Coal mine safety production intelligent decision-making method and system based on digital twinning

Abstract

The invention relates to a coal mine safety production intelligent decision-making system based on digital twinning, and the system comprises a physical sensing layer which collects coal mine environment parameters, equipment states and personnel positioning data through the deployment of a multi-mode sensor network, and generates a structured data flow; the edge calculation layer is used for operating an incremental multi-objective evolutionary algorithm, quickly generating a cache strategy in combination with a strategy cache pool preloading mechanism, uploading the processed data to the digital twinborn layer, receiving a global instruction of the intelligent decision-making layer and decomposing the global instruction into a device-level control signal; the digital twinborn layer is used for receiving the real-time data uploaded by the edge calculation layer, updating the state of a digital twinborn body and feeding back an optimization demand to the intelligent decision-making layer; and the intelligent decision-making layer is used for generating a global strategy by means of digital twin-guided hybrid optimization and a special FPGA acceleration card for a coal mine, and fusing the cache strategy of the edge calculation layer and the global strategy of the intelligent decision-making layer to generate a global instruction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coal mine safety, and in particular to a coal mine safety production intelligent decision-making method and system based on digital twinning. BACKGROUND

[0002] The traditional coal mining industry has long been faced with the deep-seated contradiction between complex geological conditions, high-risk operating environment and low-efficiency management mode. In the scene where uncontrollable factors such as variable geological structure, gas outburst and water hazard threat are intertwined, the traditional manual inspection, experience-based decision-making and post-disposal mode has been difficult to meet the safety production demand of zero tolerance for accidents. The digital twinning technology, by constructing a digital mirror that maps the physical coal mine and the virtual space in real time, fusing Internet of Things sensing, multi-source heterogeneous data fusion, high-precision simulation modeling and artificial intelligence algorithms, realizes the dynamic monitoring and full-factor visualization of mine environment parameters (such as gas concentration, ground pressure distribution, ventilation state), and its core value lies in breaking the boundary between the physical world and the digital world, through the closed-loop feedback mechanism of virtual and real interaction, not only can simulate the whole life cycle operation state of the mine in the virtual space, and predict the disaster risks such as water inrush and roof fall in advance, but also can provide dynamic optimization decision support for emergency plan making, equipment health management, personnel path planning and other scenarios based on real-time data flow driven intelligent analysis model, so as to transform passive accident response into active risk control. The innovation of this technical architecture not only significantly improves the intrinsic safety level of coal mining enterprises, but also reduces the unplanned downtime and optimizes the mining operation efficiency through data-driven intelligent decision-making, and ultimately ensures the safety of miners while promoting the leap-forward development of the coal industry towards digitalization and intelligentization.

[0003] In the prior art, a particle swarm-genetic hybrid algorithm (PSO-GA) and a multi-agent game optimization and other composite algorithm system are adopted. The advantage lies in that through the multi-objective collaborative optimization mechanism and the distributed intelligent decision-making capability, the dynamic balance problem of the three major targets of safety, efficiency and cost in the coal mine safety production is systematically solved. However, the calculation complexity of the multi-objective optimization algorithm, the particle swarm-genetic hybrid algorithm (PSO-GA) and the multi-agent game optimization and other composite algorithm system have good optimization performance, but the calculation complexity is high, and it may be difficult to meet the real-time decision-making demand, therefore, it is particularly important to propose the coal mine safety production intelligent decision-making method and system based on digital twinning. SUMMARY

[0004] The purpose of the present application is to solve the problems existing in the prior art, and to provide a coal mine safety production intelligent decision-making method and system based on digital twinning.

[0005] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: The coal mine safety production intelligent decision-making system based on digital twinning comprises: Physical perception layer: Through the deployment of a laser radar array (distance measurement accuracy ± 2 cm), a fiber grating sensor network (strain measurement accuracy 1με), a multi-parameter gas sensor (CH4 / CO / CO2 / O2 detection), an acoustic emission sensor (crack monitoring), an electronic nose (identification of odor gases), and other multi-modal sensor networks (among them, the key equipment is supplemented with a UWB positioning base station (positioning error <0.1 m) and a microseismic monitoring array (frequency response range 1-1000 Hz)), coal mine environmental parameters, equipment status, and personnel positioning data are collected (data acquisition frequency: environmental parameters ≥20 Hz, equipment status ≥100 Hz, personnel positioning ≤0.1 m accuracy), and after time alignment of the collected data by a time-sensitive network (TSN), structured data streams are generated, and the generated structured data streams are transmitted to the edge computing layer through a 5G private network; Edge computing layer: An intelligent gateway equipped with an NVIDIA Jetson AGX Orin platform runs an incremental multi-objective evolutionary algorithm (IMOEA) combined with a strategy cache pool preloading mechanism to quickly generate cache strategies (local decisions such as automatic reconstruction of the ventilation system when the gas concentration exceeds the limit and real-time correction control of belt deviation), the strategy cache pool preloads cloud optimization results to achieve fast decision matching, and the processed data is uploaded to the digital twin layer while receiving global instructions from the intelligent decision layer and decomposing them into device-level control signals; Digital twin layer: Receive real-time data uploaded by the edge computing layer, update the digital twin body state, and feed back optimization requirements to the intelligent decision layer. Based on a hybrid-dimensional geological model (structure surface uses discrete element method, rock mass uses continuous medium method, resolution 0.5m×0.5m×1m), MTConnect standard device digital twin body, and lightweight physical information neural network (PINN), through laser SLAM+IMU tight coupling, UWB+geomagnetic fingerprint fusion, and other spatio-temporal joint calibration technologies, combined with a self-encoder for scene compressed sensing, a virtual-real mapping model is constructed; Intelligent decision layer: Hybrid optimization guided by digital twin and coal mine dedicated FPGA acceleration card are used for global strategy generation, and the virtual environment pre-training model parameters are applied to real-time optimization through a transfer learning mechanism. The cache strategy of the edge computing layer and the global strategy of the intelligent decision layer are fused to generate global instructions (global optimization strategy).

