triggered, online iterative optimization is paused, and the highest priority safety protection action is executed. The safety protection action includes at least one of cutting off the power supply to the entire area, activating the emergency evacuation warning, and closing the safety isolation door. When the security status feedback data recovers to a level not exceeding the security status emergency threshold, the reinforcement learning decision model is further iteratively optimized online using the security status feedback data through incremental learning to obtain an updated reinforcement learning decision model.

[0012] In one embodiment, the step of using incremental learning to iteratively optimize the reinforcement learning decision using the security state feedback data to obtain an updated reinforcement learning decision model includes: The safety status feedback data is used as a new training sample and input into the reinforcement learning decision model to calculate the deviation between the predictive control strategy and the optimal control strategy under the current model parameters. Based on the gradient update direction of the model parameters calculated by the deviation, the initial fully connected layer parameters of the reinforcement learning decision model are locally incrementally updated. A second learning rate for incremental learning is set, and the parameters are fine-tuned by a single iteration or a small number of iterations. The fine-tuned parameters are used as the updated model parameters, wherein the second learning rate is less than the first learning rate in the offline training stage. The updated model parameters are weighted and fused with the initial fully connected layer parameters to obtain the fused model parameters. The reinforcement learning decision model is iteratively optimized online based on the fused model parameters to obtain an updated reinforcement learning decision model.

[0013] Furthermore, to achieve the above objectives, this application also proposes a mine safety interlocking control device based on reinforcement learning, wherein the reinforcement learning-based mine safety interlocking control device includes: The acquisition module is used to build a mine safety situation awareness system, collect and process multi-dimensional safety status data, and obtain pre-processed data. The multi-dimensional safety status data includes environmental parameters, equipment operating parameters and personnel location parameters. The mine safety situation awareness system adopts a three-level architecture of perception layer-transmission layer-processing layer. The strategy generation module is used to input the preprocessed data into a reinforcement learning decision model to obtain the optimal control strategy. The reinforcement learning decision model includes a feature extraction layer, a fully connected layer, and an output layer. It is also used to construct an initial decision model based on a reinforcement learning algorithm, using the preprocessed data as the safety state space and at least one interlocking control action among ventilation equipment adjustment, gas extraction equipment start / stop, mining equipment emergency stop, personnel evacuation early warning, power supply system power outage, and waterproof gate closure as the action space. A reward function is constructed based on the safety risk level, the effect of control action execution, and the emergency control result. The module acquires historical mine safety operation data, accident case data, and simulated working condition data. Based on the historical mine safety operation data, accident case data, and simulated working condition data, the initial decision model is trained offline to obtain the reinforcement learning decision model. The instruction generation module is used to generate control instructions based on the optimal control strategy; The execution module is used to send the control command to the corresponding safety interlocking device so that the safety interlocking device performs the interlocking action.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the reinforcement learning-based mine safety interlocking control method described above.

[0015] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the reinforcement learning-based mine safety interlocking control method described above.

[0016] This application employs a three-tiered architecture of perception, transmission, and processing to construct a mine safety situational awareness system. It collects and preprocesses multi-dimensional safety data; a decision-making model is built based on reinforcement learning, using safety data as the state space, interlocking actions as the action space, and safety risk as the reward function. This model is trained offline to obtain the reinforcement learning decision-making model. Then, the optimal strategy is output through a three-layer network to drive the actions of safety interlocking equipment, achieving automatic control and closed-loop optimization. This constructs an integrated mine safety interlocking control architecture and realizes emergency interlocking closed-loop control based on multi-dimensional safety data, improving the intelligence level and response speed of mine safety interlocking control, reducing reliance on manual labor, and ensuring mine production safety. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the first embodiment of the mine safety interlocking control method based on reinforcement learning in this application. Figure 2 This is a flowchart illustrating the second embodiment of the mine safety interlocking control method based on reinforcement learning in this application. Figure 3 This is a schematic diagram of the module structure of the mine safety interlocking control device based on reinforcement learning in this application. Figure 4 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the reinforcement learning-based mine safety interlocking control method in the embodiments of this application.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] Current mine safety control primarily relies on independent monitoring equipment to collect environmental, equipment, and personnel data. Simple alarms or manual interventions are triggered by preset fixed thresholds. Some systems use basic logic control to implement simple interlocking actions for single devices. A few intelligent solutions employ conventional machine learning or shallow networks for data analysis, relying on manually configured rules for safety control. Control parameters are set based on historical experience, and models are directly deployed after offline training. Existing technologies lack a unified, hierarchical situational awareness architecture, making it difficult to achieve collaborative collection and efficient preprocessing of multi-dimensional data such as environmental parameters, equipment operating parameters, and personnel location parameters. They do not employ reinforcement learning to build adaptive decision-making models, nor do they clearly define state and action spaces or construct reward functions adapted to mine safety scenarios, hindering the autonomous generation of optimal control strategies. Models do not support iterative optimization mechanisms combining offline training and online incremental learning, resulting in rigid control strategies and poor adaptability. Furthermore, they lack emergency interlocking mechanisms that automatically judge threshold exceedances, pause optimization, and execute the highest-priority protective actions in emergency situations, leading to delayed control responses, untimely risk handling, and difficulty in forming an adaptive, high-real-time safety interlocking control closed loop. Therefore, how to construct an integrated mine safety interlocking control architecture and realize emergency interlocking closed-loop control based on multi-dimensional safety data, thereby improving the intelligence level and response speed of mine safety interlocking control, has become an urgent problem to be solved.

[0023] Based on the above, this application also provides a mine safety interlocking control method based on reinforcement learning, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the mine safety interlocking control method based on reinforcement learning in this application.

[0024] In this embodiment, the mine safety interlocking control method based on reinforcement learning includes steps S10 to S40: Step S10: Construct a mine safety situation awareness system, collect and process multi-dimensional safety status data, and obtain pre-processed data.

[0025] It should be noted that the mine safety situation awareness system is an overall operational framework for comprehensively acquiring safety-related information at the mine site. This framework continuously collects various safety-related data during mine operations, enabling real-time monitoring of the overall mine safety status. Multi-dimensional safety status data is a collection of safety-related information obtained from different monitoring objects within the mine. This data includes environmental parameters, equipment operating parameters, and personnel location parameters. Environmental parameters are monitoring data reflecting the natural and operational environment conditions at the mine site. This type of data directly indicates whether the mine faces environmental safety risks such as excessive gas or dust levels. Equipment operating parameters are monitoring data recording the working status of various production and safety equipment in the mine. Personnel location parameters are location data identifying the specific area where mine workers are located. The mine safety situation awareness system adopts a three-tier architecture: perception layer, transmission layer, and processing layer. The perception layer is the most front-end component of the mine safety situation awareness framework, responsible for directly collecting various types of raw safety data and serving as the starting point for all data acquisition. The transmission layer is the intermediate link connecting the perception layer and the processing layer. The transport layer is responsible for stably and efficiently transmitting the raw data collected by the perception layer to the processing layer, ensuring smooth data transmission. The processing layer is the core component that processes and organizes the collected data. It standardizes the raw data, removes invalid information, and provides usable data for subsequent use.

