Coal yard debris cleaning device based on automatic robot

Abstract

This invention relates to the fields of automated robot technology and coal yard safety management technology in thermal power plants, and particularly to a coal yard debris cleaning device based on an automated robot. The device includes a sensing module that collects and fuses data to output a comprehensive data stream. A twin module dynamically constructs and updates a dynamic digital twin through coupled discrete element method and computational fluid dynamics simulation. A risk prediction module uses a graph neural network surrogate model and Monte Carlo simulation to generate a spontaneous combustion probability spatial distribution map as dynamic risk information. A decision-making module integrates dynamic risk information and robot state to generate a set of candidate joint action strategies. A control module verifies the strategies through parallel simulation and performs multi-objective arbitration based on dynamic risk information to select the optimal control law and decouple it into low-level control commands. A feedback module executes commands and collects actual data, feeding it back to the twin module to correct the dynamic digital twin and to the control module to optimize decision weights.

Description

Technical Field

[0001] This invention relates to the fields of automated robot technology and coal yard safety management technology in thermal power plants, and particularly to a coal yard debris cleaning device based on automated robots. Background Technology

[0002] In the coal yard environment of a thermal power plant, automated robots perform debris removal tasks. Automated robots are automated devices consisting of a mobile platform, actuators, and a control system, operating autonomously according to pre-programmed logic. The coal yard is a storage area for coal, and debris, including non-coal solids such as stones and wood, can cause blockages in the conveying system or equipment malfunctions. During the removal process, the robot uses a navigation system such as encoders or beacons for positioning, moves along a planned path, and uses a robotic arm or collection device to physically contact and pick up the debris, subsequently transporting it to a designated disposal area to maintain coal purity and the continuity of the processing flow.

[0003] Existing automated robotic debris removal technologies suffer from the following technical challenges: when automated robots perform debris removal operations in coal yards of thermal power plants, their mechanical movement physically disturbs the coal pile, unintentionally altering the oxygen diffusion channel structure between the surface and interior of the coal pile. This structural change stems from the localized compaction of the coal pile by the robot's walking device or the rearrangement of coal particles and pore reconstruction caused by the cleaning tools grabbing debris. For example, when a robot uses its robotic arm to remove large pieces of debris, the grabbing action creates pits and loose areas on the coal pile surface. The newly formed pores become pathways for rapid oxygen permeation, while the surrounding areas become denser due to pressure, hindering uniform oxygen diffusion. This uneven oxygen diffusion environment leads to a localized increase in the rate of low-temperature oxidation reactions inside the coal pile, accelerating heat accumulation and significantly increasing the risk of latent spontaneous combustion. This process involves the coupling of the porous media structure of the coal pile with oxidation kinetics, and existing cleaning technologies do not fully consider the indirect impact of robotic operations on the thermal stability of the coal pile. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a coal yard debris cleaning device based on an automated robot. This invention solves the technical problem that the automated robot unintentionally alters the oxygen diffusion channel structure on the surface and inside of the coal pile during cleaning operations, resulting in an accelerated low-temperature oxidation process and a significantly increased risk of latent spontaneous combustion.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: The coal yard debris cleaning device based on an automated robot provided by the present invention includes: The perception module is used to collect visual, temperature, geometric and environmental sensing data of the coal pile surface and output a spatiotemporally aligned comprehensive data stream, which includes the robot's state. The twin module is used to receive the integrated data stream and dynamically construct and update a dynamic digital twin reflecting the pore structure and oxygen diffusion state of the coal pile through multi-physics coupling simulation. The risk prediction module is used to receive the dynamic digital twin, extract the distribution fields of porosity, temperature and oxygen concentration, quickly infer the oxidation reaction through the surrogate model, quantify and predict the probability distribution of spontaneous combustion risk, generate dynamic risk information, and output the dynamic risk information to the decision module and control module. The decision-making module is used to receive the dynamic risk information and the robot status obtained from the integrated data stream, and output a set of joint action strategy candidates through collaboratively optimized path planning and operation trajectory generation. The control module is used to receive the candidate set of joint action strategies and the dynamic risk information, verify the impact of the strategies on the dynamic digital twin through parallel simulation, and select the optimal control law based on multi-objective arbitration in combination with the dynamic risk information, and decouple the optimal control law to generate the underlying control command. The feedback module is used to execute the underlying control commands to drive the robot's actions, collect the actual execution data generated by the robot's actions, and feed the actual execution data back to the twin module to correct the dynamic digital twin, and to the control module to optimize the arbitration of the next control cycle.

[0006] Furthermore, in the coal yard debris cleaning device based on an automated robot described in this invention, the sensing module includes: The data synchronization submodule is used to acquire the raw signals from the visual camera, infrared thermal imager, lidar, torque sensor and temperature and humidity sensor under the same clock source, timestamp the raw signals, and generate a time-aligned raw dataset. The geometric reconstruction submodule is used to receive the time-aligned raw dataset, parse the lidar point cloud data from the time-aligned raw dataset, perform point cloud registration and denoising, and reconstruct the three-dimensional geometric model of the coal pile surface. The data fusion submodule is used to receive the time-aligned original dataset and the three-dimensional geometric model, extract visual images at the same time from the time-aligned original dataset, map the texture features of the visual images to the corresponding surfaces of the three-dimensional geometric model to form a three-dimensional mesh with texture information; and simultaneously extract the temperature data of the infrared thermal imager from the time-aligned original dataset, map the temperature data to the corresponding vertices of the three-dimensional mesh to generate a temperature and geometry fusion model. The integrated stream generation submodule is used to receive the time-aligned raw dataset and the temperature and geometry fusion model, read the angle and torque sensor readings of the robot joint encoder from the time-aligned raw dataset, combine the environmental parameters of the temperature and humidity sensor read from the time-aligned raw dataset, package and encapsulate the temperature and geometry fusion model, robot pose, joint torque and environmental parameters, generate an integrated data stream, and output it to the twin module.

[0007] Furthermore, in the coal yard debris cleaning device based on an automated robot described in this invention, the twin module includes: The model initialization submodule is used to receive the integrated data stream, extract the coal pile boundary and initial pore distribution parameters from the temperature and geometry fusion model included in the integrated data stream, configure the particulate property parameters and contact model parameters of the discrete element simulation, and set the initial oxygen concentration field and boundary conditions of the computational fluid dynamics simulation according to the environmental parameters included in the integrated data stream. The coupled simulation submodule receives discrete element simulation (DEM) and computational fluid dynamics (CFD) simulation parameters configured by the model initialization submodule, and runs the DEM and CFD simulations in parallel. The DEM receives planning paths and action instructions from the decision module and calculates coal particle displacement, contact force, and pore structure evolution. The CFD, based on the current pore structure provided by the DEM, solves for the concentration distribution and diffusion flux of oxygen in the pore network. At each simulation step, the coupled simulation submodule exchanges the porosity field updated by the DEM and the oxygen concentration field calculated by the CFD. The data assimilation submodule is used to acquire the latest temperature and geometry fusion model output by the sensing module in real time during the simulation process. It compares the geometric deformation and temperature distribution of the coal pile surface predicted by the coupled simulation submodule with the actual observation data of the latest temperature and geometry fusion model, calculates the geometric residual and temperature residual, and uses the Kalman filter algorithm to adjust the interparticle friction coefficient and rolling resistance coefficient of the discrete element simulation in the coupled simulation submodule in reverse, using the geometric residual and temperature residual as the observed values, so as to correct the pore evolution prediction of the next simulation step. It outputs a dynamic digital twin including the corrected porosity field, temperature field and oxygen concentration field to the risk prediction module and the control module.

[0008] Furthermore, in the coal yard debris cleaning device based on an automated robot described in this invention, the risk prediction module includes: The field variable extraction submodule is used to receive the dynamic digital twin output by the twin module and extract the three-dimensional matrix of porosity, the three-dimensional matrix of temperature, and the three-dimensional matrix of oxygen concentration from the dynamic digital twin. The reaction inference submodule is used to receive the porosity matrix, temperature matrix, and oxygen concentration matrix extracted by the field variable extraction submodule, and input the porosity matrix, temperature matrix, and oxygen concentration matrix into a pre-trained graph neural network. The graph neural network uses the porosity matrix, temperature matrix, and oxygen concentration matrix as node features and the adjacency relationship of adjacent grid cells as edges. Through multi-layer graph convolution operation, it calculates the oxidation reaction rate of the next time step node by node to generate an oxidation reaction rate field. The risk quantification submodule receives the oxidation reaction rate field generated by the reaction deduction submodule and the temperature matrix extracted by the field variable extraction submodule. Using the oxidation reaction rate field as the heat source, and combining the temperature matrix with the probability distribution of the coal quality activity parameters of the coal pile, the submodule uses Monte Carlo simulation to randomly sample the coal quality activity parameters and calculates thousands of future temperature evolution paths in parallel. It counts the number of evolution paths in each three-dimensional grid cell where the temperature exceeds the preset spontaneous combustion critical temperature. The ratio of this number to the total number of simulated paths is defined as the spontaneous combustion probability of that grid cell. The probability values ​​of all cells constitute a spontaneous combustion probability spatial distribution map, which is then output as dynamic risk information to the decision module.

[0009] Furthermore, in the coal yard debris cleaning device based on an automated robot described in this invention, the decision module includes: The cost map construction submodule is used to integrate the spontaneous combustion probability spatial distribution map in dynamic risk information, the robot's real-time position coordinates obtained from the comprehensive data stream, and the pre-stored or identified clutter position coordinates from the comprehensive data stream. The spontaneous combustion probability value is mapped to the risk cost of spatial position, and the pre-stored obstacle occupancy information and the task attractiveness of the distance to the clutter point are superimposed to generate a multi-layer cost map including a risk cost layer, an obstacle layer, and a target layer. The path planning submodule is used to receive the multi-layer cost map generated by the cost map construction submodule. On the multi-layer cost map, starting from the robot's current position obtained from the integrated data stream and aiming to traverse all clutter positions, an improved ant colony optimization algorithm is applied to search for the K collision-free candidate paths with the minimum total cost under the constraints of the risk cost layer and the obstacle layer, where K is an integer greater than 1. The trajectory optimization submodule is used to receive K candidate paths generated by the path planning submodule. For each candidate path, with the dynamic model of the robot arm as a constraint, and with the goal of minimizing the rate of change of the normal contact force of the end effector on the coal pile during the picking up of debris, the model predictive control algorithm is applied to generate a smooth end position and attitude trajectory through rolling optimization, thereby obtaining the candidate operation trajectory corresponding to each path. The strategy integration submodule receives K candidate paths generated by the path planning submodule and K candidate operation trajectories generated by the trajectory optimization submodule. It combines the K candidate paths with the corresponding K candidate operation trajectories to form K complete joint action strategies. Through a cooperative game evaluation model, it evaluates the performance of each strategy in three dimensions: expected cleanup time, cumulative disturbance risk, and energy consumption. It then selects the Pareto optimal strategy set that is not inferior to other strategies in all dimensions and outputs the Pareto optimal strategy set as the joint action strategy candidate set to the control module.

[0010] Furthermore, in the coal yard debris cleaning device based on an automatic robot described in this invention, the control module includes: The simulation verification submodule is used to receive the dynamic digital twin output by the twin module and the joint action strategy candidate set output by the decision module, copy the dynamic digital twin to generate a copy of the current state, convert each strategy in the joint action strategy candidate set into a virtual robot joint control command sequence based on the robot inverse kinematics and dynamics model, perform advanced simulation of the control cycle in the copy of the dynamic digital twin, and predict the change gradient of the porosity field of the coal pile and the robot's predicted state sequence after the strategy is executed. The robustness testing submodule is used to inject Gaussian white noise within a preset range into the sensor input channel of the dynamic digital twin replica during the advanced simulation, and randomly perturb the model parameters of the discrete element simulation in the coupled simulation submodule. It records the execution completion degree and final state of the strategy under this disturbance, and calculates the robustness score of the strategy.

