Underground virtual reality scene dynamic generation system based on reinforcement learning

Abstract

The invention relates to the technical field of mine safety emergency drilling, and discloses an underground virtual reality scene dynamic generation system based on reinforcement learning, and the system comprises the steps: converting the interaction operation of ventilation facilities of multiple trainers into a ventilation parameter change instruction time queue; fire-ventilation alternate coupling transient simulation is executed, a change instruction is injected, and a fire smoke flow time-space coupling state data flow is generated; detecting a wind flow reversal event and constructing an operation consequence causal chain; evaluating a multi-work-type cooperation capability through a physical causal enhanced graph network; the multi-agent reinforcement learning agent group generates a differentiated interference event subjected to physical consistency verification and time sequence conflict mediation; and finally generating a multi-view differentiation VR scene rendering instruction set.

Description

Technical Field

[0001] This invention relates to the field of mine safety emergency drill technology, and more specifically, to a dynamic generation system for underground virtual reality scenes based on reinforcement learning. Background Technology

[0002] In the joint fire emergency drills involving multiple trades underground, virtual reality (VR) technology was used to create immersive drill scenarios. Trainees used VR devices to enter a virtual mine environment to conduct collaborative drills involving multiple trades such as ventilation workers, electromechanical workers, safety officers, and team leaders.

[0003] Existing methods for dynamically generating VR scenes in mines based on reinforcement learning assess trainees' abilities through a disaster evolution-controlled agent, adjust disaster evolution parameters using reinforcement learning strategies, and map the evolution results onto the VR scene for rendering. During disaster evolution, the ventilation network state obtains the baseline airflow distribution based on steady-state calculations. Disaster events are injected by the reinforcement learning agent according to a preset action space. The agent optimizes its strategy using the rate of improvement of individual trainees' abilities as the reward function. The generated disaster scene is presented to all trainees using uniform rendering parameters.

[0004] The existing methods described above have the following technical problems: First, the physical simulation of disaster evolution only considers the changes in ventilation parameters caused by mechanical failures, without introducing the dynamic feedback of the thermal buoyancy effect of fire smoke on the ventilation network. The steady-state calculation and unidirectional fault injection methods cannot reproduce dangerous state transitions such as airflow reversal, resulting in a lack of physical credibility in the fire smoke evolution in the VR scene. Second, the disaster evolution process runs unidirectionally according to a predetermined plan and does not accept real-time operation input from multiple trainees, resulting in the inability of the operations of each type of work in multi-job collaborative exercises to affect each other, thus losing the training value of collaborative exercises. Third, the reinforcement learning agent only performs ability assessment and difficulty adjustment for a single trainee, and cannot impose differentiated challenges on different types of work, nor can it perceive the temporal dependence and physical causal relationship between the operations of each type of work. The generated interference events may contradict the physical state of the fire or disrupt the reasonable multi-job operation sequence. Summary of the Invention

[0005] This invention provides a dynamic generation system for underground virtual reality scenes based on reinforcement learning, which solves the technical problems in related technologies such as the inability to dynamically generate underground multi-job joint fire emergency drill scenes based on real-time operation of trainees, the lack of physical consistency simulation support, and the difficulty in realizing the rendering of differentiated scenes of multiple jobs and the objective evaluation of collaborative capabilities.

[0006] This invention discloses a method for dynamically generating underground virtual reality scenes based on reinforcement learning, comprising: acquiring real-time operation data streams and job role labels of each trainee in an underground VR fire scene; converting ventilation facility interaction events into ventilation network parameter change instructions; generating a ventilation parameter change instruction time queue by timestamp; injecting the ventilation parameter change instruction time queue into a fire-ventilation alternating coupled transient simulation process; alternately executing fire source heat release calculation, temperature field solution, buoyancy head calculation, and ventilation network solution within each simulation time step; outputting the airflow value, temperature value, smoke concentration value, and airflow sign change marker for each branch; generating a fire smoke flow spatiotemporal coupled state data stream; and performing airflow reversal event detection and operation-consequence causal attribution analysis on the fire smoke flow spatiotemporal coupled state data stream to generate a fire scene. The system generates a key event timeline and an operation-consequence causal chain data structure. Trainee behavior data and the operation-consequence causal chain data structure are input into a physical causal-enhanced multi-skill collaborative capability assessment graph network, outputting individual capability score vectors, team collaborative effectiveness scores, and causal responsibility scores. The score results and the current fire smoke flow status are input into a physical constraint multi-skill disaster evolution game reinforcement learning agent group, outputting global disaster evolution baseline parameters and local interference event parameters for each skill. After physical consistency verification and temporal conflict detection, a multi-skill joint disaster event scheme is generated. This multi-skill joint disaster event scheme is then injected into the fire-ventilation alternating coupled transient simulation process. The simulation output is mapped to differentiated VR scene rendering parameters according to the trainee's region and skill role, generating a multi-view differentiated VR scene rendering instruction set.

[0007] Furthermore, in the alternating coupled transient simulation of fire and ventilation, the calculation of heat release from the fire source includes: based on the type of combustible material in the branch where the fire source is located and the airflow value of the branch where the fire source is located in the previous simulation time step, the heat release rate of the fire source in the current simulation time step is calculated according to the airflow-heat release rate coupling relationship. The airflow-heat release rate coupling relationship is: based on the baseline value of the standard heat release rate curve at the current moment, multiplied by a correction factor jointly determined by the dimensionless deviation ratio between the airflow value and the reference airflow value in the previous time step and the airflow-heat release rate coupling coefficient. When the airflow is greater than the reference airflow, combustion is enhanced; when the airflow is less than the reference airflow, combustion is enhanced. Combustion is weakened when the air volume is increased; the air volume-heat release rate coupling coefficient is pre-calibrated and determined by the combustible material type; the buoyancy head calculation includes: based on the temperature field solution results, for each branch containing vertical or inclined sections, integrating along the branch axis the difference between the ambient air density and the gas density corresponding to the current temperature by multiplying by the gravitational acceleration and the sine of the branch inclination angle to obtain the buoyancy head increment of each branch; superimposing the buoyancy head increment into the pressure balance equation of each loop of the ventilation network, resolving the air volume distribution of the ventilation network, and feeding back the updated air volume of the branch where the fire source is located to the fire source heat release calculation of the next simulation time step.

[0008] Furthermore, the temperature field solution includes: using a one-dimensional unsteady-state energy equation to describe the change of gas temperature with time and space for each branch. The one-dimensional unsteady-state energy equation includes a time derivative term, a convection term, a diffusion term, and a volumetric heat source term. The volumetric heat source term for the branch where the fire source is located is obtained by the distribution of the heat release rate of the fire source along the length of the fire source region, while the volumetric heat source term for the non-fire source branches is the rock wall heat transfer term. The one-dimensional unsteady-state energy equation is numerically discretized and solved using the finite difference method. Each branch is uniformly divided into several control volume elements along the axial direction. The time derivative term uses forward difference, the convection term uses the upwind difference scheme, and the diffusion term uses the central difference scheme. The discrete temperature distribution of each branch is obtained by solving the branch sequentially. The ventilation network airflow distribution is re-solved using a ventilation network solution method based on the Newton-Raphson iteration method. The airflow distribution of the previous simulation time step is used as the initial value for iteration. After iteration, the converged airflow distribution is output after the absolute value of the pressure imbalance of all loops is lower than the preset convergence threshold.

[0009] Furthermore, the airflow reversal event detection includes: traversing the airflow symbol change markers of each branch at each time step, identifying the branch and time when the airflow symbol flips, extracting the rate of change of airflow value in adjacent time steps before and after the reversal as the airflow change gradient, extracting the change amplitude of smoke concentration at the outlet of the reversal branch between adjacent time steps as the smoke concentration jump variable, traversing the ventilation network topology along the airflow direction after the reversal from the reversal branch to obtain the new path of smoke propagation after the reversal, and generating a timeline of key events in the fire scene; the operation-consequence causal attribution analysis includes: when airflow reversal or sudden change in smoke concentration is detected, starting from the branch where the change occurred, along other branches in the ventilation network that have a loop connection with this branch, retrieving ventilation network parameter change commands executed within a set time window before the change occurred, marking the trainer role and operation action that caused the change, calculating the propagation delay between the operation command timestamp and the time of the change event and the maximum value of the absolute value of the airflow change of the affected branch as the influence amplitude, and organizing the attribution results into an operation-consequence causal chain data structure.