[0026] Further, step S10 includes: acquiring mine geological exploration data, historical accident distribution data, and work area planning data, and determining the mine safety risk assessment results through a risk area identification algorithm; constructing a mine safety situation awareness system using a three-level architecture of perception layer-transmission layer-processing layer; deploying environmental sensing devices in environmental risk areas to monitor environmental safety parameters, deploying equipment sensing devices in key equipment parts to monitor equipment operating parameters, and deploying personnel sensing devices in personnel work areas to monitor personnel location parameters; collecting multi-dimensional safety status data in real time through environmental sensing devices, equipment sensing devices, and personnel sensing devices; and performing format unification, anomaly removal, and missing data filling on the multi-dimensional safety status data through edge computing nodes to obtain preprocessed data.

[0027] It's important to understand that the perception layer is used to deploy environmental sensing devices, equipment sensing devices, and personnel sensing devices in a differentiated manner based on the mine safety risk assessment results; the transmission layer is used for data transmission; and the processing layer is used to deploy edge computing nodes with data caching capabilities to complete data processing locally. The deployment density of environmental sensing devices, equipment sensing devices, and personnel sensing devices is positively correlated with the risk level in the risk assessment results.

[0028] Specifically, the process begins by acquiring mine geological exploration data, historical accident distribution data, and operational area planning data. A risk area identification algorithm is then used to determine the mine safety risk assessment results. This involves collecting information on mine geological structures, gas occurrence patterns, the spatiotemporal distribution characteristics of historical accidents, and the layout of mining faces. Cluster analysis or decision tree algorithms are then used to identify the spatial distribution of high-risk, medium-risk, and low-risk areas. This is crucial because the geological conditions and operational environments vary significantly across different areas of the mine, and only a comprehensive assessment based on multi-source data can scientifically determine the risk level, providing a basis for subsequent differentiated deployment. Secondly, a three-tiered architecture—perception layer, transmission layer, and processing layer—is used to construct the mine safety situation awareness system. The perception layer deploys environmental sensing devices, equipment sensing devices, and personnel sensing devices based on the aforementioned risk assessment results. The transmission layer uses a dual-mode communication method combining industrial Ethernet and 5G mobile communication technology for data transmission. The processing layer deploys edge computing nodes with data caching capabilities to process data locally. This three-tiered architecture decouples data acquisition, transmission, and processing functions, while the caching function of the edge computing nodes can handle underground network interruptions, ensuring no data loss. Then, based on the risk assessment results, environmental sensing devices are deployed in environmental risk areas to monitor environmental safety parameters, equipment sensing devices are deployed in key equipment parts to monitor equipment operating parameters, and personnel sensing devices are deployed in personnel work areas to monitor personnel location parameters. The deployment density of environmental sensing devices, equipment sensing devices, and personnel sensing devices is positively correlated with the risk level. That is, in high-risk areas where gas is prone to accumulate, one gas sensor is deployed at every first distance; in medium-risk areas, one gas sensor is deployed at every second distance; and in low-risk areas, one gas sensor is deployed at every third distance. The first distance is smaller than the second distance, and the second distance is smaller than the third distance. This is because the higher the risk level, the higher the monitoring accuracy requirement. By differentiating the density, the monitoring resources are optimized, avoiding resource waste or monitoring blind spots caused by applying equal effort. Next, multi-dimensional safety status data are collected in real time through environmental sensing devices, equipment sensing devices, and personnel sensing devices. Specifically, environmental sensing devices collect data on gas concentration, dust concentration, underground temperature, humidity, ventilation volume, and roof pressure; equipment sensing devices collect data on the operating speed and vibration amplitude of mining equipment, the load weight and operating speed of transportation equipment, the voltage and current of power supply equipment, and the air volume and air pressure of ventilation equipment; and personnel sensing devices collect data on the real-time coordinates, work area, movement trajectory, and dwell time of workers. This is because the environment, equipment, and personnel are the three core elements of mine safety, and only by monitoring them simultaneously can the safety situation be fully grasped.Finally, edge computing nodes are used to perform format unification, anomaly removal, and missing data imputation on the multi-dimensional security status data to obtain preprocessed data. Specifically, data from different protocols are converted into a unified structure using a preset data format, a preset anomaly removal criterion is used to identify and remove jump data that exceeds a reasonable range, and linear interpolation is used to imput missing data due to brief interruptions, resulting in a standardized multi-dimensional security status dataset. This is because the original data has problems such as format heterogeneity, noise interference, and packet loss during transmission, and must be preprocessed to meet the quality requirements of subsequent model inputs.

[0029] Step S20: Input the preprocessed data into the reinforcement learning decision model to obtain the optimal control strategy.

[0030] It should be noted that the reinforcement learning decision model includes a feature extraction layer, a fully connected layer, and an output layer. The feature extraction layer adopts a convolutional neural network (CNN) or recurrent neural network (RNN) structure. The CNN structure consists of multiple stacked convolutional layers, pooling layers, and activation layers. Each convolutional layer contains several convolutional kernels, with kernel sizes typically set to 3×3 or 5×5 and strides of 1 or 2. The pooling layers use max pooling or average pooling, with pooling window sizes typically set to 2×2, to reduce feature dimensionality while retaining key information. The activation layers use modified linear units (MRUs) or sigmoid functions to introduce nonlinear transformation capabilities. The RNN structure consists of several long short-term memory (LSTM) units or gated recurrent units (GRUs), with hidden layers set to 128 or 256 dimensions to capture the temporal dependencies of safety status data. This is because mine safety status data includes both spatial multi-sensor correlations and dynamic temporal evolution. CNNs excel at extracting local spatial features, while RNNs excel at extracting temporal dependent features; combining the two can comprehensively represent complex working conditions. Secondly, the fully connected layer consists of an input layer, several hidden layers, and an output layer. The dimension of the input layer matches the output dimension of the feature extraction layer. There are usually 2 to 4 hidden layers, with 256 or 512 neurons in each layer. Random deactivation regularization is used between layers to prevent overfitting, and the deactivation ratio is set to 0.3 or 0.5. The activation function is a modified linear unit. The number of neurons in the output layer is consistent with the size of the action space, and it is used to output the expected reward value of each candidate action. This is because the fully connected layer can map the extracted high-dimensional features to the action value space, the multi-layer structure can learn complex nonlinear decision boundaries, and the regularization technique can improve the generalization ability of the model. Finally, the output layer adopts either a fully connected structure or a policy network structure. The fully connected structure directly outputs the expected reward vector of each action in the discrete action space, with the dimension equal to the number of actions. The policy network structure outputs the action probability distribution, which is normalized using the softmax function. For continuous action control, a Gaussian distribution is used to output the mean and variance. Finally, the action with the highest expected reward is selected through a greedy policy, or the action is selected according to probability through a sampling policy. The output is the optimal control policy containing the action identifier and action execution parameters. This is because the output layer is the terminal of decision-making and needs to select value output or policy output according to the characteristics of the task. The greedy policy is suitable for deterministic decision-making scenarios, while the sampling policy is suitable for exploratory learning scenarios.

[0031] Furthermore, prior to step S20, the process includes: constructing an initial decision model based on a reinforcement learning algorithm, using the preprocessed data as the safety state space, and using at least one interlocking control action among ventilation equipment adjustment, gas extraction equipment start-up and shutdown, mining equipment emergency stop, personnel evacuation early warning, power supply system power outage, and waterproof gate closure as the action space, and constructing a reward function based on the safety risk level, the effect of control action execution, and the emergency control result. It is important to understand that a policy optimization-based reinforcement learning algorithm is chosen as the core algorithm for constructing the initial decision model. Specifically, a proximal policy optimization algorithm or a trust region policy optimization algorithm is adopted, and two parallel network structures, a policy network and a value network, are set up. The policy network is used to output the action probability distribution, and the value network is used to evaluate the state value. The two networks share the feature extraction layer parameters but have independent fully connected layer parameters. The pruning ratio parameter is set to 0.2 to limit the policy update magnitude, and the generalized advantage estimation parameter is set to 0.95 to balance bias and variance. This is because the policy optimization-based algorithm can directly optimize the policy parameters and avoid the accumulation of value function estimation errors. Proximal policy optimization prevents the policy update from being too large and causing performance collapse by pruning the objective function, which is more suitable for scenarios such as mine safety control that require stable policy iteration.

