Full-scene emergency escape training system based on artificial intelligence

Abstract

The invention relates to the technical field of emergency safety training, and discloses a full-scene emergency escape training system based on artificial intelligence, and the system comprises a multi-source sensor data collection module which is used for obtaining industrial field environment parameters and trainee space distribution information in real time, and outputting heterogeneous reference data; the dynamic danger assessment and scene generation module is used for receiving the heterogeneous reference data and constructing a danger evolution model to dynamically generate a virtual escape situation; and the individual behavior prediction and interactive training module is used for identifying the action characteristics of the trainee in the virtual escape situation and predicting the path selection trend of the trainee. The dynamic risk evaluation and scene generation technology is adopted, the effect of real-time linkage of risk evolution and virtual situations is achieved, precise simulation of secondary disaster spreading characteristics is achieved, and the defect that an existing static training mode cannot restore the transient evolution rule of risk factors of a complex industrial site is overcome.

Description

Technical Field

[0001] This invention relates to the field of emergency safety training technology, specifically to an artificial intelligence-based full-scenario emergency escape training system. Background Technology

[0002] With the accelerating pace of industrial modernization, high-risk production sites such as chemical plants, power stations, and oil storage and transportation bases are expanding in scale and becoming increasingly complex in their internal spatial structures. Potential hazards are exhibiting diversified and interconnected characteristics. In the event of sudden accidents such as fires, toxic gas leaks, or explosions, environmental parameters at industrial sites undergo drastic, non-linear evolution over time, generating numerous secondary disasters. Emergency escape training, as a means to improve the survival rate and accident response capabilities of employees, is directly related to production safety in terms of its realism and effectiveness. Traditional emergency drills mainly rely on on-site exercises or digital demonstrations based on fixed scripts, playing a certain role in popularizing basic procedures, but their limitations in dealing with extremely complex environments are becoming increasingly apparent.

[0003] Existing training methods can only achieve a static display of a single hazard factor, but cannot simulate the complex interaction between secondary disasters, environmental evolution and personnel escape behavior in industrial sites. This results in a broken logical chain in the simulated scenario, making it difficult to reproduce the dynamic characteristics of rapidly changing hazards in real accidents, and failing to meet the requirements of in-depth practical training in complex industrial environments.

[0004] Existing simulation tools focus on process demonstration and lack the ability to capture and deeply analyze trainees' path selection and behavioral decisions when facing different levels of danger in real time. Because they cannot provide immediate corrective feedback for individual decision-making biases, trainees are less likely to develop scientific escape strategies during simulation training, thus limiting their ability to improve decision-making in real emergencies.

[0005] Existing technologies typically employ fixed modeling for specific scenarios, lacking mechanisms for automated construction and dynamic adjustment across all scenarios using artificial intelligence. This fixed model prevents training systems from flexibly switching based on the actual layout and risk distribution in different high-risk environments such as chemical plants or power stations, resulting in a narrow scope of application and high scenario construction costs.

[0006] To address the aforementioned issues, this invention proposes an artificial intelligence-based full-scenario emergency escape training system. This system utilizes multi-source data fusion and reinforcement learning to simulate hazard evolution, correct individual behavior, and adaptively evolve scenarios, thereby enhancing trainees' emergency decision-making capabilities. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based full-scenario emergency escape training system to solve the problems mentioned in the background section.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an artificial intelligence-based full-scenario emergency escape training system, comprising:

[0009] The multi-source sensor data acquisition module is used to acquire industrial field environmental parameters and trainee spatial distribution information in real time, and output heterogeneous reference data.

[0010] The dynamic hazard assessment and scenario generation module is used to receive heterogeneous baseline data, construct a hazard evolution model, and dynamically generate virtual escape scenarios.

[0011] The individual behavior prediction and interactive training module is used to identify the trainees' action characteristics in virtual escape scenarios and predict the trainees' path selection trends.

[0012] The real-time feedback and automatic evaluation module is used to receive the trainee's path selection trend, perform quantitative analysis on the trainee's escape decision and output a correction guidance report; the correction guidance report includes decision response time, path deviation rate, accuracy of avoidance actions and physiological and psychological stability evaluation indicators.

[0013] The training optimization and continuous learning module is used to mine historical training data involved in the correction guidance report and autonomously update the escape scenario configuration in the dynamic hazard assessment and scenario generation module.

[0014] Preferably, the multi-source sensor data acquisition module integrates an infrared thermal imaging sensor, a gas concentration monitoring sensor, a pressure sensor, and a high-definition visual capture component.

[0015] Preferably, the dynamic hazard assessment and scenario generation module has a built-in neural network evolution model, which calculates the spread rate and concentration distribution vector of the hazard factor in three-dimensional space by performing convolution operations on environmental parameters and preset accident logic.

[0016] Preferably, the individual behavior prediction and interaction training module uses a recurrent neural network to perform time-series modeling of the trainee's skeletal joint data and movement trajectory, calculates the logical distance between the trainee's current behavior state and the preset danger boundary, and triggers dynamic obstacles in the scene according to the changing trend of the logical distance.

