Multi-person safety education training method and system for primary and secondary school students based on VR

Abstract

The invention provides a VR-based primary and secondary school student multi-person safety education training method and system, and relates to the technical field of VR education, and the method comprises the steps: obtaining the behavior data of each student in a historical safety education training process, and carrying out the evaluation of the safety literacy capability of each student according to a preset safety literacy evaluation model, the safety literacy level of each student is obtained, a safety education training project recommendation sorting result for each student is obtained, and the students are allocated to corresponding safety education training projects according to the recommendation sorting result; real-time synchronous rendering is carried out on virtual avatars of the multiple students in the target VR system through the head-mounted display device, and cooperative training of the multiple students is realized; and after each student completes the safety education training project, scoring the completion condition of the safety education training task of each student by using a preset training task evaluation model, and generating a corresponding personalized evaluation report.

Description

Technical Field

[0001] The present invention relates to VR education technology, and in particular to a VR-based multi-person safety education training method and system for primary and secondary school students. Background Art

[0002] With the development of society and the advancement of science and technology, safety education has become an important part of primary and secondary education. Traditional safety education methods are mainly based on classroom lectures and demonstrations. Students passively accept knowledge, lack of interactivity and practicality, and it is difficult to achieve good educational results. Virtual Reality (VR) technology, with its immersive, interactive and imaginative characteristics, provides a new way for safety education. Introducing VR technology into primary and secondary school safety education can create realistic dangerous situations, allowing students to learn safety knowledge and emergency skills in a virtual environment, and improve students' safety awareness and self-rescue and mutual rescue capabilities. However, most existing VR safety education systems focus on simulation training of a certain dangerous situation. Therefore, there is an urgent need for a comprehensive training platform that integrates multiple scenarios, personalization and collaboration. Summary of the invention

[0003] The embodiments of the present invention provide a VR-based multi-person safety education and training method and system for primary and secondary school students, which can solve the problems in the prior art.

[0004] According to a first aspect of the embodiments of the present invention,

[0005] A VR-based multi-person safety education and training method for primary and secondary school students is provided, including:

[0006] Obtain the behavioral data of each student during the historical safety education and training process, evaluate the safety literacy ability of each student according to the preset safety literacy evaluation model, obtain the safety literacy ability level of each student, use the safety literacy ability level of each student as the state space, use the preset safety education and training project as the action space, learn the optimal mapping strategy from the state space to the action space, obtain the recommended ranking result of the safety education and training project for each student, and assign the student to the corresponding safety education and training project according to the recommended ranking result;

[0007] In the safety education and training project, a multi-level rendering technology combining particle system and volumetric fog is used to simulate the fire effect, and the head display position data of each student is obtained to adjust the fire effect in real time, and the target VR system is obtained. In the target VR system, the virtual avatars of multiple students are rendered in real time and synchronously through the head display device, and the voice data of each student is transmitted to the corresponding virtual avatar in real time, so as to realize the collaborative training of multiple students;

[0008] The collaborative training results of multiple students are synchronized in real time to the teacher monitoring interface, supporting teachers to select and take over specific student perspectives in real time, provide guidance on handling sudden safety incidents encountered by current students in the safety education and training program, and after each student completes the safety education and training program, use the preset training task evaluation model to score each student's safety education and training task completion status and generate a corresponding personalized evaluation report.

[0009] In an optional embodiment,

[0010] Obtain the behavioral data of each student during the historical safety education and training process, and evaluate the safety literacy ability of each student according to the preset safety literacy evaluation model. The safety literacy ability level of each student is as follows:

[0011] By selecting safety textbooks, case collections, laws and regulations as knowledge sources, natural language processing technology is used to extract safety concept entities and their corresponding semantic relationships from the knowledge sources, and combined with domain knowledge to form a safety knowledge graph containing safety concept nodes and relationship edges;

[0012] The knowledge graph embedding technology is used to learn the representation of the constructed security knowledge graph, and the triple loss function is optimized through the translation model to obtain the low-dimensional vector representation of the security knowledge graph;

[0013] Obtain the behavioral data of each student during the historical safety education and training process and perform feature extraction to obtain key behavioral features that reflect the student's safety knowledge and skill level. Fuse the obtained key behavioral features with the low-dimensional vector representation of the safety knowledge graph to obtain semantically enhanced student behavioral features.

[0014] The constructed safety knowledge graph is used as the teacher model, and the probability distribution difference between the safety literacy assessment model and the safety knowledge graph on the student behavior characteristics is calculated. At the same time, a regularization term based on the safety knowledge graph is introduced into the objective function of the safety literacy assessment model. The multi-task collaborative training framework is used to construct a joint optimization loss function of the safety knowledge graph and the safety literacy assessment model. Through alternating iterative training, the knowledge graph-enhanced safety literacy assessment model is obtained.

[0015] The safety literacy assessment model enhanced by the trained knowledge graph is used to predict the safety literacy ability of each student. Combined with the semantic associations between safety concept nodes in the safety knowledge graph, the prediction results are analyzed through graph reasoning methods to finally obtain the safety literacy ability level of each student.

[0016] In an optional embodiment,

[0017] The calculation formula for the joint optimization loss function of building the security knowledge graph and the security literacy assessment model is as follows:

[0018]

[0019] Among them, L represents the joint optimization loss function, L1 represents the mean square error loss function of the safety literacy assessment model, λ1 represents the weight coefficient of the probability distribution difference, N represents the number of students, C represents the safety literacy ability level set, and p c (·) represents the probability of predicting the safety literacy level to be c in the safety literacy assessment model, x i represents the behavior characteristics of the i-th student, P C (·) represents the probability of predicting the safety literacy level c on the safety knowledge graph, λ2 represents the weight coefficient of the regularization term of the safety knowledge graph, E represents the edge set of the knowledge graph, and W u represents the parameter matrix corresponding to concept u in the safety literacy assessment model, h u represents the embedding vector of concept u in the knowledge graph, W u′ represents the parameter matrix corresponding to concept u′ in the safety literacy assessment model, h u′ represents the embedding vector of concept u′ in the knowledge graph, ||·||2 represents the L2 norm, λ3 represents the weight coefficient of the triple loss function, L KG represents the triplet loss function.

[0020] In an optional embodiment,

[0021] In the safety education and training project, a multi-level rendering technology combining particle system and volumetric fog is used to simulate the fire effect, and the head display position data of each student is obtained to adjust the fire effect in real time, and the target VR system is obtained. In the target VR system, the virtual avatars of multiple students are rendered in real time and synchronously through the head display device, and the voice data of each student is transmitted to the corresponding virtual avatar in real time, so as to realize the collaborative training of multiple students, including:

[0022] In the safety education and training project, the particle system is used to simulate the flames and smoke in the fire in real time. The initial parameters of the particle system are determined according to the location of the fire source, the intensity of the fire and the properties of the burning material. The movement trajectory of the particle system is controlled by comprehensively considering the buoyancy, wind force and turbulence factors. The property parameters of the particle system are dynamically adjusted according to the combustion stage of the particles to obtain the optimized particle system.

[0023] The optimized particle system is used to simulate the fire effect. The effect of volumetric fog on illumination is calculated by solving the light transmission equation. The absorption, scattering and emission effects of volumetric fog on light are comprehensively considered. The propagation and attenuation of light in volumetric fog are simulated to achieve volumetric fog rendering. The volumetric fog rendering is combined with the fire effect simulation results to obtain a fire simulation screen containing flames and smoke.