[0006] The above scheme further includes: Further, the time-sensitive network (TSN) time-aligns the collected data, based on IEEE802.1AS protocol, the master clock broadcasts synchronization messages to the slave clock through generalized precision time protocol (gPTP), calculates the clock offset, and the calculation formula is Where, is the master clock sending time, To receive time from the clock, To respond time to the clock, To receive time from the master clock, compensate clock drift by linear regression so that the collected data complete time alignment.

[0007] Further, the specific steps of the edge computing layer processing real-time data stream and making local decisions are: Data reception and preprocessing: receiving structured data stream uploaded by the physical perception layer, using PCA to compress state space; Incremental multi-objective evolutionary algorithm (IMOEA) running: running incremental multi-objective evolutionary algorithm (IMOEA), generating offspring through simulated binary crossover and polynomial mutation, and performing local search; Strategy cache pool preloading and matching: with the help of strategy cache pool preloading mechanism (pulling global strategies such as ventilation network optimization parameters from the cloud managed by Kubernetes and storing them in local cache pool), completing fast decision matching; Device-level control signal generation: decomposing global instructions (such as "reduce the speed of the coal mining machine") into specific device actions.

[0008] Further, the specific steps of the incremental multi-objective evolutionary algorithm running are: Initialization: load historical Pareto frontiers (such as ventilation system optimization parameter sets) in the strategy cache pool; Local search: only perform genetic operations on new data domain, formula is: Where, New individuals generated through genetic operations (such as crossover and mutation), these new individuals will replace or supplement to the parent population, pushing the population evolution, Core operations in genetic algorithm, including selection, crossover and mutation, these operations simulate natural selection and genetic mechanisms in biological evolution, The last 100 solutions selected from the Pareto optimal solution set, in multi-objective optimization problems, the Pareto front represents a set of optimal solutions, where the improvement of any one solution cannot make at least one other solution worse, by selecting these solutions for genetic operations, the diversity of the population can be maintained and the convergence speed can be accelerated, using simulated binary crossover (SBX) and polynomial mutation (PM), represented as: ; ; Where, The ith gene value of the offspring individual generated after crossover, Random variable used to control the proportion of parent genes and recombined genes in the crossover process, its value is determined by random number and distribution index co-decide, and is the ith gene value of the individual, respectively representing two parent individuals, is a random number in the interval [0, 1] used to determine the direction of crossover, is the distribution index, which controls the distribution shape of crossover, The larger the value, the closer the offspring to the parent; The smaller the value, the greater the difference between the offspring and the parent; Non-dominated sorting acceleration: filter the front solutions by Fast Non-Dominated Sort, the formula is where, is the first non-dominated front, that is, the set of all individuals in the population that are not dominated by other individuals, in multi-objective optimization, the non-dominated front is used to sort and select solutions, is the current population, that is, the set of all individuals to be sorted, indicates that individual i is not dominated by individual j, that is, individual i is not inferior to j in all objective functions, and at least one objective function is superior to j.

[0009] Further, the digital twin layer constructs a virtual-real mapping model of the coal mine, and performs dynamic simulation and optimization; Multi-source data fusion and state updating: after receiving the real-time data stream uploaded by the edge computing layer, the Extended Kalman Filter (EKF) is used for space-time calibration, and the discrete element method (DEM) is used to simulate the crack propagation of the structure surface combined with the mixed dimension geological model, represented as where, is the mass of particle i, is the acceleration of particle i, and are the normal and tangential contact forces between particle i and particle j, respectively, is the external force acting on particle i (such as ground stress), and the rock mass region is solved by the Navier-Stokes equation to construct a 0.5m×0.5m×1m resolution stress field for dynamic mapping, represented as where, is the density of the fluid, which represents the mass of the fluid per unit volume, and is a key parameter in the Navier-Stokes equation, reflecting the inertial characteristics of the fluid, is the velocity field of the fluid, which describes the motion speed and direction of the fluid at each point in space, and is an important manifestation of the fluid dynamics behavior, is the pressure of the fluid, which represents the pressure received by the fluid at each point in space, and is the result of the interaction within the fluid, The dynamic viscosity of a fluid, which reflects the fluid's ability to resist shear deformation, is a measure of the fluid's viscous properties, The external force term acting on the fluid; Device digital twin construction: Based on the MTConnect standard, device data such as hydraulic support pressure and coal mining machine cutting current are parsed and mapped to digital twin parameters through the OPC UA protocol; Scene compression sensing and lightweight modeling: The high-dimensional (300-dimensional) state space is compressed to a low-dimensional (30-dimensional) latent space using an autoencoder. The autoencoder loss function combines reconstruction error and embedded Navier-Stokes equation residual physical constraints, represented as where, is the embedded Navier-Stokes equation residual term, calculated by automatic differentiation , deploy lightweight physical information neural network (PINN), for gas diffusion field, construct neural network loss function where, F is the convection-diffusion equation residual; Virtual-real mapping error compensation: The virtual-real mapping error is compensated by the virtual-real error compensation algorithm, and the local models of each edge node are aggregated.

