Environmental hidden danger troubleshooting method and system based on AI and VR deep fusion

Abstract

The invention discloses an environmental hidden danger checking method and system based on AI and VR deep fusion, and the method comprises the steps: S1, constructing a high-fidelity virtual environment and a structured hidden danger library, S2, obtaining immersive hidden danger checking interaction and multi-modal behaviors, S3, carrying out the evaluation of hidden danger recognition and behaviors based on a parallel deep learning model, and S4, carrying out the recognition of hidden dangers. The method comprises the steps of S1, establishing a probability time sequence interval, S4, triggering hidden dangers and generating a dynamic comprehensive risk field based on the probability time sequence interval, S5, performing optimal evacuation path planning based on the dynamic comprehensive risk field, S6, generating an emergency strategy, and S7, pushing multi-modal information in real time. According to the invention, virtual reality hidden danger investigation training and artificial intelligence emergency decision guidance are seamlessly connected, so that real-time behavior data and risk perception information generated in the training process are directly converted into key input of emergency decision, seamless conversion and complete inspection from prevention ability cultivation to emergency decision execution are realized, and the training efficiency is improved. Meanwhile, dynamic risk simulation and personalized intelligent guidance based on a probabilistic model are introduced; the optimal path based on the environment is provided, ability evaluation results of individual users can be deeply fused, customized emergency strategies and guiding strength are generated, technologies such as virtual reality immersion rendering, artificial intelligence vision and behavior analysis, probabilistic graph models, dynamic path planning and case reasoning are deeply integrated into a unified framework, and the method has the advantages of being high in practicability and easy to popularize. Parameter circulation among the modules is natural, and synergistic interaction is achieved.

Description

Technical Field

[0001] This invention belongs to the field of virtual reality simulation technology, specifically relating to an environmental hazard investigation method and system based on the deep integration of AI and VR. Background Technology

[0002] Traditional environmental safety hazard inspections rely primarily on manual inspections and experience-based judgment, which suffers from drawbacks such as low efficiency, inconsistent standards, high training costs, and difficulty in simulating high-risk scenarios for practical drills. Virtual reality technology provides an immersive, repeatable, and risk-free training environment for safety training and has already been applied in emergency drills in industries such as fire protection, mining, and power. However, existing virtual reality safety training systems mostly focus on escape skills training after an accident or disaster occurs, lacking in-depth and intelligent training capabilities for the more proactive and preventative aspect of "intelligent hazard identification and correct initial handling" before an accident. Such systems typically have pre-set fixed drill procedures, making it difficult to provide real-time intelligent analysis and personalized feedback on the diverse inspection behaviors of trainees, resulting in a disconnect between the "training" and the actual "emergency response" decision-making chain.

[0003] On the other hand, in the field of emergency evacuation route planning, dynamic route planning models based on shortest path algorithms and their variants can effectively combine real-time monitoring data for calculation. However, these models are mainly applied to evacuation command at the macro level after a disaster, and their data sources are mostly sensor networks deployed in the physical world. They fail to deeply integrate with key information such as real-time risk warning information and individual skill status generated by personnel during front-end hazard investigation. This results in risk perception in the "prevention phase" and risk avoidance decision-making in the "emergency phase" being in an information silo state, making it impossible to build a complete intelligent decision-making closed loop from early risk perception to accurate dynamic risk avoidance.

[0004] Meanwhile, in recent years, artificial intelligence technology, especially deep learning, has made significant progress in the fields of computer vision and human behavior analysis, providing powerful technical tools for automatically identifying environmental hazards and ensuring the operational compliance of analysts. Organically and deeply integrating the real-time perception and intelligent decision-making capabilities of artificial intelligence with the highly immersive simulation capabilities of virtual reality and dynamic path planning algorithms to construct an integrated system covering the entire process of "intelligent hazard identification training - dynamic risk assessment - intelligent emergency guidance" has become an urgent need to improve the level of intelligent environmental safety management. Summary of the Invention

[0005] To achieve the above objectives, this invention proposes an environmental hazard investigation method based on the deep integration of AI and VR, specifically including the following steps: S1. Construct a high-fidelity virtual environment and a structured hazard library: Construct a high-fidelity virtual reality environment with a spatial topology consistent with the target physical environment, and pre-set a structured hazard model library in the virtual reality environment; the hazard model library includes static object defect models and dynamic process risk models, and each hazard model is associated with its triggering conditions, risk category, risk physical diffusion parameters and standard handling procedures.

[0006] In step S1, the virtual reality environment is constructed based on the building information model or high-precision 3D scan data of the target physical environment, and integrates a physics engine to simulate the dynamic physical effects of object interaction, fire spread, smoke diffusion, and fluid leakage. The dynamic process risk models in the hazard model library are associated with simplified risk physical diffusion algorithm model identifiers. Based on the building information model data or high-precision 3D scan data of the target physical environment, a 1:1 high-fidelity virtual reality environment with consistent geometry and topology is constructed using a 3D graphics engine. Within this environment, a structured hazard model library is preset. Hazard models are mainly divided into two categories: one is static object defect models, exemplarily including exposed wires, damaged guardrails, blocked fire exits, and insufficient pressure fire extinguishers; the other is dynamic process risk models, exemplarily including initial smoke from overheated equipment and leakage diffusion from chemical storage tanks. Each hazard model is associated with its unique triggering condition, risk category, corresponding risk physical diffusion model identifier, standard handling procedure list, and initial risk level.

[0007] S2. Acquiring Immersive Hazard Investigation Interaction and Multimodal Behavior: Users enter the virtual reality environment through virtual reality interaction devices to conduct hazard investigation and emergency response interaction; the system captures the user's pose data, operation action sequence, and gaze focus data in the virtual environment in real time.

[0008] In step S2, the user enters the virtual environment through a head-mounted display, handheld controller, and full-body positioning and tracking device, freely exploring from a first-person perspective and interacting with virtual objects. The system captures the user's head and hand poses in real time through the positioning and tracking unit, records the sequence of operation actions through button events, and obtains the user's gaze direction and focus coordinates through the eye-tracking module. All interaction events, changes in the state of environmental objects, and corresponding timestamps are recorded synchronously.

[0009] S3. Hazard identification and behavior evaluation based on parallel deep learning model: The user behavior data and environmental images captured in step S2 are analyzed in real time based on the parallel running deep learning model; the first deep learning model performs target detection on virtual objects in the user's field of vision area and identifies whether they match the preset hazards; the second deep learning model analyzes the user's operation sequence for the identified hazards and evaluates its conformity with the standard handling procedure; combining the target detection results and operation conformity, a user behavior evaluation report is generated, which includes hazard detection rate, operation standardization score and response timeliness score, and outputs the real-time status information of each identified hazard.

[0010] In step S3, the hazard identification and behavioral intelligent assessment based on the parallel deep learning model includes running two analysis threads in parallel: Thread one is the visual hazard recognition process. This process utilizes a first deep learning model to analyze the real-time rendered image in the user's current field of view. The first deep learning model is trained using a massive amount of labeled virtual hazard scene images and can identify visual feature objects in the image that match the hazard model library, outputting their category, confidence level, and location information.