[0011] Furthermore, in the coal yard debris cleaning device based on an automatic robot described in this invention, the control module further includes: The objective function construction submodule receives the porosity field change gradient and robot predicted state sequence predicted by the simulation verification submodule, the robustness score calculated by the robustness test submodule, the dynamic risk information, and the trajectory and velocity information of each strategy in the joint action strategy candidate set obtained from the simulation verification submodule. The optimization objectives are to minimize the norm of the porosity field change gradient, minimize the estimated energy consumption of the strategy calculated from the robot predicted state sequence and trajectory information, maximize the smoothness of the action calculated from the trajectory and velocity information, and maximize the robustness score. A multi-objective evaluation function is constructed, wherein the weight of the porosity field change gradient term is positively correlated with the spontaneous combustion probability value of the strategy execution area in the spontaneous combustion probability space distribution map. The strategy arbitration submodule is used to receive the multi-objective evaluation function constructed by the objective function construction submodule and the joint action strategy candidate set, and to use a non-dominated sorting genetic algorithm to calculate and sort the multi-objective evaluation function values ​​of all strategies in the joint action strategy candidate set, and select the strategy with the highest Pareto ranking and the largest crowding distance as the optimal control law. The instruction decoupling submodule is used to receive the optimal control law determined by the strategy arbitration submodule, parse the optimal control law into a differential control instruction sequence for the left and right drive wheels of the mobile platform, and an instruction sequence for the desired angle, angular velocity and torque of each joint of the robotic arm, and send it to the feedback module.

[0012] Furthermore, in the coal yard debris cleaning device based on an automatic robot described in this invention, the feedback module includes: The instruction execution submodule is used to receive the mobile platform differential control instruction sequence and the robotic arm joint instruction sequence sent by the instruction decoupling submodule of the control module, and drive the robot mobile platform and robotic arm to perform debris picking and handling actions through the current loop and position loop control of the driver. The data acquisition submodule is used to collect the actual rotation angle and rotation speed of each joint in real time through the encoder during the process of the robot being driven by the instruction execution submodule, collect the three-dimensional force and three-dimensional torque at the contact point between the end effector of the robotic arm and the coal pile through the six-dimensional force sensor, and collect the phase current of the drive motor of each joint through the current sensor. The feature calculation submodule is used to receive the actual rotation angle, rotation speed, three-dimensional force, three-dimensional torque and phase current collected by the data acquisition submodule, perform low-pass filtering on the raw signals of the actual rotation angle, rotation speed, three-dimensional force, three-dimensional torque and phase current to eliminate high-frequency noise, then calculate the tracking position error and tracking speed error of the joint, and perform fast Fourier transform on the phase current signal to extract the characteristic frequency components characterizing mechanical vibration.

[0013] Furthermore, in the coal yard debris cleaning device based on an automated robot described in this invention, the step of feeding back the actual execution data to the twin module to correct the digital twin includes: The feature calculation submodule of the feedback module packages the calculated tracking position error, tracking speed error, contact force, contact torque and vibration characteristic frequency data into a correction data package and sends it to the data assimilation submodule of the twin module. The data assimilation submodule receives the correction data packet, parses the actual contact force and actual torque from the correction data packet, compares the actual contact force and actual torque with the contact force and torque predicted by the simulation verification submodule in the advanced simulation, generates force residuals and torque residuals, parses the vibration characteristic frequency from the correction data packet, compares the vibration characteristic frequency with the system modal frequency predicted by the simulation verification submodule in the advanced simulation, generates frequency residuals; The data assimilation submodule uses the force residual, torque residual, and frequency residual as new observation values. In the next data assimilation cycle, it updates the observation equation of the Kalman filter algorithm, performs online estimation and adjustment of the interparticle contact stiffness and damping coefficient in the discrete element simulation of the coupled simulation submodule of the twin module, and simultaneously corrects the effective oxygen diffusion coefficient related to pore structure in the computational fluid dynamics simulation of the coupled simulation submodule of the twin module, so that the simulation output of the dynamic digital twin approximates the actual system response collected by the feedback module.

[0014] Furthermore, in the coal yard debris cleaning device based on an automatic robot described in this invention, the step of feeding back the actual execution data to the control module to optimize the arbitration for the next control cycle includes: The feature calculation submodule of the feedback module sends the tracking position error, tracking speed error, and the current system state vector as performance feedback data to the strategy arbitration submodule of the control module. The strategy arbitration submodule receives the performance feedback data and calculates the deviation vector between the actual execution effect of the previous control cycle and the advanced simulation prediction effect of the simulation verification submodule on multiple objectives. The strategy arbitration submodule dynamically adjusts the weight allocation coefficients of the multi-objective evaluation function in the objective function construction submodule of the control module according to the magnitude and direction of the deviation vector. Specifically, if the actual disturbance is greater than the prediction, the weight of the porosity field change gradient term is increased; if the actual energy consumption is greater than the prediction, the weight of the energy consumption term is increased.

[0015] Beneficial effects of this invention: The coal yard debris cleaning device based on an automated robot provided by this invention has the advantage of constructing a closed-loop intelligent system capable of perception, prediction, decision-making, and self-optimization, thereby fundamentally suppressing the risk of spontaneous combustion of coal piles induced by cleaning operations. Specifically, the perception module generates a comprehensive data stream including coal pile geometry, temperature, texture, and the robot's own state through spatiotemporal alignment and fusion of multiple sensors, providing a high-fidelity and consistent environmental cognition foundation for all subsequent analysis and decision-making. The twin module uses this data stream to dynamically construct and continuously assimilate and update a dynamic digital twin. This twin, through discrete element method and computational fluid dynamics coupled simulation, realizes real-time mapping and prediction of the complex chain reaction process of robot action-coal particle displacement-pore structure-oxygen diffusion, quantifying for the first time in virtual space the microscopic impact of cleaning operations on the thermal stability of coal piles. Based on the dynamic digital twin, the risk prediction module rapidly infers the oxidation reaction through a graph neural network surrogate model and generates a spontaneous combustion probability spatial distribution map by combining Monte Carlo simulation, transforming the difficult-to-observe spontaneous combustion risk into intuitive and quantifiable dynamic risk information, enabling the system to pre-identify high-risk areas. The decision-making module utilizes dynamic risk information to generate a multi-layered cost map, guiding improved ant colony optimization and model predictive control algorithms. Under the constraint of clearing debris, it collaboratively plans a Pareto-optimal joint action strategy across multiple objectives, including accumulated disturbance risk, time, and energy consumption, achieving proactive avoidance from risk perception to action generation. The control module further verifies and arbitrates the strategy in the digital twin through parallel simulation. Using a multi-objective evaluation function combined with adaptive weighting based on dynamic risk information, it selects the execution scheme with the least disturbance to the coal pile structure and the strongest robustness, ensuring that the instruction generation process fully considers and suppresses potential risks. The feedback module collects actual data after execution. On one hand, it corrects the model parameters of the twin module, making the dynamic digital twin continuously approximate the real physical system; on the other hand, it optimizes the decision weights of the control module to achieve online self-tuning, thus forming a reinforced closed loop that continuously learns and improves from actual operational results. Through the organic synergy of the aforementioned modules and the closed-loop flow of data, the device transforms the existing open-loop mechanical operation, which only focuses on the removal of debris, into an intelligent operation that can assess in real time and proactively manage the impact of its own behavior on the complex physicochemical state of the coal pile. This effectively solves the key technical problem of accelerating the low-temperature oxidation process due to the unintentional alteration of the coal pile's pore structure by the robot's operation. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0017] Figure 1 This is a system architecture diagram of the coal yard debris cleaning device based on an automatic robot according to the present invention. Detailed Implementation

[0018] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0019] Please see Figure 1 The present invention provides a coal yard debris cleaning device based on an automatic robot, comprising: The perception module is used to collect visual, temperature, geometric and environmental sensing data of the coal pile surface and output a spatiotemporally aligned comprehensive data stream, which includes the robot's state. The twin module is used to receive the integrated data stream and dynamically construct and update a dynamic digital twin reflecting the pore structure and oxygen diffusion state of the coal pile through multi-physics coupling simulation. The risk prediction module is used to receive the dynamic digital twin, extract the distribution fields of porosity, temperature and oxygen concentration, quickly infer the oxidation reaction through the surrogate model, quantify and predict the probability distribution of spontaneous combustion risk, generate dynamic risk information, and output the dynamic risk information to the decision module and control module. The decision-making module is used to receive the dynamic risk information and the robot status obtained from the integrated data stream, and output a set of joint action strategy candidates through collaboratively optimized path planning and operation trajectory generation. The control module is used to receive the candidate set of joint action strategies and the dynamic risk information, verify the impact of the strategies on the dynamic digital twin through parallel simulation, and select the optimal control law based on multi-objective arbitration in combination with the dynamic risk information, and decouple the optimal control law to generate the underlying control command. The feedback module is used to execute the underlying control commands to drive the robot's actions, collect the actual execution data generated by the robot's actions, and feed the actual execution data back to the twin module to correct the dynamic digital twin, and to the control module to optimize the arbitration of the next control cycle.

[0020] The device's workflow begins with the perception module's comprehensive data acquisition of the working environment. This module is equipped with a vision camera, an infrared thermal imager, a lidar, a torque sensor, and temperature and humidity sensors. The vision camera captures high-resolution images of the coal pile surface and debris, the infrared thermal imager simultaneously records the temperature distribution in the corresponding area, and the lidar acquires precise 3D point cloud geometric information of the coal pile surface by emitting and receiving laser beams. The torque sensor is mounted on the robot joints to measure the forces acting during operation, and the temperature and humidity sensor monitors environmental conditions. A key processing step is data synchronization. All sensors are connected to the same clock source, and the raw signals are timestamped to generate a time-aligned raw dataset. Subsequently, the processing flow registers and denoises the lidar point cloud to reconstruct a 3D geometric model of the coal pile surface. Further, visual images taken at the same moment are extracted from the time-aligned dataset, and their texture information is mapped onto the surface of the 3D geometric model to form a 3D mesh with texture information; simultaneously, temperature data from the infrared thermal imager is extracted and mapped to the corresponding vertices of the 3D mesh, generating a temperature and geometry fusion model that integrates temperature and geometric information. Finally, the module reads the robot's joint encoder angles, torque sensor readings, and temperature and humidity parameters from the dataset. It then packages and encapsulates this information with the temperature and geometry fusion model and the calculated robot pose, outputting a spatiotemporally aligned integrated data stream. This integrated data stream provides a unified and consistent perceptual input for all subsequent modules.

[0021] The twin module receives the integrated data stream, and its core task is to construct and continuously update a dynamic digital twin reflecting the pore structure and oxygen diffusion state of the coal pile. The module first initializes the model, extracting the coal pile boundary contour and estimated initial pore distribution parameters from the temperature and geometry fusion model included in the integrated data stream. Based on this, it configures the particle property parameters and interparticle contact model parameters required for discrete element simulation (DIMS), and sets the initial oxygen concentration field and boundary conditions required for computational fluid dynamics (CFD) simulation based on environmental parameters. After initialization, the coupled simulation submodule runs the DIMS and CFD simulations in parallel. The DIMS, based on the planned path and action instructions from the decision module, simulates the disturbance of coal particles caused by the movement and manipulation of a computational robot, outputting the coal particle displacement, contact force chain network, and pore structure evolution results. The CFD simulation, based on the current pore structure provided by the DIMS, solves the convection-diffusion equations for oxygen in the porous medium, outputting the oxygen concentration distribution and diffusion flux field. The two simulations exchange data at each computational step. The discrete element method (DIM) simulation transmits the updated porosity field to the computational fluid dynamics (CFD) simulation, which in turn feeds back the calculated oxygen concentration field to the DIM simulation, thus achieving thermal-fluid-structure interaction (TFI). To maintain consistency between the digital twin and the real world, the data assimilation submodule acquires the latest temperature and geometric fusion model output by the sensing module in real time during the simulation. It compares the predicted geometric deformation and temperature distribution of the coal pile surface from the coupled simulation submodule with the actual observation data, calculating the geometric and temperature residuals. Using the Kalman filter algorithm, with the residuals as observed values, the interparticle friction coefficient and rolling resistance coefficient in the DIM simulation are adjusted in reverse, thereby correcting the porosity evolution prediction for the next simulation step. After continuous assimilation and correction, the module outputs a dynamic digital twin including the corrected porosity field, temperature field, and oxygen concentration field.