[0024] Based on the fire simulation screen, the head movement of the students is tracked through the head display device, the position and orientation data of the head display device is mapped to the virtual scene, and the corresponding fire event is generated according to the position data of the head display device. Based on the fire event and the behavioral data of the students in the historical safety education and training process, the triggering conditions of the fire event and the parameters of the fire spread simulation are dynamically adjusted to obtain the target VR system;

[0025] In the target VR system, the head display position and orientation data of multiple students are synchronized through network communication, the game engine is used to update the position, posture and movement of each student's virtual avatar, and the student's facial expression data is mapped to the virtual avatar's facial model. At the same time, the student's voice data is transmitted through network communication and mapped to the corresponding virtual avatar's lip animation, thereby realizing collaborative training of multiple students.

[0026] In an optional embodiment,

[0027] The head movement of students is tracked through the head display device, and the position and orientation data of the head display device are mapped to the virtual scene. According to the position data of the head display device, the corresponding fire events are generated, including:

[0028] Use the depth camera in the head-mounted display device to collect depth image sequences during the student's head movement, and use a high-precision optical tracking system to record the real-time posture of the head-mounted display device as the true value annotation of the head posture;

[0029] Preprocess the collected depth image sequence, pair the preprocessed depth image sequence with the corresponding head pose ground truth annotation, and construct a training data set and a test data set for head pose estimation;

[0030] A head pose regression model based on a convolutional neural network is designed, which includes multiple convolutional layers and pooling layers. It is used to extract features from the input depth image sequence to obtain depth image features, and then map the extracted depth image features into parameters representing the head pose through a fully connected layer. At the same time, a loss function is constructed according to the difference between the predicted value and the true value annotation of the head pose parameters, and the head pose regression model is trained by minimizing the loss function.

[0031] The trained head pose regression model is used to predict the real-time acquired depth image sequence to obtain the three-dimensional position coordinates and orientation angle parameters of each student;

[0032] The spatial position and viewing direction of the virtual scene camera are dynamically adjusted according to the obtained three-dimensional position coordinates and orientation angle parameters of each student, so that the viewing angle of the virtual scene camera is consistent with the actual posture of the corresponding student, and the corresponding fire event is rendered and generated using the adjusted virtual scene camera.

[0033] In an optional embodiment,

[0034] Based on the fire events and the behavioral data of students in the historical safety education and training process, the triggering conditions of the fire events and the parameters of the fire spread simulation are dynamically adjusted, and the target VR system includes:

[0035] Obtain the historical behavior data of students during historical safety education and training, and record the fire state parameters and environmental parameters in the training scene synchronized with the historical behavior data;

[0036] Preprocess the acquired historical behavior data, annotate the training sample data according to the students' decision choices and action categories under different fire conditions, and construct a labeled student behavior prediction training data set;

[0037] Construct a recurrent neural network model as a student behavior prediction model, take the student's historical behavior data and current state characteristics as the model input, output the probability distribution of the student behavior category in the current state as the prediction result, and use the student behavior prediction training data set to train the student behavior prediction model through the back propagation algorithm;

[0038] Obtain the student's current status information and perform feature extraction to obtain the current status features, which are then input into the trained recurrent neural network model together with the historical behavior data. The predicted probability distribution of student behavior categories is obtained through forward reasoning calculation. Based on the prediction results, the student behavior category with the highest confidence is taken as the decision choice and behavior category that the student may take in the current state.

[0039] According to the decision choices and behavior categories that students may take, the triggering conditions of the fire event and the parameters of the fire spread simulation are dynamically adjusted to obtain the target VR system.

[0040] In an optional embodiment,

[0041] After each student completes the safety education and training program, the preset training task evaluation model is used to score the completion of each student's safety education and training task, and before generating the corresponding personalized evaluation report, it also includes:

[0042] The front camera on the head-mounted display device collects video images of students performing safety education and training operations in real time, inputs the collected video image frame sequence into a pre-built human key point detection model, extracts the student's body bone feature information sequence, and inputs the extracted body bone feature information sequence into a pre-trained action recognition model to recognize the student's training actions;

[0043] Compare the student training action recognition results with the preset safety training key action reference template, and judge whether the student's training action meets the standard requirements by calculating the similarity between the student's body bone feature sequence and the key action reference template. When the student reaches the key decision node in the virtual scene, obtain the scene environment parameters of the current node, input the decision rule library for reasoning, and obtain the optimal decision at the current node;

[0044] Based on the optimal decision obtained, virtual character dialogue text and scene instruction graphics are generated to guide students to complete the correct decision-making actions within the specified time.

[0045] According to a second aspect of the embodiments of the present invention,

[0046] A VR-based multi-person safety education and training system for primary and secondary school students is provided, including:

[0047] The first unit is used to obtain the behavior data of each student in the historical safety education and training process, evaluate the safety literacy ability of each student according to the preset safety literacy evaluation model, obtain the safety literacy ability level of each student, use the safety literacy ability level of each student as the state space, use the preset safety education and training project as the action space, learn the optimal mapping strategy from the state space to the action space, obtain the recommended ranking result of the safety education and training project for each student, and assign the student to the corresponding safety education and training project according to the recommended ranking result;

[0048] The second unit is used to simulate the fire effect in the safety education and training project by using a multi-level rendering technology combining a particle system and volume fog, and to obtain the head display position data of each student to adjust the fire effect in real time, so as to obtain the target VR system, and to perform real-time synchronous rendering of the virtual avatars of multiple students through the head display device in the target VR system, and to transmit the voice data of each student to the corresponding virtual avatar in real time, so as to realize the collaborative training of multiple students;

[0049] The third unit is used to synchronize the collaborative training results of multiple students to the teacher monitoring interface in real time, support teachers to select and take over specific student perspectives in real time, provide guidance on handling sudden safety incidents encountered by current students in the safety education and training program, and after each student completes the safety education and training program, use the preset training task evaluation model to score each student's safety education and training task completion status and generate a corresponding personalized evaluation report.

[0050] According to a third aspect of the embodiments of the present invention,

[0051] An electronic device is provided, comprising:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] A fourth aspect of the embodiments of the present invention is:

[0056] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0057] In this embodiment, by evaluating the safety literacy ability of each student and using the reinforcement learning algorithm to learn the optimal mapping strategy from state space to action space, personalized safety education and training project recommendations and allocations can be achieved for each student. By performing real-time synchronous rendering of the virtual avatars of multiple students in the VR system and transmitting the voice data of each student to the corresponding virtual avatar in real time, collaborative training of multiple people in a virtual environment can be achieved. Comprehensive analysis of the system training results of multiple students can improve the safety of the training process and strengthen the interaction between teachers and students. By collecting and analyzing various behavioral data of students during the training process, such as safety literacy ability assessment, completion of training projects, etc., the content, methods and recommendation strategies of safety education and training can be continuously optimized, data-driven training optimization can be achieved, and students can be provided with a more efficient, vivid and comprehensive safety education and training experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of a method for multi-person safety education and training for primary and secondary school students based on VR according to an embodiment of the present invention;

[0059] Figure 2 It is a structural schematic diagram of a multi-person safety education and training system for primary and secondary school students based on VR according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0062] Figure 1 FIG. 1 is a flow chart of a method for multi-person safety education and training for primary and secondary school students based on VR according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] S101. Obtain the behavioral data of each student during the historical safety education and training process, evaluate the safety literacy ability of each student according to the preset safety literacy assessment model, obtain the safety literacy ability level of each student, use the safety literacy ability level of each student as the state space, and use the preset safety education and training items as the action space, learn the optimal mapping strategy from the state space to the action space, obtain the recommended ranking results of the safety education and training items for each student, and assign the students to the corresponding safety education and training items according to the recommended ranking results.