[0010] Further, the specific steps of the virtual-real mapping error compensation are: Error space modeling: Use error vector modeling and covariance matrix analysis method, by defining multi-dimensional error vector and constructing its spatio-temporal covariance matrix, expressed as: where, is the multi-dimensional error vector at time t, which captures the state difference between the physical entity and the virtual model, is the actual state vector of the physical entity at time t, including actual measurement values such as position and temperature, is the state vector of the virtual model at time t predicted based on the model parameters θ, which is the model's estimate of the physical entity's state, is the error vector covariance matrix of the error vector, which quantifies the statistical properties of the error, such as variance and covariance, is used to calculate the statistical average of the error vector; combined with the exponential weighted moving average algorithm to extract the error propagation characteristics, the extraction formula is: where, is the error covariance matrix estimate at the kth iteration, used to track the spatio-temporal changes of the error, is the forgetting factor, with a value range of [0, 1], controlling the degree of influence of historical error information on the current estimate, is the error vector at the kth iteration, reflecting the deviation between the physical entity and the virtual model at the current moment; thus quantifying the deviation degree and dynamic characteristics of the virtual-real mapping; Error tracing and dynamic compensation: Using error propagation mechanism analysis technology, singular value decomposition is used to extract the principal components of the model sensitivity matrix, identify the dominant parameter dimensions that cause errors, and establish an error transmission chain model. ,in is the model sensitivity matrix, The process noise is identified by singular value decomposition (SVD): ,in, is the model sensitivity matrix, and its intrinsic structure is revealed by singular value decomposition. and They are the left singular vector matrix and the right singular vector matrix, providing the matrix The orthogonal basis of is a singular value matrix containing singular values , arranged in descending order, represents the matrix Energy in different directions, is the principal component direction, corresponding to the largest singular value, indicating the main direction of error transmission, that is, the direction in which the model is most sensitive to parameter changes; a recursive least squares (RLS) algorithm with a forgetting factor is designed for online parameter estimation, which is expressed as ,in, represents the parameter estimate at time t, represents the parameter estimate at time t-1, represents the prediction error, which is the difference between the actual output and the predicted output, Represents the observation matrix or regression vector, which contains the input information at the current moment. represents the change in parameter estimates, is the gain matrix, used to update parameter estimates, expressed as ,in, Represents the covariance matrix at time t, which is used to measure the uncertainty of parameter estimation, represents the covariance matrix at time t−1, and the covariance matrix update is expressed as ,in is the regularization coefficient to prevent numerical instability; Feedback control and dynamic optimization: Compensation is generated through a feedforward-feedback composite control strategy. The feedforward term directly offsets the current error, while the feedback term eliminates static deviations through integral action. The control gain matrix is ​​determined by combining the optimal control theory of linear quadratic regulators to ensure dynamic optimization of the compensation. Closed-loop verification and stability guarantee: error convergence is verified by Lyapunov stability analysis, and model prediction correction and version iteration are performed by using rolling horizon optimization and digital thread technology.

[0011] Further, the specific steps of the intelligent decision layer for generating a global strategy by means of digital twin guided hybrid optimization and coal mine special FPGA acceleration card are as follows: Hybrid optimization: a hybrid optimization strategy combining genetic algorithm (GA) and deep reinforcement learning (DRL) is adopted to simulate the global decision-making process in the digital twin environment. The genetic algorithm searches the strategy space through population evolution, and the deep reinforcement learning fits the state-action value function through a neural network. The two are combined to perform multi-objective optimization of parameters such as support scheme and ventilation network, represented as wherein, is the policy function, is the state and the reward (such as safety factor improvement) of the action is performed, and γ is the discount factor. This formula shows that the policy π needs to maximize the cumulative reward R (including safety factor, energy consumption, etc.); FPGA acceleration calculation: the FPGA special acceleration card is developed by customizing the parallel computing unit through hardware description language (HDL), and offloads the intensive computing tasks (such as matrix operation, path planning) in hybrid optimization to the coal mine special FPGA acceleration card, represented as wherein, is the CPU computing time, is the number of FPGA parallel cores, is the hardware efficiency factor (usually > 10), which uses its thousands of parallel computing cores and low-latency memory access characteristics to shorten the optimization computing time to 1 / 15 of the original CPU scheme (such as ventilation network optimization from 120 seconds to 8 seconds); Global strategy generation: based on the optimized digital twin model and the results of the hybrid optimization algorithm, a global strategy candidate set that takes into account safety and economy is generated, the effectiveness of the strategy is verified through simulation, and the strategy is ensured to improve the level of coal mine safety production.