[0010] Furthermore, the physical causal enhanced multi-job collaborative capability evaluation graph network uses each trainer as a node. The node input feature vector is composed of one-hot encoding of job role labels, movement trajectory features, statistical features of operation action time interval sequences, and voice communication command-response pair matching rate. The movement trajectory features and statistical features of operation action time interval sequences are processed using Z-score normalization. The graph network contains two types of edges: the first type is task dependency edges between jobs, established based on the task sequence dependency relationships between jobs in a predefined multi-job emergency operation process, with weights based on the cosine similarity between the actual operation sequence and the standard operation sequence; the second type is physical causal edges, based on... An operation-consequence causal chain data structure is established, connecting the trainer node performing the operation with the trainer node affected by changes in the smoke flow environment, with the weight representing the magnitude of the impact. The graph network aggregates information through a message passing mechanism. In each round of message passing, the features of each node are weighted by the sum of its own features transformed by a learnable weight matrix and the features of its neighboring nodes transformed by another learnable weight matrix, and then processed by an activation function before being updated. After multiple rounds of message passing, the output features of each node are mapped to an individual ability score vector through a fully connected layer. The output features of all nodes are averaged and then mapped to a team collaboration effectiveness score through a fully connected layer. The weighted message volume passed by each node through physical causal edges is normalized and used as the causal responsibility score.

[0011] Furthermore, the individual capability score vector includes dimensions of operational response timeliness, operational correctness, collaborative communication effectiveness, and risk perception capability, with each dimension ranging from zero to one. The physical causal enhanced multi-task collaborative capability assessment graph network adopts a supervised training method, using individual capability scores and team collaborative effectiveness scores annotated by experts in historical exercise data as supervisory labels. The loss function is the sum of the mean squared error loss of the individual capability score vector and the mean squared error loss of the team collaborative effectiveness score.

[0012] Furthermore, the physical constraint multi-task disaster evolution game reinforcement learning agent group consists of a global coordinating agent and multiple task-specific interference agents. The state space of the global coordinating agent includes the current spatiotemporal coupling state of the fire smoke flow, the team collaboration efficiency score, and the causal responsibility score of each trainer. The action space is the global disaster evolution benchmark parameters, and the reward function is the increment of the team collaboration efficiency score between adjacent evaluation periods. Each task-specific interference agent corresponds to a different task. The state space of each task-specific interference agent includes the individual ability score vector of the corresponding task trainer and the smoke flow state of the area where the trainer is located. The action space is the local interference event parameters for that task. The reward function is the increment of the mean of each dimension component of the individual ability score vector of the task trainer between adjacent evaluation periods. The action space is constrained by the global disaster evolution benchmark parameters output by the global coordinating agent. Both the global coordinating agent and each task-specific interference agent use a multi-agent proximal policy optimization algorithm for policy optimization. The policy network and value network of each agent adopt a fully connected neural network structure.

[0013] Furthermore, the physical consistency verification includes: substituting the parameters of the proposed interference event into the fire-ventilation alternating coupling transient simulation at the current time step in trial calculation mode, checking whether the air volume of each branch in the trial calculation results meets the node air volume conservation constraint and whether the absolute value of the air volume of each branch exceeds the upper limit determined by the product of the cross-sectional area of ​​the branch and the maximum allowable wind speed. If it is violated, the event is rejected and the corresponding agent resamples and outputs the replacement event parameters. The timing conflict detection includes: detecting whether the trigger time of each interference event violates the operation timing window based on a predefined multi-job emergency operation timing dependency graph. The timing dependency graph uses the standard emergency operation steps of each job as nodes and the timing constraints between operation steps as directed edges, and each directed edge is marked with the allowable time interval range. For interference events with timing conflicts, the trigger time is first shifted to the nearest feasible time after the corresponding operation timing window. If a new conflict is generated after the shift, it is replaced with a non-conflicting event sampled from the interference event candidate library of that job.

[0014] Furthermore, the online learning steps include: real-time capture of each trainee's response to disaster events and collaborative interaction among trainees; using the spatiotemporal correlation data of the operational decision sequence and the fire smoke flow spatiotemporally coupled state data stream as multi-agent joint experience samples; determining the sampling priority of each experience sample by adding a small constant to prevent the priority from being zero based on the absolute value of the temporal difference error; storing the samples in a multi-agent shared experience replay buffer; and updating the joint parameters of the policy networks of the global coordinating agent and the job-specific interference agents by sampling from the buffer according to the sampling priority. The temporal difference error of the global coordinating agent is the sum of the incremental team collaboration effectiveness score at the current decision step and the discounted value network estimate for the next decision step, minus the current value network estimate. The temporal difference error of each job-specific interference agent is the sum of the average incremental trainee ability score for the corresponding job at the current decision step and the discounted value network estimate for the next decision step, minus the current value network estimate for the job.

[0015] This invention provides a reinforcement learning-based system for dynamically generating downhole virtual reality scenes, comprising: an operation data acquisition and instruction generation module, used to acquire real-time operation data streams and job role labels of each trainee, convert ventilation facility interaction events into ventilation network parameter change instructions, and generate a ventilation parameter change instruction time queue; a fire-ventilation alternating coupling transient simulation module, used to inject the ventilation parameter change instruction time queue into the fire-ventilation alternating coupling transient simulation process, and alternately execute fire source heat release calculation, temperature field solution, buoyancy head calculation, and ventilation network solution in each simulation time step, generating a fire smoke flow spatiotemporal coupling state data stream; and a key event detection and causal attribution module, used to perform airflow reversal event detection and operation-consequence causal attribution analysis on the fire smoke flow spatiotemporal coupling state data stream, generating a fire scene key event timeline and operation-consequence causal chain data structure. The system comprises: a multi-task collaborative capability assessment module, which inputs the behavioral data of each trainee and the operation-consequence causal chain data structure into a physically causal enhanced multi-task collaborative capability assessment graph network, and outputs individual capability score vectors, team collaborative effectiveness scores, and causal responsibility scores; a disaster evolution decision and verification module, which inputs the scoring results and the current fire smoke flow status into a physically constrained multi-task disaster evolution game reinforcement learning intelligent agent group, outputs global disaster evolution benchmark parameters and local interference event parameters for each task, and generates a multi-task joint disaster event scheme after physical consistency verification and temporal conflict detection; and a differentiated VR scene rendering module, which injects the multi-task joint disaster event scheme into the fire-ventilation alternating coupling transient simulation process, maps the simulation output to differentiated VR scene rendering parameters according to the area and task role of each trainee, and generates a multi-view differentiated VR scene rendering instruction set.

[0016] This invention addresses the technical problem of lack of physical credibility in fire smoke evolution caused by the absence of bidirectional coupling between fire and ventilation in existing physical simulations by introducing a dynamic superposition of buoyancy pressure head and alternating coupling transient solution process of fire source power and air volume feedback in fire-ventilation simulation. This achieves the technical effect of reproducing dangerous state transitions such as airflow reversal in VR scenarios. Furthermore, by converting real-time interactive operations of ventilation facilities by multiple trainees into ventilation network parameter change commands and injecting them into the coupled simulation, this invention solves the technical problem of the inability of multi-task operations to influence each other due to the lack of real-time input from trainees during disaster evolution. This achieves the technical effect of restoring the training value of operational mutual influence in multi-task collaborative exercises. Finally, by employing a multi-task collaborative capability evaluation graph network with enhanced physical causality and a reinforcement learning agent group composed of a global coordinating agent and a task-specific interference agent, and by applying physical consistency verification and temporal conflict mediation to the output interference events, this invention solves the technical problem that a single reinforcement learning agent cannot handle the contradictions between interference events and the physical state of the fire caused by the differentiated challenges of multiple tasks and physical consistency, or the disruption of operational timing. This achieves the technical effect that the generated disaster events simultaneously satisfy physical constraints and the reasonable timing of multi-task emergency operations. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for dynamically generating downhole virtual reality scenes based on reinforcement learning provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the spatiotemporal coupling state evolution of fire smoke flow in branch 3 provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the radar bar comparison of the multidimensional ability scores of each trainee provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of the timing of the wind flow reversal causal chain triggered by the operation of trainer B, provided in an embodiment of the present invention; Figure 5 This is a schematic diagram comparing the impact of ventilation parameter change instructions on branch air volume provided in an embodiment of the present invention; Figure 6 This is a schematic diagram comparing the trigger times before and after the timing adjustment of multi-task interference events provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the distribution of normalized rendering parameters for each trainee's perspective at the 83rd second, provided in an embodiment of the present invention. Figure 8 This is a schematic diagram of the scatter plot of the coupling relationship between the heat release rate of the fire source and the air volume of branch 3 provided in the embodiment of the present invention. Detailed Implementation

[0018] In the virtual reality (VR) scenario of a joint fire emergency drill involving multiple work types underground, existing methods for dynamic scene generation based on reinforcement learning assess trainees' abilities through a disaster evolution-controlled agent. The agent then uses reinforcement learning strategies to adjust disaster evolution parameters (such as gas concentration and ventilation failure injection) and maps the disaster evolution results onto the VR scene for rendering. During the disaster evolution process, the ventilation network state obtains a baseline airflow distribution based on steady-state calculations. Disaster events (such as damper blockage, partial fan shutdown, and other mechanical failures) are injected by the reinforcement learning agent according to a preset action space. The agent optimizes its strategy using the rate of improvement of each trainee's abilities as a reward function. The generated disaster scene is presented to all trainees using uniform rendering parameters.

