Hazardous chemical substance storage and transportation device safety accident simulation system based on VR technology

Abstract

The invention discloses a hazardous chemical substance storage and transportation device safety accident simulation system based on a VR technology, and the system comprises a scene simulation module which is used for training a generative adversarial network, taking a real hazardous chemical substance database as input, receiving physical and chemical rule constraints of a knowledge base, synthesizing the parameters of a new virtual hazardous chemical substance, constructing a hazardous chemical substance storage and transportation scene, and simulating the hazardous chemical substance storage and transportation scene; performing rendering and scene interaction through a VR engine; the accident simulation module is used for simulating dangerous chemical accidents and defining different disaster evolution rules through a multi-agent reinforcement learning framework, and agents generate disaster scenes through self-chess-playing confrontation; the user interaction module is used for a user to interact with the virtual environment through VR equipment; the data analysis and evaluation module is used for recording user behavior data, dynamically constructing a decision tree model based on real-time disaster parameters, and generating a non-stylized scheme in combination with an expert rule base and a reinforcement learning reward function; and the knowledge base provides hazardous chemical substance safety specifications, an accident case library and training progress management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of dangerous chemical safety accident simulation management, and particularly relates to a dangerous chemical storage and transportation device safety accident simulation system based on VR technology. BACKGROUND

[0002] The dangerous chemical storage and transportation device involves flammable, explosive and toxic substances, and the accident consequences are extremely destructive. Safety simulation is in urgent need. The VR technology provides immersive interactive experience by constructing a three-dimensional virtual environment, and becomes an innovative solution for industry safety simulation. The dangerous chemical storage and transportation device safety accident simulation system based on VR technology restores the device leakage, fire, explosion and other accident processes through a virtual scene, so that the trainees can experience the disaster impact, strengthen the risk cognition and decision-making pressure, and repeatedly simulate extreme scenarios without interrupting the actual production or consuming real resources.

[0003] In the traditional VR-based dangerous chemical accident simulation system, the generation of disaster chain usually depends on preset scripts or simple conditional branching logic, which will cause the trainees to be familiar with only simple scenarios, lack experience in dealing with multi-disaster coupling in real accidents, and rare but high-risk disaster combinations cannot be simulated due to not being preset, forming a safety blind area, secondary disasters are ignored, leading to incomplete emergency plans. Therefore, a dangerous chemical storage and transportation device safety accident simulation system based on VR technology is proposed. SUMMARY

[0004] The present application relates to the technical field of dangerous chemical safety accident simulation management, and particularly relates to a dangerous chemical storage and transportation device safety accident simulation system based on VR technology.

[0005] To achieve the above purpose, the present application adopts the following technical scheme:

[0006] A dangerous chemical storage and transportation device safety accident simulation system based on VR technology, comprising:

[0007] The scene simulation module: a training generative adversarial network (StyleGAN3) is used as input with a real dangerous chemical database, and receives physical and chemical rule constraints of a knowledge base to optimize the generation quality, synthesizes the molecular structure, reaction activity and other parameters of a virtual new dangerous chemical, establishes an uncertainty quantification model, evaluates the confidence interval of the parameters of the synthesized virtual new dangerous chemical, constructs a dangerous chemical storage and transportation scene (such as a storage tank area, a pipeline, a transportation vehicle, etc.), renders and interacts with the scene through a VR engine, dynamically generates smoke, leaked liquid and other effects through an environment simulation algorithm, provides scene basic data to the accident simulation module, and receives environment modification instructions from the user interaction module;

[0008] Accident simulation module: simulate hazardous chemical accidents (leakage, fire, explosion, etc.), support custom configuration of accident type, location, scale, define different disaster evolution rules (such as the coupling logic of leakage → explosion → secondary fire) through the multi-agent reinforcement learning (MARL) framework, and generate disaster scenes through self-play against agents, output disaster parameters to the scene simulation module;

[0009] User interaction module: users interact with the virtual environment through VR devices and send user operation logs (action time, sequence) to the data analysis module, and receive operation guidance information from the knowledge base;

[0010] Data analysis and evaluation module: record user behavior data, dynamically build decision tree models based on real-time disaster parameters (temperature, pressure, toxic gas concentration, etc.), combine expert rule base and reinforcement learning reward function to generate non-programmed solutions (such as prioritizing leak source or evacuation path), obtain real-time disaster data from the accident module, push recommended solutions to the user interaction module, and feedback user decision results to the reinforcement learning framework to optimize strategies, and use Isolation Forest to identify abnormal behaviors that deviate from standard operation modes;

[0011] Knowledge base: provides hazardous chemical safety specifications, accident case library, and training progress management.