[0017] Preferably, a two-way feedback link is provided between the dynamic hazard assessment and scenario generation module and the individual behavior prediction and interactive training module to adjust the diffusion weight of hazard factors in real time according to the changes in the trainee's location.

[0018] Preferably, the multi-source sensor data acquisition module further includes:

[0019] The basic data synchronous sampling unit is used to acquire raw sensor datasets containing temperature field distribution, gas component concentration, ground pressure vector, and trainee joint coordinates through infrared thermal imaging sensors, gas concentration monitoring sensors, pressure sensors, and high-definition visual capture components; and to perform time-stamping unification processing on the raw sensor datasets using crystal oscillator clock signals to generate a spatiotemporal data array to be fused with a unified time reference.

[0020] A heterogeneous data weight allocation unit receives a spatiotemporal data array to be fused and calculates the information confidence score of each sensor using an adaptive weight allocation algorithm based on feature entropy. Simultaneously, it weights the sensor data based on the information confidence score to obtain a weighted environmental feature vector. The calculation formula is:

[0021] ,

[0022] in, For the first The characteristic entropy value of each sensor, For the first The characteristic entropy value of each sensor, This represents the total number of sensors;

[0023] The multi-dimensional feature fusion determination unit receives weighted environmental feature vectors, constructs a risk feature fusion model to calculate a comprehensive risk probability index value, and determines whether the comprehensive risk probability index value exceeds a preset risk activation threshold. To determine whether to trigger heterogeneous baseline data output; the comprehensive risk probability index value The calculation formula is:

[0024] ,

[0025] in, For the first Normalized values ​​of environmental factors The sensitivity coefficient of the preset risk impact factors, The total number of environmental factors;

[0026] Heterogeneous benchmark data integration output unit, used to ensure that the comprehensive risk probability index value meets the requirements. ≥ At that time, the weighted environmental feature vector and the spatial distribution information of the trainees are encapsulated to generate heterogeneous benchmark data containing environmental evolution trends and personnel location topology.

[0027] Preferably, the dynamic hazard assessment and scenario generation module further includes:

[0028] The physical field feature extraction unit is used to receive the output heterogeneous reference data, use a preset convolution feature extraction operator to identify the spatial distribution features of the fire thermal field and toxic gas pressure field in the heterogeneous reference data, and generate an initial physical field feature tensor containing three-dimensional spatial coordinates and initial physical quantity gradients.

[0029] The hazard evolution vector calculation unit is used to receive the initial physical field feature tensor, calculate the dynamic migration direction and rate of the hazard factor in the virtual space using a fluid dynamics evolution algorithm, and obtain the hazard evolution velocity vector field at each coordinate point through calculation.

[0030] The migration rate of the hazard factor The calculation formula is:

[0031] ,

[0032] in, The preset environmental conductivity coefficient, The spatial gradient of physical quantities in the initial physical field characteristic tensor. For virtual medium density, This is a correction factor for the external wind field. This is the preset global wind speed vector;

[0033] The concentration spatiotemporal evolution determination unit is used to calculate the concentration evolution value of the hazard factor within the target time step based on the hazard evolution velocity vector field and the material transport equation.

[0034] The concentration evolution values The calculation formula is:

[0035] ,

[0036] in, This represents the initial concentration value in the initial physical field characteristic tensor. The physical attenuation constant of the hazard factor. It represents the linear displacement of the hazard factor from its source to its target location;

[0037] By judging the concentration evolution values Has the preset escape and entry restriction threshold been reached? To determine the accessibility of this coordinate point in a virtual escape scenario;

[0038] The virtual escape scenario construction unit is used to call the graphics rendering engine in real time to adjust the smoke concentration distribution and fire light rendering brightness in the three-dimensional virtual space according to the concentration evolution value and the passage status, so as to transform the environmental evolution logic into a visualized virtual escape scenario.

[0039] Preferably, the individual behavior prediction and interaction training module further includes:

[0040] The action feature sequence extraction unit is used to receive real-time posture data of the trainee in the virtual escape scenario, extract the displacement vectors of the skeletal joints of the trainee in continuous time frames using a long short-term memory network, and generate action feature sequence information that reflects the trainee's escape plan.

[0041] The path selection trend prediction unit is used to receive the action feature sequence information, calculate the probability value of the trainee pointing to different escape exits in combination with the spatial topology in the virtual escape scenario, and obtain the path selection trend distribution of the trainee at the current moment through calculation.

[0042] The pointing probability value The calculation formula is:

[0043] ,

[0044] in, The instantaneous velocity vector in the action feature sequence information. Point the trainee's current position to the first The unit direction vector of each escape exit. This represents the total number of available escape exits.

[0045] The behavioral risk logic determination unit is used to calculate the logical distance between the trainee's current behavioral state and the preset danger boundary based on the path selection trend distribution.

[0046] The logical distance value The calculation formula is:

[0047] ,

[0048] in, The real-time coordinate vector of the trainee. This represents the boundary coordinate vector of the nearest danger source in the virtual escape scenario. For the trainee's real-time movement rate, The preset redundancy time step for the risk avoidance response;

[0049] By judging the logical distance value Whether the value is less than or equal to zero is used to determine whether the trainee is in a high-risk intrusion state;

[0050] The dynamic obstacle triggering intervention unit is used to select the escape path with the highest probability value in the trend distribution according to the path selection when outputting the high-risk intrusion status judgment result, and dynamically generate physical obstacles or visual occlusion factors in the virtual escape scenario to achieve deep two-way coupling between the trainee's behavior and the evolving environment.