[0022] The risk prediction module receives the dynamic digital twin and aims to quantitatively assess the risk of spontaneous combustion of the coal pile that may be triggered by the cleanup operation. The field variable extraction submodule first extracts the three-dimensional distribution matrices of porosity, temperature, and oxygen concentration from the dynamic digital twin. The reaction inference submodule inputs the distribution matrices into a pre-trained graph neural network surrogate model. This graph neural network uses spatial grid cells as nodes, porosity, temperature, and oxygen concentration as node features, and the connections between adjacent cells as edges. Through forward propagation and multi-layer graph convolution operations, the model can quickly and node-by-node infer the oxidation reaction rate for the next time step, generating an oxidation reaction rate field. The risk quantification submodule uses the oxidation reaction rate field as a heat source term, combines the current temperature matrix with the probability distribution of coal pile activity parameters, and employs a Monte Carlo simulation method for large-scale random sampling. For each sampling, a future temperature evolution path is calculated in parallel. The frequency with which the temperature of all three-dimensional grid cells exceeds the preset spontaneous combustion critical temperature in thousands of evolution paths is statistically analyzed, and this frequency is defined as the probability of spontaneous combustion for that grid cell. The probability values ​​of all cells together constitute a spatial distribution map of the spontaneous combustion probability. The module outputs this spontaneous combustion probability spatial distribution map as dynamic risk information, providing intuitive and quantitative risk guidance for downstream decision-making.

[0023] The decision-making module receives dynamic risk information and robot status data from the integrated data stream, and is responsible for generating a safe and efficient cleanup action plan. The cost map construction submodule integrates the spontaneous combustion probability spatial distribution map, the robot's real-time position coordinates, and pre-stored or identified debris position coordinates. It maps spontaneous combustion probability values ​​to spatial location risk costs, overlays pre-stored obstacle occupancy information, and assigns task attractiveness to distances from debris points, thereby generating a multi-layered cost map including risk cost layers, obstacle layers, and target layers. The path planning submodule operates on this multi-layered cost map, starting from the robot's current position and aiming to traverse all debris positions, applying an improved ant colony optimization algorithm for searching. Under the constraints of the risk cost layer and obstacle layer, the algorithm explores K collision-free candidate paths with the minimum total cost. The trajectory optimization submodule, for each candidate path, uses the robot arm's dynamic model as a constraint, with the optimization objective of minimizing the rate of change of the normal contact force of the end effector on the coal pile during debris pickup. It applies a model predictive control algorithm for rolling optimization to generate a smooth end effector position and attitude trajectory, obtaining a candidate operation trajectory corresponding to each path. The strategy integration submodule combines K candidate paths with K candidate operation trajectories to form K complete joint action strategies. It then evaluates each strategy from three dimensions—expected cleanup time, cumulative disturbance risk, and energy consumption—through a cooperative game evaluation model. The Pareto optimal strategy set that is not inferior to other strategies in all dimensions is selected as the output of the joint action strategy candidate set.

[0024] The control module receives a candidate set of joint action strategies and dynamic risk information, and is responsible for verifying and arbitrating the candidate strategies and generating the final execution instructions. The simulation verification submodule obtains a copy of the current state of the dynamic digital twin from the twin module, converts each strategy in the candidate set into a virtual sequence of robot joint control instructions based on the robot's inverse kinematics and dynamics model, and performs a forward simulation of the control cycle in the digital twin copy. The simulation predicts the gradient of the coal pile porosity field change and the robot's predicted state sequence after the strategy is executed. The robustness testing submodule actively injects uncertainty into the forward simulation, adds Gaussian white noise to the sensor input channels, and randomly perturbs the model parameters of the discrete element simulation in the twin module. It records the execution completion degree and final state of each strategy under this disturbance, thereby calculating the robustness score of the strategy. The objective function construction submodule integrates the porosity field change gradient predicted by the simulation verification submodule, the robot's predicted state sequence, the robustness score calculated by the robustness testing submodule, the dynamic risk information, and the trajectory and velocity information of each strategy to construct a multi-objective evaluation function. The function aims to minimize the gradient of porosity field change and the energy consumption predicted by the strategy, while maximizing the smoothness of the action and the robustness score. The weight of the porosity field change gradient term is positively correlated with the spontaneous combustion probability value of the strategy execution region in the spontaneous combustion probability space distribution map. The strategy arbitration submodule uses a non-dominated sorting genetic algorithm to calculate and sort the multi-objective evaluation function values ​​of all strategies, selecting the strategy with the highest Pareto ranking and the largest crowding distance as the optimal control law. The instruction decoupling submodule parses the optimal control law into a differential control instruction sequence for the left and right drive wheels of the mobile platform, and an instruction sequence for the desired angle, angular velocity, and torque of each joint of the robotic arm.

[0025] The feedback module executes low-level control commands and completes the system's learning loop. The command execution submodule receives control commands and, through the current and position loops of the driver, drives the robot's mobile platform and robotic arm to pick up and transport debris. During the execution of the action, the data acquisition submodule collects the actual joint rotation angle and speed in real time through encoders, collects the three-dimensional force and three-dimensional torque at the contact point between the end effector and the coal pile through a six-dimensional force sensor, and collects the phase current of the drive motor through a current sensor. The feature calculation submodule performs low-pass filtering on the raw signal to eliminate high-frequency noise, then calculates the tracking position error and tracking speed error of the joint, and performs a fast Fourier transform on the phase current signal to extract the characteristic frequency components representing mechanical vibration. The processed data constitutes the actual execution data. The actual execution data is fed back to the twin module and the control module respectively. When fed back to the twin module, the data is packaged into a correction data packet and sent to the data assimilation submodule. The data assimilation submodule parses the actual contact force, actual torque, and vibration characteristic frequencies, and compares them with the contact force, torque, and system modal frequencies predicted by the simulation verification submodule in the advanced simulation, generating force residuals, torque residuals, and frequency residuals. These residuals, as new observations, drive the Kalman filter algorithm to update the observation equations in the next cycle, adjusting the interparticle contact stiffness and damping coefficients in the discrete element simulation online, and simultaneously correcting the effective oxygen diffusion coefficient related to the pore structure in the computational fluid dynamics simulation, thereby continuously approximating the simulation output of the dynamic digital twin to the real system response. When fed back to the control module, the tracking position error, tracking velocity error, and system state vector are sent as performance feedback data to the strategy arbitration submodule. The strategy arbitration submodule calculates the deviation vector between the actual execution effect and the simulation prediction effect in the previous cycle on multiple objectives, and dynamically adjusts the weight allocation coefficients of the multi-objective evaluation function in the submodule based on the deviation. For example, when the actual disturbance is greater than the prediction, the weight of the porosity field change gradient term is increased, realizing online self-tuning and continuous optimization of the control module.

[0026] The operation of the sensing module begins with the coordinated acquisition of signals from multiple sensors by the data synchronization submodule. The vision camera, infrared thermal imager, LiDAR, torque sensor mounted on the robotic arm joint, and ambient temperature and humidity sensor are all connected to the same high-precision clock source. The data synchronization submodule assigns a unified timestamp to the raw signal streams captured by each sensor. For example, at the same moment, a frame of 3D point cloud acquired by the LiDAR, an RGB image captured by the vision camera, a temperature matrix generated by the infrared thermal imager, and readings from the torque sensor and temperature and humidity sensor are all marked with the same time stamp. Through this processing, all sensor data from different sources and with different acquisition frequencies are integrated into a time-aligned raw dataset, laying the foundation for subsequent fusion processing.

[0027] The geometric reconstruction submodule then receives the time-aligned raw dataset. The submodule first parses the LiDAR point cloud data from the dataset. Since the robot is mobile, a single frame of point cloud only reflects a local viewpoint, therefore point cloud registration is required. The submodule uses algorithms such as iterative nearest point to align multiple frames of point clouds acquired at different times and poses to the same coordinate system, stitching them together to create a complete 3D point cloud of the coal pile and its surrounding environment. Next, the stitched point cloud is denoised and filtered to remove outliers caused by dust, flying insects, or sensor noise. Finally, based on the cleaned point cloud, a 3D geometric model of the coal pile surface is reconstructed. The model is represented in the form of triangular meshes, defining the spatial shape of the coal pile surface.

[0028] The data fusion submodule's job is to enrich the physical properties of the 3D geometric model. This submodule receives both the time-aligned raw dataset and the 3D geometric model. It extracts the visual image corresponding to the 3D geometric model at the same time from the dataset. Using computer vision algorithms, it accurately maps texture features from the visual image, such as color and pattern, onto the corresponding triangular faces of the 3D geometric model, thus forming a 3D mesh model with realistic texture information. Simultaneously, the submodule extracts temperature data collected by an infrared thermal imager at the same time from the same dataset. Each pixel value in the temperature data is assigned attributes to the corresponding vertex in the 3D mesh model according to a pre-defined spatial correspondence, generating a temperature and geometry fusion model. This model simultaneously includes the geometry, visual texture, and temperature distribution of the coal pile surface.

[0029] The integrated flow generation submodule is the final step in the perception module, responsible for encapsulating all perceived information. This submodule receives a time-aligned raw dataset and a temperature and geometry fusion model. It reads the angle values ​​from the robot's joint encoders from the raw dataset, combines them with the robot's kinematic model, and calculates the robot's overall pose, including position and orientation. Simultaneously, it reads torque sensor readings to obtain the joint force state and reads environmental parameters from the temperature and humidity sensors. Finally, the submodule packages the temperature and geometry fusion model, the calculated robot pose, joint torque data, and environmental parameters together into a structured integrated data stream. This data stream serves as the single, consistent source of perceived facts for all subsequent modules' decision-making and simulation.

[0030] The model initialization submodule of the twin module receives the integrated data stream from the sensing module. The initialization submodule parses the temperature and geometry fusion model from the stream, extracts the boundary contour of the coal pile, and estimates the initial porosity distribution parameters inside the coal pile based on geometric features and prior material knowledge. These parameters are used to configure the discrete element simulation: setting particle size distribution, density, elastic modulus, and other particulate properties of the coal particles, as well as contact model parameters defining the contact and collision behavior between coal particles. Simultaneously, based on the environmental temperature and humidity parameters included in the integrated data stream, the initialization submodule sets the initial conditions for the computational fluid dynamics simulation, i.e., initializing the oxygen concentration field within the simulation domain and setting the oxygen concentration flux or fixed value conditions at the simulation boundary.

[0031] The coupled simulation submodule then loads the two simulation parameters configured by the model initialization submodule and begins running the discrete element simulation (DEM) and computational fluid dynamics (CFD) simulation in parallel. The DEM receives the planned path and action instructions from the decision module, transforming them into "forces" acting on the virtual coal particle set. The simulation progressively calculates the displacement and velocity of each coal particle under these forces, as well as the contact force network formed between the particles, thereby dynamically predicting the evolution of the coal pile's pore structure. The CFD simulation, based on the real-time porosity distribution field provided by the DEM, treats the coal pile as a porous medium, solves the convection-diffusion equations for oxygen in this medium, and outputs the spatial distribution of oxygen concentration and the diffusion flux field. The two simulation processes are not independent; at the end of each simulation step, the coupled simulation submodule performs data exchange: the DEM sends the updated porosity field to the CFD as the new medium parameters for its next calculation; the CFD feeds back the calculated new oxygen concentration field to the DEM, as oxygen concentration can affect the oxidation reaction rate on the coal particle surface, potentially fine-tuning its physical properties. This exchange enables bidirectional coupling between the pore structure and gas diffusion.

[0032] The data assimilation submodule runs continuously during the simulation process, aiming to correct the simulation model and reduce prediction errors. This submodule monitors the output of the sensing module in real time, acquiring the latest frame of the temperature and geometry fusion model as actual observation data. Simultaneously, it obtains the geometric deformation and temperature distribution of the coal pile surface at the same moment predicted by the coupled simulation submodule. It compares the predicted geometric and temperature data point-by-point with the actual observed geometric and temperature data, calculating the differences between them—the geometric residuals and temperature residuals. Subsequently, the submodule uses a Kalman filter algorithm, treating the residuals as deviations between observed and predicted values. Driven by these deviations, it adjusts key parameters such as the interparticle friction coefficient and rolling resistance coefficient used in the discrete element simulation within the coupled simulation submodule. By adjusting these parameters, the simulation model's prediction of coal particle movement is corrected, making the pore evolution prediction in the next simulation step closer to actual observations. After continuous assimilation and correction, the twin module finally outputs a dynamic digital twin that is highly synchronized with the actual physical state of the coal pile, including the corrected porosity field, temperature field, and oxygen concentration field.