[0012] The above technical solution further comprises:

[0013] Further, the scene simulation module trains a generative adversarial network (StyleGAN3) to input real hazardous chemical databases and receive physical and chemical rule constraints from the knowledge base to optimize the generation quality, and synthesizes virtual new hazardous chemical molecular structures, reaction activity, and other parameters. The specific steps are as follows:

[0014] Data preprocessing and knowledge base construction: convert hazardous chemical molecular structures into images or parameter vectors to construct training set X real , define physical and chemical rule knowledge base R = {r1, r2,..., r n};

[0015] StyleGAN3 model: conditional input layer: introduce rule condition vector c in generator G and discriminator D, encode the rules that the currently generated molecule needs to meet; rule constraint module: add physical and chemical rule constraint layer at the end of the generator, input generated molecule parameters θ gen to rule checker R, output rule compliance score s ∈ [0, 1];

[0016] Loss function: the total loss function includes three parts, namely, adversarial loss, rule constraint loss, and physical rationality loss, the adversarial loss is represented as L adv=E[log D(X)] real )]+E[log(1-D(G(z)))], where z is the latent spatial noise, and the rule constraint loss uses mean squared error (MSE) or cross-entropy to measure the deviation between the generated parameters and the rules, expressed as The physical plausibility loss is calculated using a pre-trained chemical property prediction model to determine the plausibility of the generated molecule, denoted as L. phy =-logP valid (θ gen The total loss function is expressed as L. total =L adv +λ1L rule +λ2L phy , where λ1 and λ2 are hyperparameters (usually set to 0.1 to 1.0);

[0017] Training process: Generator optimization: Minimize L total Update generator parameter θ G Discriminator optimization: Maximize L adv Update the discriminator parameters θ D After each fixed number of iterations, λ1 is dynamically adjusted based on the rule violation rate of the generated samples.

[0018] VR simulation integration: Input effective parameters into the VR engine to generate accident scenarios such as hazardous chemical leaks and fires.

[0019] Furthermore, the specific steps for establishing an uncertainty quantification model and evaluating the confidence intervals of the parameters for synthesizing a new virtual hazardous chemical are as follows:

[0020] Identify key parameters and their sources of uncertainty: Determine the key parameters that affect the simulation results and analyze the sources of their uncertainty;

[0021] Establish a probability distribution model: select or fit a probability distribution for the parameters;

[0022] Uncertainty propagation (Monte Carlo simulation): The Monte Carlo (MC) method is used to propagate parameter uncertainties to the model output (such as leakage consequences, explosion overpressure, etc.), generating N parameter samples k. i ~N(μ, σ 2 ), where μ is the mean and σ is the standard deviation. For each k, the model output y is calculated. i =g(k i (e.g., leakage rate, overpressure peak), statistical output y distribution characteristics;

[0023] Calculate the confidence interval: Calculate the confidence interval (CI) based on the output distribution;

[0024] Sensitivity analysis (Sobol index): Quantify the contribution of each parameter to the output, identify key parameters, total effect index S T Total effect of parameter i on output, formula is

[0025] Verification and iteration: Compare experimental data and simulation results to verify the reliability of the model.

[0026] Further, the specific steps of rendering and scene interaction through the VR engine are as follows:

[0027] Scene modeling and physical engine integration: Use 3D modeling tools (such as Blender) to build tank farms, pipeline networks, and transport vehicles, integrate NVIDIA PhysX or Havok engine in Unity / Unreal, and set rigid body dynamics parameters;

[0028] Fluid leakage simulation (SPH method): Liquid particle initialization, discrete leakage liquid into smooth particles, particle mass Where M leak is the total leakage, N p is the number of particles, and the particle position and velocity update formula is:

[0029]

[0030] Where P is the pressure, ρ is the density, W ij is the smoothing kernel function, g is the gravity, and f visc is the viscous force;

[0031] Smoke diffusion simulation: Discretize the scene space into a three-dimensional grid with grid size Δx = 0.5m and time step Δt = 0.01s, the evolution equation of smoke concentration c is Where u is the wind speed field, D is the diffusion coefficient, S is the source term (leakage point), and the large eddy simulation is used to close the subgrid stress, which is Where C s = 0.1, Δ is the grid scale, is the strain rate module;

[0032] Parameter-driven dynamic effect generation: Map the parameters generated by the generative adversarial network (StyleGAN3) (such as molecular weight MW, reaction activity RA) to the simulation parameters, and trigger the leakage event by VR handle operation;

[0033] Scene rendering and VR interaction optimization: Use GPU particle system to render smoke, and combine noise texture to generate turbulence details.

[0034] Further, the specific steps of defining different disaster evolution rules (such as the coupling logic of leakage → explosion → secondary fire) through the multi-agent reinforcement learning (MARL) framework are as follows:

[0035] Customizing accident initial condition configuration: the user inputs initial parameters through the interface, converts discrete inputs into initial state vectors (such as leakage rate, coordinates) recognizable by the system, and inputs them into the multi-agent reinforcement learning (MARL) framework as the starting point of the disaster evolution process;