[0051] Preferably, the real-time feedback and automatic evaluation module further includes:

[0052] The path decision deviation quantification unit receives and transmits prediction and intervention data, calculates the trajectory similarity value between the trainee's actual movement trajectory and the dynamically optimal path generated by the system using a dynamic time warping algorithm, and obtains the trainee's path deviation rate index throughout the training process. The trajectory similarity value... The calculation formula is:

[0053] ,

[0054] in, The actual movement trajectory of the trainee With dynamic optimal path The cumulative Euclidean distance between them This represents the total length of the corresponding path;

[0055] The action accuracy determination unit is used to receive the output trajectory similarity value, combine it with the limb movement information in the prediction and intervention data, and calculate the accuracy score of the trainee's execution of avoidance actions; the trainee's avoidance action accuracy index is obtained through calculation.

[0056] The accuracy score The calculation formula is:

[0057] ,

[0058] in, For trainees in The actual action feature vector at any given time. The preset standard avoidance action mode vector, These are the action weighting coefficients at different time points. The total execution time of the action;

[0059] The multi-dimensional performance evaluation unit receives the path deviation rate index and the risk avoidance action accuracy index, and calculates the trainee's final ability score using a fuzzy comprehensive evaluation model; the final ability score... The calculation formula is:

[0060] ,

[0061] in, For the trainees' decision response time, The maximum permissible emergency response time limit;

[0062] By judging the final ability score Has the preset passing threshold been reached? To determine the trainees' training and assessment status;

[0063] The deviation correction guidance report generation unit is used to retrieve a preset expert guidance strategy library based on the final capability score, the path deviation rate index, the risk avoidance action accuracy index, and the decision response time, and combine them to generate a deviation correction guidance report containing targeted improvement suggestions.

[0064] Preferably, the training optimization and continuous learning module further includes:

[0065] The performance data feature mining unit is used to receive the correction guidance report output by the real-time feedback and automatic evaluation module, extract the weakness feature vectors associated with the trainees' skill deficiencies in the correction guidance report, and use the incremental learning algorithm to perform cluster analysis on the weakness feature vectors to generate a common training needs matrix that reflects the collective cognitive biases of the trainees.

[0066] The scenario complexity assessment unit is used to receive the common training requirement matrix, calculate the training difficulty weight of the current escape scenario configuration using information entropy theory, and obtain the coverage efficiency of the current escape scenario configuration for improving the trainee's skills through calculation.

[0067] The training difficulty weight The calculation formula is:

[0068] ,

[0069] in, For the first The probability distribution of each hazard factor in a virtual escape scenario. This represents the total number of risk factor types. This is a difficulty correction factor;

[0070] The model strategy iterative update unit is used to adjust the weight parameters in the dynamic hazard assessment and scene generation module according to the training difficulty weights using a reinforcement learning algorithm;

[0071] The updated value of the weight parameter The calculation formula is:

[0072] ,

[0073] in, The preset learning rate, The reward score is based on the conversion of the corrective guidance report. As a discount factor, In the current training state The following scene adjustment actions are taken. The value function, The direction of the policy gradient;

[0074] By judging the updated value of the weight parameters Whether the logic of convergence is used to determine whether the configuration update for this escape scenario is complete;

[0075] The personalized scenario delivery unit is used to generate customized escape drill plans adapted to different industrial parks based on the updated weight parameters after the logic update is completed, and synchronize the customized escape drill plans to the dynamic hazard assessment and scenario generation module to realize the adaptive model evolution of the training system under different physical layouts.

[0076] This invention provides an artificial intelligence-based, full-scenario emergency escape training system. It has the following beneficial effects:

[0077] 1. This invention employs dynamic hazard assessment and scenario generation technology to achieve real-time linkage between hazard evolution and virtual scenarios, enabling accurate simulation of the spread characteristics of secondary disasters and addressing the shortcomings of existing static training methods that cannot reproduce the transient evolution patterns of hazardous factors in complex industrial sites.

[0078] 2. This invention employs individual behavior prediction and automatic evaluation technology to achieve real-time quantification and in-depth analysis of escape decision trends, enabling the immediate output of targeted corrective guidance reports and addressing the shortcomings of existing technologies that lack behavioral analysis capabilities, thus limiting the improvement of trainees' emergency decision-making abilities.

[0079] 3. This invention employs continuous learning and automatic generation technology to achieve the effect of autonomously adjusting the escape scenario according to industrial layout and training needs, realizing the leap from fixed modeling to personalized dynamic construction of the training mode, and solving the shortcomings of the existing technology in terms of narrow applicability and high cost of scenario construction. Attached Figure Description

[0080] Figure 1 This is a block diagram of the overall logical architecture of the system of the present invention;

[0081] Figure 2 This is a schematic diagram of the internal structure of the multi-source sensor data acquisition module of the present invention;

[0082] Figure 3 This is a flowchart illustrating the logic of hazard evolution and virtual scenario generation in this invention.