[0033] The field variable extraction submodule of the risk prediction module receives a dynamic digital twin from the twin module. This submodule accesses the twin's internal data structure, which stores the physical field data of the coal pile space in the form of three-dimensional grid cells. The field variable extraction submodule extracts three complete three-dimensional matrices from this structure: a porosity distribution matrix, a temperature distribution matrix, and an oxygen concentration distribution matrix. The element values ​​in each matrix represent the state parameters of the corresponding spatial grid cell, thus discretizing the continuous physical field into a data structure that can be processed by the computer model.

[0034] The reaction inference submodule receives three distribution matrices output by the field variable extraction submodule. A graph neural network model is pre-trained within this submodule. This model models the coal pile space as a graph structure: each 3D grid cell is considered a node in the graph, and the node's feature vector consists of the cell's porosity, temperature, and oxygen concentration values. Edges are established between adjacent grid cells to represent spatial adjacency. The reaction inference submodule inputs the three distribution matrices into the graph neural network. Through its built-in multi-layer graph convolution operations, the network can capture the mutual influence and transmission relationships of node features within their local neighborhoods. After forward propagation, the network finally outputs a scalar value for each node, which is the predicted oxidation reaction rate of that grid cell in the next time step. The predicted rate values ​​of all nodes together constitute the oxidation reaction rate field covering the coal pile space. The surrogate model enables rapid approximate calculations of complex oxidation kinetic processes.

[0035] The risk quantification submodule transforms reaction rate predictions into intuitive risk probabilities. This submodule receives the oxidation reaction rate field generated by the reaction deduction submodule and the current temperature matrix provided by the field variable extraction submodule. The submodule treats the oxidation reaction rate field as an internal heat source causing temperature increases. Considering the uncertainties in the coal pile's coal activity parameters, the submodule employs a Monte Carlo simulation method: randomly selecting a set of parameter values ​​from the probability distribution of the coal activity parameters, and then calculating the temperature evolution path over a future period, considering heat conduction and reaction exothermics, based on the current temperature field and reaction rate field. This process is repeated thousands of times, each time using different random parameter samples, thereby calculating thousands of possible future temperature evolution paths in parallel. For each three-dimensional grid cell in the coal pile, the risk quantification submodule counts the number of paths in all simulations where the cell's temperature exceeds the preset spontaneous combustion critical temperature. Dividing this number by the total number of simulation paths yields the spontaneous combustion probability of that grid cell. Traversing all grid cells yields a spatial distribution map of the spontaneous combustion probability. This map, serving as quantified dynamic risk information, clearly marks high-risk areas in the coal pile that require close monitoring.

[0036] The cost map construction submodule of the decision-making module is responsible for creating an environmental model for navigation decisions. This submodule integrates dynamic risk information from the risk prediction module, provided in the form of a spontaneous combustion probability spatial distribution map. The submodule obtains the robot's real-time position coordinates from the integrated data stream of the perception module and loads a pre-stored list of debris locations, or identifies debris coordinates in real-time by analyzing visual information in the integrated data stream. Next, the submodule performs a mapping operation: each probability value in the spontaneous combustion probability spatial distribution map is converted into a risk cost for the robot to traverse that location, according to preset mapping rules; the higher the probability, the greater the cost. Simultaneously, the submodule loads pre-stored information on fixed obstacles in the coal yard and sets the task attractiveness, typically showing a stronger attraction the closer to the debris point. Finally, the submodule generates a multi-layered cost map, where the risk cost layer reflects spatial risk, the obstacle layer marks impassable areas, and the target layer indicates the location of debris.

[0037] The path planning submodule receives the multi-layered cost map generated by the cost map construction submodule. The submodule uses the robot's current position as the starting point for path search, with the overall goal of traversing all clutter locations. The submodule applies an improved ant colony optimization algorithm for the search. The algorithm simulates the behavior of ants releasing pheromones, navigating the multi-layered cost map where risk and obstacle layers constrain the ants' movement direction. In each iteration, the algorithm tends to select the path with the lower cumulative cost and increases the pheromone concentration on that path. After multiple iterations, the algorithm converges and finds the K collision-free candidate paths with the minimum total cost. K is an integer greater than 1, providing multiple alternatives for subsequent decisions, rather than a unique solution.

[0038] The trajectory optimization submodule plans specific robotic arm operations for each candidate path. This submodule receives K candidate paths generated by the path planning submodule. For each path, the submodule considers the constraints of the robot arm's dynamic model to ensure the planned trajectory is physically feasible. The core optimization objective is to minimize the rate of change of the normal contact force of the end effector on the coal pile surface during debris pickup, aiming to achieve gentle and stable operation to reduce disturbances. The submodule applies a model predictive control algorithm, predicting the system behavior over a future time domain based on the current state in each control cycle, and continuously solving for a series of smooth end-effector position and attitude commands to form a candidate operation trajectory that matches the current candidate path.

[0039] The strategy integration submodule is the final output of the decision-making module. This submodule receives K candidate paths generated by the path planning submodule and K candidate work trajectories generated by the trajectory optimization submodule. The submodule pairs each path with its corresponding work trajectory to form K complete joint action strategies, each including a complete plan of "how to go" and "how to do it." Subsequently, the submodule uses a cooperative game evaluation model to quantitatively evaluate the performance of each strategy from three key dimensions: estimated cleanup time, cumulative disturbance risk, and energy consumption. The evaluation model does not seek a single optimal strategy but rather selects a Pareto-optimal set of strategies; that is, in this set, no strategy is superior to any other strategy in all three dimensions. This Pareto-optimal strategy set, as the final candidate set of joint action strategies, is output to the control module for further verification and arbitration.

[0040] The simulation verification submodule of the control module first receives the dynamic digital twin from the twin module and the joint action strategy candidate set output by the decision module. The submodule copies the current state of the dynamic digital twin, generating a dedicated verification copy to avoid affecting the running master twin. For each strategy in the joint action strategy candidate set, the simulation verification submodule, based on the robot's inverse kinematics and dynamics models, converts the high-level action sequence described in the strategy, such as "move to position A, grasp in posture B," into a series of virtual low-level control command sequences of robot joint angles, angular velocities, or torques. Then, in the virtual environment constructed by the dynamic digital twin copy, the submodule injects the virtual control commands and executes a full control cycle of advanced simulation. The simulation process simulates the state changes experienced by the robot after strictly executing the strategy from the current moment. After the simulation, the submodule outputs two key prediction results: first, the gradient of the coal pile porosity field compared to the start of the simulation, quantifying the potential perturbation of the coal pile structure by the strategy; and second, the predicted state sequence of the robot at the end of the simulation, reflecting the execution result of the strategy.

[0041] The robustness testing submodule, building upon the advanced simulation performed by the simulation verification submodule, further evaluates the stability of the strategies. During the advanced simulation, this submodule actively injects Gaussian white noise with a preset variance into the sensor input channels of the virtual environment to simulate random errors in sensor measurements in the real world. Simultaneously, the submodule randomly perturbs the model parameters used in the discrete element simulation within the twin module, such as fine-tuning particle stiffness, to simulate the uncertainties between the model itself and the real physical world. In this "harsh" simulation environment superimposed with noise and parameter perturbations, the robustness testing submodule re-records the execution of each strategy, including whether the predetermined task can still be completed, and the deviations between the final coal pile state and the robot state and the state without perturbation. Through comparative analysis, the submodule calculates a robustness score for each strategy; the score reflects the strategy's ability to maintain performance under uncertain environments.

[0042] The objective function construction submodule of the control module integrates multi-path information to establish evaluation criteria. This submodule receives the porosity field change gradient predicted by the simulation verification submodule and the robot's predicted state sequence, the robustness score calculated by the robustness testing submodule, dynamic risk information from the risk prediction module, and trajectory and velocity details of each strategy obtained from the simulation verification submodule. The submodule uses these to construct a multi-objective evaluation function. The function includes four optimization objectives: first, minimizing the norm of the porosity field change gradient to suppress structural disturbances; second, minimizing the strategy energy consumption estimated from the robot's predicted state sequence and trajectory information; third, maximizing the motion smoothness calculated from the trajectory and velocity information; and fourth, maximizing the robustness score. The weight of the first objective, suppressing porosity changes, is not fixed but positively correlated with the spontaneous combustion probability value of the strategy execution area in the spontaneous combustion probability space distribution map. When executing in high-risk areas, the system will tend to select strategies that cause less disturbance to the coal pile structure.

[0043] The strategy arbitration submodule receives the multi-objective evaluation function defined by the objective function construction submodule and a set of candidate joint action strategies to be evaluated. The submodule applies a non-dominated sorting genetic algorithm, a multi-objective optimization tool, to handle decision problems involving conflicting objectives. The algorithm evaluates all strategies in the candidate set, obtaining the function value of each strategy on the four objectives. Based on the values, the algorithm performs non-dominated sorting and crowding calculation on the strategies. Non-dominated sorting stratifies the strategies, with higher-ranked strata containing better strategies; crowding measures the distribution density of strategies within the same stratum. Finally, the strategy arbitration submodule selects the strategy with the highest Pareto ranking and the largest crowding distance within the stratum, determining it as the optimal control law. Selecting the strategy with the largest crowding helps maintain decision diversity and avoids premature convergence to local optima.

[0044] The instruction decoupling submodule receives the optimal control law determined by the strategy arbitration submodule. The control law is a high-level behavioral description. The submodule is responsible for "translating" it into instructions that the robot's underlying actuators can directly recognize and execute. The decoupling process consists of two parts: For the mobile platform, the submodule parses the movement-related instructions in the optimal control law into a sequence of differential speed values ​​required by the left and right drive wheels in each subsequent control cycle. For the robotic arm, the submodule parses the operation-related instructions into a sequence of angle, angular velocity, and torque values ​​expected to be achieved by each joint of the robotic arm in each subsequent control cycle. The generated sequence of underlying control instructions is sent to the feedback module to drive the robot's actual movements.

[0045] The instruction execution submodule of the feedback module serves as the interface connecting control commands and physical actions. This submodule receives the underlying control command sequence from the control module's instruction decoupling submodule, including differential control commands for the mobile platform and desired angle, angular velocity, and torque commands for each joint of the robotic arm. The submodule converts these electrical signal commands into physical actions via the robot's servo drivers. For the mobile platform, the differential command is controlled by the driver's current loop, converting it into precise torque output for the left and right hub motors, enabling the platform's movement and steering. For the robotic arm, joint commands are controlled by the driver's position loop, velocity loop, or torque loop, driving the motors of each joint to move, causing the end effector to move along a planned trajectory to complete the picking and handling of debris.

[0046] The data acquisition submodule simultaneously and at high speed acquires actual execution data while the command execution submodule drives the robot's movements. The submodule uses high-precision encoders installed at the robot joints to read the actual rotation angle and speed of each joint in real time. A six-dimensional force sensor installed between the end effector and the wrist measures the three-dimensional force and torque generated when the end effector contacts the coal pile or debris in real time. Furthermore, current sensors connected in series in the joint drive motor circuits acquire the real-time phase current waveforms of each motor. The raw sensor signals are continuously acquired with millisecond-level delays, faithfully recording all state details during the robot's execution.

[0047] The feature calculation submodule processes the massive amounts of raw signals reported by the data acquisition submodule in real time, extracting key features. First, the submodule performs low-pass digital filtering on the raw actual rotation angle, rotation speed, three-dimensional force, three-dimensional torque, and phase current signals to eliminate signal glitches caused by high-frequency electrical noise and mechanical vibration. Next, the submodule calculates the tracking position error and tracking speed error of the joints, i.e., subtracting the desired angle / speed value issued by the control command from the actual angle / speed value acquired by the encoder. These two errors directly reflect the control accuracy of trajectory tracking. Simultaneously, the submodule performs a fast Fourier transform on the filtered phase current signal, converting the time-domain current waveform to the frequency domain, analyzing its spectral components, and extracting characteristic frequency components that characterize resonance or abnormal vibration of the mechanical system. The processed data together constitute the actual execution data used for system correction and optimization.