[0012] The coal mine safety production intelligent decision-making method used by the coal mine safety production intelligent decision-making system based on digital twinning includes the following steps: Step 1: Multi-modal perception data acquisition and real-time preprocessing Real-time collection of coal mine environmental parameters (gas concentration, temperature and humidity, stress and strain), equipment status (hydraulic support pressure, coal mining machine current) and personnel positioning data through physical sensing layer devices such as laser radar, fiber optic sensor, multi-parameter gas sensor, edge computing layer cleans the raw data, filters outliers using the Isolation Forest algorithm, and synchronizes the data in time (synchronization error < 1 μs) through Time Sensitive Network (TSN), wherein the sensor network includes laser radar (± 2 cm ranging accuracy), fiber optic sensor (1 με strain measurement accuracy), and UWB positioning (0.1 m positioning error) to generate structured data streams; Step two: dynamic mapping and calibration of digital twin: a three-dimensional space-time model of the coal mine is constructed in the digital twin layer, including geological structure (mixed dimension modeling), equipment digital twin (MTConnect standard), and personnel behavior model, dynamic mapping and error compensation are performed on rigid objects (laser SLAM + IMU tight coupling), fluid field (lightweight PINN model), and personnel positioning (UWB + geomagnetic fingerprint fusion) through a space-time joint calibration engine, so that the state of the physical coal mine and the digital twin is synchronized (error < 5%), providing a reliable virtual environment for decision-making; Step three: hierarchical multi-objective optimization and decision generation: the edge computing layer (edge side) uses incremental multi-objective evolutionary algorithm (IMOEA) to locally optimize only key variables (such as gas concentration, ventilation fan speed), generating millisecond-level response strategies (such as ventilation system reconstruction), the intelligent decision-making layer (cloud) generates global optimization strategies (such as mining plan adjustment, equipment energy efficiency optimization) through the initial solution distribution pre-trained by the digital twin, combined with FPGA accelerated NSGA-III algorithm, to balance safety, efficiency, cost and other multi-objectives, generating a decision-making scheme that takes into account real-time and globality; Step four: closed-loop control and execution feedback: the edge computing layer decomposes decision-making instructions (such as ventilation fan speed regulation, hydraulic support pressure adjustment) into device-level control signals, which are issued to physical devices through OPC UA protocol, and the execution results of physical devices (such as gas concentration change, equipment state update) are fed back to the digital twin layer in real time to verify the effectiveness of the decision and trigger dynamic model update; Step five: virtual-real environment co-evolution: the digital twin continuously updates model parameters (such as geological structure changes, equipment aging) based on feedback data from the physical system, and through federated learning mechanism, aggregates virtual-real mapping experience of multiple mines to improve algorithm generalization ability.

[0013] The present application has the following advantages: In the present invention, an incremental multi-objective evolutionary algorithm is used at the edge computing layer to perform local search on dynamic working conditions, and only the elite individuals in the historical population memory are retained to participate in genetic operations to avoid global recalculation. The early termination criterion of non-dominated sorting is used to significantly reduce the computational complexity. At the same time, a coal mine-specific FPGA acceleration card is deployed at the intelligent decision-making layer to realize hardware parallelization of the NSGA-III core operations, compress the crossover / mutation operation delay, and build a strategy cache pool for fast edge-side matching of cloud-based pre-trained strategies. This breaks through the dimensional disaster bottleneck of traditional multi-objective optimization algorithms while compressing the end-to-end decision delay from sub-seconds to milliseconds. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Fig. 1 This is a system block diagram of the digital twin-based intelligent decision-making system for coal mine safety production proposed in this invention; Fig. 2 This is a method step diagram of the digital twin-based coal mine safety production intelligent decision-making system proposed in the present invention. DETAILED DESCRIPTION

[0015] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0016] See also Figs. 1-2 As shown, the present invention is an intelligent decision-making system for coal mine safety production based on digital twins, including: Physical perception layer: By deploying multimodal sensor networks such as lidar arrays (ranging accuracy ±2 cm), fiber Bragg grating sensor networks (strain measurement accuracy 1 με), multi-parameter gas sensors (CH4 / CO / CO2 / O2 detection), acoustic emission sensors (crack monitoring), and electronic noses (odor gas identification), key equipment is equipped with UWB positioning base stations (positioning error < 0.1 m) and microseismic monitoring arrays (frequency response range 1-1000 Hz) to collect coal mine environmental parameters, equipment status, and personnel positioning data (data collection frequency: environmental parameters ≥ 20 Hz, equipment status ≥ 100 Hz, personnel positioning accuracy ≤ 0.1 m). The collected data is time-aligned using time-sensitive networking (TSN) to generate structured data streams, which are then transmitted to the edge computing layer via the 5G private network. Edge computing layer: Run incremental multi-objective evolutionary algorithm (IMOEA) on smart gateway equipped with NVIDIA Jetson AGX Orin platform, combine with strategy cache pool preloading mechanism to quickly generate cache strategy (local decision such as automatic reconstruction of ventilation system when gas concentration exceeds limit, real-time deviation correction control of belt deviation, etc.), pre-load cloud optimization results of strategy cache pool, realize fast decision matching, upload processed data to digital twin layer, and receive global instructions from intelligent decision layer and decompose into device-level control signals; Digital twin layer: Receive real-time data uploaded by edge computing layer, update digital twin state, and feed back optimization requirements to intelligent decision layer, based on hybrid dimension geological model (discrete element method for structural plane, continuous medium method for rock mass, resolution 0.5m×0.5m×1m), MTConnect standard device digital twin, and lightweight physical information neural network (PINN), through laser SLAM+IMU tight coupling, UWB+geomagnetic fingerprint fusion space-time joint calibration technology, combine with self-encoder for scene compressed sensing, construct virtual-real mapping model; Intelligent decision layer: Use digital twin guided hybrid optimization and coal mine special FPGA acceleration card to generate global strategy, and apply virtual environment pre-training model parameters to real-time optimization through transfer learning mechanism, combine cache strategy of edge computing layer with global strategy of intelligent decision layer to generate global instruction (global optimization strategy).

[0017] The above scheme further includes: In one embodiment, the time-sensitive network (TSN) performs time alignment on the collected data, based on IEEE 802.1AS protocol, the master clock broadcasts synchronization messages to the slave clock through generalized precision time protocol (gPTP), calculates the clock offset, and the calculation formula is represented as Wherein, is the master clock sending time, is the slave clock receiving time, is the slave clock response time, is the master clock receiving time, and the collected data is time-aligned by compensating clock drift through linear regression.