[0036] Defining the multi-agent reinforcement learning framework: dividing the agent roles, including the disaster evolution agent (Agent_D), the environment response agent (Agent_E), and the emergency response agent (Agent_R). The disaster evolution agent (Agent_D) controls the disaster state transition (such as leakage → explosion), the environment response agent (Agent_E) simulates the feedback of the environment to the disaster (such as wind speed, temperature influence), and the emergency response agent (Agent_R) represents the inhibitory effect of emergency measures on the disaster, taking emergency measures (such as closing valves) to suppress the disaster. Defining the action space and state space, the action space represents that the disaster evolution agent (Agent_D) adjusts the disaster parameters, the environment response agent (Agent_E) sets the environment parameters, and the disaster evolution agent (Agent_D) selects the emergency strategy, and the state space represents the current disaster parameter vector. Reward function design: the reward function drives the agent to fight, the disaster evolution agent (Agent_D) rewards the disaster severity, the environment response agent (Agent_E) rewards the rationality of the environment parameters, and the emergency response agent (Agent_R) rewards the disaster suppression effect;

[0037] Disaster evolution rule modeling: physical law coupling: leakage → explosion: leakage time exceeding the threshold triggers explosion, explosion → fire: explosion shock wave ignites surrounding combustible material probability, and secondary fire is determined according to the ignition probability of surrounding combustible material; state transition equation: converting agent actions into disaster parameter changes, embedding evolution rule formulas into state transition calculation, constraining the physical feasibility of agent actions, and disaster parameter changes as reward function input, affecting agent strategy optimization.

[0038] Further, the agent generates a disaster scene through self-play confrontation, outputs disaster parameters to the scene simulation module, including the following steps:

[0039] Agent self-play against training: Agent_D and Agent_R are in confrontation, and Agent_E provides environmental constraints. In each iteration, Agent_D tries to intensify the disaster, and Agent_R tries to suppress the disaster. The confrontation results are fed back to the policy network, and the policy network is updated using the proximal policy optimization (PPO) algorithm;

[0040] Disaster parameter output and VR scene rendering: the agent action is converted into disaster parameters, and the disaster parameters are mapped into visual effects (such as leakage rate control liquid particle density and overpressure determination shock wave intensity). The output parameters directly drive the VR engine, and the user can adjust the initial configuration through the VR experience feedback.

[0041] Further, the data analysis and evaluation module records user behavior data, dynamically constructs a decision tree model based on real-time disaster parameters (temperature, pressure, toxic gas concentration, etc.), combines an expert rule base and a reinforcement learning reward function, and generates specific steps of a non-programmed scheme (such as prioritizing leakage sources or evacuation routes):

[0042] User behavior data and real-time disaster parameter collection: real-time capture of user operations (such as valve control and path selection) and environmental sensor data (temperature, pressure, toxic gas concentration, etc.) through VR equipment to form a multi-modal raw data set, standardize disaster parameters (such as normalization in the formula), eliminate dimensional differences, and improve model processing efficiency;

[0043] Dynamic decision tree model construction: taking real-time disaster parameters as root nodes, combining user historical behavior data, dynamically generating a decision tree, and using weighted information gain as the splitting criterion. The information gain calculation formula is represented as where H(D) is the entropy of the data set, A is the split attribute (such as the temperature threshold), D v is the subset of attribute A=v, and the split threshold is adjusted according to the priority of the expert rule;

[0044] Expert rule base and reinforcement learning fusion: check whether the decision tree output conforms to the pre-defined rules in the expert rule base, update the action value function Q(s,a) through Q-learning, and the update formula is represented as:

[0045]

[0046] where α is the learning rate, γ is the discount factor, and R is the reward function;

[0047] Non-programmed scheme generation: combine the decision tree, expert rules, and reinforcement learning results to generate a scheme.

[0048] Further, the data analysis and evaluation module adopts the specific steps of isolation forest (Isolation Forest) to identify abnormal behaviors deviating from the standard operation mode:

[0049] Data collection and preprocessing: capture features reflecting operation behavior (such as pressure, temperature, valve state, etc.) from the VR engine, construct quantifiable feature vectors, perform data standardization, convert continuous operation into fixed-length feature vectors (such as sliding window statistics), and adapt to the input requirements of isolation forest;

[0050] Building an isolation forest (iForest) model: randomly sampling a subset of features from the fixed-length feature vector, randomly selecting feature dimensions, randomly selecting split values within the selected feature dimension range, recursively splitting left and right subtrees until reaching the tree height limit, repeating the construction of multiple trees to form an isolation forest;

[0051] Calculate the anomaly score: the anomaly score of sample x is represented as Where E(h(x)) is the average path length of sample x in all subtrees, and c(n) = 2H(n-1)-2(n-1) / n, where H(n) = ln n + γ, where γ is the Euler constant, used to normalize the path length, and the c(n) term ensures that the scores under different subsampling scales are comparable, making the threshold setting more uniform, and the anomaly determination rule is: s(x,n)→1 indicates abnormality (short path), and s(x,n)→0 indicates normality (long path);

[0052] Threshold setting and anomaly determination: map the anomaly score to the actual safety rules (such as s>0.6 triggering an alarm), and adjust the threshold according to the scene risk level (such as reducing the threshold for high-risk operations to reduce false negatives).