[0011] Thread two is the operational behavior evaluation process. This process utilizes a second deep learning model to analyze the user's operational sequence in response to identified hazards. The second deep learning model learns the action patterns of standard handling procedures for various hazards. For each hazard initially marked by the target detection model, the continuous operational sequence of the user from discovery to completion of initial handling is analyzed, its dynamic similarity to the standard operational sequence in the time dimension is calculated, and the time nodes of key operations are recorded.

[0012] The system combines the outputs of the two models to determine whether a hazard has been effectively identified and generates a user behavior assessment report that includes the overall hazard detection rate, average operational compliance score, and average response time score. Simultaneously, based on the inherent risk level of the hazard and the user's operational compliance, the system dynamically calculates and outputs the real-time risk level and location information of each hazard point.

[0013] S4. Triggering Hidden Dangers and Generating a Dynamic Comprehensive Risk Field Based on Probabilistic Time-Series Intervals: When the triggering condition of any hidden danger model in the virtual environment is met, the system determines to enter the emergency response mode; based on the triggered hidden danger type and its risk diffusion parameters, the system uses a probabilistic time-series interval model to dynamically calculate the comprehensive risk value of each node in the virtual environment navigation network at a discrete time step; the probabilistic time-series interval model deconstructs the risk evolution process into multiple risk state atomic activities, describes the clustering characteristics of risk states in different regions through latent variables, and describes the time dependence of risk state evolution of different nodes through interval relationships, thereby simulating the dynamic and non-uniform propagation of risk with space, time and virtual personnel distribution, forming a dynamic comprehensive risk field covering the entire scene.

[0014] In step S4, hazard triggering and dynamic comprehensive risk field generation based on probabilistic time-series intervals are performed. When any hazard model in the virtual environment reaches a threshold due to its internal timer or is directly triggered by a user's erroneous operation, the current environment is determined to enter emergency response mode. First, the risk physical diffusion model corresponding to the hazard type needs to be called, such as a Gaussian plume model and its parameters. Subsequently, the dynamic and non-uniform propagation process of simulated risk in the virtual environment is quantified through a probabilistic time-series interval model. The risk state evolution of each node in the entire virtual environment navigation network is regarded as a complex risk evolution activity. Specifically: A risk state atomic activity is defined for each navigation node, representing the risk intensity level of that node at a specific time. Latent variables are also introduced to describe the clustering characteristics of risk state evolution patterns of different nodes; nodes with the same latent variable share similar probabilities of risk state occurrence.

[0015] Define a set of time-series interval relationships to describe the dependence and constraints of risk states of different nodes on the time axis. Combine the physical diffusion law of risk, environmental obstacle information and the virtual personnel density near the node estimated in real time in step S3, and generate the probability of risk state of each node at each discrete time step under the condition of satisfying the interval relationship constraints, and map it to a specific dynamic comprehensive risk value.

[0016] The above process is executed cyclically with a fixed simulation step size, thereby generating a dynamic comprehensive risk field that evolves over time and is spatially non-uniformly distributed, providing a basis for calculating network edge weights in subsequent path planning.

[0017] Specifically, the method of dynamically calculating the comprehensive risk value of a node using a probabilistic time-series interval model includes: S41: Based on the type of hazard triggered, initialize the risk source parameters, including the risk source intensity, diffusion characteristic parameters, and the dominant meteorological parameters of the virtual environment; S42: The process of risk propagation from the source point to the navigation network nodes is modeled as a complex risk evolution activity consisting of multiple risk intensity interval sequences; a risk state atomic activity is defined for each node, and latent variables are introduced to describe the clustering of nodes with similar risk evolution patterns; S43: Based on the allocation of generated node latent variables, nodes under the same latent variable share the same probability distribution of risk state occurrence; according to the allocation of node latent variables, the atomic activities of risk state of each node at the initial time are sampled; S44: Define the set of temporal interval relationships between atomic activities of different node risk states; determine the constrained interval relationships of risk state evolution between nodes based on the physical diffusion law of risk and the spatial layout of obstacles; S45: Under the aforementioned constrained interval relationship, based on the conditional probability distribution corresponding to the combination of atomic activities of risk states between node pairs, sample and generate specific interval relationships for the evolution of risk states between nodes; S46: Combining the virtual personnel density near the node estimated in real time in step S3, the personnel exposure correction is performed on the sampled node risk status; based on the semantic mapping of the risk status atomic activities and the personnel exposure correction results, the dynamic comprehensive risk value of each node at the current moment is calculated. S47: Proceed with a fixed time step, repeat steps S43 to S46, update the latent variable allocation, risk status and interval relationships of all nodes, and realize the time-series evolution simulation of the dynamic comprehensive risk field.

[0018] S5. Optimal evacuation route planning based on dynamic comprehensive risk field: Based on the dynamic comprehensive risk field generated in step S4, optimal evacuation route planning is performed for users in the virtual environment; the virtual environment is abstracted into a navigation road network composed of nodes and edges. According to the geometric attributes of the user's current source node, the preset safe destination node, the edges connecting each node, and the real-time comprehensive risk value and predicted personnel density of the nodes associated with the edges, the passage cost weight of each edge is dynamically calculated; the shortest path search algorithm is used to calculate the path with the minimum cumulative passage cost from the source node to the destination node, which is taken as the optimal evacuation route at the current moment.

[0019] In step S5, optimal evacuation route planning is performed based on the dynamic integrated risk field. In emergency response mode, the system abstracts the virtual environment as a weighted directed graph consisting of nodes and edges, where nodes represent key points in walkable areas and edges represent paths connecting nodes. The goal of route planning is to find the path with the minimum cumulative dynamic travel cost for the user from their current location (source point) to the nearest safe exit (endpoint).

[0020] First, dynamic edge weights are calculated. For an edge connecting two nodes, its dynamic weight consists of three parts: basic travel time cost, risk exposure cost, and congestion time cost. The basic travel time is determined by the geometric length of the edge and the estimated walking speed of people under the influence of the comprehensive risk value of the area associated with the current edge. This speed is calculated using a risk-speed reduction function. The risk exposure cost is directly represented by the real-time comprehensive risk value of the area associated with the edge. For critical passage edges, if the predicted personnel density exceeds a threshold, an additional queuing time cost based on density is added. Finally, the dynamic weight of the edge is a weighted sum of the above three costs, where the trade-off factor can be dynamically adjusted according to the event type.

[0021] Secondly, a shortest path search algorithm is employed on a dynamically weighted navigation graph. The algorithm uses the user's current node as the source point and iteratively calculates the shortest distance and predecessor nodes from the source point to all other nodes in the graph based on continuously updated dynamic edge weights, until the destination is visited. Finally, the optimal evacuation path from the source point to the destination is obtained by backtracking the predecessor nodes.