[0064] First, we need to obtain the behavioral data of each student during the historical safety education and training process. All the student's operation behaviors in the VR training system, such as training project selection, completion time, score, operation path, etc., will be recorded in the form of a log. At the same time, the system can also collect students' physiological data, such as heart rate, body temperature, skin electricity, etc., through VR equipment to reflect their emotional state.

[0065] The amount of raw behavioral logs and physiological data is large and complex, and they need to be preprocessed by data cleaning, data integration, etc. to remove noise and abnormal data, and convert the data into structured feature vectors. Common data preprocessing methods include data filtering, data transformation, data reduction, etc. Based on massive behavioral data, the system uses a preset safety literacy assessment model to conduct a comprehensive assessment of students' safety cognition, risk awareness, emergency response, and self-rescue and mutual rescue capabilities. The assessment model is trained using machine learning methods, taking student behavioral characteristics as input and outputting students' ability levels in various safety literacy dimensions. Commonly used assessment models include logistic regression, decision trees, support vector machines, neural networks, etc.

[0066] Using classic reinforcement learning algorithms such as Q-Learning, the optimal mapping strategy from state space to action space is learned through continuous exploration and trial and error. Specifically, the system selects an action in each state, obtains an immediate reward (the ability improvement value after the student completes the project), and moves to the next state. Through value iteration, the state-action value function is continuously updated and converges to the optimal strategy. The optimal strategy can guide the selection of the best training project based on the student's current safety literacy.

[0067] Alternatively, the idea of ​​value function approximation can be adopted to use deep neural networks to approximate the Q function, upgrading reinforcement learning to deep reinforcement learning. Mechanisms such as experience replay and dual networks can also be introduced to improve training stability and efficiency.

[0068] In an optional embodiment,

[0069] Obtain the behavioral data of each student during the historical safety education and training process, and evaluate the safety literacy ability of each student according to the preset safety literacy evaluation model. The safety literacy ability level of each student is as follows:

[0070] By selecting safety textbooks, case collections, laws and regulations as knowledge sources, natural language processing technology is used to extract safety concept entities and their corresponding semantic relationships from the knowledge sources, and combined with domain knowledge to form a safety knowledge graph containing safety concept nodes and relationship edges;

[0071] The knowledge graph embedding technology is used to learn the representation of the constructed security knowledge graph, and the triple loss function is optimized through the translation model to obtain the low-dimensional vector representation of the security knowledge graph;

[0072] Obtain the behavioral data of each student during the historical safety education and training process and perform feature extraction to obtain key behavioral features that reflect the student's safety knowledge and skill level. Fuse the obtained key behavioral features with the low-dimensional vector representation of the safety knowledge graph to obtain semantically enhanced student behavioral features.

[0073] The constructed safety knowledge graph is used as the teacher model, and the probability distribution difference between the safety literacy assessment model and the safety knowledge graph on the student behavior characteristics is calculated. At the same time, a regularization term based on the safety knowledge graph is introduced into the objective function of the safety literacy assessment model. The multi-task collaborative training framework is used to construct a joint optimization loss function of the safety knowledge graph and the safety literacy assessment model. Through alternating iterative training, the knowledge graph-enhanced safety literacy assessment model is obtained.

[0074] The safety literacy assessment model enhanced by the trained knowledge graph is used to predict the safety literacy ability of each student. Combined with the semantic associations between safety concept nodes in the safety knowledge graph, the prediction results are analyzed through graph reasoning methods to finally obtain the safety literacy ability level of each student.

[0075] For example, a bottom-up and top-down ontology construction method is first used to extract and organize safety knowledge from structured and unstructured data sources. After determining the knowledge source, natural language processing technology is used to perform semantic understanding and knowledge extraction on the source file. First, the text is preprocessed using tools such as word segmentation and part-of-speech tagging, and the named entity recognition method is used to extract safety-related conceptual entities, such as "flammable materials" and "explosion-proof devices". Then, the semantic relationship between conceptual entities is identified through dependency syntactic analysis and semantic role labeling technology, such as "flammable materials-have-flammability", "explosion-proof devices-used-to-prevent explosions", etc. On this basis, combined with the general knowledge summarized by experts in the safety field, such as "open flames-can cause-flammable materials-combustion", etc., a knowledge graph containing rich safety concept nodes and relationship edges is finally formed.

[0076] After constructing the security knowledge graph, the nodes and relationships in the graph can be mapped to a low-dimensional continuous vector space through knowledge graph embedding technology, so that concepts that are close to each other in the vector space also have strong semantic relevance in the original knowledge graph. Specifically, the current mainstream translation model TransE is used to embed the knowledge graph. Its core idea is to make the sum of the head entity vector and the relationship vector of each triple as close as possible to the tail entity vector, that is, for the triple (head entity, relationship, tail entity), the learning goal is to minimize the bi-norm of (head entity vector + relationship vector - tail entity vector). By optimizing the loss function of the translation model, the semantic vector representation corresponding to each security concept can be obtained.

[0077] Obtain detailed behavioral data of each student during the historical safety education and training process, including time series information in multiple dimensions such as movement trajectory, field of view changes, object interaction, task completion, etc. in the virtual scene. Through feature engineering methods, key features related to safety literacy can be extracted from students' behavioral sequence data. For example, in a fire escape scene, the rationality of escape route selection, identification and avoidance of dangerous areas, and correct use of fire extinguishing equipment can all reflect the students' mastery of fire safety knowledge. The students' behavioral characteristics are integrated with the safety concept vector learned from the knowledge graph to obtain a behavior representation vector containing richer semantic information.

[0078] After obtaining the semantic vector representation of student behavior characteristics and safety knowledge graph, a knowledge graph-enhanced safety literacy assessment model is designed. This model uses the safety knowledge graph as a teacher model and guides the training optimization process of the assessment model through a knowledge distillation mechanism. The core idea is to minimize the probability distribution difference between the assessment model and the knowledge graph model on student behavior characteristics, so that the assessment model can fully learn the safety knowledge in the knowledge graph. At the same time, a regularization term based on the knowledge graph is introduced into the objective function of the assessment model to make the conceptual semantic relationship learned by the assessment model as consistent as possible with the conceptual association in the knowledge graph. During the training process, this study adopted a multi-task collaborative training framework to construct a joint optimization objective of the embedding loss of the safety knowledge graph and the classification loss of the assessment model, and performed end-to-end training through alternating iterations. Finally, an enhanced assessment model that can accurately predict students' safety literacy capabilities and make full use of the prior information of the knowledge graph is obtained.

[0079] By using the trained knowledge graph enhanced evaluation model, the safety literacy ability of each student can be comprehensively predicted. For a new student, the behavioral data generated in the virtual reality safety education and training system is input into the evaluation model, and the student's safety literacy ability level can be directly obtained. At the same time, the prediction results are deeply analyzed in combination with the concept association information contained in the safety knowledge graph. By applying the graph reasoning algorithm on the knowledge graph, it is possible to trace which key concepts or relationships affect the prediction results of the student's ability level. For example, if a student performs poorly in the operation of wearing a gas mask, through the reasoning of the relationship path "gas mask-used-to-filter poison gas-located-in-toxic space", it can be known that the student's grasp of the safety knowledge point of "correct use of protective equipment in toxic environments" is insufficient, and targeted training needs to be strengthened. By integrating the semantic reasoning of the knowledge graph and the ability prediction of the evaluation model, a comprehensive and accurate evaluation of students' safety literacy is finally achieved.