[0053] The present application has the following beneficial effects:

[0054] In the present application, new physical and chemical parameters of hazardous chemicals are synthesized by a generative adversarial network, breaking through the limitations of traditional systems relying on fixed databases, supporting unlimited expansion of hazardous chemical categories, evaluating the confidence interval of synthesized parameters, providing probabilistic risk analysis for simulation, enhancing the reliability of decision-making basis, defining coupled logic, and agents dynamically adjusting behavior strategies through environmental perception, forming a nonlinear evolution path, simulating the complexity of multiple factors intertwined in real accidents, and avoiding the "acting according to the script" defect of traditional systems. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 A system block diagram of a hazardous chemical storage and transportation device safety accident simulation system based on VR technology is proposed. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0057] Please refer to Figure 1 As shown in the figure, the present application is a dangerous chemical storage and transportation device safety accident simulation system based on VR technology, which comprises:

[0058] Scene simulation module: training generative adversarial network (StyleGAN3), taking real dangerous chemical database as input, and receiving physical and chemical rule constraints of knowledge base to optimize generation quality, synthesizing virtual new dangerous chemical molecular structure, reaction activity and other parameters, establishing uncertainty quantification model, evaluating confidence interval of synthesized virtual new dangerous chemical parameters, constructing dangerous chemical storage and transportation scene (such as storage tank area, pipeline, transportation vehicle, etc.), rendering and scene interaction through VR engine, dynamically generating smoke, leaking liquid and other effects through environmental simulation algorithm, providing scene basic data to accident simulation module, receiving environmental modification instructions of user interaction module;

[0059] Accident simulation module: simulating dangerous chemical accidents (leakage, fire, explosion, etc.), supporting custom configuration of accident type, location and scale, defining different disaster evolution rules (such as coupling logic of leakage → explosion → secondary fire) through multi-agent reinforcement learning (MARL) framework, and generating disaster scene through self-play confrontation of agents, outputting disaster parameters to scene simulation module;

[0060] User interaction module: user interacts with virtual environment through VR device, sends user operation log (action time, sequence) to data analysis module, and receives operation guidance information of knowledge base;

[0061] Data analysis and evaluation module: records user behavior data, dynamically constructs decision tree model based on real-time disaster parameters (temperature, pressure, toxic gas concentration, etc.), combines expert rule base and reinforcement learning reward function, generates non-programmed scheme (such as prioritizing leakage source or evacuation path), obtains real-time disaster data from accident module, pushes recommended scheme to user interaction module, and feeds back user decision result to reinforcement learning framework to optimize strategy, and uses isolation forest to identify abnormal behavior deviating from standard operation mode;

[0062] Knowledge base: provides dangerous chemical safety specifications, accident case library and training progress management.

[0063] In one embodiment, the scene simulation module trains a generative adversarial network (StyleGAN3) to input a real hazardous chemical database and receive physical and chemical rule constraints of the knowledge base to optimize the generation quality. The specific steps of synthesizing virtual new hazardous chemical molecular structures, reaction activity, etc. parameters are as follows:

[0064] Data preprocessing and knowledge base construction: convert the hazardous chemical molecular structure into an image or a parameter vector, and construct a training set X real , define the physical and chemical rule knowledge base R = {r1, r2,..., r n};

[0065] StyleGAN3 model: conditional input layer: introduce rule condition vector c in the generator G and discriminator D, which encodes the rules that the currently generated molecule needs to meet; rule constraint module: add a physical and chemical rule constraint layer at the end of the generator, input the generated molecular parameters θ gen and the rule checker R, and output the rule compliance score s ∈ [0, 1];

[0066] Loss function: the total loss function includes three parts, namely the adversarial loss, the rule constraint loss and the physical rationality loss. The adversarial loss is represented as L adv = E[log D(x real )] + E[log(1-D(G(z)))] where z is the latent space noise, the rule constraint loss uses mean square error (MSE) or cross entropy to measure the deviation of the generated parameters from the rules, represented as The physical rationality loss uses a pre-trained chemical property prediction model to calculate the rationality of the generated molecule, represented as L phy = -log P valid (θ gen ), and the total loss function is represented as L total = L adv + λ1L rule + λ2L phy , where λ1, λ2 are hyperparameters (usually set to 0.1-1.0);

[0067] Training process: generator optimization: minimize L total , update the generator parameters θ G ; discriminator optimization: maximize L adv , update the discriminator parameters θ D , and after each fixed number of iterations, dynamically adjust λ1 according to the rule violation rate of the generated samples;

[0068] VR simulation integration: input the effective parameters into the VR engine to generate hazardous chemical leakage, combustion and other accident scenes.

[0069] In one embodiment, the specific steps of establishing an uncertainty quantification model, evaluating the confidence interval of the parameters of the synthesized virtual new hazardous chemical:

[0070] Identify key parameters and their sources of uncertainty: Determine the key parameters that affect the simulation results and analyze their sources of uncertainty;

[0071] Establish a probability distribution model: Select or fit a probability distribution for the parameters;

[0072] Uncertainty propagation (Monte Carlo simulation): Propagate the parameter uncertainty to the model output (such as leakage consequences, explosion overpressure, etc.) through the Monte Carlo (MC) method, generate N parameter samples k i ~ N(μ, σ 2 ), where μ is the mean and σ is the standard deviation, for each k i , calculate the model output y i = g(k i ) (such as leakage rate, peak overpressure), and the distribution characteristics of the output y are calculated;

[0073] Calculate the confidence interval: Calculate the confidence interval (CI) according to the output distribution;

[0074] Sensitivity analysis (Sobol index): Quantify the contribution of each parameter to the output, identify key parameters, and total effect index S T represents the total effect of parameter i on the output, and the calculation formula is

[0075] Verification and iteration: Compare experimental data and simulation results to verify the reliability of the model.