[0032] Next, the preprocessed data is standardized using the min-max normalization method to obtain standardized data, which is then used as the safety state space of the initial decision model. Specifically, historical minimum and maximum values ​​are calculated for multiple parameters, including gas concentration, dust concentration, temperature, humidity, ventilation volume, roof pressure, equipment speed, vibration amplitude, voltage, current, and personnel coordinates. The current values ​​are then mapped to the zero-to-one interval to eliminate differences in the dimensions and numerical ranges of different parameters. This is because reinforcement learning is sensitive to the scale of input values, and differences in dimensions can cause gradient update directions to deviate. Normalization ensures that the contributions of each dimension parameter to the decision are balanced, accelerating model convergence.

[0033] The requirements for mine safety interlocking control are defined as a discrete action space in the initial decision model. This action space includes at least one interlocking control action among ventilation equipment adjustment, gas extraction equipment start / stop, mining equipment emergency stop, personnel evacuation early warning, power supply system power outage, and waterproof gate closure. Each action corresponds to a unique action identifier. Specifically, a unique integer code is assigned to each interlocking control action, such as 0 for ventilation equipment adjustment, 1 for gas extraction equipment start / stop, 2 for mining equipment emergency stop, and so on. When an action is executed, the corresponding equipment control interface is called by decoding the action identifier. This is because mine safety control actions have clear discreteness and mutual exclusion; the discrete action space facilitates the output probability distribution of the strategy network; and the unique identifier enables precise routing of control commands.

[0034] A base reward value is set based on the safety risk level. A change reward value is set based on the change in safety risk level before and after the control action is executed. An emergency reward value is set based on the execution result of the emergency interlock control process. The safety risk level is divided into three levels: high risk, medium risk, and low risk, based on preset risk thresholds. Specifically, for example, a high risk is defined as a gas concentration greater than or equal to the first threshold or a roof pressure greater than or equal to the second threshold; a medium risk is defined as a gas concentration greater than or equal to the third threshold but less than the first threshold; and a low risk is defined as a gas concentration less than the third threshold. The base reward value for high risk is -50, for medium risk it is 0, and for low risk it is +50. This is because the risk level directly reflects the safety situation, the base reward value provides the model with immediate state value feedback, and the distinction between positive and negative rewards guides the model towards a safe state.

[0035] The reward function of the initial decision model is constructed by weighting and summing the base reward value, the variable reward value, and the emergency reward value. The specific formula for the reward function R is expressed as follows: R=α·R_base+β·R_change+γ·R_emergency Wherein, R_base is the base reward value corresponding to the safety risk level. According to the mine safety regulations, safety risks are divided into three levels: high, medium, and low, corresponding to base reward values ​​of -50, 0, and 50 respectively; R_change is the reward for the change in safety risk level caused by the control action. If the safety risk level decreases after the control action, R_change is 30; if the level increases, R_change is -80; if the level remains unchanged, R_change is 0; R_emergency is the reward for emergency interlock control. If the accident is successfully avoided after triggering the emergency interlock procedure, R_emergency is 100; otherwise, it is -200; α, β, and γ are weighting coefficients, which are determined according to the actual working conditions and satisfy α+β+γ=1. For example, if the base reward value weight α is set to 0.3, the variable reward value weight β to 0.5, and the emergency reward value weight γ to 0.2, the variable reward value is +30 when the control action reduces the risk level, -80 when the control action increases the risk level, and 0 when the control action does not change the risk level. The emergency reward value is +100 when the emergency interlock control successfully avoids the accident, and -200 when the emergency interlock control fails to avoid the accident. The single-step reward value is obtained by weighting and summing the three items. This is because different reward components reflect different dimensions of control objectives, and the weight allocation reflects the priority difference. The variable reward weight is the highest, which can incentivize the model to actively reduce risk, and the setting of emergency rewards can strengthen the correct response in extreme scenarios.

[0036] Then, historical safe operation data, accident case data, and simulated working condition data of the mine are acquired. It's important to understand that extracting historical safe operation data from the mine safety monitoring database involves collecting continuous time-series records of gas concentration, dust concentration, temperature, humidity, ventilation volume, roof pressure, equipment operating parameters, and personnel location parameters from various monitoring points underground over the past three to five years. These records are then aligned by timestamps to form a multi-dimensional state sequence. This is because historical normal operation data constitutes the vast majority of the samples, providing the model with a decision-making benchmark under normal working conditions and preventing the model from overfitting to extreme scenarios. Secondly, accident case data is extracted from the accident record management system. This involves collecting complete monitoring records of major accidents such as gas explosions, roof collapses, and dust explosions, including the state evolution sequence within a preset time period before the accident, the control measures taken, and the accident consequence level. Accident cases are labeled to form hazard state tags. This is because accident cases are scarce but high-value data, clearly indicating to the model which state combinations pose a fatal risk, strengthening the model's ability to identify hazard patterns. Then, simulated operating condition data is generated based on physical simulation models or digital twin platforms. This involves setting the variation range of physical parameters such as gas outburst rate, roof stress distribution, and ventilation network resistance. State evolution trajectories under normal, abnormal, and extreme operating conditions are generated through fluid dynamics simulation or discrete element simulation, supplementing boundary conditions not covered in actual monitoring. This is because real mines do not actively create hazards to collect data; simulation can generate a large number of extreme scenario samples at low cost, improving the model's generalization ability under rare operating conditions. Finally, historical safe operation data, accident case data, and simulated operating condition data are mixed according to a preset ratio. Historical safe operation data accounts for 70% to ensure routine decision-making capabilities, while accident case data and simulated operating condition data together account for 30% to enhance risk response capabilities, forming a complete training dataset. This is because the three types of data each have value and are unevenly distributed; mixed sampling can balance the model's conservatism and aggressiveness, ensuring that the model neither misses hazards nor overreacts.

[0037] Finally, the initial decision model was trained offline based on historical mine safety operation data, accident case data, and simulated working condition data to obtain a reinforcement learning decision model. Further, the historical mine safety operation data, accident case data, and simulated working condition data were standardized and preprocessed to obtain a training dataset. Specifically, the same minimum-maximum normalization method as the state space was used to map the three types of data to the [0,1] interval. The 3σ criterion was used to remove outliers caused by sensor malfunctions, and linear interpolation was used to fill in missing values ​​caused by communication interruptions. The processed data was divided into state transition sequences according to time windows. Each sequence contains the current state, the action performed, the state after the transition, and the immediate reward value, forming an experience sample unit. This is because standardized preprocessing ensures that the distribution of training data is consistent with that of real-time inference data, and the structured storage of experience sample units facilitates subsequent batch sampling and replay training.

[0038] Initialize the model parameters of the initial decision model, setting the maximum number of iterations and the initial learning rate. Specifically, use a normal distribution to randomly initialize the weight matrices of the policy network and the value network, initialize the bias term to 0, set the maximum number of iterations to 10,000, the initial learning rate to 0.001, set the experience replay pool capacity to 100,000 samples, and set the target network update frequency to 100 iterations. This is because reasonable initialization can avoid gradient vanishing or gradient exploding, a low learning rate combined with a sufficient number of iterations can ensure stable model convergence, and the experience replay pool can break the temporal correlation between samples.