[0019] The above method has the following three technical problems when applied to joint fire emergency drills involving multiple trades (ventilation workers, electromechanical workers, safety officers, and team leaders).

[0020] First, the physical simulation of disaster evolution only considers changes in ventilation parameters caused by mechanical failures, failing to incorporate the dynamic feedback of the thermal buoyancy effect of fire smoke on the ventilation network. The high-temperature smoke from the fire source decreases in density due to increased temperature, generating a buoyancy head. This buoyancy head dynamically superimposes onto the pressure distribution of the ventilation network as the fire progresses, potentially triggering airflow reversal—a dangerous state transition with a high mortality rate. Furthermore, the fire source power itself is bidirectionally coupled with the ventilation volume. Existing steady-state calculations and unidirectional fault injection methods cannot reproduce these physical processes, resulting in a lack of physical credibility in the fire smoke evolution within VR scenarios.

[0021] Second, the disaster evolution process operates unidirectionally according to the predetermined plan, and does not accept real-time input from multiple trainees during the exercise. Operations such as adjusting dampers by ventilation workers and cutting off power supply by electromechanical workers do not change the ventilation network parameters and thus affect the direction of smoke flow. As a result, in multi-skilled collaborative exercises, the operation of one trade cannot affect the safety environment faced by other trades, thus losing the training value of collaborative exercises.

[0022] Third, reinforcement learning agents can only assess the abilities and adjust the difficulty for a single trainee. They cannot impose differentiated challenges on trainees of different jobs, nor can they perceive the temporal dependencies and physical causal relationships between the operations of different jobs. The interference events generated may contradict the physical state of the fire or disrupt the reasonable multi-job operation sequence.

[0023] According to an embodiment of this invention, this invention provides a method for dynamically generating underground virtual reality scenes based on reinforcement learning. It should be understood that the VR scene dynamic generation method operates within an underground emergency drill training system configured with VR interactive devices, a ventilation network simulation computing server, and a multi-agent reinforcement learning training server. Each trainee enters the same virtual mine scene through their own VR headset and interactive controller.

[0024] At least one embodiment of the present invention discloses a method for dynamically generating downhole virtual reality scenes based on reinforcement learning, such as... Figure 1 As shown, it includes the following steps: Step 1: Acquire real-time operation data from multiple trainees and generate a time queue for ventilation parameter change instructions; Acquire real-time operation data streams and job role tags of all trainees in the same underground VR fire scene, extract the interaction events between each trainee and the ventilation system, convert the ventilation system interaction events into ventilation network parameter change instructions, and generate a time queue of ventilation parameter change instructions driven by multiple trainees' operations by arranging them according to the event timestamp.

[0025] Among them, the job title labels are non-numerical categorical data, which are converted into numerical vectors using one-hot encoding when input into the subsequent graph neural network.

[0026] It should be noted that the specific mapping rules for converting the above-mentioned ventilation facility interaction events into ventilation network parameter change instructions include the following three scenarios: First, the change in damper opening is mapped to the change in the corresponding branch's resistance coefficient. When a trainee (such as a ventilation worker) performs an opening or closing operation on a damper in a VR scene, the damper opening value is used to query a preset damper opening-resistance coefficient correspondence table to obtain the change in the resistance coefficient of the corresponding branch. ,in First, the branch number where the damper is located; second, the start / stop mapping of the local fan is to the corresponding air source pressure parameter returning to zero or recovering. When the local fan is shut down, the air source pressure of that branch... When set to zero, the local sector is restored upon startup. The pressure is restored to the rated pressure value of the local fan; thirdly, the electromechanical worker's cutting off or restoring of the power supply is mapped to the change of the air source parameters of the associated local fan. When the power supply is cut off, the air source pressure of all local fans controlled by the power supply circuit is set to zero, and is restored accordingly when the power supply is restored.

[0027] Step 2: Execute the fire-ventilation alternating coupled transient simulation and inject ventilation parameter change commands to generate a fire smoke flow spatiotemporal coupled state data stream; The ventilation parameter change command time queue is injected into the running fire-ventilation alternating coupled transient simulation process. In each simulation time step, the calculation of fire source heat release, temperature field solution, buoyancy head calculation and ventilation network re-solution are executed alternately. At the same time, the ventilation network parameter change commands to be processed are checked and executed. The air volume value, temperature value, smoke concentration value and air volume sign change marker of each branch of the entire network are output, and the fire smoke flow spatiotemporal coupling state data stream after the operation response is generated.

[0028] It should be noted that the above-mentioned fire-ventilation alternating coupled transient simulation process is performed at each simulation time step. The following sub-steps are executed internally, among which... The sequence number of the simulation time step: Step 201: Based on the type of combustible material in the branch where the fire source is located and the current ventilation volume, calculate the time step for that branch. ignition source heat release rate The rate of heat release from the fire source is affected by the current airflow. The impact is calculated as follows: in To set a standard heat release rate curve based on the type of combustible material at time [time missing] The baseline value, This represents the airflow value of the branch where the fire source was located in the previous time step. This is the reference air volume value corresponding to the standard heat release rate curve. The air volume-heat release rate coupling coefficient, determined by the type of combustible material, reflects the degree of influence of changes in oxygen supply on combustion intensity. Greater than Combustion is enhanced when the time is less than Combustion weakens over time. In the above formula... This is a dimensionless air volume deviation ratio, which eliminates the influence of air volume dimensions on the calculation of heat release rate.

[0029] Furthermore, the aforementioned airflow-heat release rate coupling coefficient The value is pre-determined based on the type of combustible material: for wood-based combustible materials, The range of values ​​is to For flammable rubber materials, The range of values ​​is to For combustible materials such as coal, The range of values ​​is to The above value range was obtained by fitting combustion test data for the corresponding combustible material type. Before system deployment, the values ​​were calibrated according to the combustible material types involved in the actual drill scenarios and written into the parameter configuration file.

[0030] Step 202: Solve for the temperature field of each branch at time step using the one-dimensional unsteady-state energy equation. The evolutionary outcome. For the branching... The one-dimensional unsteady-state energy equation is expressed as: in For branches Internal gas density, The specific heat capacity of a gas at constant pressure. For branches Internal gas temperature For time, For branches Internal airflow velocity, The coordinates are along the branch axis. The thermal conductivity of the gas is... For branches The volumetric heat source term within the branch where the fire source is located. Depend on Acquired along the length of the fire source region, in non-fire source branches It is zero or a rock wall heat exchange term.

[0031] Furthermore, the aforementioned one-dimensional unsteady energy equation is numerically discretized and solved using the finite difference method: the branches are... The system is uniformly divided into several control volume elements along the axial direction. Forward differencing is used for the time derivative term, upwind differencing is used for the convection term, and center differencing is used for the diffusion term. At each simulation time step... Solve sequentially branch by branch to obtain the discrete temperature distribution along the axial direction of each branch. ,in For the first Axial coordinates of each control unit This is the serial number of the control unit.

[0032] Step 203: Based on the temperature field Calculate the buoyancy head increment caused by changes in gas density within each branch. This applies to branches containing vertical or inclined sections. Buoyancy head increment The calculation is as follows: in This is the density of air at room temperature. For branches In position According to temperature According to the ideal gas law Calculated gas density, Atmospheric pressure, The gas constant of air. It is the acceleration due to gravity. For branches The tilt angle, For branches The length.

[0033] Furthermore, the above integration is implemented numerically using the trapezoidal integration method, based on the discrete temperature values ​​of each control unit obtained in step 202. Calculate density difference segment by segment Then, by summing the results, the buoyancy head increment of each branch can be obtained. For horizontal branches, The buoyancy head increment is zero and does not need to be included in the calculation.

[0034] Step 204: Increase the buoyancy head By superimposing this onto the pressure equations of the ventilation network nodes, the airflow distribution of the ventilation network is re-solved. The pressure balance equations for each loop in the ventilation network are updated as follows: in For branches The drag coefficient, For branches Air volume (to be determined). For branches The pressure of the wind source, For the first The set of branches contained in an independent loop. This refers to the numbering of the independent loops. Solving this system of nonlinear equations yields the updated airflow for each branch. .