[0080] In an optional embodiment,

[0081] The calculation formula for the joint optimization loss function of building the security knowledge graph and the security literacy assessment model is as follows:

[0082]

[0083] Among them, L represents the joint optimization loss function, L1 represents the mean square error loss function of the safety literacy assessment model, λ1 represents the weight coefficient of the probability distribution difference, N represents the number of students, C represents the safety literacy ability level set, and p c (·) represents the probability of predicting the safety literacy level to be c in the safety literacy assessment model, x i represents the behavior characteristics of the i-th student, P C(·) represents the probability of predicting the safety literacy level c on the safety knowledge graph, λ2 represents the weight coefficient of the regularization term of the safety knowledge graph, E represents the edge set of the knowledge graph, and W u represents the parameter matrix corresponding to concept u in the safety literacy assessment model, h u represents the embedding vector of concept u in the knowledge graph, W u′ represents the parameter matrix corresponding to concept u′ in the safety literacy assessment model, h u′ represents the embedding vector of concept u′ in the knowledge graph, ||·||2 represents the L2 norm, λ3 represents the weight coefficient of the triple loss function, L KG represents the triplet loss function.

[0084] In this embodiment, by constructing a safety knowledge graph containing rich safety concept nodes and relationship edges, a comprehensive knowledge foundation is provided for the subsequent student safety literacy assessment. The semantic association between safety concepts can be better characterized through semantic vector representation, which facilitates the subsequent feature fusion. By integrating the knowledge graph with student behavior data and constructing a knowledge graph-enhanced safety literacy assessment model, it is possible to achieve comprehensive safety knowledge representation, semantically enhanced student behavior feature extraction, and accurate safety literacy ability assessment and analysis, providing strong support for intelligent and personalized safety education and training.

[0085] S102. In the safety education and training project, a multi-level rendering technology combining a particle system and volumetric fog is used to simulate fire effects, and the head display position data of each student is obtained to adjust the fire effect in real time, so as to obtain a target VR system. In the target VR system, the virtual avatars of multiple students are rendered in real time and synchronously through the head display device, and the voice data of each student is transmitted to the corresponding virtual avatar in real time, so as to realize collaborative training of multiple students.

[0086] In an optional embodiment,

[0087] In the safety education and training project, a multi-level rendering technology combining particle system and volumetric fog is used to simulate the fire effect, and the head display position data of each student is obtained to adjust the fire effect in real time, and the target VR system is obtained. In the target VR system, the virtual avatars of multiple students are rendered in real time and synchronously through the head display device, and the voice data of each student is transmitted to the corresponding virtual avatar in real time, so as to realize the collaborative training of multiple students, including:

[0088] In the safety education and training project, the particle system is used to simulate the flames and smoke in the fire in real time. The initial parameters of the particle system are determined according to the location of the fire source, the intensity of the fire and the properties of the burning material. The movement trajectory of the particle system is controlled by comprehensively considering the buoyancy, wind force and turbulence factors. The property parameters of the particle system are dynamically adjusted according to the combustion stage of the particles to obtain the optimized particle system.

[0089] The optimized particle system is used to simulate the fire effect. The effect of volumetric fog on illumination is calculated by solving the light transmission equation. The absorption, scattering and emission effects of volumetric fog on light are comprehensively considered. The propagation and attenuation of light in volumetric fog are simulated to achieve volumetric fog rendering. The volumetric fog rendering is combined with the fire effect simulation results to obtain a fire simulation screen containing flames and smoke.

[0090] Based on the fire simulation screen, the head movement of the students is tracked through the head display device, the position and orientation data of the head display device is mapped to the virtual scene, and the corresponding fire event is generated according to the position data of the head display device. Based on the fire event and the behavioral data of the students in the historical safety education and training process, the triggering conditions of the fire event and the parameters of the fire spread simulation are dynamically adjusted to obtain the target VR system;

[0091] In the target VR system, the head display position and orientation data of multiple students are synchronized through network communication, the game engine is used to update the position, posture and movement of each student's virtual avatar, and the student's facial expression data is mapped to the virtual avatar's facial model. At the same time, the student's voice data is transmitted through network communication and mapped to the corresponding virtual avatar's lip animation, thereby realizing collaborative training of multiple students.

[0092] For example, first, a particle system is used to simulate the flames and smoke in a fire in real time. According to the location of the fire source, the intensity of the fire, and the properties of the burning material, the initial parameters of the particle system, such as the position, speed, size, and color of the particles, are determined. During the movement of the particle system, the influence of factors such as buoyancy, wind force, and turbulence on the trajectory of particle movement is comprehensively considered, and the movement behavior of particles in a fire environment is simulated by introducing random noise and physical constraints. At the same time, according to the combustion stage of the particles, the attribute parameters of the particle system, such as the transparency, brightness, and color changes of the particles, are dynamically adjusted to simulate the dynamic characteristics of flames and smoke changing over time. By optimizing and adjusting the parameters of the particle system, realistic flame and smoke simulation results are obtained.

[0093] After obtaining the fire effect simulated by the particle system, the volume fog rendering technology is further introduced to enhance the realism of the fire scene. Define the properties of the volume fog, such as the absorption coefficient, scattering coefficient, and emission coefficient. These attribute parameters describe the absorption, scattering, and emission behavior of the volume fog on light. According to the position of the light source and the observation point in the scene, the propagation path of the light is determined. This path will determine the way the light interacts with the volume fog. For each sampling point, the light transport equation is used for integration. The light transport equation describes the propagation and attenuation process of light in the volume fog. The integration process takes into account the absorption, scattering, and emission effects along the propagation path. According to the absorption coefficient, the absorption effect along the light propagation path is calculated. The absorption effect causes the energy of the light to gradually weaken. According to the scattering coefficient, the scattering effect along the light propagation path is calculated. The scattering effect causes the direction of the light to change randomly and spread in all directions. According to the emission coefficient, the contribution of the light emitted by the volume fog itself is calculated. The emission effect increases the energy of the light. For each sampling point, the energy of the light is accumulated according to the calculation results of the absorption, scattering, and emission effects. During the rendering process, the propagation and attenuation of light in the volumetric fog are simulated, and the interaction between light and volumetric fog is calculated based on the density, absorption coefficient, and scattering coefficient of the volumetric fog to obtain a realistic volumetric fog rendering effect. Combining the volumetric fog rendering with the simulation results of the fire particle system, a fire simulation screen with realistic flames and smoke is obtained.

[0094] In order to realize the interaction between students and fire simulation scenes, the head movement of students is tracked through the head display device, and the position and orientation data of the head display device are mapped to the virtual scene. According to the position data of the head display device, fire events related to the student's position are dynamically generated, such as fire spread, smoke, structural collapse, etc. At the same time, based on the behavioral data of students in the historical safety education and training process, such as students' reaction time, decision-making choices, etc., the triggering conditions of fire events and the parameters of fire spread simulation are dynamically adjusted, so that the fire scene can intelligently adapt to the students' real-time performance and learning level. Through the dynamic generation and adaptive adjustment of fire events, a target VR system related to student interaction is obtained, which improves the pertinence and effectiveness of training.