[0076] In one embodiment, the specific steps of rendering and scene interaction through the VR engine;

[0077] Scene modeling and physical engine integration: Use 3D modeling tools (such as Blender) to build tank farms, pipeline networks, and transport vehicles, integrate NVIDIA PhysX or Havok engines in Unity / Unreal, and set rigid body dynamics parameters;

[0078] Fluid leakage simulation (SPH method): Initialize liquid particles, discretize the leaking liquid into smooth particles, and the particle mass where M leak is the total leakage amount, N p is the number of particles, and the particle position and velocity update formula is:

[0079]

[0080] where P is the pressure, ρ is the density, and Wij is the smoothing kernel function, g is the gravitational force, f visc is the viscous force;

[0081] Smoke diffusion simulation: Discretize the scene space into a three-dimensional grid with grid size Δx = 0.5 m and time step Δt = 0.01 s. The evolution equation of smoke concentration c is represented as where u is the wind velocity field, D is the diffusion coefficient, and S is the source term (leak point). The sub-grid stress is closed using large eddy simulation, represented as where C s = 0.1, Δ is the grid scale, is the strain rate module;

[0082] Parameter-driven dynamic effect generation: Map the parameters generated by the generative adversarial network (StyleGAN3) (such as molecular weight MW, reaction activity RA) to the simulation parameters, and trigger the leakage event by VR handle operation;

[0083] Scene rendering and VR interaction optimization: Render smoke using GPU particle systems, and generate turbulence details combined with noise textures.

[0084] In one embodiment, the specific steps of defining different disaster evolution rules (such as the coupling logic of leakage → explosion → secondary fire) through the multi-agent reinforcement learning (MARL) framework are as follows:

[0085] Custom accident initial condition configuration: Users input initial parameters through the interface, convert discrete inputs into system-recognizable initial state vectors (such as leakage rate, coordinates), and use them as inputs for the multi-agent reinforcement learning (MARL) framework to start the disaster evolution process;

[0086] Define a multi-agent reinforcement learning framework: perform agent role division, including disaster evolution agent (Agent_D), environment response agent (Agent_E), and emergency response agent (Agent_R), the disaster evolution agent (Agent_D) controls the disaster state transition (such as leakage explosion), the environment response agent (Agent_E) simulates the feedback of the environment to the disaster (such as wind speed, temperature influence), the emergency response agent (Agent_R) represents the inhibitory effect of emergency measures on the disaster, and takes emergency measures (such as closing the valve) to suppress the disaster; define the action space and state space, the action space represents that the disaster evolution agent (Agent_D) adjusts the disaster parameter, the environment response agent (Agent_E) sets the environment parameter, the disaster evolution agent (Agent_D) selects the emergency strategy, and the state space represents the current disaster parameter vector; reward function design: the reward function drives the agent to fight, the disaster evolution agent (Agent_D) rewards the disaster severity, the environment response agent (Agent_E) rewards the rationality of the environment parameter, and the emergency response agent (Agent_R) rewards the disaster suppression effect;

[0087] Disaster evolution rule modeling: physical law coupling: leakage explosion: leakage time exceeding threshold triggers explosion, explosion fire: explosion shock wave ignites surrounding combustible material probability, according to the ignition probability of surrounding combustible material to judge secondary fire; state transition equation: convert agent action into disaster parameter change, evolution rule formula is embedded in state transition calculation, constrain the physical feasibility of agent action, disaster parameter change as reward function input, influence agent strategy optimization.

[0088] In one embodiment, the agent generates a disaster scene by self-play against, outputs disaster parameters to the scene simulation module, including the following steps:

[0089] Agent self-play against training: disaster evolution agent (Agent_D) and emergency response agent (Agent_R) against, environment response agent (Agent_E) provides environmental constraints, in each round of iteration, the disaster evolution agent (Agent_D) tries to aggravate the disaster, the emergency response agent (Agent_R) tries to suppress the disaster, the result of the confrontation is fed back to the strategy network, and the strategy network is updated using the proximal policy optimization (PPO) algorithm;

[0090] Disaster parameter output and VR scene rendering: convert agent action into disaster parameter, map disaster parameter to visual special effect (such as leakage rate control liquid particle density, overpressure determine shock wave intensity), output parameter directly drive VR engine, user feedback through VR experience can adjust initial configuration.