[0022] Specifically, in step S5, the dynamic calculation of the passage cost weight for each edge includes: S51: For the edge e connecting node u and node v, obtain its length. ; S52: Obtain the dynamic comprehensive risk values ​​of nodes u and v at both ends of edge e at the current time. and Calculate the reference risk intensity of the edge The average of the risk values ​​of the two nodes; S53: Based on reference risk intensity The estimated walking speed of a person on edge e is calculated using a risk-velocity reduction function. The risk-velocity reduction function simulates the panic and deceleration behavior caused by increased risk. S54: Calculate the basic travel time cost of edge e For length Compared with the estimated walking speed The ratio; S55: Obtain the predicted population density in the region where edge e is located. ,like If the congestion density exceeds a preset threshold, additional congestion time costs are calculated based on the degree of exceedance. ; S56: Comprehensive Basic Travel Time Cost Reference risk level and the cost of congestion time The final dynamic weight of edge e is calculated using a weighted sum formula. Among them, the priority between rapid withdrawal and risk avoidance is adjusted by using trade-off factors.

[0023] S6. Generate emergency strategy: Based on the user behavior assessment report generated in step S3, the types of potential hazards and real-time risk status determined in step S4, and the optimal evacuation route planned in step S5, a decision-making model combining rule reasoning and case retrieval is used to generate a personalized emergency strategy instruction set for the current user, which includes specific action instructions and resource utilization suggestions.

[0024] In step S6, the emergency strategy generation takes into account three dimensions of input: the user profile generated from the user behavior assessment report in step S3; the event context generated from the hazard type, severity, and real-time spread trend; and the environmental situation generated from the optimal path planned in step S5 and the real-time risk parameters of its nodes. A hybrid decision-making model combining a rule engine and case-based reasoning is employed. The rule engine matches predefined logical rules, while the case-based reasoning engine retrieves successful handling cases similar to the current scenario from a historical exercise case library. Through weighted fusion and logical conflict resolution, an ordered and personalized emergency strategy instruction set is generated, which may include specific action guidance, prompts for utilizing emergency resources along the route, and suggestions for division of labor in collaborative exercises.

[0025] S7. Real-time Multimodal Information Push: The optimal evacuation route planned in step S5 and the emergency strategy instruction set generated in step S6 are presented to the user in real time and in a multimodal manner through enhanced visual overlay, three-dimensional spatial audio guidance and virtual interactive interface in the virtual reality environment, guiding the user to perform evacuation or initial response operations; the system continuously monitors changes in the environment and user status, and executes steps S4 to S6 in a loop to realize dynamic updates of the guidance strategy until the user reaches a safe point or the exercise ends.

[0026] In step S7, the optimal path and strategy instructions are transformed into intuitive multi-sensory guidance information through the virtual reality immersive interaction unit and presented to the user in real time: visual enhancement overlays the optimal path onto the user's field of vision as dynamically highlighted arrows or colored light strips on the ground; spatial audio guidance uses three-dimensional sound field technology to provide directional voice prompts from the corresponding spatial location ahead of the path; the virtual interactive interface allows users to view a simplified electronic map and a list of current priority instructions on their handheld devices. Based on this guidance, the user performs evacuation or initial response in a continuously evolving virtual risk environment. During this process, the system continuously loops through steps S4 to S6, monitoring changes in the risk situation and user location, dynamically updating the risk field, replanning the path, and adjusting strategy instructions until the user successfully reaches the virtual safety point or the preset exercise time ends.

[0027] Meanwhile, the present invention also provides an environmental hazard investigation system based on the deep integration of AI and VR, and a virtual reality immersive training platform, including: a scene model unit, used to construct, manage and render the high-fidelity virtual reality environment; The hazard interaction unit is used to manage the hazard model library and handle the interaction logic and status changes between users and hazard models. The training management unit is used to configure training courses, exercise scenarios, and assessment models. The immersive interaction unit integrates a head-mounted display, control handle, positioning tracker, and eye-tracking module to provide an immersive experience and capture user interaction data. Artificial intelligence intelligent analysis and decision-making platform, including: The hazard identification unit deploys the first deep learning model and the second deep learning model to analyze user behavior and virtual environment images in real time, identify hazards, and generate user behavior assessment reports. The dynamic risk field calculation unit is used to run the probabilistic time series interval model in emergency response mode, dynamically simulate and update the comprehensive risk value of the navigation road network nodes in the virtual environment, and generate a dynamic comprehensive risk field. The dynamic path planning unit is used to calculate the dynamic weights of the navigation road network edges based on the dynamic comprehensive risk field and real-time personnel density prediction, and to run the shortest path search algorithm to solve for the optimal evacuation path. The emergency strategy unit has a built-in rule engine and case inference engine, which is used to generate personalized emergency strategy instruction sets based on multi-dimensional input information; The network unit is used to enable data communication between the virtual reality immersive training platform and the artificial intelligence intelligent analysis and decision-making platform; The database unit is used to store virtual environment models, hazard model libraries, user behavior data, algorithm parameters, historical cases, and exercise process data.

[0028] The dynamic risk field calculation unit includes a latent variable allocation module, a risk state sampling module, an interval relationship constraint generation module, and a risk value calculation module. Its inputs are hazard triggering event parameters, environmental obstacle information, and personnel density prediction data output in real time by the hazard identification unit. Its output dynamic comprehensive risk field is directly provided to the dynamic path planning unit for edge weight calculation.

[0029] The system supports multi-user collaborative training and drill modes; the network unit is responsible for synchronizing the virtual status and behavior data of multiple users; in emergency response mode, the dynamic path planning unit plans globally coordinated evacuation paths for multiple collaborative users, and the emergency strategy unit generates a joint emergency strategy that includes division of labor and cooperation instructions.

[0030] Meanwhile, the present invention also provides an electronic device, the electronic device including a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method described above.

[0031] The present invention also provides a computer-readable storage medium, characterized in that it includes: a computer program stored on the computer-readable storage medium, wherein the computer program, when executed by the processor, implements the steps of the method described above.

[0032] The beneficial effects of this invention are as follows: (1) This invention seamlessly integrates virtual reality hazard investigation training with artificial intelligence emergency decision guidance through the “training-combat integration” approach, so that the real-time behavioral data and risk perception information generated during the training process can be directly transformed into key inputs for emergency decision-making, realizing a seamless transition and complete verification from prevention capability cultivation to emergency decision execution.

[0033] (2) This invention introduces dynamic risk simulation based on probabilistic models: applying probabilistic time series interval models to the generation of virtual risk fields can simulate the probabilistic evolution of risk with space, time and human behavior in a more refined and realistic way, overcoming the shortcomings of traditional deterministic models in characterizing the uncertainty of complex systems, and greatly improving the realism of the exercise and the challenge of decision-making.

[0034] (3) This invention provides highly personalized intelligent guidance: the system not only provides the optimal path based on the environment, but also deeply integrates the individual user's ability assessment results to generate tailored emergency strategies and guidance intensity, realizing the leap from "one path for a thousand people" to "one policy for one person", and improving the accuracy and effectiveness of training and guidance.