[0048] The feature calculation submodule of the feedback module packages and integrates a series of key data obtained from calculations, including tracking position error, tracking velocity error, end contact force, end contact torque, and vibration characteristic frequencies extracted from the current signal, into a structured correction data packet. The data packet is sent to the data assimilation submodule of the twin module through the high-speed communication bus inside the system, initiating the correction process for the dynamic digital twin.

[0049] The data assimilation submodule of the twin module receives the correction data packet. The submodule parses two core measured data points from the packet: actual contact force and actual torque. Simultaneously, the submodule accesses the simulation verification submodule of the control module, which predicted the contact force and torque for the selected optimal strategy in the previous advanced simulation. The data assimilation submodule compares the parsed actual contact force / torque with the simulation-predicted contact force / torque item by item, calculates the difference between the two, and generates force residuals and torque residuals. Similarly, the submodule parses the vibration characteristic frequency from the correction data packet and compares it with the system structural modal frequencies predicted by the simulation verification submodule in the advanced simulation, generating frequency residuals.

[0050] The data assimilation submodule takes the newly generated force residuals, torque residuals, and frequency residuals as a new and richer set of observations as input. In the next data assimilation cycle, the submodule updates the observation equations of its Kalman filter algorithm, incorporating the residuals into the state estimation process. Based on the observed residuals, the filtering algorithm back-estimates and adjusts the key parameters of the discrete element simulation in the coupled simulation submodule within the twin module, primarily the contact stiffness coefficient and damping coefficient between particles. Simultaneously, since pore structure characteristics are related to gas diffusion capacity, the algorithm also synchronously corrects the effective oxygen diffusion coefficient related to pore structure in the computational fluid dynamics simulation. Through this continuous online parameter adjustment, the simulation output behavior of the dynamic digital twin is constantly "pulled closer," enabling it to more realistically approximate the actual response collected by the feedback module from the real physical system, thereby maintaining the high fidelity of the virtual model.

[0051] The feature calculation submodule of the feedback module packages another set of data, mainly including tracking position error, tracking velocity error, and a state vector representing the current overall state of the system, into performance feedback data. This data is not used to calibrate the model, but rather to optimize the decision logic itself. The performance feedback data is then sent to the policy arbitration submodule of the control module.

[0052] The strategy arbitration submodule of the control module receives performance feedback data. Based on this data, the submodule calculates the actual execution effect of the control cycle. Specifically, it calculates the actual achieved index values ​​for multiple objectives such as porosity field change and energy consumption. Simultaneously, the submodule retrieves the parallel simulation prediction effect of the simulation verification submodule for the selected strategy in the previous cycle, i.e., the predicted values ​​for each index. The strategy arbitration submodule compares the actual effect with the predicted effect, calculates the deviation vector, and quantifies the specific manifestation of the gap between the simulation prediction and the real world on multiple objectives.

[0053] The strategy arbitration submodule dynamically adjusts the weighting coefficients of each term in the multi-objective evaluation function defined by the objective function construction submodule within the control module based on the magnitude and direction of the calculated deviation vector. The adjustment rules are empirical and adaptive. For example, if the actual disturbance to the coal pile caused by execution is greater than the prediction in parallel simulation, the strategy arbitration submodule increases the weighting coefficient of the porosity field change gradient term in the objective function. This makes the system more inclined to choose a strategy with less structural disturbance in subsequent decisions, thereby compensating for the model's prediction bias. If the actual energy consumption is higher than the prediction, the weight of the energy consumption term is increased accordingly. Through this online, real-time weight adjustment mechanism, the control module's decision logic possesses self-tuning and continuous optimization capabilities. It can adaptively adjust its decision-making tendencies based on feedback from actual execution results to achieve better overall control performance in constantly changing environments.

[0054] The execution of the data synchronization submodule within the perception module involves rigorous time calculations. Raw signals from the visual camera, infrared thermal imager, LiDAR, torque sensor, and temperature / humidity sensor are connected to the same high-precision clock source. The data synchronization submodule timestamps each input signal stream based on the current clock source. The marking process involves the submodule reading the current clock count value whenever a sensor outputs a new data packet, binding this clock value with the packet content, and generating a timestamped raw data unit. This operation is repeated for all sensor channels, ultimately aggregating time-aligned timestamped data units from all channels to generate a time-aligned raw dataset. The geometric reconstruction submodule's computation begins with the parsing of LiDAR point cloud data from the time-aligned raw dataset. The submodule extracts multiple frames of point cloud data belonging to the LiDAR channels and with consecutive timestamps from the dataset. The core of the computation is point cloud registration. The submodule employs an iterative nearest-point algorithm, calculating the minimum sum of squared distances between corresponding points in two consecutive frames of point clouds, iteratively solving for the coordinate transformation matrix that minimizes this sum, thereby transforming the point clouds of subsequent frames into the coordinate system of the first frame. After multi-frame registration, statistical filtering is applied to the stitched point cloud for denoising. The average distance from each point to its neighbors is calculated, and outliers exceeding the standard deviation threshold are removed. Based on the denoised clean point cloud, a Poisson reconstruction or triangulation algorithm is used to calculate the triangular mesh surface of the connection points, reconstructing the 3D geometric model of the coal pile surface. The computation path of the data fusion submodule is parallel. The submodule extracts visual image data with the same timestamp as the 3D geometric model from the time-aligned original dataset. The projection relationship between image pixels and points on the 3D geometric model surface is calculated using camera calibration parameters. The color texture values ​​in the image are then bilinearly interpolated based on the projection relationship and assigned to the vertices of the corresponding triangular facets of the 3D geometric model, forming a 3D mesh with texture information. Simultaneously, the submodule extracts infrared thermal imager temperature matrix data with the same timestamp. Based on the spatial calibration relationship between the infrared thermal imager and the 3D geometric model, each temperature value in the temperature matrix is ​​assigned to the temperature attribute of the corresponding vertex of the 3D mesh through coordinate mapping and interpolation, ultimately generating a temperature and geometry fusion model. The computation of the integrated data stream generation submodule includes state calculation and data encapsulation. The submodule reads the angle value sequence from the robot joint encoders in the time-aligned raw dataset. Using the forward kinematics formula, it calculates the link transformation matrix product from the robot base to the end effector, step by step, to determine the robot's end effector pose in the global coordinate system. Simultaneously, it reads torque sensor readings and temperature and humidity parameters. Finally, the submodule arranges and packages the temperature and geometry fusion model, the calculated robot pose matrix, the joint torque array, and the environmental parameter array according to a predefined structured format to generate the integrated data stream.

[0055] The model initialization submodule in the twin module performs parameter settings based on the integrated data stream. The submodule extracts the 3D coordinates of coal pile boundary points from the temperature and geometry fusion model included in the integrated data stream using an edge detection algorithm. Based on the coal pile geometry and empirical formulas, it estimates the initial porosity distribution parameters within the coal pile. Based on these parameters, it sets the particle size distribution function, particle density, Poisson's ratio, elastic modulus, and other particle properties parameters for the discrete element simulation, as well as the Hertz-Mundling contact model parameters defining the force-displacement relationship during particle collisions. Simultaneously, based on the ambient oxygen concentration and temperature / humidity parameters in the integrated data stream, it sets the initial oxygen concentration values ​​for all grid cells in the solution domain and the oxygen inflow / outflow flux conditions at the boundaries for the computational fluid dynamics simulation. The coupled simulation submodule performs calculations in parallel with dual threads and data exchange. The submodule loads the initialization parameters and starts the discrete element simulation thread and the computational fluid dynamics simulation thread in parallel. The discrete element simulation thread receives the planned path and action instructions from the decision module and converts them into external forces acting on the simulated coal particle group. At each simulation step, the thread calculates the acceleration of each coal particle according to Newton's second law, integrates to obtain velocity and displacement, and calculates the contact force between particles based on the overlap, thereby updating the position and pore structure of all coal particles. The computational fluid dynamics simulation thread reads the current porosity field provided by the discrete element simulation and substitutes it as a property of the porous medium into the convection-diffusion equation. The thread discretizes the computational domain spatially and temporally using the finite volume method, iteratively solves the linear equation system, and obtains the oxygen concentration value and diffusion flux of each grid cell. At each synchronization step, the coupled simulation submodule performs data exchange operations: reads the updated porosity field data from the discrete element simulation memory space and writes it into the medium parameter memory area of ​​the computational fluid dynamics simulation; simultaneously, it reads the oxygen concentration field data from the computational fluid dynamics simulation memory space and writes it into the environmental parameter memory area of ​​the discrete element simulation. The calculation of the data assimilation submodule is a closed-loop correction. The submodule obtains the latest temperature and geometric fusion model output by the sensing module in real time as observation values. At the same time, it obtains the simulation prediction of the geometric deformation and temperature distribution of the coal pile surface at the corresponding time from the coupled simulation submodule. The submodule calculates the coordinate difference between observed and predicted values ​​at each surface grid point, and takes the root mean square of all point differences as the geometric residual; it also calculates the difference between observed and predicted temperature values, and takes the root mean square as the temperature residual. The submodule uses a Kalman filter algorithm, defining the system state as the interparticle friction coefficient and rolling resistance coefficient in the discrete element simulation. The algorithm predicts observed values ​​based on the current state, compares the predicted observed values ​​with the actual observed geometric and temperature residuals, and calculates the Kalman gain. Then, it uses the Kalman gain to update the state variables, generating optimal estimates of the friction coefficient and rolling resistance coefficient. The submodule writes the updated parameter values ​​back to the discrete element simulation thread of the coupled simulation submodule, correcting the calculation basis for its next simulation step.The assimilated simulation output includes the corrected porosity field, temperature field, and oxygen concentration field, forming a dynamic digital twin.

[0056] The field variable extraction submodule within the risk prediction module performs data access and rearrangement computations. This submodule receives a dynamic digital twin, whose internal data is stored in memory as three-dimensional arrays. Using pointers or indices, the field variable extraction submodule directly accesses the three three-dimensional arrays representing porosity, temperature, and oxygen concentration, and copies them completely to its independent working memory space, forming three-dimensional matrices for porosity, temperature, and oxygen concentration. The reaction inference submodule's computations are based on graph neural network inference. This submodule inputs the three distribution matrices into a pre-trained graph neural network model. When constructing the computational graph, the network maps each three-dimensional grid cell to a graph node. The node's feature vector is a three-dimensional vector composed of the corresponding porosity, temperature, and oxygen concentration values. Edges between nodes are established according to the six-neighbor or twenty-six-neighbor connection rules of the three-dimensional grid. The network performs forward propagation computation. Each layer of graph convolution operation can be described as follows: For each node, firstly, the feature vectors of its neighboring nodes are aggregated, which can be done by taking the mean, summing, or maximizing the value; then, the aggregated neighbor features are linearly transformed with the node's own features and weighted summed, and then passed through a nonlinear activation function; finally, new node features are output. After multiple layers of such graph convolution operations, the final output layer calculates a scalar value for each node, which is the oxidation reaction rate inferred by the model for the next time step. The output scalars of all nodes are arranged in their original spatial positions to generate an oxidation reaction rate field. The calculation of the risk quantification submodule involves large-scale stochastic simulation. The submodule uses the oxidation reaction rate field as the volumetric heat source intensity distribution, combined with the current temperature matrix as the initial condition. Before the simulation begins, the submodule loads a probability distribution model of coal activity parameters, such as the normal distribution of activation energy. For each Monte Carlo sampling, the submodule randomly selects a set of activation energy parameters from this distribution. Based on these parameters, the current temperature field, and the reaction rate field, partial differential equations considering heat conduction and exothermic reactions are solved numerically to calculate the time-varying temperature sequence of each grid cell over a future period—that is, a temperature evolution path. The submodule initiates thousands of such computational tasks in parallel, each using different random sampling parameters to generate thousands of independent future temperature evolution paths. After the simulation is complete, for each 3D grid cell, the risk quantification submodule traverses all thousands of evolution paths, checking whether the cell's temperature curve exceeds the preset spontaneous combustion critical temperature at any given time. The number of paths exceeding the critical temperature is counted, and this number is divided by the total number of simulated paths; the result is the probability of spontaneous combustion for that cell. This calculation is performed on all grid cells, generating a spatial distribution map of the spontaneous combustion probability.