[0091] In one embodiment, the data analysis and evaluation module records user behavior data, dynamically constructs a decision tree model based on real-time disaster parameters (temperature, pressure, toxic gas concentration, etc.), combines expert rule base and reinforcement learning reward function, and generates specific steps of non-programmed scheme (such as prioritizing leakage source or evacuation path):

[0092] User behavior data and real-time disaster parameter collection: real-time capture user operation (such as valve control, path selection) and environmental sensor data (temperature, pressure, toxic gas concentration, etc.) through VR device, form multi-modal raw data set, standardize disaster parameters (such as normalization in formula), eliminate dimension difference, and improve model processing efficiency;

[0093] Dynamic decision tree model construction: taking real-time disaster parameters as root node, combining user historical behavior data, dynamically generating decision tree, splitting criterion using weighted information gain, the information gain calculation formula is represented as Wherein, H(D) is the entropy of data set, A is the split attribute (such as temperature threshold), D v is the subset of attribute A=v, adjust the split threshold according to the priority of expert rules;

[0094] Expert rule base and reinforcement learning fusion: check whether the output of decision tree meets the pre-defined rules in expert rule base, update action value function Q(s,a) through Q-learning, the update formula is represented as:

[0095]

[0096] Wherein, α is the learning rate, γ is the discount factor, and R is the reward function;

[0097] Non-programmed scheme generation: generate scheme by integrating decision tree, expert rules and reinforcement learning results.

[0098] Scenario: tank leakage, real-time parameters T=45℃, P=1.8MPa, C=0.6, user historical preference is "priority control of leakage source".

[0099] Decision-making process:

[0100] Dynamic decision tree splitting:

[0101] Root node: P<2MPa→ enter the child node.

[0102] Child node: T<50℃→ split into "control leakage" or "evacuation".

[0103] Expert rule triggering:

[0104] Rule 1: if C≥0.5 and user does not choose evacuation→ forced to recommend evacuation path.

[0105] Reinforcement learning optimization:

[0106] In historical data, the reward value Q=85 for "first control leakage" and Q=70 for "first evacuation" in a similar scenario.

[0107] Final solution generation:

[0108] The system combines rules and suggestions to generate a hybrid solution:

[0109] "Suggestion: prioritize controlling the leakage source (reward +85), but need to start the backup evacuation channel simultaneously (rule trigger).

[0110] In one embodiment, the data analysis and evaluation module adopts Isolation Forest to identify specific steps of abnormal behavior deviating from standard operating mode:

[0111] Data collection and preprocessing: Capture features reflecting operating behavior (such as pressure, temperature, valve status, etc.) from the VR engine, construct quantifiable feature vectors, perform data standardization, convert continuous operations into fixed-length feature vectors (such as sliding window statistics), and adapt to the input requirements of Isolation Forest;

[0112] Build Isolation Forest (iForest) model: Randomly sample a subset of features from the fixed-length feature vector, randomly select feature dimensions, randomly select split values within the selected feature dimension range, recursively split left and right sub-trees until the tree height limit is reached, repeat the construction of multiple sub-trees to form an isolation forest;

[0113] Calculate anomaly score: The anomaly score of sample x is represented as Where E(h(x)) is the average path length of sample x in all sub-trees, c(n) = 2H(n-1)-2(n-1) / n, where H(n) = ln n + γ, where γ is the Euler constant, used to standardize the path length, and the c(n) term ensures the comparability of scores under different sub-sampling scales, making the threshold setting more uniform, and the anomaly determination rule is: s(x,n)→1 indicates abnormality (short path), s(x,n)→0 indicates normality (long path);

[0114] Threshold setting and anomaly determination: Map the anomaly score to the actual safety rules (such as s>0.6 triggers an alarm), adjust the threshold according to the risk level of the scene (such as reducing the threshold for high-risk operations to reduce false negatives).

[0115] Scenario: Abnormal pressure rise in a hazardous chemical storage tank (simulated operation error).

[0116] Steps:

[0117] Data extraction:

[0118] Normal operation: pressure stabilized at 0.5-0.8 MPa, temperature 25-30℃.

[0119] Abnormal operation: pressure surged to 1.2 MPa, temperature rose to 40℃.

[0120] Feature vector:

[0121] Normal sample = [0.6, 0.75, 0, 1, 5] (pressure, temperature, valve status, operation sequence, time interval)

[0122] Abnormal sample = [1.0, 1.0, 1, 0, 2]

[0123] Isolation forest calculation:

[0124] Abnormal sample has shorter path length (e.g. E(h(x)) = 3), normal sample has longer path length (e.g. E(h(x)) = 8).

[0125] Substitute formula:

[0126] s ≈ 0.85 (abnormal)

[0127] s ≈ 0.2 (normal)

[0128] Trigger alarm: system marks abnormal operation, real-time prompt in VR simulation "pressure exceeds threshold! Check valve status".

[0129] Model optimization

[0130] Feature engineering: add operation context (e.g. consecutive error times) to improve accuracy.

[0131] Integrated learning: combine LSTM to handle time series dynamics, enhance capture of time series anomalies.