[0035] (4) This invention promotes the deep integration of multiple cutting-edge technologies, integrating virtual reality immersive rendering, artificial intelligence visual and behavioral analysis, probabilistic graphical models, dynamic path planning, case reasoning, and other technologies into a unified framework. The parameters flow naturally and synergistically between the various technical modules, forming a system-level intelligent solution. This invention has broad application prospects and scalability: the method and system framework can be widely applied to safety education and training, emergency plan simulation, and emergency decision support in high-risk industries such as energy and chemical industry, mining and metallurgy, large public buildings, transportation hubs, and construction. By combining with the Internet of Things and digital twin technologies, it can further realize intelligent safety management that links the virtual and real worlds, and has significant social and economic benefits. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating an environmental hazard investigation method based on the deep integration of AI and VR. Figure 2 This is a schematic diagram of a module for an environmental hazard investigation system based on the deep integration of AI and VR. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Reference Figure 1 , Figure 1 This is a schematic diagram of an environmental hazard investigation method based on the deep integration of AI and VR according to the present invention. Specifically, the method includes the following steps: S1. Construct a high-fidelity virtual environment and a structured hazard library: Construct a high-fidelity virtual reality environment with a spatial topology consistent with the target physical environment, and pre-set a structured hazard model library in the virtual reality environment; the hazard model library includes static object defect models and dynamic process risk models, and each hazard model is associated with its triggering conditions, risk category, risk physical diffusion parameters and standard handling procedures.

[0039] Understandably, in step S1, the virtual reality environment is constructed based on the building information model of the target physical environment or high-precision 3D point cloud data obtained through laser scanning, ensuring that the virtual environment remains consistent with the real environment in terms of geometric dimensions and spatial topology. The hazard model library is managed using relational database tables. Each record contains, but is not limited to, the following fields: a unique hazard identifier, name, type, associated 3D model resource index, trigger condition logical expression, risk category code, corresponding risk physical diffusion model identifier, an ordered list of standard handling procedure steps, and an initial risk level value. The physics engine integrated into the 3D graphics engine is used to simulate interactive behaviors such as collisions, grabbing, and switching between objects, as well as dynamic physical effects such as fire spread, smoke diffusion, and fluid leakage. These effects provide the physical basis and behavioral triggering conditions for the risk dynamic evolution simulation in the subsequent step S4.

[0040] S2. Acquiring Immersive Hazard Investigation Interaction and Multimodal Behavior: Users enter the virtual reality environment through virtual reality interaction devices to conduct hazard investigation and emergency response interaction; the system captures the user's pose data, operation action sequence, and gaze focus data in the virtual environment in real time.

[0041] Understandably, in step S2, multiple users enter the virtual environment through virtual reality head-mounted displays, handheld motion controllers, and indoor optical or laser positioning systems. The immersive interaction unit continuously captures and generates multiple data streams, including: the six-DOF pose of the head-mounted display, which includes three-dimensional position and rotation orientation; the six-DOF pose of the left and right hand controllers and all button press and / or release events; the gaze direction vector obtained through an integrated or external eye tracker; and the focal coordinates in the virtual world. All captured raw data is tagged with high-precision timestamps and transformed into a unified virtual world coordinate system. Intentional interactions between users and virtual objects, exemplarily including clicking on a hazard icon, grabbing a fire extinguisher, and turning a valve, are converted into structured operation sequence logs. Each log entry includes the operation type, a unique identifier for the target object, and the operation start and end times.

[0042] S3. Hazard identification and behavior evaluation based on parallel deep learning model: The user behavior data and environmental images captured in step S2 are analyzed in real time based on the parallel running deep learning model; the first deep learning model performs target detection on virtual objects in the user's field of vision area and identifies whether they match the preset hazards; the second deep learning model analyzes the user's operation sequence for the identified hazards and evaluates its conformity with the standard handling procedure; combining the target detection results and operation conformity, a user behavior evaluation report is generated, which includes hazard detection rate, operation standardization score and response timeliness score, and outputs the real-time status information of each identified hazard.

[0043] Understandably, in step S3, the hazard identification unit deploys two parallel deep learning models, forming a dual-threaded analysis pipeline, including a first deep learning model and a second deep learning model. The first deep learning model is an object detection model, employing a single-stage or two-stage object detector based on a convolutional neural network, exemplified by YOLOv5 or Faster R-CNN. This first deep learning model is trained using a large number of precisely annotated images of hazard objects rendered in virtual scenes. The model input is a real-time rendered image frame within the user's field of vision, and the output is the bounding box coordinates, category labels, and confidence scores of all detected objects in the image. The detection results are matched against a hazard model library. If a match is successful and the confidence score is higher than a preset visual threshold, the object is initially marked as a "suspected hazard."

[0044] Furthermore, the second deep learning model is a behavior recognition model: it employs a temporal deep learning model, exemplified by a 3D CNN, a Long Short-Term Memory (LSTM) network, or a Transformer encoder, to model and analyze the user's operation sequence. This second deep learning model can learn the action patterns of various standard hazard handling procedures. Its input is a feature vector of the user's operation sequence within a certain time window, exemplified by the start of visual detection of a suspected hazard until the current moment or the end of the handling action. This feature vector is composed of features fused from continuous frames, such as changes in handle pose, button event sequences, proximity to specific virtual objects, and interaction states. The output of the second deep learning model is a scalar value representing the similarity score between the user's current operation sequence and the standard operation sequence in terms of temporal pattern.

[0045] Understandably, the final determination of hazard identification results follows a fusion logic. The system determines a hazard is "effectively identified" only when the target detection confidence exceeds the visual threshold and the behavior recognition model analysis indicates that the user performed a sequence of key inspection or handling actions in the standard procedure for that hazard. Simultaneously, the system records the accurate identification time. Furthermore, for each hazard i determined to be identified, the user behavior evaluation sub-item is calculated as follows: Operational Standardization Score The operation sequence similarity score output by the second deep learning model is directly taken. Response timeliness score The completion time for the user to complete all key steps of the standard procedure for handling this hazard is [time]. The maximum permissible time limit for handling this potential hazard is... The response time rating The calculation process is as follows:

[0046] Understandably, the scoring for operational standardization and the response time rating All scores are in the [0,1] interval, and the operational standardization score is calculated accordingly. In other words, the closer the action pattern is to the standard handling procedure, the closer the score is to 1, which affects the response timeliness score. In other words, the more timely the handling, the closer the score is to 1. Furthermore, this is achieved through the aforementioned operational standardization scoring. Compared with the number of potential hazards identified by users The ratio determines the average operational standardization score, which is obtained through the response timeliness score. Compared with the number of potential hazards identified by users The ratio is used to determine the average response time score.

[0047] Furthermore, the user behavior evaluation report for the entire virtual training scenario also summarizes and generates the following core indicators: Hazard detection rate ,in This represents the total number of potential hazards preset in the scenario.

[0048] Furthermore, based on the inherent risk level of the pre-set hazards in the hazard database... And user operation standardization score Dynamically calculate the remaining risk level of the potential hazard at the current moment. The specific calculation process is as follows:

[0049] in, The attenuation coefficient of the treatment effect must meet the following requirements. It is understandable that the more standardized the user's operation, the higher the score for the standardization of the user operation. , The larger the value, the higher the remaining risk level at the current moment. The lower.