[0018] In one embodiment, the edge computing layer performs real-time data stream processing and local decision-making in the following specific steps: Data reception and preprocessing: Receive structured data stream uploaded by physical perception layer, and use PCA to compress state space; Incremental multi-objective evolutionary algorithm (IMOEA) running: Run incremental multi-objective evolutionary algorithm (IMOEA), generate offspring through simulated binary crossover and polynomial mutation, and perform local search; Policy cache pool preloading and matching: fast decision matching with policy cache pool preloading mechanism (pulling global policies such as ventilation network tuning parameters from the cloud managed by Kubernetes and storing them in the local cache pool); Device-level control signal generation: decomposing global instructions (such as "reduce the speed of the coal mining machine") into specific device actions.

[0019] In one embodiment, the specific steps of the incremental multi-objective evolutionary algorithm are as follows: Initialization: load the historical Pareto front (such as the set of ventilation system optimization parameters) in the policy cache pool; Local search: only perform genetic operations on the new data domain, formula: where, is the new individual generated by genetic operations (such as crossover and mutation), which will replace or supplement the parent population to promote population evolution, is the core operation in genetic algorithms, including selection, crossover and mutation, which simulates the natural selection and genetic mechanism in biological evolution, is the last 100 solutions selected from the Pareto optimal solution set, in multi-objective optimization problems, the Pareto front represents a set of optimal solutions, where the improvement of any one solution cannot make at least one other solution worse, by selecting these solutions for genetic operations, the diversity of the population can be maintained and convergence can be accelerated, using simulated binary crossover (SBX) and polynomial mutation (PM), represented as: ; ; where, is the i-th gene value of the offspring individual after crossover, is a random variable used to control the proportion of parent genes and recombined genes in the crossover process, its value is determined by the random number and the distribution index , and are the i-th gene values of the two parent individuals respectively, is a random number in the interval [0, 1] used to determine the direction of crossover, is the distribution index that controls the shape of the crossover distribution, the larger the value, the closer the offspring to the parent; the smaller the value, the greater the difference between the offspring and the parent; Non-dominated sorting acceleration: filtering the front solutions by fast non-dominated sorting (Fast Non-Dominated Sort), formula where, the first non-dominated front, i.e., the set of all individuals in the population that are not dominated by any other individual, is used in multi-objective optimization to rank and select solutions, the current population, i.e., the set of all individuals to be ranked, denotes that individual i is not dominated by individual j, i.e., individual i is not worse than j in all objective functions and at least better than j in one objective function.

[0020] In one embodiment, the digital twin layer constructs a virtual-real mapping model of the coal mine, performs dynamic simulation and optimization; Multi-source data fusion and state updating: after receiving the real-time data stream uploaded by the edge computing layer, extended Kalman filtering (EKF) is used for space-time calibration, and the hybrid dimension geological model is used to simulate the crack expansion of the structural surface by the discrete element method (DEM), which is represented as wherein, is the mass of particle i, is the acceleration of particle i, and are the normal and tangential contact forces between particle i and particle j, respectively, is the external force acting on particle i (such as ground stress), and the rock mass region is solved by the Navier-Stokes equation to construct a 0.5m×0.5m×1m resolution stress field for dynamic mapping, which is represented as wherein, is the density of the fluid, which represents the mass of the fluid per unit volume, and is a key parameter in the Navier-Stokes equation, reflecting the inertial characteristics of the fluid, is the velocity field of the fluid, which describes the motion speed and direction of the fluid at each point in space, and is an important manifestation of the fluid dynamics behavior, is the pressure of the fluid, which represents the pressure exerted by the fluid at each point in space, and is the result of the interaction within the fluid, is the dynamic viscosity of the fluid, which reflects the ability of the fluid to resist shear deformation, and is a measure of the viscosity characteristics of the fluid, is the external force term acting on the fluid; Device digital twin construction: based on the MTConnect standard, device data (such as hydraulic support pressure and coal mining machine cutting current) is analyzed, and mapped to digital twin parameters through the OPC UA protocol; Scene compression sensing and lightweight modeling: a self-encoder is used to compress the high-dimensional (300-dimensional) state space to a low-dimensional (30-dimensional) latent space, and the self-encoder loss function is fused with the reconstruction error and the physical constraint of embedding the Navier-Stokes equation residual, which is represented as wherein, For the embedded Navier-Stokes equation residual term, the automatic differentiation is calculated , a lightweight physical information neural network (PINN) is deployed, and a neural network loss function is constructed for the gas diffusion field , wherein F is the residual of the convection-diffusion equation; Virtual-real mapping error compensation: compensate the virtual-real mapping error through the virtual-real error compensation algorithm, and aggregate the local models of each edge node.