[0132] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and modifications can be made to these embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1.A VR technology-based safety accident simulation system for a hazardous chemical storage and transportation device, characterized in that, The application comprises: a scene simulation module: training a generative adversarial network to input a real hazardous chemical database and receive physical and chemical rule constraints of a knowledge base, synthesize parameters of a virtual new hazardous chemical, establish an uncertainty quantification model, evaluate the confidence interval of the synthesized parameters of the virtual new hazardous chemical, construct a hazardous chemical storage and transportation scene, render and interact with the scene through a VR engine, provide scene basic data to an accident simulation module, and receive environment modification instructions from a user interaction module; an accident simulation module: simulating a hazardous chemical accident, supporting custom configuration of accident type, location, and scale, defining different disaster evolution rules through a multi-agent reinforcement learning framework, and generating a disaster scene through self-play and confrontation of agents, and outputting disaster parameters to the scene simulation module; a user interaction module: a user interacts with a virtual environment through a VR device, sends user operation logs to a data analysis module, and receives operation guidance information from a knowledge base; a data analysis and evaluation module: recording user behavior data, dynamically constructing a decision tree model based on real-time disaster parameters, combining an expert rule base and a reinforcement learning reward function to generate a non-programmed solution, obtaining real-time disaster data from the accident module, pushing a recommended solution to the user interaction module, and feeding back user decision results to the reinforcement learning framework to optimize the strategy, and using an isolation forest to identify abnormal behaviors deviating from standard operating modes; a knowledge base: providing hazardous chemical safety specifications, an accident case library, and training progress management. 2.The VR technology-based dangerous chemical storage and transportation device safety accident simulation system according to claim 1, characterized in that, The scene simulation module trains a generative adversarial network to input a real hazardous chemical database and receive physical and chemical rule constraints of a knowledge base, and the specific steps of synthesizing parameters of a virtual new hazardous chemical are as follows: Data preprocessing and knowledge base construction: convert the structure of hazardous chemicals into images or parameter vectors, and construct the training set X real , define the physical and chemical rule knowledge base R = {r1, r2,..., r n} StyleGAN3 model: conditional input layer: introduce a rule condition vector c in the generator G and the discriminator D, which encodes the rule subset that the current generated molecule needs to meet; rule constraint module: add a physical and chemical rule constraint layer at the end of the generator, which will constrain the generated molecule parameters θ gen to the input rule checker R, and output a rule compliance score s∈[0,1]; loss function: the total loss function includes three parts, namely, adversarial loss, rule constraint loss, and physical rationality loss, and the adversarial loss is represented as L adv = E[log D(X real )] + E[log(1-D(G(z)))], where z is the latent space noise, the regularity constraint loss measures the deviation of the generated parameters from the rules using mean square error or cross-entropy, denoted as The physical plausibility loss calculates the plausibility of the generated molecules using a pre-trained chemical property prediction model, denoted as L phy = -log P valid (θ gen ), and the total loss function is denoted as L total = L adv + λ1L rule + λ2L phy , where λ1, λ2 are hyperparameters. Training procedure: Generator optimization: minimize L total , update generator parameters θ G ; Discriminator optimization: maximize L adv , update discriminator parameters θ D After every fixed number of iterations, dynamically adjust λ1 according to the rule violation rate of generated samples; VR simulation integration: inputting effective parameters into a VR engine. 3.The VR technology-based dangerous chemical storage and transportation device safety accident simulation system according to claim 1, characterized in that, The specific steps of establishing an uncertainty quantification model and evaluating the confidence interval of the synthesized parameters of the virtual new hazardous chemical are as follows: identifying key parameters and their uncertainty sources: determining key parameters that affect simulation results and analyzing their uncertainty sources; establishing a probability distribution model: selecting or fitting a probability distribution for the parameters; Uncertainty propagation: propagate parameter uncertainty to model output by Monte Carlo method, generate N parameter samples k i ~ N(μ, σ 2 ), where μ is the mean and σ is the standard deviation, for each k i , calculate the model output y i = g(k i ), and statistics the distribution characteristics of the output y; calculating the confidence interval: calculating the confidence interval according to the output distribution; Sensitivity analysis: Quantify the contribution of each parameter to the output, identify key parameters, total effect index The total effect of parameter i on the output is represented, and the calculation formula is represented as verification and iteration: comparing experimental data and simulation results to verify the reliability of the model. 4.The VR technology-based dangerous chemical storage and transportation device safety accident simulation system according to claim 2, characterized in that, The specific steps of rendering and scene interaction through a VR engine are as follows: scene modeling and physical engine integration: using a 3D modeling tool to construct a storage tank area, a pipeline network, and a transportation vehicle, integrating NVIDIA PhysX or Havok engine in Unity / Unreal, and setting rigid body dynamics parameters; Fluid leakage simulation: liquid particle initialization, leakage liquid is discretized into smooth particles, particle mass where M leak is the total leakage amount, N p is the number of particles, the particle position and velocity update formula is expressed by the SPH control equation: where P is pressure, p is density, W ij is a smoothing kernel function, g is gravity, f visc is a viscous force; Smoke diffusion simulation: The scene space is discretized into a three-dimensional grid, and the evolution equation for smoke concentration c is given by where u is the wind velocity field, D is the diffusion coefficient, S is the source term, and the subgrid stresses are closed using large eddy simulation, given by where C s = 0.1, Δ is the grid scale, is the rate of strain tensor; parameter-driven dynamic effect generation: mapping the parameters generated by the generative adversarial network to simulation parameters, and triggering a leakage event through VR handle operation; scene rendering and VR interaction optimization: using a GPU particle system to render smoke, and combining noise texture to generate turbulence details. 