[0057] The cost map construction submodule in the decision-making module performs layered map generation. This submodule reads the probability value of each grid cell in the spontaneous combustion probability spatial distribution map and converts it into a positive cost value using a linear or nonlinear mapping function; higher probability means higher cost, generating a risk cost layer. The submodule loads pre-stored obstacle polygon vertex coordinates and marks obstacle areas as having infinite cost using a rasterization algorithm, generating an obstacle layer. Based on the robot's real-time position obtained from the integrated data stream and pre-stored or identified debris positions, the submodule calculates the reciprocal of the Euclidean distance from each location on the map to the nearest debris point; closer distances result in higher attraction values, generating a target layer. Finally, the three layers of data are overlaid to generate a multi-layered cost map. The path planning submodule performs calculations based on an iterative search using an improved ant colony optimization algorithm. During algorithm initialization, a large number of virtual ants are randomly placed on the multi-layered cost map, each starting from the robot's current position. The ants' movement decisions are probabilistic; they tend to choose paths with lower integrated costs and higher pheromone concentrations in adjacent grid cells. The overall cost is a weighted sum of the grid values ​​corresponding to the risk cost layer, obstacle layer, and target layer. Whenever an ant completes a closed path traversing all debris points from the starting point back to the starting point, the algorithm calculates the total cost of that path. Then, the pheromone concentration is updated along this path; paths with lower total costs receive a greater pheromone boost. Simultaneously, pheromones evaporate at a certain rate. After multiple iterations, the pheromone concentration on low-cost paths is significantly higher than on other paths. From the final pheromone distribution, the algorithm extracts the K collision-free paths with the minimum total cost as candidate paths. The trajectory optimization submodule uses a rolling model predictive control framework. For a candidate path, the submodule discretizes it into a series of path points. Starting from the current moment, it looks forward to a prediction time domain at fixed time intervals. In each control cycle, the submodule uses the robotic arm dynamics model as constraints and minimizes the rate of change of the end effector's normal contact force with the coal pile as the optimization objective to construct an optimal control problem in the finite time domain. By solving this optimization problem, the optimal joint acceleration sequence for a future prediction time domain starting from the current moment is obtained. However, only the first control variable of this sequence is implemented. In the next control cycle, the submodule re-predicts and optimizes based on the new system state, repeating this process until a smooth end-position and attitude trajectory is generated as a candidate operation trajectory. The strategy integration submodule performs multi-objective evaluation and selection. It pairs each of the K candidate paths with one of the K candidate operation trajectories. For each pair forming a joint action strategy, the submodule predicts its total cleanup time using deterministic calculations, calculates the cumulative disturbance risk of its movement on the risk cost layer using integration, and calculates its estimated energy consumption using a dynamic model. Subsequently, the submodule applies a cooperative game evaluation model to evaluate the K strategies. The evaluation process includes normalizing the three objective values ​​and constructing a dominance relationship matrix between strategies.A strategy considered Pareto-dominated means that it is no worse than the latter in terms of cleanup time, cumulative disturbance risk, and energy consumption, and is at least strictly superior to the latter in terms of these objectives. The submodule, through comparison, selects all strategies not dominated by any other strategy, forming a Pareto-optimal strategy set, which serves as a candidate set for joint action strategies.

[0058] The simulation verification submodule within the control module performs calculations including model conversion and virtual execution. The submodule replicates the dynamic digital twin. For each strategy in the candidate set of joint action strategies, the submodule reads its path point sequence and work trajectory point sequence. Through robot inverse kinematics calculations, the end-effector pose at each trajectory point is solved into a corresponding set of robot joint angle values. Then, through inverse dynamics calculations, combined with parameters such as robot mass and inertia matrix, the joint torque sequence required to track this joint angle sequence is calculated. This series of joint angle and torque sequences constitutes a virtual robot joint control command sequence. The submodule injects this command sequence into the dynamic digital twin replica and runs the simulation of the control cycle. During the simulation, the discrete element simulation in the digital twin replica calculates the coal particle motion based on the virtual robot's actions, and the computational fluid dynamics simulation simultaneously calculates oxygen diffusion. After the simulation, the submodule extracts the porosity field of the final simulation state from the replica, subtracts it from the initial porosity field, divides it by the grid spacing, and calculates the change of the porosity field in three spatial directions using the central difference method. It then calculates the norm of the gradient as a measure of the porosity field change gradient. Simultaneously, it records the joint states of the robot at the end of the simulation as the robot's predicted state sequence. The robustness testing submodule repeats the above verification in a perturbed simulation environment. While the simulation verification submodule performs advanced simulation, the robustness testing submodule intervenes. At each simulation step, it superimposes noise values ​​randomly sampled from a Gaussian distribution onto the virtual sensor readings. Simultaneously, at the start of the simulation, it randomly selects model parameters from the discrete element simulation, such as particle stiffness, and adds a random perturbation ratio to their nominal values. Under this perturbed simulation environment, the execution of the strategy is re-recorded. The strategy execution completion rate is calculated by comparing the positional deviation between the actual reached path point and the target path point. The final state includes the porosity change under perturbation and the robot state. The robustness score is calculated by weighting the deviations of various results between the unperturbed and perturbed simulations. The objective function construction submodule constructs a mathematical evaluation model. This submodule receives the porosity field change gradient, the robot's predicted state sequence, the robustness score, the spontaneous combustion probability value from the dynamic risk information, and the trajectory and velocity information of each strategy. The submodule calculates the strategy's estimated energy consumption by integrating the dot product of joint torques and joint velocities in the robot's predicted state sequence. It calculates motion smoothness by taking the second difference of the position and velocity sequences of the trajectory points and then calculating their norm. The submodule defines the multi-objective evaluation function as a weighted sum of four components: the porosity field change gradient norm term, the strategy's estimated energy consumption term, the reciprocal of the motion smoothness term, and the reciprocal of the robustness score. The weight coefficient of the porosity field change gradient term is obtained by multiplying the average probability value of the strategy execution region in the spontaneous combustion probability space distribution map by a positive proportionality coefficient. The strategy arbitration submodule executes a multi-objective optimization algorithm.The submodule inputs all strategies and their corresponding four objective function values ​​from the candidate set of joint action strategies into a non-dominated sorting genetic algorithm. The algorithm first performs fast non-dominated sorting, dividing the strategies into multiple front layers by comparing their strengths and weaknesses across all objectives. The first front layer includes all strategies not dominated by any other strategy. Then, it calculates the crowding distance between strategies within the same front layer, which is the normalized sum of the differences between a strategy and its neighboring strategies at each objective function value. Based on the sorting and crowding, the algorithm performs selection, crossover, and mutation operations to generate a new generation of strategies. After multiple generations of evolution, the algorithm selects the strategy with the largest crowding distance in the front layer with the highest non-dominated sorting from the final population, determining it as the optimal control law. The instruction decoupling submodule calculates the decomposition and serialization of control instructions. The submodule parses the optimal control law, extracting the expected linear and angular velocities of the mobile platform at each control point on the path. Based on the robot's kinematic model, the linear and angular velocities are converted into independent rotational speeds of the left and right drive wheels, generating a differential control instruction sequence. Simultaneously, the expected joint angles, angular velocities, and torques corresponding to each trajectory point of the robotic arm are analyzed, arranged in chronological order, and the instruction sequence for each joint of the robotic arm is generated.

[0059] The data acquisition submodule in the feedback module performs analog-to-digital signal conversion and reading. The encoder outputs quadrature pulse signals; the submodule measures the number of pulses within a fixed time window using a counter to calculate the actual joint angle and rotational speed. The six-dimensional force sensor outputs multiple analog voltage signals; the submodule acquires these signals at a high sampling rate using an analog-to-digital converter and converts the voltage values ​​into physical quantities of three-dimensional force and torque based on the sensor's calibration matrix. The current sensor obtains voltage signals through Hall elements or sampling resistors, which are also converted to instantaneous phase current values ​​via analog-to-digital conversion. The feature calculation submodule performs signal processing and feature extraction. It applies digital low-pass filters, such as Butterworth filters, to the raw actual angle, rotational speed, three-dimensional force, three-dimensional torque, and phase current signals, using recursive calculations via difference equations to filter out high-frequency noise above the cutoff frequency. For calculating joint tracking position and speed errors, the submodule reads the desired joint angle and angular velocity commands currently sent by the instruction execution submodule, subtracts the actual angle and velocity values ​​acquired by the data acquisition submodule, and obtains the instantaneous error. The calculation of the Fast Fourier Transform (FFT) for the phase current signal involves the submodule applying a butterfly algorithm to a discrete sequence of phase currents over a period of time, transforming it from the time domain to the frequency domain to obtain the spectrum. Peak frequency points with amplitudes significantly higher than the background noise are then identified in the spectrum, and characteristic frequency components are extracted.

[0060] In the process of feeding back actual execution data to the twin module, the computational path involves residual generation and model parameter updates. The feature calculation submodule of the feedback module packages tracking position error, tracking velocity error, contact force, contact torque, and vibration characteristic frequency data. The data assimilation submodule of the twin module receives and parses the correction data packet. This submodule reads the actual contact force and actual torque vectors, and simultaneously obtains the contact force and torque vectors predicted for the same strategy in the advanced simulation from the simulation verification submodule of the control module. It calculates the force residual, i.e., the actual contact force vector minus the predicted contact force vector. It also calculates the torque residual, i.e., the actual torque vector minus the predicted torque vector. The submodule reads the vibration characteristic frequency values, and simultaneously obtains the predicted system dominant mode frequency values ​​from the simulation verification submodule. It calculates the frequency residual, i.e., the difference between the actual characteristic frequency and the predicted mode frequency. The data assimilation submodule uses the force residual, torque residual, and frequency residual as new observation vectors to update the observation equations of the Kalman filter algorithm. Based on the new observation values, the algorithm recalculates the Kalman gain and estimates the state variables. The state variables here are expanded to include the interparticle contact stiffness coefficient and damping coefficient from the discrete element simulation. The algorithm outputs the optimal estimates of the coefficients. The submodule writes the updated coefficient values ​​into the discrete element simulation parameter memory area of ​​the coupled simulation submodule. Simultaneously, based on the correlation model between pore structure and gas diffusion, the algorithm derives the correction amount of the effective oxygen diffusion coefficient from the updated particle parameters and writes it synchronously into the parameter memory area of ​​the computational fluid dynamics simulation.

[0061] In the process of feeding back actual execution data to the control module, the calculation path involves performance evaluation and parameter self-tuning. The feature calculation submodule of the feedback module packages and sends the tracking position error, tracking velocity error, and the current system state vector. The strategy arbitration submodule of the control module receives this performance feedback data. The submodule calculates the actual execution effect of the previous control cycle. The actual disturbance is evaluated by comparing the porosity change calculated by the newly reconstructed coal pile model of the sensing module before and after the current operation. The actual energy consumption is calculated by multiplying and integrating the joint torque and actual joint velocity collected by the data acquisition submodule. At the same time, the submodule retrieves the parallel simulation prediction effect of the simulation verification submodule on the selected optimal strategy in the previous cycle, that is, the predicted porosity change gradient and the predicted energy consumption. The strategy arbitration submodule calculates the deviation vector, that is, the actual disturbance minus the predicted disturbance, and the actual energy consumption minus the predicted energy consumption. The submodule adjusts the weight allocation coefficients according to the deviation vector. If the deviation value is positive (the actual disturbance is greater than the predicted disturbance), the weight coefficient of the porosity field change gradient term in the objective function construction submodule is increased by a certain proportion. If the deviation between actual energy consumption and predicted energy consumption is positive, the weighting coefficient of the energy consumption item is increased by a certain proportion. This proportion is usually a preset small positive number to ensure that the weighting adjustment is gradual. This calculation enables online adaptive adjustment of the control module's evaluation criteria.