[0095] In order to support the collaborative training of multiple students, in the target VR system, the position and orientation data of each student's head display are synchronized through network communication, and the position, posture and movement of each student's virtual avatar are updated using the game engine to achieve real-time synchronization and interaction of students in the virtual scene. At the same time, through facial expression recognition technology, the facial expression data of students is captured and mapped to the facial model of the virtual avatar, so that the virtual avatar can reflect the real-time expression changes of students. At the same time, using speech recognition and synthesis technology, the student's voice data is converted into the lip animation of the virtual avatar, so that the virtual avatar can be synchronized with the student's voice, enhancing the student's immersion and interactivity in the virtual environment. Through multi-channel voice transmission and synchronization, real-time voice collaborative training of multiple students in the target VR system is achieved.

[0096] In this embodiment, a realistic and highly interactive VR fire simulation training system is realized by comprehensively utilizing technologies such as particle system, volumetric fog rendering, head motion tracking, facial expression mapping, and voice synchronization. By dynamically generating fire events related to student interaction and supporting multi-person collaborative training, the realism, pertinence, and effectiveness of VR safety education and training are improved. This method provides a new idea and implementation approach for improving the quality and effectiveness of VR safety education and training.

[0097] In an optional embodiment,

[0098] The head movement of students is tracked through the head display device, and the position and orientation data of the head display device are mapped to the virtual scene. According to the position data of the head display device, the corresponding fire events are generated, including:

[0099] Use the depth camera in the head-mounted display device to collect depth image sequences during the student's head movement, and use a high-precision optical tracking system to record the real-time posture of the head-mounted display device as the true value annotation of the head posture;

[0100] Preprocess the collected depth image sequence, pair the preprocessed depth image sequence with the corresponding head pose ground truth annotation, and construct a training data set and a test data set for head pose estimation;

[0101] A head pose regression model based on a convolutional neural network is designed, which includes multiple convolutional layers and pooling layers. It is used to extract features from the input depth image sequence to obtain depth image features, and then map the extracted depth image features into parameters representing the head pose through a fully connected layer. At the same time, a loss function is constructed according to the difference between the predicted value and the true value annotation of the head pose parameters, and the head pose regression model is trained by minimizing the loss function.

[0102] The trained head pose regression model is used to predict the real-time acquired depth image sequence to obtain the three-dimensional position coordinates and orientation angle parameters of each student;

[0103] The spatial position and viewing direction of the virtual scene camera are dynamically adjusted according to the obtained three-dimensional position coordinates and orientation angle parameters of each student, so that the viewing angle of the virtual scene camera is consistent with the actual posture of the corresponding student, and the corresponding fire event is rendered and generated using the adjusted virtual scene camera.

[0104] For example, first, the depth camera built into the head display device is used to collect a sequence of depth images of students during head movement. Since the depth camera can obtain the distance information from the surface of the object in the scene to the camera, the three-dimensional movement trajectory of the head can be accurately perceived by analyzing the changes in continuous depth images. At the same time, a high-precision optical tracking system is also used to accurately measure the real-time posture of the head display device. By arranging reflective ball markers on the surface of the head display device, using multiple infrared cameras to shoot the marker balls from different angles, and combining the triangulation principle to calculate the spatial coordinates and rotation matrix parameters of the head display device, the true value annotation data of the head posture is obtained.

[0105] The Gaussian filter algorithm is used to smooth the depth image and remove isolated noise points. At the same time, the bilateral filter algorithm is used to smooth the image details while retaining the edge of the depth image, thereby improving the signal-to-noise ratio of the image. After preprocessing, the generated depth image sequence can more accurately reflect the movement of the head. By aligning the timestamps of the processed depth image sequence with the synchronously recorded head pose true value annotations, a dataset for training and testing the head pose estimation model can be constructed.

[0106] After constructing the head pose estimation dataset, a pose regression model based on convolutional neural network is designed for end-to-end prediction of head pose. The model takes continuous depth image frames as input, extracts spatial features in the depth image through multiple convolutional layers and pooling layers at the front end, and captures the motion change information of the head in three-dimensional space. Each convolution kernel in the convolutional layer encodes the features of the local depth image area, and scans the entire image by sliding window to obtain a feature map containing rich spatial structure information. The pooling layer reduces the size of the feature map by downsampling operation, which effectively alleviates the risk of overfitting of the model. After completing the front-end feature extraction, the fully connected layer is used to map the extracted depth features into head pose parameters, including the x, y, z coordinate values ​​representing the three-dimensional spatial position of the head and the yaw, pitch, and roll angle values ​​representing the head orientation. In the model training stage, the mean square error between the predicted head pose parameters and the true value annotation is used as the loss function, and the model parameters are optimized using the stochastic gradient descent algorithm to minimize the error of head pose estimation.

[0107] Using the trained head pose estimation model, the head movement of students during virtual reality fire escape training can be tracked in real time. By inputting the depth image sequence collected by the head display device in real time into the pose estimation model, the spatial coordinates and orientation angle of the student's head at the current moment can be predicted. In order to synchronize the fire events in the virtual scene with the changes in the students' perspective, the parameters of the virtual scene camera can be updated in real time according to the results of the head pose estimation. By mapping the head coordinates to the spatial position of the camera and the head orientation angle to the viewing direction of the camera, the perspective of the virtual scene camera can be dynamically kept consistent with the actual perspective of the students. The picture observed by the students in the virtual environment is generated by the rendering of the adjusted virtual scene camera, which includes realistic disaster events such as fire spread and smoke that change synchronously with the students' head movements.

[0108] In this embodiment, accurate, real-time, and marker-free head pose estimation is achieved through deep image sequences and convolutional neural networks, and the realism, immersion, and interactivity of VR safety education and training can be significantly improved by dynamically adjusting the virtual scene camera and personalized fire event rendering. This technical path allows students to gain an immersive fire experience in a virtual environment, improve their learning interest and training effects, and provides strong support for the promotion and application of VR safety education.

[0109] In an optional embodiment,

[0110] Based on the fire events and the behavioral data of students in the historical safety education and training process, the triggering conditions of the fire events and the parameters of the fire spread simulation are dynamically adjusted, and the target VR system includes:

[0111] Obtain the historical behavior data of students during historical safety education and training, and record the fire state parameters and environmental parameters in the training scene synchronized with the historical behavior data;

[0112] Preprocess the acquired historical behavior data, annotate the training sample data according to the students' decision choices and action categories under different fire conditions, and construct a labeled student behavior prediction training data set;

[0113] Construct a recurrent neural network model as a student behavior prediction model, take the student's historical behavior data and current state characteristics as the model input, output the probability distribution of the student behavior category in the current state as the prediction result, and use the student behavior prediction training data set to train the student behavior prediction model through the back propagation algorithm;

[0114] Obtain the student's current status information and perform feature extraction to obtain the current status features, which are then input into the trained recurrent neural network model together with the historical behavior data. The predicted probability distribution of student behavior categories is obtained through forward reasoning calculation. Based on the prediction results, the student behavior category with the highest confidence is taken as the decision choice and behavior category that the student may take in the current state.

[0115] According to the decision choices and behavior categories that students may take, the triggering conditions of the fire event and the parameters of the fire spread simulation are dynamically adjusted to obtain the target VR system.