[0035] Furthermore, the aforementioned nonlinear equations are solved using the Scott-Hinsley method (i.e., the Newton-Raphson iterative method in ventilation network calculation): the airflow distribution at the previous time step. As the initial value for iteration, a loop pressure imbalance vector and the corresponding Jacobian matrix are constructed in each iteration. The airflow correction is obtained by solving a system of linear equations, and the airflow distribution is updated until the absolute value of the pressure imbalance in all loops is lower than the preset convergence threshold. The converged airflow of each branch is then output. .

[0036] Step 205: Update the airflow of the branch where the fire source is located. The feedback is fed back to the calculation of the heat release rate of the fire source in step 201, completing one simulation time step of the fire-ventilation bidirectional coupling.

[0037] Step 206: At the current simulation time step In the process, check if there are any timestamps in the ventilation parameter change instruction time queue. If any instructions within the specified interval exist, their resistance coefficients will be modified sequentially according to their timestamps. Or air source pressure After the modification is completed, the fire-ventilation alternating coupled transient simulation solution for the current time step is executed again starting from step 202.

[0038] Step 207: Record the airflow value of each branch at the current time step. Temperature value flue gas concentration value and air volume symbol change markers ,in By comparison and The positive and negative signs are obtained, and if the sign is flipped, it is marked as a reversal event. The above data is summarized to form a spatiotemporal coupled state data stream of fire smoke flow.

[0039] Furthermore, the above-mentioned flue gas concentration values The solution method is as follows: After completing the air volume distribution solution in step 204, the air volume of each branch is used as the basis for the solution. and temperature field As input, the one-dimensional convection-diffusion equation of flue gas components along the ventilation network is numerically solved to obtain the flue gas concentration distribution along the axial direction of each branch. The solution format is consistent with the finite difference format used in the temperature field solution in step 202.

[0040] Step 3: Perform airflow reversal event detection on the spatiotemporal coupled state data stream of fire smoke flow to generate a timeline of key events in the fire scene; Traverse the airflow sign change markers of each branch at each time step in the fire smoke flow spatiotemporal coupling state data stream, identify the branch and time when the airflow sign flips, extract the airflow change gradient and smoke concentration jump variable before and after the reversal, mark the branch identifier, occurrence time, reversal duration and new smoke propagation path after the reversal of the airflow reversal event as key event sequence, and generate a key event timeline of the fire scene.

[0041] It should be noted that the above-mentioned wind volume change gradient refers to the rate of change of wind volume values ​​within adjacent time steps before and after the occurrence of wind reversal, i.e. ,in To detect the sequence number of the reversed simulation time step, To detect the moment of reversal, The previous time step before the reversal occurred. This is the simulation time step. The flue gas concentration jump variable mentioned above refers to the magnitude of the change in flue gas concentration at the outlet of the reversal branch after the reversal occurs. ,in For branches At any moment Export end (location) The flue gas concentration value at (location) This represents the flue gas concentration at the same location in the previous time step. The new flue gas propagation path after the reversal is obtained by traversing the ventilation network topology along the direction of the reversed airflow, starting from the reversed branch.

[0042] In this embodiment, to clarify the triggering source of each key flue gas change event, the following operation-effect causal attribution analysis step is added on top of step 3: When a reversal of airflow or a sudden change in flue gas concentration is detected in a certain branch, the ventilation network parameter change command that caused the change is traced back along the ventilation network topology. Specifically, starting from the branch where the change occurred, other branches in the ventilation network that have a loop connection with that branch are retrieved, and ventilation network parameter change commands executed within a certain time window before the change occurred are retrieved. The trainer role and operation action that triggered the change are marked. The propagation delay from operation execution to effect manifestation is calculated. and the extent of the impact ,in The time difference between the operation command timestamp and the moment the wind reversal event occurs. This represents the maximum absolute value of the change in airflow in the affected branches before and after the operation. The above attribution results are organized into an operation-consequence causal chain data structure, recording the triggering operation source, the set of affected branches, and the time window of impact for each key flue gas change event.

[0043] Step 4: Obtain behavioral data from each trainee, combine it with the timeline of key events in the fire scenario, input it into the multi-task collaborative capability assessment graph network with physical causal reinforcement, and output the capability scores of each trainee and the team collaborative effectiveness scores. The system acquires the movement trajectory, time interval sequence of operation actions, and voice communication command-response pair matching rate of each trainee. Combining the operation-consequence causal chain data structure and the timeline of key events in the fire scenario, it inputs these data into a multi-task collaborative capability assessment graph network with physical causal enhancement. The output includes individual capability score vectors, team collaborative effectiveness scores, and causal responsibility scores for each trainee.

[0044] Before inputting the physical causal enhancement multi-task collaborative capability evaluation graph network, the input features of each trainer are preprocessed as follows: the statistical features of movement trajectory features and operation action time interval sequences are Z-score standardized to eliminate the influence of differences in the dimensions of different physical quantities on feature concatenation operations; the voice communication command-response pair matching rate itself is... Dimensionless ratios within the specified range do not require additional normalization; job / role labels are converted into numerical vectors using one-hot encoding. The preprocessed feature vectors are then concatenated to form the node input feature vector.

[0045] It should be noted that the aforementioned physical causal-enhanced multi-job collaborative capability assessment graph network is a graph neural network model, with each trainer as a node. The input feature vector of each node is composed of one-hot encoding of the corresponding trainer's job role label, movement trajectory features, statistical features of the operation action time interval sequence, and the matching rate of voice communication command-response pairs. The physical causal-enhanced multi-job collaborative capability assessment graph network contains two types of edges: the first type is task dependency relationship edges between jobs, established based on the task sequence dependency relationship between each job in the predefined multi-job emergency operation process, such as the task dependency between ventilation workers and electromechanical workers due to the power supply relationship of ventilation equipment; the second type is physical causal edges, established based on the operation-consequence causal chain data structure, connecting the trainer node performing the operation with the trainer node affected by the change in the smoke flow environment due to the operation. The weight of the first type of edge is the operation timing consistency, that is, the similarity between the actual operation timing and the standard operation process timing; the weight of the second type of edge is the physical impact magnitude. .

[0046] Furthermore, the calculation method for the above-mentioned operation timing consistency is as follows: extract the timestamp sequence of each operation step actually performed by the trainees of each job type in the current exercise, compare it with the standard time interval sequence of the corresponding operation steps in the predefined standard operation process, use cosine similarity to measure the similarity between the two, and use the obtained cosine similarity value as the weight of the task dependency relationship edge between the corresponding jobs.

[0047] Furthermore, the aforementioned physical causal-enhanced multi-task collaborative capability assessment graph network aggregates information through a message passing mechanism. In each round of message passing, nodes... The feature update method is as follows: in For nodes In the Feature vectors after round message passing For nodes In the Feature vectors after round message passing For nodes The set of neighboring nodes, For nodes With nodes The weights of the edges between them (taken from the weight definitions of the two types of edges mentioned above). and For the first The learnable weight matrix of the round, The activation function. After multiple rounds of message passing, the output features of each node... The fully connected layer maps the data to an individual ability score vector, where... The total number of rounds of message passing is represented by the output features of all nodes. After being aggregated by graph-level pooling, the output features of all nodes are mapped to the team collaboration efficiency score through a fully connected layer. The information contribution of each node on the physical causal edge is normalized and used as the causal responsibility score.

[0048] Furthermore, the above graph-level pooling operation uses mean pooling to output features from all nodes. The average value of each node is used to obtain a graph-level representation vector, which is then mapped to a team collaboration effectiveness score via a fully connected layer. The information contribution of each node on the physical causal edge is calculated as follows: [Statistics on each node are then performed.] The weighted amount of messages transmitted to its neighboring nodes through physical causal edges, i.e. ,in For nodes The set of neighboring nodes connected by physical causal edges The vector norm is used to normalize the above contributions of all nodes to obtain the causal responsibility score of each trainer.

[0049] Furthermore, the aforementioned physical causal enhanced multi-task collaborative capability assessment graph network adopts a supervised training method, using individual capability scores and team collaborative effectiveness scores labeled by experts in historical exercise data as supervision labels. The loss function is the sum of the mean squared error loss of the individual capability score vector and the mean squared error loss of the team collaborative effectiveness score, and the Adam optimization algorithm is used for parameter updates.

[0050] In this embodiment, to enable the physical causal-enhanced multi-task collaborative ability assessment graph network to distinguish the performance differences of each trainee across different ability dimensions, the aforementioned individual ability score vector includes multiple dimensional components, corresponding to operational response timeliness, operational correctness, collaborative communication effectiveness, and risk perception ability, respectively. The value range of each dimensional component is as follows: The above causal responsibility score reflects the degree to which each trainer's actions affect the safety environment of other trainers, and the value range is [value range missing]. A higher value indicates that the trainee's actions have a greater impact on the team's overall safety status.