[0050] Understandably, the user behavior evaluation report for the hazard identification results in step 3 includes the overall hazard detection rate, average operational compliance score, and average response time score. Simultaneously, based on the inherent risk level of the hazard and the user's operational compliance, the real-time risk level and location information of each hazard point are dynamically calculated and output to prepare for subsequent steps.

[0051] S4. Triggering Hidden Dangers and Generating a Dynamic Comprehensive Risk Field Based on Probabilistic Time-Series Intervals: When the triggering condition of any hidden danger model in the virtual environment is met, the system determines to enter the emergency response mode; based on the triggered hidden danger type and its risk diffusion parameters, the system uses a probabilistic time-series interval model to dynamically calculate the comprehensive risk value of each node in the virtual environment navigation network at a discrete time step; the probabilistic time-series interval model deconstructs the risk evolution process into multiple risk state atomic activities, describes the clustering characteristics of risk states in different regions through latent variables, and describes the time dependence of risk state evolution of different nodes through interval relationships, thereby simulating the dynamic and non-uniform propagation of risk with space, time and virtual personnel distribution, forming a dynamic comprehensive risk field covering the entire scene.

[0052] Specifically, step S4 includes: S41: Based on the type of hazard triggered, initialize the risk source parameters, including the risk source intensity, diffusion characteristic parameters, and the dominant meteorological parameters of the virtual environment; Understandably, in step S41, when the system determines that a certain hidden danger k has been triggered, it determines the location of the hidden danger k based on its coordinates in the virtual world. The dynamic risk field calculation is started immediately. In step S4, the abstract hidden danger event can be quantified into a spatial risk distribution that covers the entire virtual environment navigation road network and evolves probabilistically over time. This is specifically achieved through a probabilistic time series interval model.

[0053] Understandably, the first step is to call the corresponding risk physics diffusion model based on the type of hazard to obtain the basic parameters. Taking a toxic gas leak as an example, a Gaussian plume model is used as the physical basis. The model involves parameters including: the intensity of the risk source release, the severity level of the hazard, the intensity conversion coefficient, the dominant wind speed vector of the virtual environment, and the atmospheric stability level.

[0054] S42: The process of risk propagation from the source point to the navigation network nodes is modeled as a complex risk evolution activity consisting of multiple risk intensity interval sequences; a risk state atomic activity is defined for each node, and latent variables are introduced to describe the clustering of nodes with similar risk evolution patterns; Understandably, in step S42, when entering the probabilistic temporal interval modeling process, the set of nodes included in the virtual environment navigation road network is set as follows: , each node The risk intensity level at any time t is defined as a risk state atomic activity. ,in The value of comes from a discrete set of risk states.

[0055] S43: Based on the allocation of generated node latent variables, nodes under the same latent variable share the same probability distribution of risk state occurrence; according to the allocation of node latent variables, the atomic activities of risk state of each node at the initial time are sampled; Understandably, in step S43, in order to describe the similarity and clustering characteristics of the risk state evolution among different nodes, a latent variable is introduced for each node j. The hidden variable The allocation is performed through a nonparametric Bayesian process, and the latent variables are... The allocation process is as follows: The first node is determined according to its distance from the risk source. The first latent variable is always selected with probability 1, denoted as . For the nth node, select an existing latent variable. The probability and the current allocation Number of nodes The probability of choosing a new latent variable is directly proportional to the concentration parameter. Proportional, the specific process of probability allocation is as follows:

[0056] in, This is a hyperparameter that controls the rate at which new latent variables are generated. Assignment via a nonparametric Bayesian process can automatically cluster nodes with similar risk evolution patterns—for example, nodes downstream of a ventilation duct and similarly obscured by a wall—to the same latent variable. Furthermore, each latent variable t is associated with a multinomial distribution. This distribution defines the prior probability of a node exhibiting various risk state atomic activities under this latent variable. Furthermore, the multinomial distribution... The risk state distribution parameters are obtained by sampling from a symmetric Dirichlet distribution, therefore for each latent variable t, the risk state distribution parameters are the same for all latent variables. For nodes, they share the same probability of occurrence of risky states. .

[0057] S44: Define the set of temporal interval relationships between atomic activities of different node risk states; determine the constrained interval relationships of risk state evolution between nodes based on the physical diffusion law of risk and the spatial layout of obstacles; In step S44, the temporal constraints between risk state intervals of different nodes are described by interval relationships. For any pair of nodes (i,j), there are interval relationship generation constraints between their risk state evolution. , where R is the set of the seven basic relations, defined as: The constraints generated based on interval relationships originate from physical diffusion laws and spatial barriers. For example, if it takes time for a risk to diffuse from i to j, then a relationship may exist. If the two are completely isolated, then there may only be a relationship. or This is a non-overlapping relationship. Therefore, initially, the constraint is most relaxed for directly adjacent node pairs (n, n+1), i.e. For non-directly adjacent node pairs (i,j), the constraints are... The derivation is made through the transitive combination of relationships between intermediate nodes along its path, specifically as follows: ,in The compound operation representing interval relationships describes the mutual constraints between risk state intervals of different nodes in time, ensuring that the generated entire interval network is consistent in time relationships.

[0058] S45: Under the aforementioned constrained interval relationship, based on the conditional probability distribution corresponding to the combination of atomic activities of risk states between node pairs, sample and generate specific interval relationships for the evolution of risk states between nodes; In step S45, it can be understood that for each node j, based on its latent variables... From its corresponding risk status distribution Sampling was performed to obtain the initial risk state atomic activity. For each pair of existence constraints For each node, define a set of conditional probability distributions. This is in the constraint set The multinomial distribution on the conditional probability distribution has the following parameters: It indicates that the state of a given node i is The state of node j is Furthermore, the relationship between the two must belong to a constraint set. Under what conditions, what specific type of interval relationship will occur? The probability of [the outcome]. Furthermore, after the initial state is determined, the process continues in discrete time steps. At each time step, based on the current risk state of each node and the constraints between nodes... and conditional probability distribution parameters The system samples and generates or updates the interval relationships of risk states between nodes. Simultaneously, nodes may also transition to new risk states with a preset probability, based on their latent variable distribution and the influence of neighboring nodes. This process simulates the probabilistic propagation of risk and state transitions in space.

[0059] S46: Combining the virtual personnel density near the node estimated in real time in step S3, the personnel exposure correction is performed on the sampled node risk status; based on the semantic mapping of the risk status atomic activities and the personnel exposure correction results, the dynamic comprehensive risk value of each node at the current moment is calculated. Specifically, in step S46, each risk state atomic activity Mapped to a base risk intensity value The mapping relationship can be predefined. Further considering the mitigation of risk by environmental obstacles, an obstacle influence factor is preset for each node j. ,in It is understandable that when the obstacle influence factor A value of 1 indicates that the obstacle completely blocks the risk; the adjusted risk intensity is:

[0060] Furthermore, by incorporating personnel exposure factors, the real-time estimated virtual personnel density of the area near node j is obtained from the visual analysis module in step S3. The dynamic comprehensive risk value of node j at time t The calculation is as follows:

[0061] in, It is the personnel exposure risk coefficient, which is a comprehensive coefficient that combines theoretical intensity and personnel density to transform into the degree of danger to personnel.