[0039] The training dataset is input into the initial decision model, and random perturbations are added to simulate the random changes in mining conditions, resulting in a cumulative reward value. Specifically, in each training batch, Gaussian noise or a random mask is applied to the input state features with a preset probability to simulate sensor measurement errors, equipment parameter drift, and sudden environmental changes. The action probability distribution is output through the policy network, and the immediate reward value is calculated according to the reward function after the action is sampled and executed. The discounted cumulative reward value is obtained by accumulating over time steps. This is because mining conditions have inherent randomness and uncertainty, and adding perturbations during training can enhance the robustness of the model, making it tolerant of data noise during actual deployment.

[0040] The optimized model is obtained by updating the model parameters based on the first learning rate. Specifically, stochastic gradient descent or adaptive moment estimation algorithm is used. The policy gradient is calculated on the policy network parameters based on the cumulative reward value, and the value gradient is calculated on the value network parameters based on the temporal difference error. The weights are updated in the opposite direction of the gradient with the first learning rate. After every 10 iterations of the target network update frequency, the main network parameters are copied to the target network. This is because the policy gradient method can directly optimize the expected cumulative reward, and the delayed update of the target network can reduce the bias of the bootstrap estimation and improve the training stability.

[0041] When the cumulative reward value stabilizes or the maximum number of iterations is reached, the optimized model at this point is used as the reinforcement learning decision model. Specifically, when the fluctuation range of the cumulative reward value within a preset number of iterations is less than a preset threshold or the maximum number of iterations is reached, the optimized model at this point is used as the reinforcement learning decision model. This means saving the weight parameters, network structure configuration, and hyperparameter settings of the current policy network and value network to form a deployable model file. This is because the convergence of the cumulative reward value indicates that the model has learned a stable decision policy, the preset termination condition can prevent overfitting caused by overtraining, and saving the complete model facilitates subsequent loading, deployment, and incremental optimization.

[0042] Step S30: Generate control commands based on the optimal control strategy.

[0043] Specifically, the process begins by parsing the action identifiers in the optimal control strategy to determine the target control object and control type. This involves reading the action code field from the strategy output and using a lookup table to determine the type of interlocking device corresponding to that code. For example, action code 0 corresponds to a local ventilation fan controller, action code 1 to a gas extraction pump controller, and action code 2 to a tunneling machine power-off relay. This is because the action identifier is an abstract representation of the model decision and must be converted into a specific device address to execute physical control. Secondly, the action execution parameters in the optimal control strategy are parsed to generate control parameter frames that the device can recognize. This involves extracting numerical parameters such as the percentage of airflow adjustment, the warning duration, and the coordinates of the power cut-off area based on the action type. These parameters are then encapsulated into Modbus messages or CAN bus data frames according to the device communication protocol. For instance, increasing ventilation by 20% is converted into a frequency converter setting value, and personnel evacuation warnings are converted into the trigger duration and broadcast content of an audible and visual alarm. This is because different interlocking devices use different communication protocols and control interfaces, and standardized encapsulation of parameter frames enables unified scheduling of heterogeneous devices. Then, the action identifier, target device address, and control parameter frames are assembled into a complete control command, with a timestamp, priority identifier, and checksum added to form a control command packet to be sent. This involves setting either an emergency priority or a regular priority for the control command packet. Emergency priority is used in hazardous scenarios such as gas exceeding limits, ensuring the command skips the queue and is sent directly. The checksum uses cyclic redundancy check or parity check to verify data integrity at the receiving end. This is because underground mine communication is subject to interference and packet loss risks; the integrity verification and priority marking of the command packet ensure the reliability and real-time performance of critical controls. Finally, the control command packet is sent through the transport layer to the corresponding safety interlocking device, triggering the device's actuators to complete physical actions. Specifically, industrial Ethernet or 5G mobile communication technology routes the command packet to the target device's programmable logic controller (PLC). After parsing the command, the controller drives actuators such as the ventilation fan inverter, gas extraction pump contactor, or tunneling machine circuit breaker to complete state switching and feeds the execution results back to the safety interlocking control system. This is because control commands must ultimately be translated into physical device actions; the closed-loop feedback of the execution results confirms the control's effectiveness and provides a basis for subsequent state assessment and model optimization.

[0044] Step S40: Send control commands to the corresponding safety interlocking device so that the safety interlocking device can perform interlocking actions.

[0045] Specifically, firstly, a communication session is established between the safety interlock control system and each safety interlock device. This involves establishing a connection with the programmable logic controller or remote terminal unit of the target device through a preset communication protocol stack, activating the device's listening state, and preparing to receive control commands. This is because there are a large number of heterogeneous devices in underground mines, and a stable communication link must be established in advance to ensure reliable command transmission and avoid command loss due to incomplete connection during transmission. Secondly, the control commands are sent to the corresponding safety interlock devices through the transport layer. This involves routing the control commands to the device controller in a fixed area underground via industrial Ethernet, or using 5G mobile communication technology to fill in the gaps in the transmission of control commands to portable devices in the mobile work area. Before transmission, the control commands are encoded, modulated, and error-controlled. After transmission, the system waits for a reception confirmation response from the device. If no confirmation is received, the system automatically retransmits within a preset number of retransmissions. This is because the underground communication environment is complex, and electromagnetic interference and signal attenuation may cause transmission failures. Dual-mode communication and automatic retransmission mechanisms can improve the command arrival rate. Then, the safety interlocking equipment receives and parses the control commands, verifying their integrity and authorization. Specifically, the programmable logic controller (PLC) at the equipment end extracts the checksum from the command packet and compares it with a locally calculated checksum to confirm the data has not been tampered with. It also compares the sender's identifier with a preset authorization list to confirm the command's legal origin. After successful verification, the action identifier and action execution parameters are extracted. This is because mine safety control involves critical infrastructure, and it is essential to prevent safety accidents caused by misoperation or malicious commands. Integrity verification and authorization verification are necessary safety barriers before equipment execution. Finally, the safety interlocking equipment drives the actuator to complete the interlocking action based on the parsed action identifier and action execution parameters. This involves the controller outputting a switch signal to drive a relay to cut off the tunneling machine's power supply, or an analog signal to adjust the frequency of the ventilation fan's inverter, or a digital signal to trigger the audible and visual alarm and broadcast system. After execution, the equipment status change and execution timestamp are fed back to the safety interlocking control system. This is because control commands must ultimately be converted into physical actions to produce actual safety effects. Status feedback forms a control closed loop, providing data support for subsequent safety status assessments and model optimization.