[0051] Step 5: Input the evaluation results and the current fire smoke flow status into the physical constraint multi-operation disaster evolution game reinforcement learning intelligent agent group, and output the physically consistent multi-operation local disturbance event parameters; The individual ability score vector, team collaboration effectiveness score, causal responsibility score, and the latest state data in the current fire smoke spatiotemporal coupling state data stream are input into the physical constraint multi-job disaster evolution game reinforcement learning intelligent agent group, and the global disaster evolution benchmark parameters and local disturbance event parameters of each job are output.

[0052] Before inputting the physical constraints of the multi-task disaster evolution game reinforcement learning agent group, the various scoring and state data are preprocessed as follows: the components of each dimension of the individual ability scoring vector, the team collaboration effectiveness score, and the causal responsibility score are all preprocessed. Within the specified range, no additional normalization is required; however, the air volume, temperature, and smoke concentration values ​​in the fire smoke flow spatiotemporal coupled state data stream have different dimensions. A mean normalization based on the range is used to uniformly scale each physical quantity to the specified value. The range is then concatenated into a state vector input physical constraint multi-task disaster evolution game reinforcement learning agent group.

[0053] It should be noted that the aforementioned multi-task disaster evolution game reinforcement learning agent group with physical constraints consists of a global coordinating agent and multiple task-specific disturbance agents. The state space of the global coordinating agent includes the current spatiotemporal coupling state of the fire and smoke flow, the team's collaborative effectiveness score, and the causal responsibility scores of each trainer. The action space of the global coordinating agent consists of global disaster evolution benchmark parameters (including the fire source power adjustment coefficient, the global ventilation system disturbance amplitude, etc.). The reward function of the global coordinating agent is the rate of improvement of the team's collaborative effectiveness under the operation-smoke flow coupling scenario, that is, the increment of the team's collaborative effectiveness score between adjacent evaluation periods. Each job-specific interference agent corresponds to a ventilation worker, electromechanical worker, safety officer, and team leader, respectively. The state space of each agent includes the individual ability score vector of the corresponding trainee and the current smoke flow state of the area where the trainee is located. The action space of each agent consists of local interference event parameters specific to that job (e.g., injecting a damper jamming event into the ventilation worker, or injecting an equipment fault alarm into the electromechanical worker). The reward function for each agent is the ability improvement rate of the corresponding trainee, i.e., the increment of the mean of each dimension component of the trainee's individual ability score vector between adjacent evaluation periods. The action space of each agent is constrained by the global disaster evolution baseline parameters output by the global coordinating agent. Both the global coordinating agent and the job-specific interference agents employ a multi-agent proximal policy optimization algorithm for policy optimization.

[0054] Furthermore, in the aforementioned multi-agent proximal policy optimization algorithm, both the policy network and value network of each agent adopt a fully connected neural network structure. The policy network takes the state vector of the corresponding agent as input and outputs the probability distribution in the action space (for discrete action spaces) or the action mean and standard deviation (for continuous action spaces). The value network takes the state vector as input and outputs an estimate of the value of the current state. Within each policy update cycle, each agent uses the truncation ratio between the probability ratio of the current policy and the old policy as a policy update constraint to prevent the policy update step from being too large, leading to training instability. The upper and lower bounds of the truncation ratio are the standard hyperparameter settings for the multi-agent proximal policy optimization algorithm.

[0055] The global disaster evolution baseline parameters and local disturbance event parameters of each task output by the physical constraint multi-task disaster evolution game reinforcement learning agent group are continuous or discrete action values. It is necessary to decode the global disaster evolution baseline parameters and local disturbance event parameters of each task into specific event parameters that can be injected into the fire-ventilation alternating coupled transient simulation: for continuous actions (such as fire source power adjustment coefficient), the normalized action value is back-mapped back to the actual value range of the corresponding physical quantity; for discrete actions (such as disturbance event type number), the specific event type and its parameter configuration are converted by looking up the table according to the preset action number and event type correspondence table to form an executable disturbance event description.

[0056] It should be noted that all the above-mentioned output interference event parameters must be submitted to the fire-ventilation alternating coupled transient simulation for physical consistency verification before taking effect. Specifically, the interference event parameters to be injected are substituted into the fire-ventilation alternating coupled transient simulation at the current time step in trial calculation mode. The trial calculation results are checked for any violations of physical constraints (e.g., the injected damper jamming event causes the airflow of a certain branch to exceed the physically feasible range). If physical inconsistencies exist, the event is rejected and fed back to the corresponding agent, which then resamples and outputs replacement event parameters.

[0057] Furthermore, the basis for determining the aforementioned physical feasibility range is: the air volume of each branch. The nodal airflow conservation constraint of the ventilation network must be satisfied, that is, for any node... The sum of the airflow of all branches connected to this node must be zero; simultaneously, the absolute value of the airflow of each branch must not exceed the upper limit determined by the product of the branch's cross-sectional area and the maximum permissible air velocity, which is determined by the mine safety regulations for this type of roadway. If any branch in the trial calculation violates the above constraints, it is considered physically inconsistent.

[0058] In this embodiment, to avoid temporal conflicts between interference events output by the specialized interference agents of each job, after physical consistency verification is passed, the following collaborative temporal conflict detection and mediation steps are included: detecting whether the triggering time of each interference event violates a reasonable operation timing window based on a predefined multi-job emergency operation timing dependency graph. The aforementioned multi-job emergency operation timing dependency graph uses the standard emergency operation steps of each job as nodes and the temporal sequence constraints between operation steps as directed edges, with each directed edge labeled with an allowed time interval range. The proposed triggering time of each interference event is compared with the current operation steps of each trainee. If the triggering of a certain interference event will cause a trainee of a certain job to face subsequent interference before completing its preceding operation steps, it is determined to be a temporal conflict. For interference events with temporal conflicts, mediation is performed according to the following priority: firstly, the triggering time of the event is shifted to the nearest feasible time after the corresponding operation timing window; if the shift causes new conflicts with other events, it is replaced with a non-conflicting event randomly sampled from the interference event candidate library for that job. After mediation is completed, a multi-task joint disaster event plan with time-coordinated procedures is generated.

[0059] Step 6: Inject the multi-task joint disaster event plan into the coupled simulation and generate a multi-view differentiated VR scene rendering instruction set; The time-coordinated multi-task joint disaster event plan is injected into the fire-ventilation alternating coupled transient simulation process running in step 2 for synchronous execution. The smoke concentration and temperature values ​​of each branch output by the fire-ventilation alternating coupled transient simulation are mapped to differentiated VR scene rendering parameters according to the region where each trainee is located, generating a multi-view differentiated underground VR fire scene rendering instruction set.

[0060] It should be noted that the aforementioned differentiated VR scene rendering parameters are distinguished according to the trainee's job role label: the VR scene rendering parameters for the ventilation worker's area focus on the visual and sound effects of ventilation anomalies, including the dynamic updates of airflow direction indicator signs, the visual rendering of the direction and density of smoke flow in the roadway, and the changes in the sound effects of fan operation; the VR scene rendering parameters for the electromechanical worker's area focus on changes in equipment status indicators, including the status of the power distribution cabinet indicator lights, cable temperature alarm signs, and visual feedback on the operation status of the local fan; the VR scene rendering parameters from the safety officer's perspective focus on the updates of personnel distribution and evacuation route status, including the overlay display of personnel location markers in each roadway and the indication of the unobstructedness of evacuation routes; the VR scene rendering parameters from the team leader's perspective focus on the refresh of global situation information, including the real-time numerical labeling of the air volume and smoke concentration of each branch on the simplified diagram of the mine ventilation network.

[0061] Before mapping the smoke concentration and temperature values ​​output from the alternating fire-ventilation transient simulation to VR scene rendering parameters, a mean normalization based on the range is performed to scale each physical quantity to the requirements of the rendering engine. The parameter range is then converted into specific executable rendering instructions such as visual effect intensity, color depth, and sound effect amplitude according to the rendering mapping rules corresponding to each job role.

[0062] It should be noted that when a wind reversal event occurs, the wind direction indicator flips and the smoke movement direction changes abruptly in the VR scene of all trainees in the area where the reversal branch is located, allowing trainees to intuitively perceive the wind reversal state.