[0062] S47: Proceed with a fixed time step, repeat steps S43 to S46, update the latent variable allocation, risk status and interval relationship of all nodes, and realize the time-series evolution simulation of the dynamic comprehensive risk field. Specifically, in step S47, the system cyclically executes S4.4 and S4.5 with a fixed time step. Within each time step, based on the latest risk source status, the sampling results of relationships between nodes, and the changing personnel density, the system updates the risk status of all nodes and calculates the dynamic comprehensive risk value. This generates and continuously updates a set of nodes covering all virtual environment navigation networks. Dynamic integrated risk field evolving over time t .

[0063] S5. Optimal evacuation route planning based on dynamic comprehensive risk field: Based on the dynamic comprehensive risk field generated in step S4, optimal evacuation route planning is performed for users in the virtual environment; the virtual environment is abstracted into a navigation road network map composed of nodes and edges. According to the geometric attributes of the user's current source node, the preset safe destination node, the edges connecting each node, and the real-time comprehensive risk value and predicted personnel density of the nodes associated with the edges, the passage cost weight of each edge is dynamically calculated; the shortest path search algorithm is used to calculate the path with the minimum cumulative passage cost from the source node to the destination node, which is taken as the optimal evacuation route at the current moment; It is understandable that during the path planning process, in emergency response mode, the virtual environment is abstracted as a weighted directed graph. ,in It is a set of nodes covering all virtual environment navigation networks. It is the set of edges that form walkable paths connecting adjacent nodes. It uses dynamic weights that change over time, and its goal is to find a path P from the user's current source node s to a specified safe destination node t, such that the cumulative sum of the weights of all edges on the path is minimized. Furthermore, in step S5, the dynamic calculation of the dynamic weight of the passage cost for each edge specifically includes: S51: For the edge e connecting node u and node v, obtain its length. ; S52: Obtain the dynamic comprehensive risk values ​​of nodes u and v at both ends of edge e at the current time. and Calculate the reference risk intensity of the edge The average of the risk values ​​of the two nodes; S53: Based on reference risk intensity The estimated walking speed of a person on edge e is calculated using a risk-velocity reduction function. ; The risk - speed reduction function simulates the behavior of people panicking and decelerating due to increased risk. S54: Calculate the basic travel time cost of edge e is the length and the estimated walking speed ratio; S55: Obtain the predicted population density in the area where edge e is located , if exceeds the preset congestion density threshold, calculate the additional congestion time cost according to the exceeding degree ; S56: Synthesize the basic travel time cost , the reference risk intensity and the congestion time cost , calculate the final dynamic weight of edge e through the weighted sum formula ; Among them, the trade - off factor is used to adjust the priority between rapid evacuation and risk avoidance.

[0064] Furthermore, after obtaining the dynamic weights, solve the shortest path on the weighted directed graph G. It can be understood that since the edge weights change with time, the algorithm uses the snapshot of the edge weights at the current moment for calculation at each path planning request moment , and the specific steps are as follows: Initialization: Set the distance of the source point s , and the distance of all other nodes is , mark all nodes as unvisited, create a priority queue Q, usually based on a minimum heap, add all nodes, or initialize the set of unvisited nodes .

[0065] Loop and execute the following operations until the end point t is marked as visited or the priority queue Q is empty. This process includes the following steps: Take out the node with the smallest current dist value from Q, denoted as current, traverse all adjacent nodes neighbor of current, and calculate the temporary distance: , if tempDist < dist[neighbor], then update dist[neighbor] = tempDist, record the predecessor node of neighbor as current, and at the same time update the priority of neighbor in the priority queue Q, and mark the node current as visited.

[0066] It can be understood that after the process ends, starting from the end point t, trace back along the recorded predecessor nodes until the source point s, that is, obtain the optimal evacuation path at the moment ; .

[0067] S6. Generate emergency strategy: Based on the user behavior assessment report generated in step S3, the types of hidden danger events and real-time risk situation determined in step S4, and the optimal evacuation route planned in step S5, a decision-making model combining rule reasoning and case retrieval is used to generate a set of emergency strategy instructions that are personalized for the current user and include specific action instructions and resource utilization suggestions. Specifically, in step S6, the user profile is generated based on the user behavior assessment report generated in step S3; the trigger event context is input, including the type of hazard triggered, its severity, and its real-time spread; and the optimal path planned in step S5 and the real-time risk parameters of each node along the path are input. and population density The rule engine employs a hybrid decision-making model that combines rule engines with case-based reasoning. This means the rule engine includes logical combinations of evaluation metrics corresponding to user profiles and event attributes, such as electrical fires, gas leaks, and sudden increases in risk at nodes ahead along the path. The instruction section provides specific action suggestions, resource prompts, or warnings, with each rule having a corresponding weight.

[0068] Furthermore, cases similar to the current scenario are retrieved from the historical exercise case library. Case similarity is calculated by weighting event type similarity, user profile similarity based on the distance of each indicator in the evaluation report, and environmental situation similarity based on the matching of key features of the risk field. The top K cases with the highest similarity are retrieved, and each case includes its historically successfully executed strategy instruction sequence and historical utility evaluation.

[0069] Furthermore, the current input information is matched with the rule base to generate a preliminary set of instruction candidates. Each instruction is assigned a corresponding rule weight. For each candidate instruction, it is checked whether it appears in the retrieved similar cases. If it does, the instruction weight is enhanced based on the similarity and utility of the case, and the summation is performed on all similar cases containing the candidate instruction.

[0070] Simultaneously, conflict resolution and prioritization are also required during this process. For instructions that may have logical or spatiotemporal conflicts, such as the instruction "run fast" and the instruction "crawl forward" being mutually conflicting, the instruction with higher weight after enhancement should be selected, or a decision should be made based on the physical characteristics of the event. All selected instructions should be arranged in descending order according to their enhanced weights, and adjustments should be made in conjunction with the temporal and spatial logic of emergency response to generate the final ordered and personalized emergency strategy instruction set.

[0071] S7. Real-time Multimodal Information Push: The optimal evacuation route planned in step S5 and the emergency strategy instruction set generated in step S6 are presented to the user in real time and in a multimodal manner through enhanced visual overlay, three-dimensional spatial audio guidance and virtual interactive interface in the virtual reality environment, guiding the user to perform evacuation or initial response operations; the system continuously monitors changes in the environment and user status, and executes steps S4 to S6 in a loop to realize dynamic updates of the guidance strategy until the user reaches a safe point or the exercise ends.

[0072] Understandably, in step S7, decision-making information is transformed into multi-sensory guidance that users can intuitively and naturally perceive through immersive virtual reality interaction. This includes: Visual Enhancement: Within the user's field of vision, the optimal path is clearly presented as a continuously highlighted arrow on the ground, a light strip, or a color gradient channel. Key decision points and emergency resource points, such as fire extinguishers and escape doors, are highlighted or have glowing icons. Important text instructions are presented as semi-transparent information boxes, floating next to relevant objects or in non-core areas above or to the side of the field of vision.