[0021] In one embodiment, the specific steps of the virtual-real mapping error compensation are: Error space modeling: adopt error vector modeling and covariance matrix analysis method, define multi-dimensional error vector and construct its space-time covariance matrix, expressed as: , wherein is the multi-dimensional error vector at time t, which captures the state difference between the physical entity and the virtual model, is the actual state vector of the physical entity at time t, including actual measurement values such as position and temperature, is the state vector of the virtual model at time t predicted based on the model parameters θ, which is the model's estimate of the physical entity's state, is the error vector Covariance matrix of error vector, quantifying the statistical properties of error such as variance and covariance, is used to calculate the statistical average of the error vector; combined with the exponential weighted moving average algorithm to extract the error propagation characteristics, the extraction formula is expressed as: , wherein is the error covariance matrix estimate at the kth iteration, used to track the space-time variation of the error, is the forgetting factor, with a value range of [0, 1], controlling the influence degree of historical error information on the current estimate, is the error vector at the kth iteration, reflecting the deviation between the physical entity and the virtual model at the current time; further quantifying the deviation degree and dynamic characteristics of virtual-real mapping; Error tracing and dynamic compensation: using error propagation mechanism analysis technology, principal component extraction is performed on the model sensitivity matrix through singular value decomposition, identifying the dominant parameter dimension that causes the error, and establishing an error transmission chain model , wherein is the model sensitivity matrix, is the process noise, and the dominant error source is identified through singular value decomposition (SVD): , wherein is the model sensitivity matrix, and its internal structure is revealed through singular value decomposition, and are the left singular vector matrix and the right singular vector matrix respectively, which provide the matrix orthogonal basis of is a singular value matrix, containing singular values , arranged in descending order, representing the matrix energy in different directions, is the principal component direction, corresponding to the largest singular value, indicating the main direction of error transmission, i.e. the direction in which the model is most sensitive to parameter changes; the recursive least squares (RLS) algorithm with forgetting factor is designed for online parameter estimation, denoted as , where represents the parameter estimation value at time t, represents the parameter estimation value at time t-1, represents the prediction error, i.e. the difference between the actual output and the predicted output, represents the observation matrix or regression vector, containing the input information at the current time, represents the change amount of parameter estimation, is the gain matrix, used to update the parameter estimation, denoted as , where represents the covariance matrix at time t, used to measure the uncertainty of parameter estimation, represents the covariance matrix at time t-1, and the covariance matrix update is denoted as , where is the regularization coefficient, to prevent numerical instability; feedback control and dynamic optimization: the compensation quantity is generated through the feedforward-feedback compound control strategy, where the feedforward term directly offsets the current error, and the feedback term eliminates the static deviation through integral action, while the control gain matrix is determined by combining the linear quadratic regulator optimal control theory to ensure the dynamic optimization of the compensation quantity; closed-loop verification and stability guarantee: the error convergence is verified through Lyapunov stability analysis, and model prediction correction and version iteration are performed using rolling horizon optimization and digital thread technology.

[0022] In one embodiment, the intelligent decision-making layer generates global strategies through the following specific steps with the help of digital twin guided hybrid optimization and coal mine special FPGA acceleration card: Hybrid optimization: a hybrid optimization strategy combining genetic algorithm (GA) and deep reinforcement learning (DRL) is adopted to simulate the global decision-making process in the digital twin environment, the genetic algorithm searches the strategy space through population evolution, and the deep reinforcement learning fits the state-action value function through neural network, both of which are combined to perform multi-objective optimization of support scheme, ventilation network and other parameters, denoted as , where is the strategy function, is the state performs the action where R is the cumulative reward (including safety factor, energy consumption, etc.), γ is the discount factor, and the formula indicates that the strategy π needs to maximize the cumulative reward R (including safety factor, energy consumption, etc.); FPGA acceleration calculation: FPGA special acceleration card is developed by hardware description language (HDL) to customize parallel computing unit, and the intensive calculation task (such as matrix operation, path planning) in hybrid optimization is unloaded to coal mine special FPGA acceleration card, which is represented as wherein, is the CPU calculation time, is the number of FPGA parallel cores, is the hardware efficiency factor (usually > 10), which uses its thousands of parallel computing cores and low delay memory access characteristics to shorten the optimization calculation time to 1 / 15 of the original CPU scheme (such as ventilation network optimization from 120 seconds to 8 seconds); Global strategy generation: based on the results of the optimized digital twin model and hybrid optimization algorithm, a global strategy candidate set that takes into account safety and economy is generated, and the effectiveness of the strategy is verified through simulation to ensure that the strategy can improve the level of coal mine safety production.

[0023] The coal mine safety production intelligent decision-making method used by the coal mine safety production intelligent decision-making system based on digital twinning includes the following steps: Step one: multi-modal perception data acquisition and real-time preprocessing: Real-time acquisition of coal mine environmental parameters (gas concentration, temperature and humidity, stress and strain), equipment status (hydraulic support pressure, coal mining machine current) and personnel positioning data through physical perception layer devices such as laser radar, fiber optic sensor, multi-parameter gas sensor, etc. The edge computing layer cleans the original data, filters outliers using the Isolation Forest algorithm, and synchronizes the data in time through Time Sensitive Network (TSN) (synchronization error < 1 μs), wherein the sensor network includes laser radar (±2 cm ranging accuracy), fiber optic sensor (1 με strain measurement accuracy), UWB positioning (0.1 m positioning error), and generates structured data stream; Step two: dynamic mapping and calibration of digital twin: a three-dimensional space-time model of the coal mine is constructed in the digital twin layer, including geological structure (mixed dimension modeling), equipment digital twin (MTConnect standard) and personnel behavior model, and through the space-time joint calibration engine, rigid objects (laser SLAM + IMU tight coupling), fluid field (lightweight PINN model) and personnel positioning (UWB + geomagnetic fingerprint fusion) are dynamically mapped and error compensated, so that the state of the physical coal mine and the digital twin is synchronized (error < 5%), providing a reliable virtual environment for decision-making; Step three: hierarchical multi-objective optimization and decision generation: the edge computing layer (edge side) uses incremental multi-objective evolutionary algorithm (IMOEA) to only optimize key variables (such as gas concentration, fan speed) locally, and generates millisecond-level response strategies (such as ventilation system reconstruction), and the intelligent decision layer (cloud) generates global optimization strategies (such as mining plan adjustment, equipment energy efficiency optimization) through the initial solution distribution of the pre-trained digital twin, combined with the FPGA accelerated NSGA-III algorithm, to balance safety, efficiency, cost and other multi-objectives, and generate a decision scheme that takes into account real-time and global considerations; Step four: closed-loop control and execution feedback: the edge computing layer decomposes decision instructions (such as fan speed regulation, hydraulic support pressure adjustment) into device-level control signals, which are issued to physical devices through the OPC UA protocol, and the execution results of physical devices (such as gas concentration changes, device state updates) are fed back to the digital twin layer in real time to verify the effectiveness of the decision and trigger dynamic model updates; Step five: virtual-real environment co-evolution: the digital twin continuously updates model parameters (such as changes in geological structure, equipment aging) based on feedback data from the physical system, and aggregates virtual-real mapping experience from multiple mines through federated learning to improve algorithm generalization ability.