[0063] In this embodiment, to continuously improve the policy quality of the reinforcement learning agent group in the physical constraint multi-task disaster evolution game, the following online learning steps are included concurrently with step 6: Real-time capture of each trainer's response behavior to disaster events and collaborative interaction behavior among trainers; using the spatiotemporal correlation data of each trainer's operational decision sequence and the fire smoke flow spatiotemporally coupled state data stream as multi-agent joint experience samples. For each multi-agent joint experience sample, the temporal difference error is calculated based on the physical coupling collaborative effectiveness. Weighted sampling priority determination ,in For the first The temporal difference error of a empirical sample This is the index number for the empirical sample. To prevent the use of tiny constants with zero priority, the joint experience samples of the multi-agent agents are stored in a shared experience replay buffer. Joint experience samples are then sampled from the shared experience replay buffer according to sampling priority, and the joint parameters of the policy networks of the global coordinating agent and the task-specific interference agents are updated.

[0064] Furthermore, the aforementioned timing difference error The calculation method is as follows: for a globally coordinating agent, The difference between the sum of the team collaboration effectiveness score at the current decision-making step and the value network estimate for the next decision-making step after discounting, and the value network estimate at the current decision-making step, is... ,in To enhance the index of learning decision-making steps, For decision-making steps The increase in team collaboration effectiveness score As a discount factor, For decision-making steps Value network estimate, For decision-making steps Value network estimate, For decision-making steps state, For decision-making steps The state; for each job-specific interference agent, The difference between the sum of the trainee's skill improvement rate for the current decision-making step and the discounted value network estimate for the next decision-making step, and the value network estimate for the current decision-making step, is: ,in For decision-making steps No. The increment of the mean of each dimension component of the individual ability score vector for each type of job trainee. Assign job titles For this job-specific interference agent, steps are taken during decision-making. Value network estimate, For this job-specific interference agent, steps are taken during decision-making. Value network estimate, For this job-specific interference agent, steps are taken during decision-making. state, For this job-specific interference agent, steps are taken during decision-making. The state of the above discount factor. Standard hyperparameter settings for multi-agent proximal policy optimization algorithms.

[0065] This implementation method introduces a transient solution process based on dynamic superposition of buoyancy head and feedback of fire source power and air volume in the fire-ventilation simulation. This incorporates the buoyancy head term generated by the change in the density of high-temperature smoke in the ventilation network into the calculation of the air volume distribution. Furthermore, the heat release rate of the fire source at each simulation time step is adjusted by the feedback of the ventilation air volume in the previous time step. Therefore, the evolution of fire smoke flow in the VR scene follows the thermal buoyancy effect and the two-way coupling relationship between fire and ventilation in thermodynamic physics. It can reproduce the dangerous state transition process of airflow reversal, which cannot be presented in steady-state calculations. This overcomes the problem of lack of physical credibility in the evolution of fire smoke flow caused by the lack of two-way coupling between fire and ventilation in the physical simulation of existing methods.

[0066] This implementation method converts the interactive operations of multiple trainees with ventilation facilities into ventilation network parameter change commands in real time and injects the fire-ventilation alternating coupled transient simulation process into each simulation time step. This allows the trainees' operations to change the parameters of the ventilation network branches and trigger the re-solution of the fire-ventilation alternating coupled transient simulation. Therefore, the change in the resistance coefficient caused by the ventilation worker adjusting the damper can be propagated to the airflow and smoke concentration distribution of other branches through the fire-ventilation alternating coupled transient simulation, thereby changing the smoke flow environment in the areas where other types of trainees are located. This overcomes the problem in existing methods where disaster evolution does not accept real-time operation input from trainees, resulting in the inability of multiple types of operations to influence each other. It restores the training value of the mutual influence of operations in multi-type collaborative exercises.

[0067] This implementation method employs a multi-task collaborative capability assessment graph network with enhanced physical causality. By introducing the causal relationship between trainee operations and changes in the smoke flow environment through physical causal edges, the assessment results include quantification of causal responsibility for each trainee's operations. Furthermore, it utilizes a physical constraint multi-task disaster evolution game reinforcement learning agent group composed of a global coordinating agent and a task-specific interference agent. Physical consistency verification and temporal conflict mediation are applied to the output interference events. Therefore, the generated disaster events satisfy the physical constraints of the fire-ventilation alternating coupled transient simulation without disrupting the reasonable timing of multi-task emergency operations. This overcomes the problem in existing methods where a single reinforcement learning agent cannot handle the contradictions between interference events and the physical state of the fire caused by the differentiated challenges and physical consistency of multiple tasks, or disrupt the timing of operations.

[0068] The following is an example of an application of the present invention, such as Figure 2-8 As shown, the implementation process is as follows: In March 20XX, a coal mine training center organized a joint underground fire emergency drill involving multiple work types using its VR emergency drill system. Four trainees participated in the drill, serving as a ventilation worker (Trainee A), an electromechanical worker (Trainee B), a safety officer (Trainee C), and a team leader (Trainee D). All four simultaneously wore VR headsets and entered the same virtual mine scenario. The virtual mine ventilation network consisted of 7 branches (Branch 1 to Branch 7) and 5 nodes (Node 1 to Node 5). The fire source was located in Branch 3 (transport roadway), the combustible material was wood, and the initial fire source power was 120kW. At the start of the drill, the ventilation network was in normal ventilation mode, the airflow in each branch was stable, and the system clock started counting from simulation time 0s.

[0069] At the 38th second of the simulation, trainee A (ventilation worker) partially opened damper #2 on branch 2 in the VR scene (adjusting the opening from 100% to 50%). The system queried the damper opening-resistance coefficient correspondence table to obtain the change in the resistance coefficient of branch 2. At the 61st second of the simulation, trainee B (mechanical and electrical worker) cut off the power supply to the local fan control panel in the VR scene, triggering the shutdown event of local fan #5 on branch 5. Both operation events were extracted as ventilation network parameter change instructions and queued according to their timestamps.

[0070] Table 1: Ventilation Parameter Change Instruction Time Queue The unique hot-spot encoding result of the job role label is: Ventilation worker Mechanical and electrical engineering Safety officer Team leader This is used for subsequent splicing of graph network node features.

[0071] The simulation runs in 5-second time steps. Taking the 38th second (the 8th time step) as an example, the execution process of each sub-step is explained.

[0072] Branch 3 (Fire Source Branch): Combustible material type is wood, coupling coefficient... Reference air volume The previous step air volume Reference heat release rate ,but: Branch 3 is uniformly divided into 10 control volume elements along the axial direction. The one-dimensional unsteady energy equation is solved using the upwind differential scheme to obtain the temperature distribution of each control volume element.

[0073] Branch 3 tilt angle , Branch length Density at room temperature Taking a temperature of 348K at control unit 5 (axial position 42.5m) as an example, the corresponding density is: After summing the buoyancy head increments of each branch, and using the air volume of the previous time step as the initial value for iteration, the updated air volume of each branch is obtained by iteratively solving using the Scott-Hinsley method.

[0074] The instruction at time 38s (queue 001) timestamp falls within the time interval of steps 7 to 8, so the resistance coefficient of branch 2 is modified to the original value plus 3.2 N·s. 2 / m 8 Then, the solution is executed again from step 202 to solve the current time step.

[0075] Table 2 Simulation spatiotemporal coupling state data flow (partial time steps, branch 3): At time 68s, the branch 5 fan stopped operating (injected by instruction 002 at time 61s), causing the air volume of branch 3 to continue to decrease. Combined with the increase in buoyancy head due to the fire, this eventually triggered airflow reversal, and the air volume of branch 3 changed from a positive value to a negative value.

[0076] By iterating through the airflow symbol change markers at each time step, it was detected that airflow reversal occurred in branch 3 at step 14 (68s).

[0077] Calculation of air volume change gradient: Flue gas concentration jumps (at the outlet of branch 3) ): Operation-Consequence Causal Attribution Analysis: Looking back at the ventilation parameter change command executed within the time window prior to the reversal (56s to 68s), a loop connection was found between queue 002 (61s, when trainee B cut off power to branch 5's local fan) and branch 3. This is attributed to trainee B's operation triggering the airflow reversal in branch 3. The propagation delay is 68−61=7s, and the impact magnitude is... .

[0078] Table 3 Timeline of Key Events in the Fire Scene: The raw behavioral data of each trainee are preprocessed to form the node input feature vector: the movement trajectory features and the statistical features of the time interval sequence of operation actions are standardized by Z-score, the voice communication command-response pair matching rate is directly used, and the job role label is represented by one-hot encoding. After concatenation, the vector is input into the graph network.

[0079] The graph network is constructed with two types of edges: task dependency edges between job types (e.g., between trainer A and trainer B, due to the dependency on the power supply of ventilation equipment, the cosine similarity weight is 0.73); and physical causal edges (trainer B → trainer C, due to the deterioration of the smoke flow environment in the roadway where the safety officer is located caused by EVT-01, the weight is taken as the influence amplitude value of 0.78).