[0062] In the specific application scenario of a coal yard in a thermal power plant, coal is piled up, and debris such as stones and wood mixed in between needs to be removed to ensure the smooth operation of the coal conveying system. Existing automated cleaning methods inevitably cause compression, trampling, and rakeing of the coal pile during movement and grabbing. This physical disturbance changes the packing state between coal particles and reshapes the pore network. The change in pore structure directly affects the diffusion path and rate of oxygen in the coal pile, which may lead to oxygen enrichment or stagnation in local areas, thereby disrupting the original thermal equilibrium, potentially accelerating the low-temperature oxidation process, and creating a risk of spontaneous combustion. The core design concept of the device in this invention is to transform the cleaning operation from a simple mechanical task of "removing debris" into a complex system control problem that requires real-time assessment and proactive management of the "impact of the operation on the thermal state of the coal pile".

[0063] After the device is activated, its perception module begins a multi-dimensional scan of the work area. Visual cameras, infrared thermal imagers, and LiDAR deployed at fixed locations on the robot and at the site work synchronously, while torque sensors at the robot's joints and temperature and humidity sensors on-site continuously collect data. The raw signals from all sensors are connected to a unified precision clock source. The data synchronization submodule within the perception module adds a timestamp accurate to the microsecond to every incoming data point—whether it's a point cloud frame from the LiDAR, an image from the camera, or a sensor reading. This processing ensures that all subsequent calculations are based on spatiotemporally aligned data, avoiding decision-making errors caused by data asynchrony. The geometric reconstruction submodule specifically processes the LiDAR point cloud from this time-stamped data, using a point cloud registration algorithm to stitch together multi-view fragments collected during robot movement into a complete three-dimensional surface model of the coal pile, and filtering out noise points. The data fusion submodule performs information enhancement. It extracts ordinary visible light images from the data at the same time and "wraps" their color and texture information onto the surface of the 3D geometric model, making the model more realistic. Simultaneously, it extracts infrared thermal imager data and assigns temperature values ​​to the corresponding vertices of the 3D model, generating a temperature-geometry fusion model that integrates geometry, texture, and temperature. Finally, the comprehensive flow generation submodule gathers all information: including the aforementioned fusion model, the robot's pose and joint forces calculated from the encoder and kinematic model, and environmental temperature and humidity, packaging them into a structured comprehensive data stream. This data stream is the sole source of factual information about the current state of the world, trusted by all subsequent advanced cognitive and decision-making modules.

[0064] The twin module's job is to create a virtual replica that can evolve synchronously with the real coal pile and allows for "if-then" scenario simulations. The model initialization submodule receives the integrated data stream and parses the physical outline and initial state of the coal pile. Based on this, it configures the discrete element simulation (DEM) to set the physical properties and interaction rules of millions of virtual coal particles; simultaneously, it configures the computational fluid dynamics (CFD) simulation to set the initial oxygen distribution in space. The coupled simulation submodule then runs these two simulations in parallel. The DEM simulation handles the "solid" part: upon receiving the planned path, it calculates the impact, compression, and grasping effects of each action of the virtual robot on the virtual coal particle swarm, predicting how the coal particles will displace, how contact forces will be transmitted, and how the final pore structure will be reshaped. The CFD simulation handles the "fluid" part: based on the constantly changing pore structure provided by the DEM simulation, it calculates how oxygen flows and diffuses through the winding channels. The two simulations are not isolated; they are tightly coupled at each computational step, exchanging the latest data: updated pore structures immediately affect the calculation of oxygen diffusion, while new oxygen concentrations may feed back and affect the simulation of oxidation reactions on the coal particle surface. To ensure the virtual world doesn't deviate from reality, the data assimilation submodule acts as a calibrator. It continuously compares the differences between the virtual simulation's predicted coal pile surface deformation and temperature and the real-time observation data transmitted back by the sensing module. Using the Kalman filter algorithm, the submodule reverse-engineers which parameters in the discrete element simulation need fine-tuning, ensuring that the simulator's predicted output continuously approximates the sensor's measurements. This dynamically calibrated digital twin is no longer a static model, but a dynamic digital twin capable of reflecting and even predicting the dynamic changes in porosity, temperature, and oxygen concentration within the coal pile in real time.

[0065] The risk prediction module's task is to make probabilistic predictions about the future. The field variable extraction submodule extracts complete three-dimensional data of three key physical fields—porosity, temperature, and oxygen concentration—from the dynamic digital twin. The reaction inference submodule inputs the data into a pre-trained graph neural network surrogate model. The model treats the coal pile space as a vast network composed of countless interconnected grid cells. Through graph convolution operations, the network efficiently captures the physical interactions between each cell and its surrounding cells, quickly inferring the possible oxidation reaction rate at each location within a short period, thus generating an oxidation reaction rate field distribution map. The risk quantification submodule then performs large-scale stochastic simulations based on this. Considering the inherent uncertainty of the coal quality, the submodule uses the Monte Carlo method to randomly select thousands of different combinations of coal activity parameters. For each parameter combination, using the current oxidation reaction rate field as a heat source, it calculates thousands of possible temperature evolution paths within the coal pile over a future period. Finally, it counts how many of these future paths at each location will exceed the critical temperature for spontaneous combustion, defining the proportion as the probability of spontaneous combustion at that location. The final output is a spatial distribution map of the probability of spontaneous combustion, which visually indicates which areas on the coal pile are "high-risk areas" that require extra caution during the cleanup operation in the form of a probability cloud.

[0066] The core challenge facing the decision-making module is how to plan a route and action sequence with the "least disturbance and safety" while traversing risky areas to clear debris. The cost map construction submodule converts the spontaneous combustion probability spatial distribution map into navigation costs that the robot can understand. It maps probability values ​​to passage costs, with the passage cost increasing exponentially for areas with higher probabilities. Simultaneously, fixed obstacle information and task attractiveness to the location of debris points are overlaid on the map. This generates a multi-layered cost map, including information on risk, obstacles, and the objective. The path planning submodule runs an improved ant colony optimization algorithm on this special map. The algorithm simulates ant colony exploration, guiding the virtual "ants" to try to connect all debris points while avoiding high-cost risk areas and bypassing obstacles. After multiple iterations, the algorithm converges, outputting several candidate global paths with the minimum total cost, providing the robot with multiple feasible macroscopic walking schemes. The trajectory optimization submodule then designs specific picking actions for the robot arm for each macroscopic path. Constrained by robotic arm dynamics, its core optimization objective is to ensure that the vertical contact force on the coal pile is as smooth as possible when the end effector contacts and grasps debris, avoiding sudden punctures or pulls. Through model predictive control algorithms, the submodule generates a series of smooth and stable end-effector trajectories through rolling optimization. The strategy integration submodule then pairs several macroscopic paths with their corresponding fine-grained trajectories, forming several complete "movement + operation" joint strategy packages. Using a cooperative game theory evaluation model, it comprehensively evaluates these strategy packages from three dimensions: estimated time, cumulative risk of the traversed areas, and overall energy consumption, selecting the "Pareto optimal" strategy set. None of the strategies in this set is absolutely superior to others in terms of time, risk, and energy consumption; they represent optimal compromises under different emphases and are submitted as a candidate set of joint action strategies to the control module for final decision.

[0067] The control module is responsible for the final decision-making and instruction conversion. The simulation verification submodule first performs a "pre-run" for each candidate strategy. It replicates the current dynamic digital twin as a sandbox, converting the movements and actions described in the strategy into a series of detailed control instructions for virtual joint motors through the robot's inverse kinematics and dynamics model. Then, it rigorously runs the complete control cycle in the digital twin copy. This advanced simulation predicts the extent to which the porosity field of the coal pile will change after executing the strategy, and what state the robot itself will reach. The robustness testing submodule introduces "uncertainty" into the pre-run, actively adding noise to sensor readings and randomly perturbing some parameters of the simulation model to test the robustness of each strategy under imperfect real-world conditions and provide a robustness score. The objective function construction submodule integrates all information to construct a multi-objective evaluation function to quantify the merits of each strategy. The function considers four objectives simultaneously: minimizing the perturbation gradient on the coal pile porosity structure, minimizing the estimated energy consumption, maximizing the smoothness of the action, and maximizing the robustness score. Crucially, the weights for suppressing pore disturbances are not fixed, but rather positively correlated with the risk value of the main execution area of ​​the strategy in the spontaneous combustion probability space distribution map. In high-risk areas, the system will instinctively be more cautious. The strategy arbitration submodule applies a non-dominated sorting genetic algorithm to handle multi-objective optimization problems. The algorithm sorts all candidate strategies and ultimately selects the strategy that performs best in all aspects and maintains decision diversity, which is then determined as the optimal control law. The instruction decoupling submodule translates high-level behavioral descriptions into instructions that the robot's underlying hardware can directly execute: including the speed difference instruction sequence of the left and right wheels of the mobile platform, and the angle, speed, and torque instruction sequences of each joint of the robotic arm.

[0068] The feedback module is a crucial link in the system's transition from theory to practice and its learning from practice. The instruction execution submodule receives low-level control commands and, through servo control via current loops and position loops, precisely drives the robot's moving platform and robotic arm to complete the actual picking and handling of debris. Simultaneously, the data acquisition submodule records the actual execution at high speed and synchronously: the actual joint angles and speeds fed back by the encoders, the actual force and torque measured by the six-dimensional force sensor when the end effector contacts the coal pile, and the real-time current waveform of the motor driver. The feature calculation submodule filters and denoises the raw signal, calculates the tracking error between the actual trajectory and the desired trajectory, and performs spectral analysis on the current signal to extract characteristic frequencies representing mechanical vibration. The processed actual execution data is then fed back in two separate paths.

[0069] The data is fed back to the data assimilation submodule of the digital twin module. This submodule compares the actual collected contact force, torque, and vibration frequency with the corresponding values ​​predicted in the advanced simulation, calculating the force residual, torque residual, and frequency residual. These residuals, as new evidence, drive the Kalman filter algorithm to update its state estimate again, thereby adjusting the fine parameters such as the contact stiffness and damping between coal particles in the discrete element simulation online, and simultaneously correcting the oxygen diffusion coefficient in the computational fluid dynamics simulation. Through this process, the simulation behavior of the dynamic digital twin is continuously fine-tuned, making its mapping of the physical world increasingly realistic.

[0070] Another feedback loop leads to the strategy arbitration submodule of the control module. Based on the actual execution results, this submodule calculates the deviation between the actual pore disturbance and energy consumption of the previous cycle and the predicted values ​​from the advanced simulation. Based on the deviation vector, the submodule dynamically adjusts the weight allocation coefficients of the multi-objective evaluation function within the submodule. For example, if the actual disturbance is greater than the prediction, the weight of the objective "suppressing pore disturbance" is automatically increased in the next decision, making the system more conservative and cautious in subsequent decisions. This online self-tuning mechanism based on actual effect feedback enables the control system's decision logic to continuously learn and self-optimize, adapting to changes in coal pile characteristics and model uncertainties over long-term operation, ultimately achieving a long-term dynamic balance between safety and efficiency.

Claims

1. A coal yard debris cleaning device based on an automated robot, characterized in that, include: The perception module is used to collect visual, temperature, geometric and environmental sensing data of the coal pile surface and output a spatiotemporally aligned comprehensive data stream, which includes the robot's state. The twin module is used to receive the integrated data stream and dynamically construct and update a dynamic digital twin reflecting the pore structure and oxygen diffusion state of the coal pile through multi-physics coupling simulation. The risk prediction module is used to receive the dynamic digital twin, extract the distribution fields of porosity, temperature and oxygen concentration, quickly infer the oxidation reaction through the surrogate model, quantify and predict the probability distribution of spontaneous combustion risk, generate dynamic risk information, and output the dynamic risk information to the decision module and control module. The decision-making module is used to receive the dynamic risk information and the robot status obtained from the integrated data stream, and output a set of joint action strategy candidates through collaboratively optimized path planning and operation trajectory generation. The control module is used to receive the candidate set of joint action strategies and the dynamic risk information, verify the impact of the strategies on the dynamic digital twin through parallel simulation, and select the optimal control law based on multi-objective arbitration in combination with the dynamic risk information, and decouple the optimal control law to generate the underlying control command. The feedback module is used to execute the underlying control commands to drive the robot's actions, collect the actual execution data generated by the robot's actions, and feed the actual execution data back to the twin module to correct the dynamic digital twin, and to the control module to optimize the arbitration of the next control cycle.