[0024] Although embodiments of the present application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. The intelligent decision-making system for coal mine safety production based on digital twins is characterized by: include: Physical perception layer: By deploying a multimodal sensor network, it collects coal mine environmental parameters, equipment status, and personnel location data. This data is time-aligned using a time-sensitive network to generate a structured data stream. This structured data stream is then transmitted to the edge computing layer via a 5G private network. Edge computing layer: Utilizing an intelligent gateway powered by the NVIDIA Jetson AGX Orin platform, it runs an incremental multi-objective evolutionary algorithm, combined with a policy cache pool preloading mechanism to quickly generate cache policies. This layer then uploads processed data to the digital twin layer, receiving global instructions from the intelligent decision-making layer and breaking them down into device-level control signals. The digital twin layer receives real-time data uploaded by the edge computing layer, updates the digital twin status, and feeds optimization requirements back to the intelligent decision-making layer. Based on the mixed-dimensional geological model, MTConnect standard equipment digital twins, and lightweight physical information neural networks, it uses spatiotemporal joint calibration technology and autoencoders for scene compression sensing to build a virtual-reality mapping model. Intelligent decision-making layer: Global strategies are generated with the help of digital twin-guided hybrid optimization and coal mine-specific FPGA acceleration cards. The pre-trained model parameters of the virtual environment are applied to real-time optimization through the transfer learning mechanism. The cache strategy of the edge computing layer is integrated with the global strategy of the intelligent decision-making layer to generate global instructions.

2. The digital twin-based intelligent decision-making system for coal mine safety production according to claim 1 is characterized in that: The time-sensitive network performs time alignment on the collected data. Based on the IEEE 802.1AS protocol, the master clock broadcasts synchronization messages to the slave clocks through the generalized precision time protocol (gPTP) and calculates the clock offset. The calculation formula is expressed as ,in, Send time to the master clock, To receive time from a clock, is the slave clock response time, The master clock receives time and compensates for clock drift through linear regression to achieve time alignment of the collected data.

3. The coal mine safety production intelligent decision-making system based on digital twin according to claim 1 is characterized in that: The specific steps of the edge computing layer to process real-time data streams and make local decisions are as follows: Data reception and preprocessing: Receives the structured data stream uploaded by the physical perception layer and uses PCA to compress the state space; Incremental multi-objective evolutionary algorithm operation: Run the incremental multi-objective evolutionary algorithm to generate offspring by simulating binary crossover and polynomial mutation, and perform local search; Policy cache pool preloading and matching: Rapid decision matching is achieved with the help of the policy cache pool preloading mechanism; Device-level control signal generation: decomposes global instructions into specific device actions.

4. The coal mine safety production intelligent decision-making system based on digital twin according to claim 3 is characterized in that: The specific steps of running the incremental multi-objective evolutionary algorithm are as follows: Initialization: Load the historical Pareto frontier in the strategy cache pool; Local search: Genetic operations are performed only on the new data domain. The formula is: ,in, is a new individual generated by genetic manipulation, The core operations in genetic algorithms include selection, crossover, and mutation. The last 100 solutions selected from the Pareto optimal solution set are represented by simulated binary crossover and polynomial mutation: ； ； in, is the gene value of the i-th offspring individual produced after crossover, is a random variable used to control the ratio of parental genes and recombinant genes during the crossover process. and are the i-th gene values ​​representing the two parent individuals, is a random number in the interval [0,1], used to determine the direction of the crossover. is the distribution index, which controls the shape of the crossover distribution; Non-dominated sorting acceleration: Screen the frontier solutions through fast non-dominated sorting, the formula is ,in, is the first non-dominated frontier, that is, the set of all individuals in the population that are not dominated by other individuals, is the current population, that is, the set of all individuals to be sorted, It means that individual i is not dominated by individual j, that is, individual i is not inferior to j in all objective functions and is better than j in at least one objective function.

5. The digital twin-based intelligent decision-making system for coal mine safety production according to claim 1 is characterized in that: The digital twin layer constructs a virtual-reality mapping model of the coal mine and performs dynamic simulation and optimization; Multi-source data fusion and status update: After receiving the real-time data stream uploaded by the edge computing layer, the extended Kalman filter is used for time and space calibration. The discrete element method is used to simulate the crack expansion on the structural surface in combination with the mixed-dimensional geological model, which is expressed as ,in, is the mass of particle i, is the acceleration of particle i, and are the normal and tangential contact forces between particles i and j, respectively, is the external force acting on particle i, and the rock mass area uses the continuum method to solve the Navier-Stokes equation to construct the stress field for dynamic mapping, which is expressed as ,in, is the density of the fluid, is the velocity field of the fluid, is the pressure of the fluid, is the dynamic viscosity of the fluid, is the external force acting on the fluid; Equipment digital twin construction: Analyze equipment data based on the MTConnect standard and map it to digital twin parameters via the OPC UA protocol; Scene compression sensing and lightweight modeling: An autoencoder is used to compress the high-dimensional state space into a low-dimensional latent space. The autoencoder loss function integrates the reconstruction error and the physical constraints embedded in the residual of the Navier-Stokes equation, expressed as ,in, is the residual term of the embedded Navier-Stokes equation, calculated by automatic differentiation , deploy lightweight physical information neural network, build neural network loss function for gas diffusion field , where F is the residual of the convection-diffusion equation; Virtual-real mapping error compensation: The virtual-real mapping error is compensated by the virtual-real error compensation algorithm, and the local model of each edge node is aggregated.