[0080] During message passing, taking the feature update of trainer node B after the first round of message passing as an example (simplified scalar illustration), its updated feature is transformed from its previous feature by a weight matrix, superimposed with weighted neighbor messages from trainers A and D, and then output by an activation function. After three rounds of message passing, the output features of each node are mapped to an individual ability score vector through a fully connected layer, and after graph-level mean pooling, mapped to a team collaboration effectiveness score.

[0081] Table 4: Individual Trainee Ability Rating Vectors and Team Rating Outputs Trainee B had the highest causal responsibility score (0.83), consistent with the attribution conclusion in step 3: their operation of cutting off the power supply to the local fan was the direct cause of triggering airflow reversal and worsening the safety environment for other trainees.

[0082] Input the scores output from step 4 and the latest status of the fire smoke flow (air volume, temperature, and smoke concentration are concatenated into a state vector after normalization of range and mean) into the intelligent agent group. The global coordination intelligent agent's perception team collaboration efficiency score (0.63) is low, and the output fire source power adjustment coefficient is 1.12 (i.e., an increase of 12% on the current baseline), and the global ventilation system disturbance amplitude is low; the ventilation worker's dedicated interference intelligent agent perception trainer A's operation correctness (0.65) and risk perception ability (0.62) are relatively weak, and the output is the damper jamming interference event for branch 2; the electromechanical worker's dedicated interference intelligent agent perception trainer B's scores in all dimensions are low, and the output is the equipment fault alarm interference event for branch 4.

[0083] Physical consistency verification: Substitute the two disturbance events mentioned above into the simulation at the current time step in trial mode to verify whether the air volume of each branch meets the node conservation constraints and the maximum allowable wind speed constraints. The verification is successful.

[0084] Cooperative timing conflict detection: The ventilation worker's exclusive interference agent is detected to be attempting to trigger the damper jam at 75s, while trainer A has not yet completed the preceding "confirm reverse branch" operation step. This is determined to be a timing conflict, and the trigger time is shifted to 83s (the most recent feasible time), thus resolving the conflict.

[0085] Table 5: Multi-task joint disaster event plan after time-series coordination: After the fire-ventilation alternating coupled transient simulation in step 2 of DIS-01 and DIS-02 injection, the simulation outputs the smoke concentration and temperature values ​​of each branch at the 83rd second, which are then normalized and scaled to the mean. Subsequently, differentiated rendering parameters were mapped according to the job roles of each trainee: In the area where Trainee A (ventilation worker) was located, an animation of a sudden change in the direction of smoke flow in the alleyway and a flipping of the airflow direction indicator were triggered. Simultaneously, due to the EVT-01 airflow reversal event, a flipping animation of the airflow direction indicator was overlaid in his view. In the area where Trainee B (mechanical and electrical worker) was located, the overload indicator light on the distribution cabinet flashed red, and sound effects of changes in the operating status of the local fan were rendered. In the view of Trainee C (safety officer), updates to the personnel position markers in each alleyway and indications of the evacuation route's accessibility were overlaid (after the reversal of branch 3, the corresponding evacuation route was marked as obstructed). In the view of Trainee D (team leader), the airflow of branch 3 (−0.58m³) was marked in real-time on the simplified ventilation network diagram. 3 The values ​​for ( / s) and flue gas concentration (15.1%).

[0086] Table 6. Multi-view Differentiated VR Scene Rendering Instruction Set (Snapshot 83s): The online learning process is executed synchronously: capturing the response behaviors of each trainee to DIS-01 and DIS-02 to form a multi-agent joint experience sample. Taking a certain decision-making step of the electromechanical-specific interference agent (corresponding to trainee B) as an example, immediate rewards are provided. Discount factor Next-state value estimation Current state value estimation Then the timing difference error is: The priority of this sample is The data is stored in the shared experience replay buffer for subsequent joint parameter updates.

[0087] The data flow throughout the implementation process reflects a rigorous logical chain: Step 1 converts the VR interaction between trainee A and trainee B into two ventilation parameter change commands and queues them. These two commands are injected sequentially in the simulation time step of Step 2, driving the fire-ventilation alternating coupled transient simulation to generate a spatiotemporal coupled state data stream containing the airflow reversal event; Step 3 detects the airflow reversal at 68s in branch 3 from this data stream and attributes the triggering source to trainee B's operation, forming a causal chain data; Step 4 constructs a graph network using the causal chain data as edge weights and outputs the differentiated score result with the highest causal responsibility score (0.83) for trainee B; Step 5 generates targeted interference events based on the weak points of each trainee's score. After physical consistency verification and temporal conflict mediation, the triggering time of the ventilation worker interference is postponed from 75s to 83s; Step 6 injects the mediated solution into the simulation, generates four differentiated rendering commands according to job roles, and stores the experience samples composed of response behaviors in the playback buffer by weighting them according to the temporal differential error, forming a closed-loop online learning data stream. The data from each step are seamlessly connected, and the physical simulation results, evaluation outputs, and rendering commands remain consistent throughout the same training scenario.

[0088] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A dynamic generation system for downhole virtual reality scenes based on reinforcement learning, characterized in that, The system executes the following method: Acquire real-time operation data streams and job role tags of each trainee in the VR fire scene in the mine, convert ventilation facility interaction events into ventilation network parameter change instructions, and generate a ventilation parameter change instruction time queue by timestamp; The ventilation parameter change command time queue is injected into the fire-ventilation alternating coupled transient simulation process. In each simulation time step, the calculation of fire source heat release, temperature field solution, buoyancy head calculation and ventilation network solution are executed alternately. The air volume value, temperature value, smoke concentration value and air volume sign change mark of each branch are output to generate the fire smoke flow spatiotemporal coupled state data stream. Perform airflow reversal event detection and operation-consequence causal attribution analysis on the fire smoke flow spatiotemporal coupled state data stream to generate a timeline of key events in the fire scene and an operation-consequence causal chain data structure. The behavioral data of each trainee and the operation-consequence causal chain data structure are input into the multi-task collaborative ability assessment graph network with physical causality enhancement, and the output is an individual ability score vector, a team collaborative effectiveness score and a causal responsibility score. The scoring results and the current fire smoke flow status are input into a physical constraint multi-task disaster evolution game reinforcement learning intelligent agent group, which outputs global disaster evolution benchmark parameters and local interference event parameters of each task. After physical consistency verification and temporal conflict detection, a multi-task joint disaster event scheme is generated. The multi-job joint disaster event scheme is injected into the fire-ventilation alternating coupled transient simulation process, and the simulation output is mapped into differentiated VR scene rendering parameters according to the region and job role of each trainee, generating a multi-view differentiated VR scene rendering instruction set.

2. The reinforcement learning-based dynamic generation system for downhole virtual reality scenes according to claim 1, characterized in that, In the transient simulation of alternating fire and ventilation, the calculation of heat release from the fire source includes: Based on the combustible material type of the branch where the fire source is located and the airflow value of the branch where the fire source is located in the previous simulation time step, the heat release rate of the fire source in the current simulation time step is calculated according to the airflow-heat release rate coupling relationship. The airflow-heat release rate coupling relationship is as follows: Based on the baseline value of the standard heat release rate curve at the current moment, multiply by a correction factor determined by the dimensionless deviation ratio between the air volume value of the previous time step and the reference air volume value and the air volume-heat release rate coupling coefficient. When the air volume is greater than the reference air volume, combustion is enhanced; when the air volume is less than the reference air volume, combustion is weakened. The air volume-heat release rate coupling coefficient is pre-calibrated and determined by the combustible material type; the buoyancy head calculation includes: based on the temperature field solution results, for each branch containing vertical or inclined sections, integrating along the branch axis the difference between the ambient air density and the gas density corresponding to the current temperature by multiplying by the gravitational acceleration and the sine of the branch inclination angle to obtain the buoyancy head increment of each branch; The buoyancy head increment is superimposed on the pressure balance equation of each loop in the ventilation network, the air volume distribution of the ventilation network is re-solved, and the updated air volume of the branch where the fire source is located is fed back to the fire source heat release calculation in the next simulation time step.

3. The reinforcement learning-based downhole virtual reality scene dynamic generation system according to claim 2, characterized in that, The temperature field solution includes: The gas temperature changes with time and space are described by a one-dimensional unsteady-state energy equation for each branch. The one-dimensional unsteady-state energy equation includes a time derivative term, a convection term, a diffusion term, and a volumetric heat source term. The volumetric heat source term of the branch where the fire source is located is obtained by the distribution of the fire source heat release rate along the length of the fire source region. The volumetric heat source term of the non-fire source branch is the rock wall heat transfer term. The one-dimensional unsteady energy equation is numerically discretized and solved using the finite difference method. Each branch is uniformly divided into several control volume elements along the axial direction. The time derivative term uses forward difference, the convection term uses the upwind difference scheme, and the diffusion term uses the center difference scheme. The discrete temperature distribution of each branch is obtained by solving the branch sequentially. The ventilation network airflow distribution is re-solved using a ventilation network solution method based on the Newton-Raphson iteration method. The airflow distribution of the previous simulation time step is used as the initial value for iteration. After iterating until the absolute value of the pressure imbalance of all loops is lower than the preset convergence threshold, the converged airflow distribution is output.