[0073] Spatial audio guidance: Directional voice prompts, such as "Turn left" or "Go straight for 20 meters," are processed by the 3D audio system of the headset and delivered from the corresponding spatial location to enhance the sense of direction. Emergency alarms and key action instructions, such as "Lie down! Cover your mouth and nose!", can also be emphasized with different tones, speeds, and intensities of voice.

[0074] Haptic and interactive feedback: Users can view a simplified electronic map and a list of current priority commands on the terminal screen on their virtual wrist. When simulating key operations, such as forcefully pushing open an escape door or using a fire extinguisher, the handle provides vibration feedback of different modes and intensities to enhance the realism and immersion of the operation.

[0075] Furthermore, based on these multimodal guidance, users perform evacuation or initial response actions within a continuously evolving dynamic risk field. It is understood that this process is not static, but rather a continuous loop of steps S4 to S6: monitoring whether the hazard has spread or new secondary hazards have emerged, thereby updating the risk field, and planning the optimal path in real time based on the updated risk field and the user's latest location; updating emergency strategy instructions in conjunction with the new path and environmental situation. This process continues until the user successfully reaches the virtual safe point, or the preset exercise time ends. All process data is recorded and used to generate the final exercise report and accumulate a case library.

[0076] In multi-user mode, multiple users access the same virtual environment. The system needs to handle more complex global state synchronization and collaborative decision-making. During this process, hazard identification and behavior assessment require integrating the behavioral data of all users. The dynamic risk field calculation unit needs to consider the location information of all online users when calculating personnel density. The dynamic path planning unit, when planning a path for a single user, considers the locations of other users as dynamic "personnel obstacles" in real time. It's understandable that in edge weight calculation, the personnel density of other users' current nodes or edges will increase significantly, leading to a greater congestion time cost for that edge. This causes the above methods to tend to plan paths for the current user that avoid areas where other users congregate, thus optimizing and avoiding global evacuation flows.

[0077] Reference Figure 2 , Figure 2 This is an environmental hazard investigation system based on the deep integration of AI and VR, and a virtual reality immersive training platform, including: a scene model unit, used to construct, manage and render the high-fidelity virtual reality environment; The hazard interaction unit is used to manage the hazard model library and handle the interaction logic and status changes between users and hazard models. The training management unit is used to configure training courses, exercise scenarios, and assessment models. The immersive interaction unit integrates a head-mounted display, control handle, positioning tracker, and eye-tracking module to provide an immersive experience and capture user interaction data. Artificial intelligence intelligent analysis and decision-making platform, including: The hazard identification unit deploys the first deep learning model and the second deep learning model to analyze user behavior and virtual environment images in real time, identify hazards, and generate user behavior assessment reports. The dynamic risk field calculation unit is used to run the probabilistic time series interval model in emergency response mode, dynamically simulate and update the comprehensive risk value of the navigation road network nodes in the virtual environment, and generate a dynamic comprehensive risk field. The dynamic path planning unit is used to calculate the dynamic weights of the navigation road network edges based on the dynamic comprehensive risk field and real-time personnel density prediction, and to run the shortest path search algorithm to solve for the optimal evacuation path. The emergency strategy unit has a built-in rule engine and case inference engine, which is used to generate personalized emergency strategy instruction sets based on multi-dimensional input information; The network unit is used to enable data communication between the virtual reality immersive training platform and the artificial intelligence intelligent analysis and decision-making platform; The database unit is used to store virtual environment models, hazard model libraries, user behavior data, algorithm parameters, historical cases, and exercise process data.

[0078] The dynamic risk field calculation unit includes a latent variable allocation module, a risk state sampling module, an interval relationship constraint generation module, and a risk value calculation module. Its inputs are hazard triggering event parameters, environmental obstacle information, and personnel density prediction data output in real time by the hazard identification unit. Its output dynamic comprehensive risk field is directly provided to the dynamic path planning unit for edge weight calculation.

[0079] The system supports multi-user collaborative training and drill modes; the network unit is responsible for synchronizing the virtual status and behavior data of multiple users; in emergency response mode, the dynamic path planning unit plans globally coordinated evacuation paths for multiple collaborative users, and the emergency strategy unit generates a joint emergency strategy that includes division of labor and cooperation instructions.

[0080] Meanwhile, the present invention also provides an electronic device, the electronic device including a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method described above.

[0081] It should be noted that, in this invention, the terms "comprising," "including," or any other variations are intended to cover a non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0082] The sequence numbers of the above embodiments of the present invention are merely for description and do not represent the superiority or inferiority of the embodiments. Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments may be implemented by means of software plus necessary general-purpose hardware platforms, and of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0083] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for investigating environmental hazards based on the deep integration of AI and VR, characterized in that, Includes the following steps: S1. Construct a high-fidelity virtual environment and a structured hazard library: Construct a high-fidelity virtual reality environment with a spatial topology consistent with the target physical environment, and pre-set a structured hazard model library in the virtual reality environment; the hazard model library includes static object defect models and dynamic process risk models, and each hazard model is associated with its triggering conditions, risk category, risk physical diffusion parameters and standard handling procedures; S2. Acquiring Immersive Hazard Investigation Interaction and Multimodal Behavior: Users enter the virtual reality environment through virtual reality interaction devices to conduct hazard investigation and emergency response interaction; the system captures the user's pose data, operation action sequence, and gaze focus data in the virtual environment in real time; S3. Hazard identification and behavior assessment based on parallel deep learning model: Real-time analysis of user behavior data and environmental images captured in step S2 based on parallel deep learning model; The first deep learning model performs target detection on virtual objects in the user's field of vision area and identifies whether they match the preset hazards; The second deep learning model analyzes the user's operational sequence for identified hazards and assesses its conformity with standard handling procedures. Based on the combined target detection results and operational compliance, a user behavior evaluation report is generated, which includes the hazard detection rate, operational standardization score, and response timeliness score, and outputs the real-time status information of each identified hazard. S4. Triggering hidden dangers and generating a dynamic comprehensive risk field based on probabilistic time-series intervals: When the triggering condition of any hidden danger model in the virtual environment is met, the system determines to enter the emergency response mode; based on the triggered hidden danger type and its risk diffusion parameters, the comprehensive risk value of each node in the virtual environment navigation road network at the discrete time step is dynamically calculated using a probabilistic time-series interval model. The probabilistic time series interval model deconstructs the risk evolution process into multiple risk state atomic activities, describes the clustering characteristics of risk states in different regions through latent variables, and describes the time dependence of risk state evolution at different nodes through interval relationships. In this way, it simulates the dynamic and non-uniform propagation of risk with spatial, temporal and virtual personnel distribution, forming a dynamic comprehensive risk field covering the entire scenario. S5. Optimal evacuation route planning based on dynamic comprehensive risk field: Based on the dynamic comprehensive risk field generated in step S4, optimal evacuation route planning is performed for users in the virtual environment; the virtual environment is abstracted into a navigation road network map composed of nodes and edges. According to the geometric attributes of the user's current source node, the preset safe destination node, the edges connecting each node, and the real-time comprehensive risk value and predicted personnel density of the nodes associated with the edges, the passage cost weight of each edge is dynamically calculated; the shortest path search algorithm is used to calculate the path with the minimum cumulative passage cost from the source node to the destination node, which is taken as the optimal evacuation route at the current moment; S6. Generate emergency strategy: Based on the user behavior assessment report generated in step S3, the types of hidden danger events and real-time risk situation determined in step S4, and the optimal evacuation route planned in step S5, a decision-making model combining rule reasoning and case retrieval is used to generate a set of emergency strategy instructions that are personalized for the current user and include specific action instructions and resource utilization suggestions. S7. Real-time Multimodal Information Push: The optimal evacuation route planned in step S5 and the emergency strategy instruction set generated in step S6 are presented to the user in real time and in a multimodal manner through enhanced visual overlay, three-dimensional spatial audio guidance and virtual interactive interface in the virtual reality environment, guiding the user to perform evacuation or initial response operations; the system continuously monitors changes in the environment and user status, and executes steps S4 to S6 in a loop to realize dynamic updates of the guidance strategy until the user reaches a safe point or the exercise ends.