[0046] Further, step S40 includes: collecting safety status feedback data after the safety interlocking equipment performs its actions. Specifically, firstly, safety status feedback data after the safety interlocking equipment performs its actions is collected. This involves rereading data from various sensors through the perception layer of the mine safety situation awareness system, including gas concentration, dust concentration, roof pressure, equipment operating parameters, and personnel location parameters. The collection time is aligned with the action execution time to form a state transition pair. This is because feedback data is the foundation for evaluating control effectiveness and model optimization; a clear time correspondence must be established with action execution to accurately calculate the reward value. Secondly, when the safety status feedback data does not exceed the emergency threshold for safety status, incremental learning is used to iteratively optimize the reinforcement learning decision model online using the safety status feedback data, resulting in an updated reinforcement learning decision model. It is important to understand that the safety status feedback data is used as new training samples input into the reinforcement learning decision model to calculate the deviation between the predictive control strategy and the optimal control strategy under the current model parameters. Specifically, the state features from the feedback data are input into the current reinforcement learning decision model. Forward propagation yields the action probability distribution or value estimate of the model's output in that state. This is compared with the optimal control policy determined through Monte Carlo tree search or expert experience. Cross-entropy loss or mean squared error is calculated as a deviation metric. This deviation calculation is a prerequisite for gradient updates; only by clearly identifying the gap between the model's prediction and the optimal policy can parameters be adjusted in a targeted manner. Next, based on the gradient update direction of the model parameters calculated by the deviation, the initial fully connected layer parameters of the reinforcement learning decision model are locally incrementally updated. A second learning rate is set for incremental learning, and parameter fine-tuning is performed using a single iteration or a small number of iterations. The fine-tuned parameters are used as the updated model parameters, where the second learning rate is less than the first learning rate during offline training. Specifically, based on the gradient update direction of the bias calculation model parameters, the initial fully connected layer parameters of the reinforcement learning decision model are locally incrementally updated. This involves calculating the partial derivatives of the bias with respect to the weights of each neuron in the fully connected layer using the backpropagation algorithm to obtain the gradient vector. The feature extraction layer parameters are frozen to maintain the stability of the underlying feature representation. Only the gradient is applied to the weight matrix and bias term of the fully connected layer. A second learning rate of 0.001 is set for incremental learning, and parameter fine-tuning is performed using a single iteration or a small number of iterations. The fine-tuned parameters are used as the updated model parameters. The second learning rate is lower than the first learning rate during offline training because the fully connected layer directly determines action selection. Local updates can quickly adapt to changes in operating conditions, while a low learning rate and a small number of iterations prevent a single sample from excessively affecting the model, avoiding policy oscillations under normal operating conditions. Then, the updated model parameters are weighted and fused with the initial fully connected layer parameters to obtain the fused model parameters.Specifically, the fusion weights are set to 0.2 and 0.8. The updated parameters are multiplied by 0.2, and the initial fully connected layer parameters are multiplied by 0.8. The two are then added element-wise to obtain the fusion parameters, which serve as the new fully connected layer weights. This is because completely replacing the parameters may lead to the forgetting of historical experience, while weighted fusion can retain most of the historical knowledge while absorbing new feedback, ensuring the continuity and stability of the model's behavior. Finally, the reinforcement learning decision model is iteratively optimized online based on the fused model parameters to obtain an updated reinforcement learning decision model. Specifically, the fused parameters are loaded into the model's fully connected layer, replacing the original parameters, forming an updated reinforcement learning decision model for the next real-time decision. The updated parameters, corresponding state samples, and timestamps are recorded in the model version log. This is because online iterative optimization is the core mechanism for the continuous evolution of the model, and the version log can trace the model's change history, facilitating anomaly diagnosis and rollback recovery.

[0047] When the safety status feedback data exceeds the emergency threshold, the emergency interlock control process is triggered, online iterative optimization is paused, and the highest priority safety protection action is executed. This action includes at least one of the following: cutting off power to the entire area, activating emergency evacuation warnings, and closing safety isolation doors. When the safety status feedback data recovers and does not exceed the emergency threshold, incremental learning is used to iteratively optimize the reinforcement learning decision model online using the safety status feedback data, resulting in an updated reinforcement learning decision model. Specifically, the safety status feedback data is compared item by item with the emergency threshold to determine if any emergency risk parameters exceed the threshold. This involves reading key parameters such as gas concentration, roof pressure, and ventilation volume from the feedback data and comparing them in real time with preset emergency thresholds for gas concentration and roof pressure. If any parameter exceeds the corresponding threshold, an emergency state is triggered. This is because mine safety follows the strict principle of alarming when a single parameter exceeds its limit; parallel comparison of multiple parameters ensures that no single hazard indicator is missed, achieving comprehensive coverage of hazard identification. Secondly, when the safety status feedback data exceeds the emergency threshold, the emergency interlock control process is immediately triggered, online iterative optimization is paused, and the highest priority safety protection action is executed. This means interrupting the ongoing model inference calculation and incremental learning process, switching all computing resources to the emergency response channel, and directly sending hard-coded emergency control commands to the corresponding safety interlock devices. Priority is given to executing at least one of the following actions: cutting off the power supply to the entire area to eliminate ignition sources, activating emergency evacuation warnings to guide personnel to escape, and closing safety isolation doors to prevent the spread of disaster. This is because millisecond-level response is crucial in an emergency. The inference process of the intelligent model has computational delays, and the preset hard-coded emergency process can bypass complex decision-making and directly trigger the highest level of protection. Pausing optimization can prevent emergency data from polluting model parameters and causing subsequent decision-making biases. Then, the recovery of safety status feedback data is continuously monitored to determine whether each parameter has fallen back below the emergency threshold of the safety status. That is, sensor data is continuously read at a higher sampling frequency, and the average value of parameters within the sliding window is calculated to smooth out instantaneous fluctuations. When the gas concentration is continuously below 1.5% and the roof pressure is continuously below 2MPa for a preset safe duration, the dangerous situation is determined to be lifted. This is because the instantaneous drop in parameters may have measurement noise. Verification of the average value of the sliding window and the duration can prevent frequent fluctuations near the threshold from causing the system to switch repeatedly, ensuring the stability of the status determination.Finally, when the safety status feedback data is determined to have recovered to below the emergency threshold, incremental learning is used to iteratively optimize the reinforcement learning decision model online using the safety status feedback data, resulting in an updated reinforcement learning decision model. This restarts the model inference and incremental learning process, marking high-risk state samples collected during the emergency as special training data and assigning them higher reward weights for inclusion in subsequent incremental learning. This strengthens the model's understanding of such extreme conditions while resuming regular optimization iterations at a normal frequency. This is because emergency data contains valuable information on extreme conditions, and delaying its inclusion in training can improve the model's ability to identify and respond to dangerous patterns while ensuring safety, achieving a balance between safety and model evolution.

[0048] This embodiment employs a three-tiered architecture of perception, transmission, and processing to construct a mine safety situation awareness system. It collects and preprocesses multi-dimensional safety data; a decision-making model is built based on reinforcement learning, using safety data as the state space, interlocking actions as the action space, and safety risk as the reward function. This model is trained offline to obtain the reinforcement learning decision-making model. Then, the optimal strategy is output through a three-layer network to drive the actions of safety interlocking equipment, achieving automatic control and closed-loop optimization. This constructs an integrated mine safety interlocking control architecture and realizes emergency interlocking closed-loop control based on multi-dimensional safety data, improving the intelligence level and response speed of mine safety interlocking control, reducing reliance on manual labor, and ensuring mine production safety.

[0049] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The reinforcement learning-based mine safety interlocking control method further includes steps S201-S203 in step S20: Step S201: Input the preprocessed data into the feature extraction layer of the reinforcement learning decision model to obtain the current security state.

[0050] It should be noted that step S201 includes: extracting features from environmental parameters in the preprocessed data through a feature extraction layer to obtain environmental safety features; extracting features from equipment operation parameters in the preprocessed data through a feature extraction layer to obtain equipment operation features; extracting features from personnel location parameters in the preprocessed data through a feature extraction layer to obtain personnel location features; performing multi-dimensional feature fusion on environmental safety features, equipment operation features, and personnel location features to obtain comprehensive safety features; and determining the comprehensive safety features as the current safety status used to characterize the overall safety status of the mine.