4. The reinforcement learning-based dynamic generation system for downhole virtual reality scenes according to claim 1, characterized in that, The airflow reversal event detection includes: Traverse the airflow sign change markers of each branch at each time step, identify the branch and time when the airflow sign flips, extract the rate of change of airflow value in adjacent time steps before and after reversal as the airflow change gradient, extract the change amplitude of smoke concentration at the outlet of the reversal branch between adjacent time steps after reversal as the smoke concentration jump variable, traverse the ventilation network topology along the airflow direction after reversal from the reversal branch to obtain the new path of smoke propagation after reversal, and generate the timeline of key events in the fire scene. The operation-consequence causal attribution analysis includes: when airflow reversal or sudden change in flue gas concentration is detected, starting from the branch where the change occurred, along other branches in the ventilation network that have loop connections with that branch, retrieving ventilation network parameter change commands executed within a set time window before the change occurred, labeling the trainer role and operation action that caused the change, calculating the propagation delay between the operation command timestamp and the time of the change event, and the maximum value of the absolute value of the airflow change in the affected branch as the impact magnitude, and organizing the attribution results into an operation-consequence causal chain data structure.

5. The reinforcement learning-based dynamic generation system for downhole virtual reality scenes according to claim 1, characterized in that, The physical causal enhancement multi-job collaborative ability evaluation graph network uses each trainer as a node. The node input feature vector is composed of one-hot encoding of job role labels, movement trajectory features, statistical features of operation action time interval sequences, and voice communication command-response pair matching rate. Among them, the movement trajectory features and the statistical features of operation action time interval sequences are processed by Z-score normalization. The graph network contains two types of edges: the first type is task dependency relationship edges between job types, which are established based on the task sequence dependency relationship between each job type in the predefined multi-job emergency operation process, and the weight is the cosine similarity between the actual operation sequence and the standard operation process sequence. The second type is physical causal edges, which are built based on the operation-consequence causal chain data structure, connecting the trainer node that performs the operation with the trainer node affected by the change in the smoke flow environment, with the weight being the magnitude of the influence. The graph network aggregates information through a message passing mechanism. In each round of message passing, the features of each node are weighted by the sum of its own features after being transformed by a learnable weight matrix and the features of its neighboring nodes after being transformed by another learnable weight matrix, and then processed by an activation function before being updated. After multiple rounds of message passing, the output features of each node are mapped to an individual ability score vector through a fully connected layer. The output features of all nodes are averaged and then mapped to a team collaboration effectiveness score through a fully connected layer. The weighted message volume passed by each node through physical causal edges is normalized and used as the causal responsibility score.

6. The reinforcement learning-based dynamic generation system for downhole virtual reality scenes according to claim 5, characterized in that, The individual competency scoring vector includes dimensions of operational response timeliness, operational correctness, collaborative communication effectiveness, and risk perception ability, with each dimension ranging from zero to one. The physical causal enhanced multi-task collaborative capability assessment graph network adopts a supervised training method, using individual capability scores and team collaborative effectiveness scores labeled by experts in historical exercise data as supervision labels, and the loss function is the sum of the mean squared error loss of the individual capability score vector and the mean squared error loss of the team collaborative effectiveness score.

7. The reinforcement learning-based dynamic generation system for downhole virtual reality scenes according to claim 1, characterized in that, The physical constraint multi-task disaster evolution game reinforcement learning agent group consists of a global coordinating agent and multiple task-specific interference agents. The state space of the global coordinating agent includes the current spatiotemporal coupling state of the fire smoke flow, the team collaboration efficiency score, and the causal responsibility score of each trainer. The action space is the global disaster evolution benchmark parameter, and the reward function is the increment of the team collaboration efficiency score between adjacent evaluation periods. Each job-specific interference agent corresponds to a different job. The state space of each job-specific interference agent includes the individual ability score vector of the trainee for the corresponding job and the smoke flow state of the area where the trainee is located. The action space consists of the local interference event parameters for that job. The reward function is the increment of the mean of each dimension component of the individual ability score vector of the trainee for that job between adjacent evaluation periods. The action space is constrained by the global disaster evolution benchmark parameters output by the global coordination agent. Both the global coordinating agent and the task-specific interference agents employ a multi-agent proximal policy optimization algorithm for policy optimization, and the policy network and value network of each agent adopt a fully connected neural network structure.

8. The reinforcement learning-based dynamic generation system for downhole virtual reality scenes according to claim 7, characterized in that, The physical consistency verification includes: The parameters of the proposed interference event are substituted into the fire-ventilation alternating coupling transient simulation at the current time step in trial mode. The results of the trial are checked to see if the air volume of each branch meets the node air volume conservation constraint and whether the absolute value of the air volume of each branch exceeds the upper limit determined by the product of the cross-sectional area of ​​the branch and the maximum allowable wind speed. If the violation occurs, the event is rejected and the corresponding agent resamples and outputs the replacement event parameters. The timing conflict detection includes: detecting whether the trigger time of each interference event violates the operation timing window based on a predefined multi-job emergency operation timing dependency graph. The timing dependency graph uses the standard emergency operation steps of each job as nodes and the timing constraints between operation steps as directed edges, with each directed edge marked with an allowed time interval range. For interference events with timing conflicts, the trigger time is first shifted to the nearest feasible time after the corresponding operation timing window. If a new conflict occurs after the shift, it is replaced with a non-conflicting event sampled from the interference event candidate library of that job.

9. The reinforcement learning-based dynamic generation system for downhole virtual reality scenes according to claim 7, characterized in that, It also includes online learning steps: Real-time capture of each trainee's response to disaster events and collaborative interaction among trainees; use the spatiotemporal correlation data of the operational decision sequence and the fire smoke flow spatiotemporally coupled state data stream as multi-agent joint experience samples; determine the sampling priority of each experience sample by adding a small constant to prevent the priority from being zero according to the absolute value of the time difference error; store it in the multi-agent shared experience playback buffer; and perform joint parameter updates on the policy network of the global coordinating agent and the special interference agents of each job according to the sampling priority. The temporal difference error of the global coordinating agent is the sum of the team collaboration effectiveness score increment at the current decision step and the discounted value network estimate at the next decision step, minus the value network estimate at the current decision step. The temporal difference error of each job-specific interference agent is the sum of the mean increment of the trainer's ability score for the corresponding job at the current decision step and the discounted value network estimate for that job at the next decision step, minus the value network estimate for that job at the current decision step.

10. The downhole virtual reality scene dynamic generation system based on reinforcement learning according to claim 1, characterized in that, Its features are, The system includes: The operation data acquisition and instruction generation module is used to acquire the real-time operation data stream and job role tags of each trainee, convert ventilation facility interaction events into ventilation network parameter change instructions, and generate a ventilation parameter change instruction time queue. The fire-ventilation alternating coupling transient simulation module is used to inject the ventilation parameter change instruction time queue into the fire-ventilation alternating coupling transient simulation process. In each simulation time step, it alternately executes the fire source heat release calculation, temperature field solution, buoyancy head calculation and ventilation network solution to generate a fire smoke flow spatiotemporal coupling state data stream. The critical event detection and causal attribution module is used to perform airflow reversal event detection and operation-consequence causal attribution analysis on the spatiotemporal coupled state data stream of the fire smoke flow, and generate a timeline of critical events in the fire scene and an operation-consequence causal chain data structure. The multi-task collaborative ability assessment module is used to input the behavioral data of each trainee and the operation-consequence causal chain data structure into the physical causal enhancement multi-task collaborative ability assessment graph network, and output individual ability score vectors, team collaborative effectiveness scores and causal responsibility scores. The disaster evolution decision and verification module is used to input the scoring results and the current fire smoke flow status into a physical constraint multi-operation disaster evolution game reinforcement learning intelligent agent group, output global disaster evolution benchmark parameters and local interference event parameters of each operation, and generate a multi-operation joint disaster event scheme after physical consistency verification and temporal conflict detection. The differentiated VR scene rendering module is used to inject the multi-job joint disaster event scheme into the fire-ventilation alternating coupled transient simulation process, and map the simulation output to differentiated VR scene rendering parameters according to the area and job role of each trainee, and generate a multi-view differentiated VR scene rendering instruction set.

Images