[0083] Figure 4 This is a schematic diagram illustrating the bidirectional coupling of individual behavior prediction and dynamic intervention in this invention. Detailed Implementation

[0084] To enable those skilled in the art to understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0085] The present invention will now be described in detail with reference to the accompanying drawings:

[0086] Example 1:

[0087] Please see the appendix Figure 1 -Appendix Figure 4 This invention provides an artificial intelligence-based full-scenario emergency escape training system, comprising:

[0088] The multi-source sensor data acquisition module is used to acquire industrial field environmental parameters and trainee spatial distribution information in real time, and output heterogeneous reference data.

[0089] The dynamic hazard assessment and scenario generation module is used to receive heterogeneous baseline data, construct a hazard evolution model, and dynamically generate virtual escape scenarios.

[0090] The individual behavior prediction and interactive training module is used to identify the trainees' action characteristics in virtual escape scenarios and predict the trainees' path selection trends.

[0091] The real-time feedback and automatic evaluation module is used to receive trainees’ path selection trends, perform quantitative analysis on trainees’ escape decisions and output a correction guidance report; the correction guidance report includes decision response time, path deviation rate, accuracy of avoidance actions and physiological and psychological stability evaluation indicators.

[0092] The training optimization and continuous learning module is used to mine historical training data involved in the correction guidance report and autonomously update the escape scenario configuration in the dynamic hazard assessment and scenario generation module.

[0093] The multi-source sensor data acquisition module further includes:

[0094] The basic data synchronous sampling unit is used to acquire raw sensor datasets containing temperature field distribution, gas component concentration, ground pressure vector, and trainee joint coordinates through infrared thermal imaging sensors, gas concentration monitoring sensors, pressure sensors, and high-definition visual capture components; and to perform time-stamping unification processing on the raw sensor datasets using crystal oscillator clock signals to generate a spatiotemporal data array to be fused with a unified time reference.

[0095] The heterogeneous data weight allocation unit receives the spatiotemporal data array to be fused and calculates the information confidence score of each sensor using an adaptive weight allocation algorithm based on feature entropy. Simultaneously, it weights the sensor data based on the information confidence score to obtain a weighted environmental feature vector. The calculation formula is:

[0096] ,

[0097] in, For the first The characteristic entropy value of each sensor, For the first The characteristic entropy value of each sensor, This represents the total number of sensors;

[0098] The multi-dimensional feature fusion determination unit receives weighted environmental feature vectors, constructs a risk feature fusion model to calculate a comprehensive risk probability index value, and determines whether the comprehensive risk probability index value exceeds a preset risk activation threshold. To determine whether to trigger heterogeneous baseline data output; comprehensive risk probability index value. The calculation formula is:

[0099] ,

[0100] in, For the first Normalized values ​​of environmental factors The sensitivity coefficient of the preset risk impact factors, The total number of environmental factors;

[0101] Heterogeneous benchmark data integration output unit, used to ensure that the comprehensive risk probability index value meets the requirements. ≥ At that time, the weighted environmental feature vector and the spatial distribution information of the trainees are encapsulated to generate heterogeneous benchmark data containing environmental evolution trends and personnel location topology.

[0102] The dynamic hazard assessment and scenario generation module further includes:

[0103] The physical field feature extraction unit is used to receive the output heterogeneous reference data, use a preset convolution feature extraction operator to identify the spatial distribution features of the fire thermal field and toxic gas pressure field in the heterogeneous reference data, and generate an initial physical field feature tensor containing three-dimensional spatial coordinates and initial physical quantity gradients.

[0104] The hazard evolution vector calculation unit is used to receive the initial physical field feature tensor, calculate the dynamic migration direction and rate of hazard factors in the virtual space using a fluid dynamics evolution algorithm, and obtain the hazard evolution velocity vector field at each coordinate point through calculation.

[0105] migration rate of risk factors The calculation formula is:

[0106] ,

[0107] in, The preset environmental conductivity coefficient, The spatial gradient of physical quantities in the initial physical field characteristic tensor. For virtual medium density, This is a correction factor for the external wind field. This is the preset global wind speed vector;

[0108] The concentration spatiotemporal evolution determination unit is used to calculate the concentration evolution value of the hazard factor within the target time step based on the hazard evolution velocity vector field and the material transport equation.

[0109] Concentration evolution values The calculation formula is:

[0110] ,

[0111] in, This represents the initial concentration value in the initial physical field characteristic tensor. The physical attenuation constant of the hazard factor. It represents the linear displacement of the hazard factor from its source to its target location;

[0112] By judging the concentration evolution values Has the preset escape and entry restriction threshold been reached? To determine the accessibility of this coordinate point in a virtual escape scenario;

[0113] The virtual escape scenario construction unit is used to call the graphics rendering engine in real time to adjust the smoke concentration distribution and fire light rendering brightness in the three-dimensional virtual space based on the concentration evolution value and passage status, and transform the environmental evolution logic into a visualized virtual escape scenario.