6. The digital twin-based intelligent decision-making system for coal mine safety production according to claim 5 is characterized in that: The specific steps of the virtual-real mapping error compensation are: Error space modeling: Using error vector modeling and covariance matrix analysis methods, we define a multi-dimensional error vector and construct its spatiotemporal covariance matrix, which can be expressed as: ,in, is the multidimensional error vector at time t, is the actual state vector of the physical entity at time t, is the state vector of the virtual model at time t predicted based on the model parameters θ, which is the model’s estimate of the state of the physical entity. is the error vector The covariance matrix of Used to calculate the statistical average of the error vector; combined with the exponentially weighted moving average algorithm to extract the error propagation characteristics, the extraction formula is expressed as: ,in, is the error covariance matrix estimate at the kth iteration, which is used to track the spatiotemporal variation of the error. is the forgetting factor, which ranges from [0,1] and controls the influence of historical error information on the current estimate. is the error vector at the kth iteration; thus quantifying the deviation degree and dynamic characteristics of the virtual-real mapping; Error tracing and dynamic compensation: Using error propagation mechanism analysis technology, singular value decomposition is used to extract the principal components of the model sensitivity matrix, identify the dominant parameter dimensions that cause errors, and establish an error transmission chain model. ,in is the model sensitivity matrix, is the process noise, and the dominant error source is identified by singular value decomposition: ,in, is the model sensitivity matrix, and are the left singular vector matrix and the right singular vector matrix, is a singular value matrix containing singular values , arranged in descending order, represents the matrix Energy in different directions, is the principal component direction, corresponding to the largest singular value; a recursive least squares algorithm with a forgetting factor is designed for online parameter estimation, which is expressed as ,in, represents the parameter estimate at time t, represents the parameter estimate at time t-1, represents the prediction error, Represents the observation matrix or regression vector, which contains the input information at the current moment. represents the change in parameter estimates, is the gain matrix, used to update parameter estimates, expressed as ,in, Represents the covariance matrix at time t, which is used to measure the uncertainty of parameter estimation, represents the covariance matrix at time t−1, and the covariance matrix update is expressed as ,in is the regularization coefficient; Feedback control and dynamic optimization: Compensation is generated through a feedforward-feedback composite control strategy, where the feedforward term directly offsets the current error, and the feedback term eliminates static deviations through integral action. The control gain matrix is ​​determined by combining the optimal control theory of linear quadratic regulators. Closed-loop verification and stability assurance: Error convergence is verified through Lyapunov stability analysis, and model prediction correction and version iteration are performed using rolling horizon optimization and digital threading technology.

7. The coal mine safety production intelligent decision-making system based on digital twin according to claim 1 is characterized in that: The specific steps of the intelligent decision-making layer to generate global strategies with the help of hybrid optimization guided by digital twins and coal mine-specific FPGA acceleration cards are as follows: Hybrid optimization: A hybrid optimization strategy that integrates genetic algorithm and deep reinforcement learning is used to simulate the global decision-making process in the digital twin environment. The genetic algorithm searches the strategy space through population evolution, and the deep reinforcement learning fits the state-action value function through a neural network. The two are combined to perform multi-objective optimization of parameters, which is expressed as ,in, is the policy function, Status Next action The reward is , γ is the discount factor; FPGA accelerated computing: offload the intensive computing tasks in hybrid optimization to the coal mine-specific FPGA acceleration card, expressed as ,in, Calculates time for the CPU, is the number of FPGA parallel cores, is the hardware efficiency factor; Global strategy generation: Based on the optimized digital twin model and the results of the hybrid optimization algorithm, a global strategy candidate set that takes into account both safety and economy is generated.

8. The coal mine safety production intelligent decision-making method used in the coal mine safety production intelligent decision-making system based on digital twin according to claim 1 is characterized in that: The following steps are involved: Step 1: Multimodal perception data collection and real-time preprocessing: The physical perception layer collects coal mine environmental parameters, equipment status, and personnel location data in real time. The edge computing layer cleans the raw data, uses the isolation forest algorithm to filter outliers, and synchronizes the data through a time-sensitive network to generate a structured data stream. Step 2: Dynamic Mapping and Calibration of Digital Twins: A 3D spatiotemporal model of the coal mine is constructed on the digital twin layer, including geological structures, equipment digital twins, and personnel behavior models. Through a spatiotemporal joint calibration engine, dynamic mapping and error compensation are performed on rigid objects, fluid fields, and personnel positioning. This synchronizes the states of the physical coal mine and the digital twin, providing a virtual environment for decision-making. Step 3: Hierarchical Multi-Objective Optimization and Decision Generation: The edge computing layer uses an incremental multi-objective evolutionary algorithm to locally optimize key variables and generate response strategies. The intelligent decision-making layer uses the initial solution distribution pre-trained by the digital twin and the FPGA-accelerated NSGA-III algorithm to generate a global optimization strategy and a decision-making solution that balances real-time and global requirements. Step 4: Closed-loop control and execution feedback: The edge computing layer decomposes the decision instructions into device-level control signals and sends them to the physical devices via the OPC UA protocol. The execution results of the physical devices are fed back to the digital twin layer in real time to verify the effectiveness of the decision and trigger dynamic model updates. Step 5: Co-evolution of virtual and real environments: The digital twin continuously updates model parameters based on feedback data from the physical system.

Images