[0116] Exemplarily, first obtain detailed behavioral data of students during historical safety education and training. By recording students' motion trajectories, object interactions, task completion status and other information in the virtual reality system, the students' action decision-making process in different fire scenarios can be restored. At the same time, fire state parameters and environmental parameters synchronized with students' behavioral data are also collected. Among them, fire state parameters include the location of the fire, the size of the fire, the speed of spread, etc., and environmental parameters include building layout, available escape routes, and fire-fighting facility configuration. By comprehensively considering students' behavioral performance and changes in fire status, we can have a more comprehensive understanding of students' emergency response capabilities when facing fire hazards.

[0117] The acquired historical behavior data is preprocessed. First, the original behavior sequence data is processed by data cleaning and data normalization methods, outliers and missing values ​​are removed, and the values ​​of different feature dimensions are unified into a similar numerical range, which improves the quality and availability of the data. Then, the behavior data of each time step is semantically annotated according to the students' behavior performance under different fire conditions. By pre-defining a behavior category system, it includes various action decisions that students may take, such as going to the safe exit, avoiding the fire source, using a fire extinguisher, calling the police for help, etc. By analyzing the original behaviors such as the students' position movement and the use of items, the actual behavior category of the students at each time step can be determined. Finally, a student behavior dataset with behavior labels is formed. Each sample consists of a historical behavior sequence and a corresponding behavior category label, which is used to train the student behavior prediction model.

[0118] A student behavior prediction model is constructed using a recurrent neural network. By mining the temporal features in the student's historical behavior sequence, the student's behavior in a new fire scenario can be predicted. The input of the model consists of two parts, one is the student's historical behavior sequence data, and the other is the current fire state characteristics of the student. Among them, the historical behavior sequence is segmented by a sliding time window, and each time window contains student behavior data of a certain time step length. The current state characteristics are obtained by feature engineering extraction of fire parameters such as the fire area and smoke concentration and environmental parameters such as building layout and safety exit location. Inside the model, this study uses a multi-layer long short-term memory network LSTM to learn the features of the input time series data. The gating mechanism in LSTM can effectively alleviate the gradient disappearance problem, so that long-distance time dependencies can be modeled. By gradually inputting the historical behavior data into the hidden state of LSTM and splicing the current state features with the hidden state of the last time step, a feature representation that integrates historical information and the current state can be obtained. The output layer of the model uses the Softmax function to multi-classify the student's behavior categories and predict the possibility probability of each behavior category. When training the model, the cross entropy loss function of the predicted behavior category is used as the optimization target, and the back propagation algorithm is used to update the model parameters to continuously improve the accuracy of behavior prediction.

[0119] When students are conducting fire escape training in a virtual reality system, the current fire state information of the students is first obtained, including the students' spatial position, the development of the fire event, etc., and the feature vector representing the current state is obtained through predefined feature extraction rules. At the same time, the student's behavior sequence data in the recent period is obtained as the historical information of the student's behavior. The current state features and historical behavior sequences are input into the trained student behavior prediction model, and the model can obtain the predicted probability distribution of the student's current possible behavior category through forward reasoning calculation. According to the behavior category with the highest confidence in the prediction result, the most likely action decision for the student in the current fire state can be determined. For example, if the prediction result shows that the student has a high probability of using a fire extinguisher, it means that the student may have discovered the source of the fire and is ready to take initial fire-fighting measures.

[0120] According to the results of student behavior prediction, the generation logic of the virtual reality fire scene was adaptively adjusted. On the one hand, by dynamically modifying the triggering conditions of the fire event, the process of fire development is matched with the rhythm of students' actions. For example, if it is predicted that the student may go to the safety exit to escape next, the speed of fire spread near the safety exit is appropriately accelerated to hinder the student's escape behavior. On the other hand, this study also optimizes the simulation parameters of fire spread based on the results of student behavior prediction, so that the direction of fire development and the scope of influence are consistent with the possible decision path of students. For example, if it is predicted that the student may go to a certain area to extinguish the fire next, the density of combustibles and the intensity of combustion in the area are reduced to provide more possibilities for students' fire-fighting actions.

[0121] In this embodiment, personalized and intelligent VR safety education and training can be achieved by constructing a student behavior prediction model and dynamically adjusting the fire event trigger conditions and spread simulation parameters based on the prediction results. This adaptive training method can generate fire scenes with moderate difficulty and challenge according to the characteristics and performance of each student, thereby improving student participation and training effects. At the same time, dynamic fire simulation can also improve the authenticity and diversity of the scene, providing students with a more vivid and realistic training experience. This technical path provides new ideas and directions for the intelligent and personalized development of VR safety education and training.

[0122] S103. Synchronize the collaborative training results of multiple students to the teacher monitoring interface in real time, support teachers to select and take over specific student perspectives in real time, provide guidance on handling sudden safety incidents encountered by current students in the safety education and training program, and after each student completes the safety education and training program, use the preset training task evaluation model to score each student's safety education and training task completion status, and generate a corresponding personalized evaluation report.

[0123] In an optional embodiment,

[0124] After each student completes the safety education and training program, the preset training task evaluation model is used to score the completion of each student's safety education and training task, and before generating the corresponding personalized evaluation report, it also includes:

[0125] The front camera on the head-mounted display device collects video images of students performing safety education and training operations in real time, inputs the collected video image frame sequence into a pre-built human key point detection model, extracts the student's body bone feature information sequence, and inputs the extracted body bone feature information sequence into a pre-trained action recognition model to recognize the student's training actions;

[0126] Compare the student training action recognition results with the preset safety training key action reference template, and judge whether the student's training action meets the standard requirements by calculating the similarity between the student's body bone feature sequence and the key action reference template. When the student reaches the key decision node in the virtual scene, obtain the scene environment parameters of the current node, input the decision rule library for reasoning, and obtain the optimal decision at the current node;

[0127] Based on the optimal decision obtained, virtual character dialogue text and scene instruction graphics are generated to guide students to complete the correct decision-making actions within the specified time.

[0128] For example, first, the front camera of the head display device is used to collect real-time video images of students performing firefighting training operations. Since the front camera maintains a fixed relative position relationship with the student's head posture, the captured video images can accurately reflect the student's perspective changes and body movements in the virtual training environment. By reading the video image data frame by frame, a continuous image frame sequence reflecting the student's training process can be obtained.

[0129] Then, the collected video image frame sequence is sequentially input into the pre-built human key point detection model for processing. The model uses a convolutional neural network to locate and segment the human body area in the image, and obtains the three-dimensional spatial coordinates of the human skeletal structure through a regression analysis algorithm. By detecting the changes in the student's body posture frame by frame, a human skeletal motion sequence synchronized with the video can be obtained. The sequence is expressed in the form of three-dimensional coordinates of human joints, and contains the motion information of major body parts such as the head, torso, and limbs.

[0130] After obtaining the student's body skeletal movement sequence, it is input into the pre-trained action recognition model to classify the student's specific training movements. The model uses a time series classification algorithm based on the long short-term memory network LSTM, which can model the time domain dependency in the skeletal movement sequence, thereby accurately identifying key actions made by students, such as using a fire extinguisher and opening an escape door. By counting the frequency and duration of various action categories within a period of time, the student's training behavior can be quantitatively described.

[0131] In order to determine whether the students' training actions meet the fire safety operation specifications, the similarity between the students' action category sequence and the pre-recorded safety training key action reference template is compared. By using the dynamic time warping algorithm to calculate the optimal matching distance between the two time series data, the degree of deviation between the students' actual operations and the standard actions can be obtained. If the similarity is lower than a certain threshold, it means that the students have made mistakes or irregular operations, which need to be reminded and corrected in time.