2. The coal yard debris cleaning device based on an automatic robot according to claim 1, characterized in that, The sensing module includes: The data synchronization submodule is used to acquire the raw signals from the visual camera, infrared thermal imager, lidar, torque sensor and temperature and humidity sensor under the same clock source, timestamp the raw signals, and generate a time-aligned raw dataset. The geometric reconstruction submodule is used to receive the time-aligned raw dataset, parse the lidar point cloud data from the time-aligned raw dataset, perform point cloud registration and denoising, and reconstruct the three-dimensional geometric model of the coal pile surface. The data fusion submodule is used to receive the time-aligned original dataset and the three-dimensional geometric model, extract visual images at the same time from the time-aligned original dataset, map the texture features of the visual images to the corresponding surfaces of the three-dimensional geometric model to form a three-dimensional mesh with texture information; and simultaneously extract the temperature data of the infrared thermal imager from the time-aligned original dataset, map the temperature data to the corresponding vertices of the three-dimensional mesh to generate a temperature and geometry fusion model. The integrated stream generation submodule is used to receive the time-aligned raw dataset and the temperature and geometry fusion model, read the angle and torque sensor readings of the robot joint encoder from the time-aligned raw dataset, combine the environmental parameters of the temperature and humidity sensor read from the time-aligned raw dataset, package and encapsulate the temperature and geometry fusion model, robot pose, joint torque and environmental parameters, generate an integrated data stream, and output it to the twin module.

3. The coal yard debris cleaning device based on an automatic robot according to claim 2, characterized in that, The twin module includes: The model initialization submodule is used to receive the integrated data stream, extract the coal pile boundary and initial pore distribution parameters from the temperature and geometry fusion model included in the integrated data stream, configure the particulate property parameters and contact model parameters of the discrete element simulation, and set the initial oxygen concentration field and boundary conditions of the computational fluid dynamics simulation according to the environmental parameters included in the integrated data stream. The coupled simulation submodule receives discrete element simulation (DEM) and computational fluid dynamics (CFD) simulation parameters configured by the model initialization submodule, and runs the DEM and CFD simulations in parallel. The DEM receives planning paths and action instructions from the decision module and calculates coal particle displacement, contact force, and pore structure evolution. The CFD, based on the current pore structure provided by the DEM, solves for the concentration distribution and diffusion flux of oxygen in the pore network. At each simulation step, the coupled simulation submodule exchanges the porosity field updated by the DEM and the oxygen concentration field calculated by the CFD. The data assimilation submodule is used to acquire the latest temperature and geometry fusion model output by the sensing module in real time during the simulation process. It compares the geometric deformation and temperature distribution of the coal pile surface predicted by the coupled simulation submodule with the actual observation data of the latest temperature and geometry fusion model, calculates the geometric residual and temperature residual, and uses the Kalman filter algorithm to adjust the interparticle friction coefficient and rolling resistance coefficient of the discrete element simulation in the coupled simulation submodule in reverse, using the geometric residual and temperature residual as the observed values, so as to correct the pore evolution prediction of the next simulation step. It outputs a dynamic digital twin including the corrected porosity field, temperature field and oxygen concentration field to the risk prediction module and the control module.

4. The coal yard debris cleaning device based on an automatic robot according to claim 3, characterized in that, The risk prediction module includes: The field variable extraction submodule is used to receive the dynamic digital twin output by the twin module and extract the three-dimensional matrix of porosity, the three-dimensional matrix of temperature, and the three-dimensional matrix of oxygen concentration from the dynamic digital twin. The reaction inference submodule is used to receive the porosity matrix, temperature matrix, and oxygen concentration matrix extracted by the field variable extraction submodule, and input the porosity matrix, temperature matrix, and oxygen concentration matrix into a pre-trained graph neural network. The graph neural network uses the porosity matrix, temperature matrix, and oxygen concentration matrix as node features and the adjacency relationship of adjacent grid cells as edges. Through multi-layer graph convolution operation, it calculates the oxidation reaction rate of the next time step node by node to generate an oxidation reaction rate field. The risk quantification submodule receives the oxidation reaction rate field generated by the reaction deduction submodule and the temperature matrix extracted by the field variable extraction submodule. Using the oxidation reaction rate field as the heat source, and combining the temperature matrix with the probability distribution of the coal quality activity parameters of the coal pile, the submodule uses Monte Carlo simulation to randomly sample the coal quality activity parameters and calculates thousands of future temperature evolution paths in parallel. It counts the number of evolution paths in each three-dimensional grid cell where the temperature exceeds the preset spontaneous combustion critical temperature. The ratio of this number to the total number of simulated paths is defined as the spontaneous combustion probability of that grid cell. The probability values ​​of all cells constitute a spontaneous combustion probability spatial distribution map, which is then output as dynamic risk information to the decision module.

5. The coal yard debris cleaning device based on an automatic robot according to claim 4, characterized in that, The decision-making module includes: The cost map construction submodule is used to integrate the spontaneous combustion probability spatial distribution map in dynamic risk information, the robot's real-time position coordinates obtained from the comprehensive data stream, and the pre-stored or identified clutter position coordinates from the comprehensive data stream. The spontaneous combustion probability value is mapped to the risk cost of spatial position, and the pre-stored obstacle occupancy information and the task attractiveness of the distance to the clutter point are superimposed to generate a multi-layer cost map including a risk cost layer, an obstacle layer, and a target layer. The path planning submodule is used to receive the multi-layer cost map generated by the cost map construction submodule. On the multi-layer cost map, starting from the robot's current position obtained from the integrated data stream and aiming to traverse all clutter positions, an improved ant colony optimization algorithm is applied to search for the K collision-free candidate paths with the minimum total cost under the constraints of the risk cost layer and the obstacle layer, where K is an integer greater than 1. The trajectory optimization submodule is used to receive K candidate paths generated by the path planning submodule. For each candidate path, with the dynamic model of the robot arm as a constraint, and with the goal of minimizing the rate of change of the normal contact force of the end effector on the coal pile during the picking up of debris, the model predictive control algorithm is applied to generate a smooth end position and attitude trajectory through rolling optimization, thereby obtaining the candidate operation trajectory corresponding to each path. The strategy integration submodule receives K candidate paths generated by the path planning submodule and K candidate operation trajectories generated by the trajectory optimization submodule. It combines the K candidate paths with the corresponding K candidate operation trajectories to form K complete joint action strategies. Through a cooperative game evaluation model, it evaluates the performance of each strategy in three dimensions: expected cleanup time, cumulative disturbance risk, and energy consumption. It then selects the Pareto optimal strategy set that is not inferior to other strategies in all dimensions and outputs the Pareto optimal strategy set as the joint action strategy candidate set to the control module.

6. The coal yard debris cleaning device based on an automatic robot according to claim 5, characterized in that, The control module includes: The simulation verification submodule is used to receive the dynamic digital twin output by the twin module and the joint action strategy candidate set output by the decision module, copy the dynamic digital twin to generate a copy of the current state, convert each strategy in the joint action strategy candidate set into a virtual robot joint control command sequence based on the robot inverse kinematics and dynamics model, perform advanced simulation of the control cycle in the copy of the dynamic digital twin, and predict the change gradient of the porosity field of the coal pile and the robot's predicted state sequence after the strategy is executed. The robustness testing submodule is used to inject Gaussian white noise within a preset range into the sensor input channel of the dynamic digital twin replica during the advanced simulation, and randomly perturb the model parameters of the discrete element simulation in the coupled simulation submodule. It records the execution completion degree and final state of the strategy under this disturbance, and calculates the robustness score of the strategy.

7. The coal yard debris cleaning device based on an automatic robot according to claim 6, characterized in that, The control module also includes: The objective function construction submodule receives the porosity field change gradient and robot predicted state sequence predicted by the simulation verification submodule, the robustness score calculated by the robustness test submodule, the dynamic risk information, and the trajectory and velocity information of each strategy in the joint action strategy candidate set obtained from the simulation verification submodule. The optimization objectives are to minimize the norm of the porosity field change gradient, minimize the estimated energy consumption of the strategy calculated from the robot predicted state sequence and trajectory information, maximize the smoothness of the action calculated from the trajectory and velocity information, and maximize the robustness score. A multi-objective evaluation function is constructed, wherein the weight of the porosity field change gradient term is positively correlated with the spontaneous combustion probability value of the strategy execution area in the spontaneous combustion probability space distribution map. The strategy arbitration submodule is used to receive the multi-objective evaluation function constructed by the objective function construction submodule and the joint action strategy candidate set, and to use a non-dominated sorting genetic algorithm to calculate and sort the multi-objective evaluation function values ​​of all strategies in the joint action strategy candidate set, and select the strategy with the highest Pareto ranking and the largest crowding distance as the optimal control law. The instruction decoupling submodule is used to receive the optimal control law determined by the strategy arbitration submodule, parse the optimal control law into a differential control instruction sequence for the left and right drive wheels of the mobile platform, and an instruction sequence for the desired angle, angular velocity and torque of each joint of the robotic arm, and send it to the feedback module.

8. The coal yard debris cleaning device based on an automatic robot according to claim 7, characterized in that, The feedback module includes: The instruction execution submodule is used to receive the mobile platform differential control instruction sequence and the robotic arm joint instruction sequence sent by the instruction decoupling submodule of the control module, and drive the robot mobile platform and robotic arm to perform debris picking and handling actions through the current loop and position loop control of the driver. The data acquisition submodule is used to collect the actual rotation angle and rotation speed of each joint in real time through the encoder during the process of the robot being driven by the instruction execution submodule, collect the three-dimensional force and three-dimensional torque at the contact point between the end effector of the robotic arm and the coal pile through the six-dimensional force sensor, and collect the phase current of the drive motor of each joint through the current sensor. The feature calculation submodule is used to receive the actual rotation angle, rotation speed, three-dimensional force, three-dimensional torque and phase current collected by the data acquisition submodule, perform low-pass filtering on the raw signals of the actual rotation angle, rotation speed, three-dimensional force, three-dimensional torque and phase current to eliminate high-frequency noise, then calculate the tracking position error and tracking speed error of the joint, and perform fast Fourier transform on the phase current signal to extract the characteristic frequency components characterizing mechanical vibration.

9. The coal yard debris cleaning device based on an automatic robot according to claim 8, characterized in that, The step of feeding back the actual execution data to the twin module to correct the digital twin includes: The feature calculation submodule of the feedback module packages the calculated tracking position error, tracking speed error, contact force, contact torque and vibration characteristic frequency data into a correction data package and sends it to the data assimilation submodule of the twin module. The data assimilation submodule receives the correction data packet, parses the actual contact force and actual torque from the correction data packet, compares the actual contact force and actual torque with the contact force and torque predicted by the simulation verification submodule in the advanced simulation, generates force residuals and torque residuals, parses the vibration characteristic frequency from the correction data packet, compares the vibration characteristic frequency with the system modal frequency predicted by the simulation verification submodule in the advanced simulation, generates frequency residuals; The data assimilation submodule uses the force residual, torque residual, and frequency residual as new observation values. In the next data assimilation cycle, it updates the observation equation of the Kalman filter algorithm, performs online estimation and adjustment of the interparticle contact stiffness and damping coefficient in the discrete element simulation of the coupled simulation submodule of the twin module, and simultaneously corrects the effective oxygen diffusion coefficient related to pore structure in the computational fluid dynamics simulation of the coupled simulation submodule of the twin module, so that the simulation output of the dynamic digital twin approximates the actual system response collected by the feedback module.

10. The coal yard debris cleaning device based on an automatic robot according to claim 9, characterized in that, The step of feeding back the actual execution data to the control module to optimize the arbitration for the next control cycle includes: The feature calculation submodule of the feedback module sends the tracking position error, tracking speed error, and the current system state vector as performance feedback data to the strategy arbitration submodule of the control module. The strategy arbitration submodule receives the performance feedback data and calculates the deviation vector between the actual execution effect of the previous control cycle and the advanced simulation prediction effect of the simulation verification submodule on multiple objectives. The strategy arbitration submodule dynamically adjusts the weight allocation coefficients of the multi-objective evaluation function in the objective function construction submodule of the control module according to the magnitude and direction of the deviation vector. Specifically, if the actual disturbance is greater than the prediction, the weight of the porosity field change gradient term is increased; if the actual energy consumption is greater than the prediction, the weight of the energy consumption term is increased.

Images