[0114] The individual behavior prediction and interactive training module further includes:

[0115] The motion feature sequence extraction unit is used to receive real-time posture data of trainees in virtual escape scenarios, and use a long short-term memory network to extract the displacement vectors of skeletal joints of trainees in continuous time frames to generate motion feature sequence information that reflects the trainees' escape plan.

[0116] The path selection trend prediction unit is used to receive action feature sequence information, combine it with the spatial topology in the virtual escape scenario to calculate the probability value of the trainee pointing to different escape exits; and obtain the path selection trend distribution of the trainee at the current moment through calculation.

[0117] Pointing probability value The calculation formula is:

[0118] ,

[0119] in, The instantaneous velocity vector in the action feature sequence information. Point the trainee's current position to the first The unit direction vector of each escape exit. This represents the total number of available escape exits.

[0120] The behavioral risk logic determination unit is used to calculate the logical distance between the trainee's current behavioral state and the preset danger boundary based on the path selection trend distribution.

[0121] Logical distance value The calculation formula is:

[0122] ,

[0123] in, The real-time coordinate vector of the trainee. This represents the boundary coordinate vector of the nearest danger source in the virtual escape scenario. For the trainee's real-time movement rate, The preset redundancy time step for the risk avoidance response;

[0124] By judging the logical distance value Whether the value is less than or equal to zero is used to determine whether the trainee is in a high-risk intrusion state;

[0125] The dynamic obstacle triggering intervention unit is used to select the escape path with the highest probability value in the trend distribution of the path selection when outputting the high-risk intrusion status judgment result. It dynamically generates physical obstacles or visual occlusion factors in the virtual escape scenario to achieve deep two-way coupling between the trainee's behavior and the evolving environment.

[0126] The real-time feedback and automatic evaluation module further includes:

[0127] The path decision deviation quantification unit receives and transmits prediction and intervention data, and uses a dynamic time warping algorithm to calculate the trajectory similarity value between the trainee's actual movement trajectory and the dynamically optimal path generated by the system; it also calculates the trainee's path deviation rate index throughout the training process; and the trajectory similarity value. The calculation formula is:

[0128] ,

[0129] in, The actual movement trajectory of the trainee With dynamic optimal path The cumulative Euclidean distance between them This represents the total length of the corresponding path;

[0130] The action accuracy determination unit is used to receive the output trajectory similarity value, combine it with the limb movement information in the prediction and intervention data, and calculate the accuracy score of the trainee's execution of avoidance actions; the trainee's avoidance action accuracy index is obtained through calculation.

[0131] Accuracy rating The calculation formula is:

[0132] ,

[0133] in, For trainees in The actual action feature vector at any given time. The preset standard avoidance action mode vector, These are the action weighting coefficients at different time points. The total execution time of the action;

[0134] The multi-dimensional performance evaluation unit receives path deviation rate and avoidance action accuracy indicators, and uses a fuzzy comprehensive evaluation model to calculate the trainee's final ability score; the final ability score... The calculation formula is:

[0135] ,

[0136] in, For the trainees' decision response time, The maximum permissible emergency response time limit;

[0137] By judging the final ability score Has the preset passing threshold been reached? To determine the trainees' training and assessment status;

[0138] The deviation correction guidance report generation unit is used to retrieve a preset expert guidance strategy library based on the final capability score, path deviation rate index, risk avoidance action accuracy index, and decision response time, and combine them to generate a deviation correction guidance report containing targeted improvement suggestions.

[0139] The training optimization and continuous learning module further includes:

[0140] The performance data feature mining unit is used to receive the correction guidance report output by the real-time feedback and automatic evaluation module, extract the weakness feature vectors associated with the trainees' skill deficiencies in the correction guidance report, and use the incremental learning algorithm to perform cluster analysis on the weakness feature vectors to generate a common training needs matrix that reflects the collective cognitive biases of the trainees.

[0141] The scenario complexity assessment unit receives the common training requirement matrix and uses information entropy theory to calculate the training difficulty weights of the current escape scenario configuration; it then calculates the coverage effectiveness of the current escape scenario configuration for improving the trainees' skills.

[0142] Training difficulty weights The calculation formula is:

[0143] ,

[0144] in, For the first The probability distribution of each hazard factor in a virtual escape scenario. This represents the total number of risk factor types. This is a difficulty correction factor;

[0145] The model strategy iterative update unit is used to adjust the weight parameters in the dynamic hazard assessment and scene generation module according to the training difficulty weights using a reinforcement learning algorithm;

[0146] Weight parameter update value The calculation formula is:

[0147] ,

[0148] in, The preset learning rate, The reward score is based on the conversion of the corrective guidance report. As a discount factor, In the current training state The following scene adjustment actions are taken. The value function, The direction of the policy gradient;

[0149] By judging the updated value of the weight parameters Whether the logic of convergence is used to determine whether the configuration update for this escape scenario is complete;

[0150] The personalized scenario delivery unit is used to generate customized escape drill plans adapted to different industrial parks based on the updated weight parameters after the logic update is completed. The customized escape drill plans are also synchronized to the dynamic hazard assessment and scenario generation module to realize the adaptive model evolution of the training system under different physical layouts.