[0051] Specifically, firstly, a feature extraction layer uses a convolutional neural network or a fully connected network to perform nonlinear transformations on environmental parameters such as gas concentration, dust concentration, temperature, humidity, ventilation volume, and roof pressure. This extracts deep features such as gas accumulation trends, dust diffusion patterns, and ventilation effectiveness, outputting a fixed-dimensional environmental safety feature vector. This is because environmental parameters are the most direct representation of mine safety, but raw sensor data contains noise and redundancy. Feature extraction can extract key indicators sensitive to safety risks; for example, the spatial gradient of gas concentration reflects accumulation risk better than a single-point value. Secondly, a recurrent neural network or a one-dimensional convolutional network dynamically encodes time-series parameters such as mining equipment rotation speed, vibration amplitude, transport equipment load, and power supply voltage and current. This extracts deep features such as abnormal vibration modes, load mutation characteristics, and power supply stability trends, outputting a fixed-dimensional equipment operation feature vector. This is because equipment failure is often the cause of safety accidents, and time-series feature extraction can capture the dynamic evolution of parameters; for example, abnormal harmonics in the vibration spectrum are better at predicting mechanical failures than instantaneous amplitude. Then, the embedded layer or graph neural network encodes the real-time coordinates of personnel, work area, movement trajectory, and dwell time, extracting deep features such as personnel density distribution, proximity to dangerous areas, and accessibility of evacuation routes, and outputting a fixed-dimensional personnel location feature vector. This is because personnel safety is the core objective of mine safety, and location features can quantify the spatiotemporal relationship between personnel and danger. For example, the trend of personnel movement towards dangerous areas is a better predictor of collision risk than static coordinates. Next, the attention mechanism or gating fusion network calculates the weight coefficients of the three types of features, and the feature vectors are weighted, summed, or concatenated before being dimensionality reduced through a fully connected layer to generate a unified comprehensive safety feature. This is because mine safety is the result of the coupling of multiple factors such as environment, equipment, and personnel. Isolated analysis of any dimension may miss associated risks. For example, when gas exceedance and personnel being stranded in dangerous areas occur simultaneously, the risk level increases exponentially. Fusion features can characterize this non-linear coupling relationship. Finally, the integrated safety features are defined as the current safety state used to characterize the overall safety status of the mine. The integrated safety features are used as an instantiation vector of the state space and input into the subsequent decision network for action selection. This is because the current safety state is the input basis for reinforcement learning decision-making. Only by comprehensively reflecting the multi-dimensional safety situation can the model output the globally optimal control strategy and avoid local optima or secondary risks caused by single parameter optimization.

[0052] Step S202: Input the current security state into the fully connected layer to calculate the expected reward value of each interlocking control action, and obtain the expected reward value of each interlocking control action.

[0053] It should be noted that step S202 includes: performing feature weighting calculation on the current safety state to obtain weighted safety feature information; matching and associating the weighted safety feature information with various interlocking control actions to obtain a set of interlocking control actions to be evaluated; calculating the evaluation score of each interlocking control action by comparing it with the set of interlocking control actions to be evaluated according to the reward function; and performing numerical conversion based on the evaluation score of each interlocking control action to obtain the expected reward value of each interlocking control action.

[0054] Specifically, firstly, the weight coefficients of each dimension of features in the current safety state are calculated using an attention mechanism. For example, the weight of gas concentration is set to 0.35, roof pressure to 0.25, personnel density to 0.20, and equipment vibration to 0.20. The feature vector is multiplied element-by-element by the weight coefficients and then summed to obtain a weighted feature representation highlighting key risk factors. This is because different safety parameters have different degrees of impact on the overall risk; the direct lethality of excessive gas is higher than that of abnormal humidity. Weighted calculation can strengthen the decision-making influence of high-risk factors and avoid secondary parameters interfering with core judgments. Secondly, a state-action mapping matrix is ​​established. Candidate actions are screened based on the dominant risk type in the weighted features. For example, when gas is dominant, four actions are associated: ventilation adjustment, gas extraction, power outage, and evacuation. When the roof is dominant, three actions are associated: support reinforcement, personnel evacuation, and area sealing. Waterproof gate actions that are irrelevant to the current risk are excluded, forming a simplified set to be evaluated. This is because the computational cost of full-action space evaluation is large and most actions are irrelevant to the current state. Matching and association can narrow the evaluation scope and improve the real-time nature of decision-making. Then, each candidate action is substituted into the reward function to calculate the base reward value, variable reward value, and emergency reward value. For example, the base reward value for ventilation adjustment is 0 points, corresponding to medium risk; the variable reward value for reducing the gas concentration from 0.8% to 0.4% is +30 points; and since there is no emergency scenario, the emergency reward value is 0 points. The weighted sum of the three values ​​yields the total evaluation score for the action. This is because the reward function is the core guide for strategy optimization, and benchmarking each action individually quantifies the expected benefit, providing a numerical basis for subsequent selection. Finally, the evaluation scores are normalized using softmax or linearly scaled to convert them into a probability distribution form or standardized values. For example, the ventilation adjustment action has a score of 60 points, the gas extraction action has a score of 45 points, and the power outage action has a score of 25 points. After softmax transformation, the probability values ​​are 0.52, 0.35, and 0.13, respectively. Alternatively, the original scores can be directly output as Q-values. This is because numerical transformation unifies the units of measurement for easier comparison; the probabilistic form is suitable for random strategy sampling, while the Q-value form is suitable for deterministic greedy selection, meeting the needs of different decision-making scenarios.

[0055] Step S203: Input the expected reward value into the output layer, select the action with the highest expected reward value as the optimal control action, and output the corresponding optimal control strategy.

[0056] It should be noted that step S203 includes: sorting all reward expectation values ​​numerically to obtain a reward expectation value sequence arranged from high to low; selecting the item with the largest value in the reward expectation value sequence as the optimal reward expectation value; associating and matching the optimal reward expectation value with the corresponding interlocking control action to obtain the optimal control action; generating a complete control execution plan based on the optimal control action to obtain the corresponding optimal control strategy. The optimal control strategy includes the action identifier and action execution parameters of the optimal control action. The action execution parameters include at least one of the following: ventilation equipment airflow adjustment value, warning duration, and power supply cut-off range.

[0057] Specifically, firstly, quicksort or heapsort algorithms are used to sort the expected reward values ​​of each interlocking control action in descending order, forming an ordered sequence, such as ventilation adjustment 0.52, gas extraction 0.35, personnel evacuation 0.08, and power outage 0.05. This is because sorting can intuitively present the value hierarchy of each action, providing a structured index for quickly locating the optimal action and avoiding computational redundancy caused by unordered comparisons. Secondly, the item with the largest value in the expected reward value sequence is selected as the optimal expected reward value, that is, the first element of the sorted sequence is read. For example, 0.52 corresponding to ventilation adjustment is directly extracted as the optimal expected reward value. This is because the core logic of the greedy strategy is to select the action with the highest value in the current state. Extracting the first element ensures the local optimality of the decision and maximizes the expected return of a single step. Then, the optimal reward expectation value is associated and matched with the corresponding interlocking control action to obtain the optimal control action. That is, through the action identifier index table, the value 0.52 is back-mapped to the ventilation adjustment action, confirming that the action code of this action is A2, the equipment address is the local ventilation fan frequency converter, and the control type is air volume increase. This is because the reward expectation value is only an abstract value and must be associated with a specific action entity to form an executable instruction. Index matching can achieve accurate conversion from value to physical action. Finally, a complete control execution plan is generated based on the optimal control action to obtain the corresponding optimal control strategy. The optimal control strategy includes the action identifier of the optimal control action and the action execution parameters. The action execution parameters include at least one of the following: ventilation equipment air volume adjustment value, warning duration, and power supply cut-off range. That is, parameters such as action code A2, equipment address, air volume increase percentage of 20%, and adjustment duration of 30 minutes are encapsulated into a structured data frame, and a priority mark of normal level, timestamp, and cyclic redundancy check code are added to form a complete control instruction package to be issued. This is because only action selection without execution details cannot drive physical equipment. The refinement of action execution parameters can ensure the accurate implementation of the control strategy, and the check code ensures the reliability of transmission.