[0132] When students move to key decision nodes in the virtual training scene, such as escape route selection and initial fire fighting, they will further analyze the environmental parameters of the current scene, including fire size, smoke density, building structure, etc., and input them into the pre-built decision rule knowledge base for reasoning. The knowledge base presets the optimal decision choices under various fire scenarios. By matching the student's scene with the conditions in the knowledge base, the standard response measures for the current situation can be obtained.

[0133] Based on the results of the decision knowledge base reasoning, the dialogue text and scene indication graphics of the virtual instructor role are automatically generated to guide students to make correct safety decisions within the specified time. For example, if the reasoning results of the knowledge base show that the fire and smoke density in the current area are high and it is not suitable to stay and extinguish the fire, the virtual instructor will tell the students to evacuate as soon as possible through voice prompts, and use arrows in the scene to indicate the safe evacuation route. Students need to complete the corresponding operations according to the system prompts within the specified time to successfully pass the decision node.

[0134] When all students have completed the entire virtual reality firefighting training process, the system will call the preset training evaluation model to comprehensively score each student's training data. The evaluation model scores students' performance from multiple dimensions such as scene adaptability, key action standardization, decision-making correctness, training time, etc., and weighted averages the scores of each dimension to obtain the student's overall training score. Students with scores above 80 are considered to have passed the assessment, and students with scores below 60 need to re-participate in training.

[0135] Finally, the system automatically generates a personalized training evaluation report for each student. The report lists in detail the students' scores at each key decision-making node and the normative analysis of key action operations, and puts forward improvement suggestions for the problems exposed by students during training, such as strengthening the understanding of escape route selection and improving the correct use of fire extinguishers.

[0136] In this embodiment, through real-time monitoring and perspective takeover, automatic action evaluation, intelligent decision-making guidance, personalized evaluation reports and other technical means, the standardization, standardization, intelligence and personalization of VR safety education and training can be comprehensively improved. It can ensure that each student receives high-quality and standardized training, and obtains targeted guidance and feedback based on personal performance, thereby maximizing the advantages and potential of VR technology in the field of safety education and improving students' safety awareness and emergency response capabilities.

[0137] Figure 2 FIG. 1 is a schematic diagram of the structure of a multi-person safety education and training system for primary and secondary school students based on VR according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0138] The first unit is used to obtain the behavior data of each student in the historical safety education and training process, evaluate the safety literacy ability of each student according to the preset safety literacy evaluation model, obtain the safety literacy ability level of each student, use the safety literacy ability level of each student as the state space, use the preset safety education and training project as the action space, learn the optimal mapping strategy from the state space to the action space, obtain the recommended ranking result of the safety education and training project for each student, and assign the student to the corresponding safety education and training project according to the recommended ranking result;

[0139] The second unit is used to simulate the fire effect in the safety education and training project by using a multi-level rendering technology combining a particle system and volume fog, and to obtain the head display position data of each student to adjust the fire effect in real time, so as to obtain the target VR system, and to perform real-time synchronous rendering of the virtual avatars of multiple students through the head display device in the target VR system, and to transmit the voice data of each student to the corresponding virtual avatar in real time, so as to realize the collaborative training of multiple students;

[0140] The third unit is used to synchronize the collaborative training results of multiple students to the teacher monitoring interface in real time, support teachers to select and take over specific student perspectives in real time, provide guidance on handling sudden safety incidents encountered by current students in the safety education and training program, and after each student completes the safety education and training program, use the preset training task evaluation model to score each student's safety education and training task completion status and generate a corresponding personalized evaluation report.

[0141] According to a third aspect of the embodiments of the present invention,

[0142] An electronic device is provided, comprising:

[0143] processor;

[0144] a memory for storing processor-executable instructions;

[0145] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0146] A fourth aspect of the embodiments of the present invention is:

[0147] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0148] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A VR-based multi-person safety education and training method for primary and secondary school students, characterized in that: include: Obtain the behavioral data of each student during the historical safety education and training process, evaluate the safety literacy ability of each student according to the preset safety literacy evaluation model, obtain the safety literacy ability level of each student, use the safety literacy ability level of each student as the state space, use the preset safety education and training project as the action space, learn the optimal mapping strategy from the state space to the action space, obtain the recommended ranking result of the safety education and training project for each student, and assign the student to the corresponding safety education and training project according to the recommended ranking result; In the safety education and training project, a multi-level rendering technology combining particle system and volumetric fog is used to simulate the fire effect, and the head display position data of each student is obtained to adjust the fire effect in real time, and the target VR system is obtained. In the target VR system, the virtual avatars of multiple students are rendered in real time and synchronously through the head display device, and the voice data of each student is transmitted to the corresponding virtual avatar in real time, so as to realize the collaborative training of multiple students; The collaborative training results of multiple students are synchronized in real time to the teacher monitoring interface, supporting teachers to select and take over specific student perspectives in real time, provide guidance on handling sudden safety incidents encountered by current students in the safety education and training program, and after each student completes the safety education and training program, use the preset training task evaluation model to score each student's safety education and training task completion status and generate a corresponding personalized evaluation report.

2. The method according to claim 1, characterized in that Obtain the behavioral data of each student during the historical safety education and training process, and evaluate the safety literacy ability of each student according to the preset safety literacy evaluation model. The safety literacy ability level of each student is as follows: By selecting safety textbooks, case collections, laws and regulations as knowledge sources, natural language processing technology is used to extract safety concept entities and their corresponding semantic relationships from the knowledge sources, and combined with domain knowledge to form a safety knowledge graph containing safety concept nodes and relationship edges; The knowledge graph embedding technology is used to learn the representation of the constructed security knowledge graph. The triple loss function is optimized by the translation model to obtain the low-dimensional vector representation of the security knowledge graph: Obtain the behavioral data of each student during the historical safety education and training process and perform feature extraction to obtain key behavioral features that reflect the student's safety knowledge and skill level. Fuse the obtained key behavioral features with the low-dimensional vector representation of the safety knowledge graph to obtain semantically enhanced student behavioral features. The constructed safety knowledge graph is used as the teacher model, and the probability distribution difference between the safety literacy assessment model and the safety knowledge graph on the student behavior characteristics is calculated. At the same time, a regularization term based on the safety knowledge graph is introduced into the objective function of the safety literacy assessment model. The multi-task collaborative training framework is used to construct a joint optimization loss function of the safety knowledge graph and the safety literacy assessment model. Through alternating iterative training, the knowledge graph-enhanced safety literacy assessment model is obtained. The safety literacy assessment model enhanced by the trained knowledge graph is used to predict the safety literacy ability of each student. Combined with the semantic associations between safety concept nodes in the safety knowledge graph, the prediction results are analyzed through graph reasoning methods to finally obtain the safety literacy ability level of each student.

3. The method according to claim 2, characterized in that The calculation formula for the joint optimization loss function of building the security knowledge graph and the security literacy assessment model is as follows: Among them, L represents the joint optimization loss function, L1 represents the mean square error loss function of the safety literacy assessment model, λ1 represents the weight coefficient of the probability distribution difference, N represents the number of students, C represents the safety literacy ability level set, and p c (·) represents the probability of predicting the safety literacy level c in the safety literacy assessment model, x i represents the behavior characteristics of the i-th student, P C (·) represents the probability of predicting the safety literacy level c on the safety knowledge graph, λ2 represents the weight coefficient of the regularization term of the safety knowledge graph, E represents the edge set of the knowledge graph, and W u represents the parameter matrix corresponding to concept u in the safety literacy assessment model, h u represents the embedding vector of concept u in the knowledge graph, W u′ represents the parameter matrix corresponding to concept u′ in the safety literacy assessment model, h u′ represents the embedding vector of concept u′ in the knowledge graph, ||·||2 represents the L2 norm, λ3 represents the weight coefficient of the triple loss function, L KG represents the triplet loss function.