5.The VR technology-based dangerous chemical storage and transportation device safety accident simulation system according to claim 4, characterized in that, The specific steps of defining different disaster evolution rules through a multi-agent reinforcement learning framework are as follows: Customized accident initial condition configuration: the user inputs initial parameters through the interface, converts discrete inputs into a system-recognizable initial state vector, and inputs it into the multi-agent reinforcement learning framework to start the disaster evolution process; Define the multi-agent reinforcement learning framework: divide the agent roles, including disaster evolution agents, environmental response agents, and emergency countermeasure agents, the disaster evolution agents control the disaster state transition, the environmental response agents simulate the feedback of the environment to the disaster, and the emergency countermeasure agents represent the inhibitory effect of emergency measures on the disaster, and take emergency measures to suppress the disaster; define the action space and state space, the action space represents that the disaster evolution agent adjusts the disaster parameters, the environmental response agent sets the environmental parameters, the disaster evolution agent selects the emergency strategy, and the state space represents the current disaster parameter vector; reward function design: the reward function drives the agent confrontation, the disaster evolution agent rewards the disaster severity, the environmental response agent rewards the rationality of the environmental parameters, and the emergency countermeasure agent rewards the disaster suppression effect; Disaster evolution rule modeling: physical law coupling: leakage triggers explosion when the leakage time exceeds the threshold, explosion triggers fire: explosion shock wave ignites surrounding combustible material with a certain probability, and the secondary fire is determined according to the ignition probability of surrounding combustible material; state transition equation: convert agent action to disaster parameter change, embed evolution rule formula in state transition calculation, constrain physical feasibility of agent action, and input disaster parameter change as reward function to affect agent strategy optimization. 6.The VR technology-based dangerous chemical storage and transportation device safety accident simulation system according to claim 5, characterized in that, The agent generates a disaster scene through self-play confrontation, and outputs disaster parameters to the scene simulation module, including the following steps: Agent self-play confrontation training: the disaster evolution agent and the emergency countermeasure agent confront each other, the environmental response agent provides environmental constraints, in each iteration, the disaster evolution agent tries to aggravate the disaster, the emergency countermeasure agent tries to suppress the disaster, and the confrontation result is fed back to the strategy network, and the strategy network is updated using the proximal policy optimization algorithm; Disaster parameter output and VR scene rendering: convert agent action to disaster parameter, map disaster parameter to visual special effect, output parameter directly drives VR engine, and user feedback can adjust initial configuration through VR experience. 7.The VR technology-based dangerous chemical storage and transportation device safety accident simulation system according to claim 6, characterized in that, The data analysis and evaluation module records user behavior data, dynamically constructs a decision tree model based on real-time disaster parameters, combines expert rule library and reinforcement learning reward function, and generates specific steps of non-programmed scheme: User behavior data and real-time disaster parameter collection: capture user operation and environmental sensor data in real time through VR equipment to form a multi-modal raw data set, and standardize disaster parameters; Dynamic decision tree model construction: taking real-time disaster parameters as root nodes, combining user historical behavior data, dynamically generating a decision tree, and using weighted information gain as the splitting criterion, the information gain calculation formula is represented as Wherein, H(D) is the entropy of the data set, A is the splitting attribute, D v is the subset of attribute A=v, and the splitting threshold is adjusted according to the priority of the expert rule; Expert rule library and reinforcement learning fusion: check whether the decision tree output meets the pre-defined rules in the expert rule library, update the action value function Q(s,a) through Q-learning, and the update formula is represented as: Where, alpha is the learning rate, gamma is the discount factor, and R is the reward function; Non-programmed scheme generation: generate a scheme by integrating the decision tree, expert rules, and reinforcement learning results. 8.The VR technology-based dangerous chemical storage and transportation device safety accident simulation system according to claim 7, characterized in that, The specific steps of the data analysis and evaluation module for identifying abnormal behaviors deviating from the standard operation mode by using the isolation forest are as follows: Data collection and preprocessing: Capture features reflecting operation behaviors from the VR engine, construct quantifiable feature vectors, standardize data, and convert continuous operations into fixed-length feature vectors; Building an isolation forest model: Randomly extract a sample subset from the fixed-length feature vectors, randomly select feature dimensions, randomly select split values within the selected feature dimension range, recursively split the left and right sub-trees until the tree height limit is reached, and repeat the construction of multiple sub-trees to form an isolation forest; The anomaly score of a sample x is denoted as where E(h(x)) is the average path length of sample x in all sub-trees, c(n) = 2H(n - 1) - 2(n - 1) / n, where H(n) = ln n + γ, where γ is the Euler constant, for normalizing the path length, by converting the isolated path in the tree structure into a comparable numerical value, and the anomaly determination rule is: s(x, n) → 1 indicates anomaly, s(x, n) → 0 indicates normal. Threshold setting and anomaly determination: Map the anomaly score to the actual safety rules and adjust the threshold according to the scene risk level.

Images