2. The method according to claim 1, characterized in that, In step S1, the virtual reality environment is constructed based on the building information model or high-precision 3D scan data of the target physical environment, and integrates a physics engine to simulate the dynamic physical effects of object interaction, fire spread, smoke diffusion, and fluid leakage; the dynamic process risk model in the hazard model library is associated with a simplified risk physical diffusion algorithm model identifier.

3. The method according to claim 1, characterized in that, In step S3, the second deep learning model is a time series model, which is used to analyze the continuous operation sequence of the user from identifying the hidden danger to completing the initial treatment process, and calculate its similarity score with the standard operation sequence in the time dimension, as the basis for the operation standardization score.

4. The method according to claim 1, characterized in that, In step S4, the dynamic calculation of the comprehensive risk value of the node using a probabilistic time-series interval model specifically includes: S41: Based on the type of hazard triggered, initialize the risk source parameters, including the risk source intensity, diffusion characteristic parameters, and the dominant meteorological parameters of the virtual environment; S42: The process of risk propagation from the source point to the navigation network nodes is modeled as a complex risk evolution activity consisting of multiple risk intensity interval sequences; a risk state atomic activity is defined for each node, and latent variables are introduced to describe the clustering of nodes with similar risk evolution patterns; S43: Based on the allocation of generated node latent variables, nodes under the same latent variable share the same probability distribution of risk state occurrence; according to the allocation of node latent variables, the atomic activities of risk state of each node at the initial time are sampled; S44: Define the set of temporal interval relationships between atomic activities of different node risk states; determine the constrained interval relationships of risk state evolution between nodes based on the physical diffusion law of risk and the spatial layout of obstacles; S45: Under the aforementioned constrained interval relationship, based on the conditional probability distribution corresponding to the combination of atomic activities of risk states between node pairs, sample and generate specific interval relationships for the evolution of risk states between nodes; S46: Combining the virtual personnel density near the node estimated in real time in step S3, the personnel exposure correction is performed on the sampled node risk status; based on the semantic mapping of the risk status atomic activities and the personnel exposure correction results, the dynamic comprehensive risk value of each node at the current moment is calculated. S47: Proceed with a fixed time step, repeat steps S43 to S46, update the latent variable allocation, risk status and interval relationships of all nodes, and realize the time-series evolution simulation of the dynamic comprehensive risk field.

5. The method according to claim 1, characterized in that, In step S5, the dynamic calculation of the passage cost weight for each edge specifically includes: S51: For the edge e connecting node u and node v, obtain its length. ; S52: Obtain the dynamic comprehensive risk values ​​of nodes u and v at both ends of edge e at the current time. and Calculate the reference risk intensity of the edge The average of the risk values ​​of the two nodes; S53: Based on reference risk intensity The estimated walking speed of a person on edge e is calculated using a risk-velocity reduction function. The risk-velocity reduction function simulates the panic and deceleration behavior caused by increased risk. S54: Calculate the basic travel time cost of edge e For length Compared with the estimated walking speed The ratio; S55: Obtain the predicted population density in the region where edge e is located. ,like If the congestion density exceeds a preset threshold, additional congestion time costs are calculated based on the degree of exceedance. ; S56: Comprehensive Basic Travel Time Cost Reference risk level and the cost of congestion time The final dynamic weight of edge e is calculated using a weighted sum formula. Among them, the priority between rapid withdrawal and risk avoidance is adjusted by using trade-off factors.

6. An environmental hazard investigation system based on deep integration of AI and VR for implementing the method of any one of claims 1 to 5, characterized in that, The system includes: The virtual reality immersive training platform includes: a scene model unit for building, managing, and rendering the high-fidelity virtual reality environment; The hazard interaction unit is used to manage the hazard model library and handle the interaction logic and status changes between users and hazard models. The training management unit is used to configure training courses, exercise scenarios, and assessment models. The immersive interaction unit integrates a head-mounted display, control handle, positioning tracker, and eye-tracking module to provide an immersive experience and capture user interaction data. Artificial intelligence intelligent analysis and decision-making platform, including: The hazard identification unit deploys the first deep learning model and the second deep learning model to analyze user behavior and virtual environment images in real time, identify hazards, and generate user behavior assessment reports. The dynamic risk field calculation unit is used to run the probabilistic time series interval model in emergency response mode, dynamically simulate and update the comprehensive risk value of the navigation road network nodes in the virtual environment, and generate a dynamic comprehensive risk field. The dynamic path planning unit is used to calculate the dynamic weights of the navigation road network edges based on the dynamic comprehensive risk field and real-time personnel density prediction, and to run the shortest path search algorithm to solve for the optimal evacuation path. The emergency strategy unit has a built-in rule engine and case inference engine, which is used to generate personalized emergency strategy instruction sets based on multi-dimensional input information; The network unit is used to enable data communication between the virtual reality immersive training platform and the artificial intelligence intelligent analysis and decision-making platform; The database unit is used to store virtual environment models, hazard model libraries, user behavior data, algorithm parameters, historical cases, and exercise process data.

7. The system according to claim 6, characterized in that, The dynamic risk field calculation unit includes a latent variable allocation module, a risk state sampling module, an interval relationship constraint generation module, and a risk value calculation module. Its inputs are hazard triggering event parameters, environmental obstacle information, and personnel density prediction data output in real time by the hazard identification unit. Its output dynamic comprehensive risk field is directly provided to the dynamic path planning unit for edge weight calculation.

8. The system according to claim 6, characterized in that, The system supports multi-user collaborative training and drill modes; the network unit is responsible for synchronizing the virtual status and behavior data of multiple users; in emergency response mode, the dynamic path planning unit plans globally coordinated evacuation paths for multiple collaborative users, and the emergency strategy unit generates a joint emergency strategy that includes division of labor and cooperation instructions.

9. An electronic device, characterized in that, include: The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, include: The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 5.

Images