[0058] This embodiment sequentially inputs preprocessed data into the feature extraction layer, fully connected layer, and output layer of a reinforcement learning decision model. Through layered processing, the current safety state and the expected reward values ​​of each interlocking control action are obtained. The action with the highest expected reward value is then selected to determine the optimal control strategy. The strategy includes action identifiers and action execution parameters, and can accurately adapt to scenarios such as ventilation adjustment, personnel early warning, and power cut-off. This achieves automated and precise safety decision-making, improves response speed, reduces manual intervention, and enhances the reliability and adaptability of mine safety control.

[0059] Based on the first embodiment of this application, this application also provides a mine safety interlocking control device based on reinforcement learning, please refer to... Figure 3 The device includes: The acquisition module 10 is used to build a mine safety situation awareness system, collect and process multi-dimensional safety status data, and obtain pre-processed data. The multi-dimensional safety status data includes environmental parameters, equipment operating parameters and personnel location parameters. The mine safety situation awareness system adopts a three-level architecture of perception layer-transmission layer-processing layer.

[0060] The strategy generation module 20 is used to input preprocessed data into a reinforcement learning decision model to obtain the optimal control strategy. The reinforcement learning decision model includes a feature extraction layer, a fully connected layer, and an output layer. It is also used to construct an initial decision model based on a reinforcement learning algorithm. The preprocessed data is used as the safety state space, and at least one interlocking control action among ventilation equipment adjustment, gas extraction equipment start / stop, mining equipment emergency stop, personnel evacuation early warning, power supply system power outage, and waterproof gate closure is used as the action space. A reward function is constructed based on the safety risk level, the effect of control action execution, and the emergency control result. Historical mine safety operation data, accident case data, and simulated working condition data are acquired. The initial decision model is trained offline based on the historical mine safety operation data, accident case data, and simulated working condition data to obtain the reinforcement learning decision model.

[0061] The instruction generation module 30 is used to generate control instructions based on the optimal control strategy.

[0062] The execution module 40 is used to send control commands to the corresponding safety interlocking device so that the safety interlocking device can perform interlocking actions.

[0063] The mine safety interlocking control device based on reinforcement learning provided in this application, employing the mine safety interlocking control method based on reinforcement learning in the above embodiments, can solve the technical problem of how to improve the intelligence level and response speed of mine safety interlocking control. Compared with the prior art, the beneficial effects of the mine safety interlocking control device based on reinforcement learning provided in this application are the same as those of the mine safety interlocking control method based on reinforcement learning provided in the above embodiments, and other technical features in the mine safety interlocking control device based on reinforcement learning are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0064] This application provides a mine safety interlocking control device based on reinforcement learning. The mine safety interlocking control device based on reinforcement learning includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the mine safety interlocking control method based on reinforcement learning in the first embodiment described above.

[0065] The following is for reference. Figure 4 The diagram illustrates a structural schematic of a reinforcement learning-based mine safety interlocking control device suitable for implementing embodiments of this application. The reinforcement learning-based mine safety interlocking control device in this application can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and vehicle terminals (e.g., vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 4 The reinforcement learning-based mine safety interlocking control device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments in this application.

[0066] like Figure 4As shown, the reinforcement learning-based mine safety interlocking control device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the reinforcement learning-based mine safety interlocking control device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the reinforcement learning-based mine safety interlocking control equipment to communicate wirelessly or wiredly with other devices to exchange data. Although various reinforcement learning-based mine safety interlocking control devices are shown in the figures, it should be understood that implementation or possession of all of them is not required. More or fewer may be implemented alternatively.

[0067] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0068] The mine safety interlocking control device based on reinforcement learning provided in this application, employing the mine safety interlocking control method based on reinforcement learning in the above embodiments, can solve the technical problem of how to improve the intelligence level and response speed of mine safety interlocking control. Compared with the prior art, the beneficial effects of the mine safety interlocking control device based on reinforcement learning provided in this application are the same as the beneficial effects of the mine safety interlocking control method based on reinforcement learning provided in the above embodiments, and other technical features in this mine safety interlocking control device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0069] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0070] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0071] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the reinforcement learning-based mine safety interlocking control method in the above embodiments.

[0072] The computer-readable medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0073] The aforementioned computer-readable medium may be included in a reinforcement learning-based mine safety interlocking control device; or it may exist independently and not be assembled into a reinforcement learning-based mine safety interlocking control device.

[0074] The aforementioned computer-readable medium carries one or more programs that, when executed by a reinforcement learning-based mine safety interlocking control device, enable the device to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0075] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, all blocks in the flowcharts or block diagrams may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that all blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0076] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0077] The readable medium provided in this application is a computer-readable medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-described reinforcement learning-based mine safety interlocking control method. This addresses the technical problem of how to improve the intelligence level and response speed of mine safety interlocking control. Compared with the prior art, the beneficial effects of the computer-readable medium provided in this application are the same as those of the reinforcement learning-based mine safety interlocking control method provided in the above embodiments, and will not be elaborated upon here.

[0078] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the reinforcement learning-based mine safety interlocking control method described above.

[0079] The computer program product provided in this application can solve the technical problem of how to improve the intelligence level and response speed of mine safety interlock control. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the reinforcement learning-based mine safety interlock control method provided in the above embodiments, and will not be repeated here.

[0080] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A mine safety interlocking control method based on reinforcement learning, characterized in that, The method includes: A mine safety situation awareness system is constructed to collect and process multi-dimensional safety status data to obtain pre-processed data. The multi-dimensional safety status data includes environmental parameters, equipment operating parameters, and personnel location parameters. The mine safety situation awareness system adopts a three-level architecture of perception layer, transmission layer, and processing layer. The preprocessed data is input into a reinforcement learning decision model to obtain the optimal control strategy, wherein the reinforcement learning decision model includes a feature extraction layer, a fully connected layer and an output layer. Generate control commands based on the optimal control strategy; The control command is sent to the corresponding safety interlock device so that the safety interlock device performs the interlock action; Before the step of inputting the preprocessed data into the reinforcement learning decision model to obtain the optimal control strategy, the following steps are included: An initial decision model is constructed based on a reinforcement learning algorithm. The preprocessed data is used as the safety state space, and at least one interlocking control action among ventilation equipment adjustment, gas extraction equipment start-up and shutdown, mining equipment emergency stop, personnel evacuation early warning, power supply system power outage and waterproof gate closure is used as the action space. A reward function is constructed based on the safety risk level, the effect of control action execution and the emergency control result. Acquire historical safe operation data, accident case data, and simulated working condition data of the mine; The initial decision model is trained offline based on the mine's historical safe operation data, accident case data, and simulated working condition data to obtain a reinforcement learning decision model.