[0151] The multi-source sensor data acquisition module integrates various physical quantity monitoring devices and combines clock signal synchronization technology and adaptive weight allocation algorithm. It can extract high-confidence environmental features and personnel location topology information from complex industrial environments, solving the problem of inconsistency in the spatiotemporal dimensions of multi-source heterogeneous data from the source, and providing a reliable data foundation for risk assessment and scenario generation.

[0152] The dynamic hazard assessment and scenario generation module uses fluid dynamics evolution algorithms and material transport equations to perform physical-level modeling of the spatial migration of hazard factors, transforming abstract environmental parameters into virtual escape scenarios that evolve in real time in three-dimensional space. This achieves a high degree of coupling between hazard evolution logic and visual rendering effects, ensuring that the simulated environment can realistically reproduce the transient evolution laws of the fire thermal field and toxic gas pressure field.

[0153] The individual behavior prediction and interactive training module identifies the skeletal displacement characteristics of trainees and calculates the pointing probability through time-series modeling technology. It can predict the trainees' path selection trends in advance and dynamically trigger virtual obstacle intervention based on the logical distance judgment results. This enables real-time response and deep bidirectional coupling of the evolving environment to the trainees' escape behavior, greatly improving the interactive depth and practical attributes of the simulation training.

[0154] The real-time feedback and automatic evaluation module uses a dynamic time warping algorithm and a fuzzy comprehensive evaluation model to perform multi-dimensional quantitative analysis of the trainees' movement trajectory similarity, risk avoidance action standardization, and decision response efficiency. By retrieving the expert guidance strategy library, it outputs an analysis report containing targeted improvement suggestions, providing scientific and objective data support for the quantitative assessment and closed-loop correction of trainees' emergency decision-making capabilities.

[0155] The training optimization and continuous learning module uses incremental learning and reinforcement learning algorithms to deeply mine historical performance data. It can autonomously iterate the distribution logic of risk factors and update scenario configuration parameters according to the cognitive weaknesses of trainees, enabling the training system to achieve model self-evolution and customized solution distribution under different industrial layouts, significantly reducing the construction cost of full-scenario training solutions.

[0156] Example 2: This example uses a simulated training scenario where a leak occurs in the isobutane storage tank area of ​​a large chemical plant, causing a fire.

[0157] During the training initiation phase, the multi-source sensor data acquisition module acquires raw datasets in real time, including temperature distribution, harmful gas concentration, and trainee keypoint coordinates, using sensor components deployed in the industrial park. The basic data synchronization sampling unit uses a clock signal to perform time-stamping on the raw data, generating a spatiotemporal data array. The heterogeneous data weight allocation unit automatically assigns weights based on the information confidence level of each sensor in the current environment, and the multi-dimensional feature fusion judgment unit calculates the comprehensive risk probability index. When the risk index exceeds a preset threshold, the heterogeneous baseline data integration output unit outputs the encapsulated environmental evolution trend and personnel location topology data to the next module.

[0158] After receiving heterogeneous baseline data, the dynamic hazard assessment and scenario generation module identifies the spatial distribution characteristics of the fire thermal field and toxic gas pressure field by the physical field feature extraction unit; the hazard evolution vector calculation unit uses an evolution algorithm to calculate the dynamic migration direction and rate of hazard factors in the virtual space, generating an evolution velocity vector field; the concentration spatiotemporal evolution determination unit calculates the concentration evolution values ​​of each coordinate point through the mass transport equation and determines whether the point is in an escape restricted state; the virtual escape scenario construction unit adjusts the smoke distribution and light and shadow brightness in the virtual space in real time according to the calculation results to construct a highly realistic virtual escape scenario;

[0159] After the trainee enters the virtual environment, the action feature sequence extraction unit identifies the trainee's posture data in real time and generates action feature sequences; the path selection trend prediction unit calculates the probability of the trainee pointing to different escape exits based on the spatial topology; the behavior risk logic judgment unit monitors the logical distance between the trainee and the boundary of the hazard source in real time and determines whether the trainee has entered a high-risk intrusion state; if the judgment result is a high-risk intrusion, the dynamic obstacle triggering intervention unit will dynamically generate physical obstacles on the escape path with the highest probability value, realizing a deep coupling between environmental evolution and trainee behavior;

[0160] During training, the path decision deviation quantification unit calculates the trajectory similarity between the trainee's actual trajectory and the dynamic optimal path to obtain a path deviation rate index; the action accuracy judgment unit combines the trainee's limb movement information with standard risk avoidance modes to output a risk avoidance action accuracy index. The multi-dimensional performance comprehensive evaluation unit summarizes the deviation rate, action accuracy, and decision response time, uses an evaluation model to calculate the trainee's final ability score, and determines the assessment status. The deviation correction guidance report generation unit searches the expert strategy database and combines them to generate a guidance report containing targeted improvement suggestions.

[0161] The training optimization and continuous learning module performs in-depth mining of historical training data, while the performance data feature mining unit uses learning algorithms to analyze common cognitive biases among trainees. The scene complexity assessment unit calculates the currently configured training difficulty weights based on information entropy theory and evaluates their coverage effectiveness. The model strategy iteration and update unit uses reinforcement learning algorithms to adjust the scene generation parameters until the weight update values ​​converge. The personalized scene delivery unit generates customized escape drill plans adapted to different physical layouts and synchronizes the plans to the front end, completing the system's adaptive evolution.