4. The method according to claim 1, characterized in that: In the safety education and training project, a multi-level rendering technology combining particle system and volumetric fog is used to simulate the fire effect, and the head display position data of each student is obtained to adjust the fire effect in real time, and the target VR system is obtained. In the target VR system, the virtual avatars of multiple students are rendered in real time and synchronously through the head display device, and the voice data of each student is transmitted to the corresponding virtual avatar in real time, so as to realize the collaborative training of multiple students, including: In the safety education and training project, the particle system is used to simulate the flames and smoke in the fire in real time. The initial parameters of the particle system are determined according to the location of the fire source, the intensity of the fire and the properties of the burning material. The movement trajectory of the particle system is controlled by comprehensively considering the buoyancy, wind force and turbulence factors. The property parameters of the particle system are dynamically adjusted according to the combustion stage of the particles to obtain the optimized particle system. The optimized particle system is used to simulate the fire effect. The effect of volumetric fog on illumination is calculated by solving the light transmission equation. The absorption, scattering and emission effects of volumetric fog on light are comprehensively considered. The propagation and attenuation of light in volumetric fog are simulated to achieve volumetric fog rendering. The volumetric fog rendering is combined with the fire effect simulation results to obtain a fire simulation screen containing flames and smoke. Based on the fire simulation screen, the head movement of the students is tracked through the head display device, the position and orientation data of the head display device is mapped to the virtual scene, and the corresponding fire event is generated according to the position data of the head display device. Based on the fire event and the behavioral data of the students in the historical safety education and training process, the triggering conditions of the fire event and the parameters of the fire spread simulation are dynamically adjusted to obtain the target VR system; In the target VR system, the head display position and orientation data of multiple students are synchronized through network communication, the game engine is used to update the position, posture and movement of each student's virtual avatar, and the student's facial expression data is mapped to the virtual avatar's facial model. At the same time, the student's voice data is transmitted through network communication and mapped to the corresponding virtual avatar's lip animation, thereby realizing collaborative training of multiple students.

5. The method according to claim 4, characterized in that The head movement of students is tracked through the head display device, and the position and orientation data of the head display device are mapped to the virtual scene. According to the position data of the head display device, the corresponding fire events are generated, including: Use the depth camera in the head-mounted display device to collect depth image sequences during the student's head movement, and use a high-precision optical tracking system to record the real-time posture of the head-mounted display device as the true value annotation of the head posture; Preprocess the collected depth image sequence, pair the preprocessed depth image sequence with the corresponding head pose ground truth annotation, and construct a training data set and a test data set for head pose estimation; A head pose regression model based on a convolutional neural network is designed, which includes multiple convolutional layers and pooling layers. It is used to extract features from the input depth image sequence to obtain depth image features, and then map the extracted depth image features into parameters representing the head pose through a fully connected layer. At the same time, a loss function is constructed according to the difference between the predicted value and the true value annotation of the head pose parameters, and the head pose regression model is trained by minimizing the loss function. The trained head pose regression model is used to predict the real-time acquired depth image sequence to obtain the three-dimensional position coordinates and orientation angle parameters of each student; The spatial position and viewing direction of the virtual scene camera are dynamically adjusted according to the obtained three-dimensional position coordinates and orientation angle parameters of each student, so that the viewing angle of the virtual scene camera is consistent with the actual posture of the corresponding student, and the corresponding fire event is rendered and generated using the adjusted virtual scene camera.

6. The method according to claim 4, characterized in that Based on the fire events and the behavioral data of students in the historical safety education and training process, the triggering conditions of the fire events and the parameters of the fire spread simulation are dynamically adjusted, and the target VR system includes: Obtain the historical behavior data of students during historical safety education and training, and record the fire state parameters and environmental parameters in the training scene synchronized with the historical behavior data; The acquired historical behavior data is preprocessed, and the training sample data is labeled according to the students' decision choices and action categories under different fire conditions to construct a labeled student behavior prediction training dataset: Construct a recurrent neural network model as a student behavior prediction model, take the student's historical behavior data and current state characteristics as the model input, output the probability distribution of the student behavior category in the current state as the prediction result, and use the student behavior prediction training data set to train the student behavior prediction model through the back propagation algorithm: Obtain the student's current status information and perform feature extraction to obtain the current status features, which are then input into the trained recurrent neural network model together with the historical behavior data. The predicted probability distribution of student behavior categories is obtained through forward reasoning calculation. Based on the prediction results, the student behavior category with the highest confidence is taken as the decision choice and behavior category that the student may take in the current state. According to the decision choices and behavior categories that students may take, the triggering conditions of the fire event and the parameters of the fire spread simulation are dynamically adjusted to obtain the target VR system.

7. The method according to claim 1, characterized in that After each student completes the safety education and training program, the preset training task evaluation model is used to score the completion of each student's safety education and training task, and before generating the corresponding personalized evaluation report, it also includes: The front camera on the head-mounted display device collects video images of students performing safety education and training operations in real time, inputs the collected video image frame sequence into a pre-built human key point detection model, extracts the student's body bone feature information sequence, and inputs the extracted body bone feature information sequence into a pre-trained action recognition model to recognize the student's training actions; Compare the student training action recognition results with the preset safety training key action reference template, and judge whether the student's training action meets the standard requirements by calculating the similarity between the student's body bone feature sequence and the key action reference template. When the student reaches the key decision node in the virtual scene, obtain the scene environment parameters of the current node, input the decision rule library for reasoning, and obtain the optimal decision at the current node; Based on the optimal decision obtained, virtual character dialogue text and scene instruction graphics are generated to guide students to complete the correct decision-making actions within the specified time.

8. A VR-based multi-person safety education and training system for primary and secondary school students, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain the behavior data of each student in the historical safety education and training process, evaluate the safety literacy ability of each student according to the preset safety literacy evaluation model, obtain the safety literacy ability level of each student, use the safety literacy ability level of each student as the state space, use the preset safety education and training project as the action space, learn the optimal mapping strategy from the state space to the action space, obtain the recommended ranking result of the safety education and training project for each student, and assign the student to the corresponding safety education and training project according to the recommended ranking result; The second unit is used to simulate the fire effect in the safety education and training project by using a multi-level rendering technology combining a particle system and volume fog, and to obtain the head display position data of each student to adjust the fire effect in real time, so as to obtain the target VR system, and to perform real-time synchronous rendering of the virtual avatars of multiple students through the head display device in the target VR system, and to transmit the voice data of each student to the corresponding virtual avatar in real time, so as to realize the collaborative training of multiple students; The third unit is used to synchronize the collaborative training results of multiple students to the teacher monitoring interface in real time, support teachers to select and take over specific student perspectives in real time, provide guidance on handling sudden safety incidents encountered by current students in the safety education and training program, and after each student completes the safety education and training program, use the preset training task evaluation model to score each student's safety education and training task completion status and generate a corresponding personalized evaluation report.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images