2. The method as described in claim 1, characterized in that, The steps for constructing a mine safety situation awareness system, collecting and processing multi-dimensional safety status data, and obtaining preprocessed data include: Acquire mine geological exploration data, historical accident distribution data, and operational area planning data, and determine the mine safety risk assessment results through risk area identification algorithms; A three-tier architecture of perception layer, transmission layer, and processing layer is adopted to construct a mine safety situation awareness system. The perception layer is used to deploy environmental perception devices, equipment perception devices, and personnel perception devices in a differentiated manner according to the mine safety risk assessment results. The transmission layer is used for data transmission. The processing layer is used to deploy edge computing nodes with data caching function to complete data processing nearby. Based on the risk assessment results, environmental sensing devices are deployed in environmental risk areas to monitor environmental safety parameters, equipment sensing devices are deployed in key equipment parts to monitor equipment operating parameters, and personnel sensing devices are deployed in personnel work areas to monitor personnel location parameters. The deployment density of the environmental sensing devices, equipment sensing devices, and personnel sensing devices is positively correlated with the risk level in the risk assessment results. The environmental sensing device, equipment sensing device, and personnel sensing device collect multi-dimensional safety status data in real time. The multi-dimensional security status data is processed by edge computing nodes to unify the format, remove anomalies, and fill in missing data, resulting in preprocessed data.

3. The method as described in claim 1, characterized in that, The steps for constructing the initial decision model based on the reinforcement learning algorithm include: The initial decision model is constructed by selecting a policy optimization-based reinforcement learning algorithm as the core algorithm. The preprocessed data is standardized using the min-max normalization method to obtain standardized data, and the standardized data is used as the safe state space of the initial decision model. The requirements for mine safety interlocking control are defined as the discrete action space of the initial decision model. The action space includes at least one interlocking control action among ventilation equipment adjustment, gas extraction equipment start-up and shutdown, mining equipment emergency stop, personnel evacuation early warning, power supply system power failure and waterproof gate closure. Each action corresponds to a unique action identifier. A base reward value is set based on the safety risk level, a change reward value is set based on the change in safety risk level before and after the control action is executed, and an emergency reward value is set based on the execution result of the emergency interlock control process. The safety risk level is divided into three levels: high risk, medium risk, and low risk based on a preset risk threshold. The reward function of the initial decision model is constructed by weighting and summing the base reward value, the variable reward value, and the emergency reward value.

4. The method as described in claim 1, characterized in that, The step of offline training the initial decision model based on the mine's historical safe operation data, accident case data, and simulated working condition data to obtain a reinforcement learning decision model includes: The historical safe operation data, accident case data, and simulated working condition data of the mine are standardized and preprocessed to obtain a training dataset. Initialize the model parameters of the initial decision model, and set the maximum number of iterations and the first learning rate; The training dataset is input into the initial decision model, and random perturbations are added to simulate random changes in mining conditions to obtain the cumulative reward value. The model parameters are updated based on the first learning rate to obtain the optimized model; When the cumulative reward value stabilizes or the number of iterations reaches the maximum number of iterations, the optimized model at this point is used as the reinforcement learning decision model.

5. The method as described in claim 1, characterized in that, The step of inputting the preprocessed data into the reinforcement learning decision model to obtain the optimal control strategy includes: The preprocessed data is input into the feature extraction layer of the reinforcement learning decision model to obtain the current security state. The current security state is input into the fully connected layer to calculate the expected reward value of each interlocking control action, thus obtaining the expected reward value of each interlocking control action; The expected reward value is input into the output layer, the action with the highest expected reward value is selected as the optimal control action, and the corresponding optimal control strategy is output. The optimal control strategy includes the action identifier and action execution parameters of the optimal control action. The action execution parameters include at least one of the following: ventilation equipment air volume adjustment value, warning duration, and power supply cut-off range.

6. The method as described in claim 1, characterized in that, After the step of sending the control command to the corresponding safety interlock device to cause the safety interlock device to perform the interlocking action, the following steps are included: Collect safety status feedback data after the safety interlocking device performs its actions; When the safety status feedback data does not exceed the safety status emergency threshold, the reinforcement learning decision model is iteratively optimized online using the safety status feedback data through incremental learning to obtain an updated reinforcement learning decision model. When the safety status feedback data exceeds the safety status emergency threshold, the emergency interlock control process is triggered, online iterative optimization is paused, and the highest priority safety protection action is executed. The safety protection action includes at least one of cutting off the power supply to the entire area, activating the emergency evacuation warning, and closing the safety isolation door. When the security status feedback data recovers to a level not exceeding the security status emergency threshold, the reinforcement learning decision model is further iteratively optimized online using the security status feedback data through incremental learning to obtain an updated reinforcement learning decision model.

7. The method as described in claim 6, characterized in that, The step of using incremental learning to iteratively optimize the reinforcement learning decision using the security state feedback data online to obtain an updated reinforcement learning decision model includes: The safety status feedback data is used as a new training sample and input into the reinforcement learning decision model to calculate the deviation between the predictive control strategy and the optimal control strategy under the current model parameters. Based on the gradient update direction of the model parameters calculated by the deviation, the initial fully connected layer parameters of the reinforcement learning decision model are locally incrementally updated. A second learning rate for incremental learning is set, and the parameters are fine-tuned by a single iteration or a small number of iterations. The fine-tuned parameters are used as the updated model parameters, wherein the second learning rate is less than the first learning rate in the offline training stage. The updated model parameters are weighted and fused with the initial fully connected layer parameters to obtain the fused model parameters. The reinforcement learning decision model is iteratively optimized online based on the fused model parameters to obtain an updated reinforcement learning decision model.

8. A mine safety interlocking control device based on reinforcement learning, characterized in that, The device includes: The acquisition module is used to build a mine safety situation awareness system, collect and process multi-dimensional safety status data, and obtain pre-processed data. The multi-dimensional safety status data includes environmental parameters, equipment operating parameters and personnel location parameters. The mine safety situation awareness system adopts a three-level architecture of perception layer-transmission layer-processing layer. The strategy generation module is used to input the preprocessed data into a reinforcement learning decision model to obtain the optimal control strategy. The reinforcement learning decision model includes a feature extraction layer, a fully connected layer, and an output layer. It is also used to construct an initial decision model based on a reinforcement learning algorithm, using the preprocessed data as the safety state space and at least one interlocking control action among ventilation equipment adjustment, gas extraction equipment start / stop, mining equipment emergency stop, personnel evacuation early warning, power supply system power outage, and waterproof gate closure as the action space. A reward function is constructed based on the safety risk level, the effect of control action execution, and the emergency control result. The module acquires historical mine safety operation data, accident case data, and simulated working condition data. Based on the historical mine safety operation data, accident case data, and simulated working condition data, the initial decision model is trained offline to obtain the reinforcement learning decision model. The instruction generation module is used to generate control instructions based on the optimal control strategy; The execution module is used to send the control command to the corresponding safety interlocking device so that the safety interlocking device performs the interlocking action.

9. A mine safety interlocking control device based on reinforcement learning, characterized in that, The device includes: a memory, a processor, and a reinforcement learning-based mine safety interlocking control program stored in the memory and running on the processor, the reinforcement learning-based mine safety interlocking control program being configured to implement the steps of the reinforcement learning-based mine safety interlocking control method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores a mine safety interlocking control program based on reinforcement learning. When the processor executes the mine safety interlocking control program based on reinforcement learning, it implements the steps of the mine safety interlocking control method based on reinforcement learning as described in any one of claims 1-7.

Images