[0162] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An all-scene emergency escape training system based on artificial intelligence, characterized in that, The application relates to a multi-source sensor data acquisition module for acquiring industrial field environment parameters and trainee spatial distribution information in real time and outputting heterogeneous reference data. A dynamic danger assessment and scene generation module is used for receiving the heterogeneous reference data, constructing a danger evolution model and dynamically generating a virtual escape situation. An individual behavior prediction and interactive training module is used for identifying the action features of the trainee in the virtual escape situation, and predicting the trainee path selection trend. A real-time feedback and automatic evaluation module is used for receiving the trainee path selection trend, quantitatively analyzing the trainee escape decision and outputting a rectification guidance report. The training optimization and continuous learning module is used for mining historical training data involved in the rectification guidance report, and autonomously updating the escape scene configuration in the dynamic danger assessment and scene generation module. The multi-source sensor data acquisition module is integrated with an infrared thermal imaging sensor, a gas concentration monitoring sensor, a pressure sensor and a high-definition visual capture component. 2.The all-scene emergency escape training system based on artificial intelligence according to claim 1, wherein The dynamic danger assessment and scene generation module is internally provided with a neural network evolution model. 3.The all-scene emergency escape training system based on artificial intelligence according to claim 1, wherein The individual behavior prediction and interactive training module adopts a recurrent neural network to perform time series modeling on the trainee's skeletal joint data and movement trajectory. 4.The all-scene emergency escape training system based on artificial intelligence according to claim 1, wherein, The dynamic danger assessment and scene generation module and the individual behavior prediction and interactive training module are provided with a bidirectional feedback link.

5. The full-scene emergency escape training system based on artificial intelligence according to claim 4, characterized in that, The multi-source sensor data acquisition module further comprises a basic data synchronous sampling unit. 6.The all-scene emergency escape training system based on artificial intelligence according to claim 1, wherein, The dynamic danger assessment and scene generation module further comprises a physical field feature extraction unit and a danger evolution vector calculation unit. The physical field feature extraction unit is used for receiving the output heterogeneous reference data, identifying the spatial distribution features of the fire thermal field and the toxic gas pressure field in the heterogeneous reference data by using a preset convolution feature extraction operator, and generating an initial physical field feature tensor comprising three-dimensional space coordinates and initial physical quantity gradients. The isomerous data weight distribution unit is used for receiving a to-be-fused space-time data array, calculating information confidence of each sensor by using a feature entropy-based adaptive weight distribution algorithm, and weighting sensor data by using the information confidence to obtain a weighted environmental feature vector. The calculation formula of the information confidence is as follows: ， wherein, is the feature entropy value of the th sensor, is the feature entropy value of the th sensor, is the total number of sensors; The multi-dimensional feature fusion determination unit receives weighted environmental feature vectors, constructs a risk feature fusion model to calculate a comprehensive risk probability index value, and determines whether the comprehensive risk probability index value exceeds a preset risk activation threshold. To determine whether to trigger heterogeneous baseline data output; the comprehensive risk probability index value The calculation formula is: ， in, For the first Normalized values ​​of environmental factors The sensitivity coefficient of the preset risk impact factors, The total number of environmental factors; The integrated output unit of the heterogeneous reference data is used to encapsulate the weighted environmental feature vector and the trainee spatial distribution information to generate the heterogeneous reference data containing the environmental evolution trend and the personnel position topology when the comprehensive risk probability index value meets ≥ ​ 7.The all-scene emergency escape training system based on artificial intelligence according to claim 1, wherein, The danger evolution vector calculation unit is used for receiving the initial physical field feature tensor, calculating the dynamic migration direction and rate of the danger factor in the virtual space by using a fluid dynamics evolution algorithm, and obtaining a danger evolution velocity vector field of each coordinate point. ​ ​ The rate of migration of the risk factor The formula for calculating the rate of migration is: ， wherein, is a preset environmental conduction coefficient, is a spatial gradient of a physical quantity in the initial physical field characteristic tensor, is a virtual medium density, is an external wind field correction factor, is a preset global wind speed vector; A concentration space-time evolution determination unit is configured to calculate a concentration evolution value of the hazard factor in a target time step by using a material transport equation according to the hazard evolution velocity vector field. The concentration evolution numerical The formula for calculating is: ， wherein, is an initial concentration value in the initial physical field characteristic tensor, is a physical decay constant of the hazard factor, is a straight-line displacement amount of the hazard factor from the source to the target position; By judging the concentration evolution values Has the preset escape and entry restriction threshold been reached? To determine the accessibility of this coordinate point in a virtual escape scenario; A virtual escape scenario construction unit is configured to adjust, in real time, a smoke concentration distribution in a three-dimensional virtual space and a firelight rendering brightness by calling a graphic rendering engine according to the concentration evolution value and the passing state, and to convert an environment evolution logic into a visualized virtual escape scenario. 8.The all-scene emergency escape training system based on artificial intelligence according to claim 1, wherein, The individual behavior prediction and interaction training module further includes: An action feature sequence extraction unit is configured to receive real-time posture data of a trainee in the virtual escape scenario, extract a skeletal joint displacement vector of the trainee in a continuous time frame by using a long short-term memory network, and generate action feature sequence information reflecting an escape intention of the trainee. A path selection trend prediction unit is configured to receive the action feature sequence information, calculate a pointing probability value of the trainee for different escape exits in combination with a spatial topological structure in the virtual escape scenario, and obtain a path selection trend distribution of the trainee at a current moment by calculation. The pointing probability value The calculation formula is: ， wherein, is an instantaneous velocity vector in the motion feature sequence information, is a unit direction vector pointing to the first escape exit from the current position of the trainee, is the total number of optional escape exits; A behavior risk logic determination unit is configured to calculate a logic distance value between a current behavior state of the trainee and a preset hazard boundary according to the path selection trend distribution. the logical distance value The formula for calculating the logical distance value is: ， wherein, is a real-time coordinate vector of the trainee, is a boundary coordinate vector of the nearest hazard in the virtual escape scenario, is a real-time movement speed of the trainee, is a pre-set hazard-avoidance reaction redundancy time step; By judging whether the logical distance value is less than or equal to zero whether the trainee is in a high-risk intrusion state; A dynamic obstacle triggering intervention unit is configured to dynamically generate a physical obstacle or a visual shielding factor in the virtual escape scenario according to an escape path with the highest probability value in the path selection trend distribution when a high-risk intrusion state determination result is output, so as to realize a deep bidirectional coupling between a behavior of the trainee and an evolving environment. 9.The all-scene emergency escape training system based on artificial intelligence according to claim 1, wherein, The real-time feedback and automatic evaluation module further includes: The path decision bias quantification unit is configured to receive the transmitted prediction and intervention data, calculate a trajectory similarity value between an actual moving trajectory of the trainee and a dynamically optimal path generated by the system by using a dynamic time warping algorithm, and obtain a path deviation rate index of the trainee in the whole training process by calculation. The calculation formula of the trajectory similarity value is: ， wherein, is the cumulative Euclidean distance between the actual moving trajectory of the trainee and the dynamic optimal path, is the total length of the corresponding path;​ An action accuracy determination unit is configured to receive the output trajectory similarity value, calculate an accuracy score of an evasive action performed by the trainee in combination with limb action information in the prediction and intervention data, and obtain an evasive action accuracy index of the trainee by calculation. The accuracy score The formula for calculating the accuracy score is: ， wherein, is the actual action feature vector of the trainee at the moment, is a preset standard risk-avoiding action modal vector, is an action weight coefficient at different time nodes, is the total duration of action execution; The multi-dimensional performance comprehensive evaluation unit is configured to receive the path deviation rate index and the hedging action accuracy index, and calculate a final ability score value of the trainee by using a fuzzy comprehensive evaluation model; and the final ability score value of the trainee is calculated according to the following formula: The calculation formula is as follows: ， wherein, a decision response time length for the trainee, a maximum emergency response time limit allowed; by judging whether the final ability score value reaches a preset passing threshold value to determine the training examination state of the trainee; A rectification guidance report generation unit is configured to retrieve a preset expert guidance strategy library according to the final ability score value, the path deviation rate index, the evasive action accuracy index, and the decision response time length, and generate a rectification guidance report containing a targeted improvement suggestion by combination. 10.The all-scene emergency escape training system based on artificial intelligence according to claim 1, wherein, The training optimization and continuous learning module further includes: A performance data feature mining unit is configured to receive a rectification guidance report output by the real-time feedback and automatic evaluation module, extract a weak feature vector associated with a skill defect of the trainee in the rectification guidance report, and generate a common training demand matrix reflecting a collective cognitive bias of the trainee by cluster analysis of the weak feature vector by using an incremental learning algorithm. A scene complexity evaluation unit is configured to receive the common training demand matrix, calculate a training difficulty weight value of a current escape scene configuration by using an information entropy theory, and obtain a coverage efficiency of the current escape scene configuration for skill improvement of the trainee by calculation. The training difficulty weight The calculation formula is: ， wherein, is the number of the first hazard factor in the distribution probability of the virtual escape situation, is the total number of hazard factor species, is the difficulty correction coefficient; A model strategy iteration updating unit is configured to adjust a weight parameter in the dynamic hazard assessment and scene generation module by using a reinforcement learning algorithm according to the training difficulty weight value. the weight parameter update value The calculation formula is: ， wherein, is a preset learning rate, is a reward score value converted based on the rectification guidance report, is a discount factor, is a value function of taking a scene adjustment action at a current training state , is a policy gradient direction; determining whether the logic update of the current escape scenario configuration is completed by judging whether the weight parameter update value converges The personalized scene issuing unit is configured to generate a customized escape drill scheme adapted to different industrial parks according to the updated weight parameters after the logical update is completed, and synchronize the customized escape drill scheme to the dynamic risk assessment and scene generation module, so as to realize adaptive model evolution of the training system under different physical layouts.

Images