Fire training examination intelligent evaluation management method based on big data

Abstract

This invention relates to the field of intelligent assessment technology for fire training, specifically a big data-based intelligent assessment and management method for fire training, comprising: obtaining initial capability scores for each assessment dimension of the trainees, inputting them into a state-inversion-based inference calculation model, and completing score correction and inversion deduction by simulating the state evolution of different training environments and emergencies to generate a comprehensive capability profile; identifying the strengths and weaknesses of the trainees, locating specific behavioral segments or physiological abnormality periods corresponding to the weaknesses, and generating personalized reinforcement training plans that include key skills, training scenarios, and target ranges for physiological indicators; integrating historical assessment data to construct a fire training big data knowledge base and iteratively optimizing the inference calculation model. This method enables dynamic deduction of the assessment process and accurate tracing of weaknesses, improves the adaptability of assessment results to real rescue scenarios, and enhances the pertinence of personalized training plans.

Description

Technical Field

[0001] This invention relates to the field of intelligent assessment technology for fire training, and in particular to an intelligent assessment and management method for fire training assessment based on big data. Background Technology

[0002] Current fire training assessment and management methods mostly adopt manual evaluation combined with basic data statistics. They collect operational data of trainees during the assessment process to form initial capability scores for each assessment dimension. Based on preset assessment standards, they complete score calculation and capability level classification. Some existing technologies use basic big data tools to store and simply summarize historical training data. The assessment process only processes static assessment data and does not set up technical processes related to dynamic scenario simulation and state inversion.

[0003] Existing assessment technologies can only integrate and analyze static data under fixed assessment conditions. They cannot simulate the evolution of states under different training environments and emergencies, nor can they dynamically correct or extrapolate initial ability scores. Consequently, the assessment results deviate from the actual ability states in real fire and rescue scenarios. Current technologies can only classify the strengths and weaknesses of trainees across various ability dimensions, but cannot trace the causes of weakness patterns, nor can they pinpoint the specific behavioral segments or physiological abnormalities corresponding to weakness patterns. Therefore, the development of personalized training programs lacks precise guidance. Furthermore, parameter adjustments in the assessment model rely on manual operation, and a dedicated knowledge base has not been built based on complete historical assessment data. Model iteration and optimization lack systematic data support.

[0004] It is necessary to simulate the evolution of scene states through a state inversion-based inference calculation model, correct and inversely extrapolate the initial ability scores, and at the same time, locate the specific behavioral segments or physiological abnormal periods corresponding to the shortcomings based on the identification of ability patterns, so as to provide precise guidance for the generation of personalized reinforcement training programs. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a big data-based intelligent assessment and management method for fire training and assessment.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a big data-based intelligent assessment and management method for fire training and assessment, comprising: Obtain the initial competency scores of the trainees in each assessment dimension of the fire training assessment; The initial capability score is input into the inference calculation model based on state inversion. The inference calculation model corrects and inverses the initial capability score by simulating the state evolution under different training environments and sudden situations. The inference and calculation model of state inversion generates a comprehensive ability profile of the training object under the preset assessment criteria. The comprehensive ability profile includes scores of multiple interrelated ability dimensions. Based on the comprehensive capability profile, the strengths and weaknesses of the training subjects in different capability dimensions are identified, and the specific behavioral segments or physiological abnormal periods that cause the weakness patterns to appear are located. Based on the identified weakness patterns and their location information, a personalized reinforcement training plan suggestion is generated for the training object. The personalized reinforcement training plan suggestion includes the skills that need to be focused on training, the suggested training scenarios, and the expected physiological indicator target range. By integrating all assessment data of historical training subjects, a fire training big data knowledge base is constructed, and the deep assessment network and the inference calculation model are iteratively optimized using the fire training big data knowledge base.

[0007] As a further aspect of the present invention, obtaining the initial ability scores of the training subjects in each assessment dimension of the fire training assessment includes: Continuously collect real-time training data streams from fire training subjects, including multi-sensor motion data, physiological index data, task scene video streams, and operation instruction logs; The real-time training data stream is synchronized and fused with multi-source heterogeneous data to construct the continuous behavioral trajectory of the training object in three-dimensional space and its accompanying physiological state time series. Based on the continuous behavioral trajectory and the physiological state time series, a multi-dimensional evaluation feature set reflecting the training object's operational standardization, action efficiency, responsiveness and physiological stability is extracted; Construct a deep evaluation network containing multiple hidden layers, and input the multidimensional evaluation feature set into the deep evaluation network; The deep evaluation network performs layer-by-layer nonlinear transformation and feature abstraction on the multidimensional evaluation feature set through its multiple hidden layers, and calculates the initial ability scores of the training object in each evaluation dimension in the fire training assessment. The real-time training data stream is synchronized and fused with multi-source heterogeneous data to construct the continuous behavioral trajectory of the training object in three-dimensional space and its accompanying physiological state time series, specifically including: The motion data from the multi-sensor sensors, the physiological index data, the task scene video stream, and the operation instruction log from different sensors are all timestamped to achieve millisecond-level time alignment. From the video stream of the task scene, the three-dimensional coordinates of the key points of the human skeleton of the training object are extracted in real time using visual recognition technology to form a visual behavior trajectory sequence. The attitude calculation is performed on the inertial measurement unit data in the multi-sensor motion data, and spatial registration and Kalman filtering are performed on the visual behavior trajectory sequence to generate the continuous behavior trajectory. Synchronously, the heart rate, blood oxygen, and body temperature data in the physiological indicators are aligned with each point in time of the continuous behavioral trajectory according to the timestamp, forming a synchronized time series of the physiological state.

[0008] As a further aspect of the present invention, based on the continuous behavioral trajectory and the physiological state time series, a multi-dimensional evaluation feature set reflecting the training subject's operational standardization, action efficiency, immediate response, and physiological stability is extracted, including: From the continuous behavioral trajectory, calculate the joint angles, movement speeds, and path lengths of the training subject when performing standard fire-fighting operations, and perform dynamic time warping matching with the preset standard action template to extract the quantitative deviation features of the operational standardization. By analyzing the continuous behavioral trajectory, including the time taken from the start of the task to the completion of the key sub-task, the total length of the movement path, and the proportion of invalid movements, the speed and path optimization characteristics of the action efficiency are calculated. The time delay between the appearance of the emergency signal and the first effective response action of the training object is detected in the task scene video stream and the operation instruction log, as well as the accuracy of the response action, and the delay and accuracy features of the immediate response are extracted. By statistically analyzing the fluctuation range of the physiological state time series throughout the entire training process and the duration of exceeding the set threshold, the heart rate variability index is calculated to obtain the fluctuation characteristics of the physiological stability.

[0009] As a further aspect of the present invention, the deep evaluation network performs layer-by-layer nonlinear transformation and feature abstraction on the multidimensional evaluation feature set through its multiple hidden layers to calculate the initial capability scores of the training object in each evaluation dimension of the fire training assessment, including: The first layer of the deep evaluation network standardizes and performs preliminary feature dimensionality reduction on the input multidimensional evaluation feature set; The deep evaluation network uses multiple hidden layers in the middle to learn complex high-order relationships between the quantitative deviation characteristics of operational norms, the speed and path optimization characteristics of action efficiency, the delay and accuracy characteristics of strain immediacy, and the fluctuation characteristics of physiological stability through nonlinear activation functions. The last layer of the deep evaluation network is a multi-task output layer, which maps the deeply abstracted features to operational standardization scores, action efficiency scores, responsiveness scores, and physiological stability scores. The operational standardization scores, action efficiency scores, responsiveness scores, and physiological stability scores together constitute the initial capability score.

[0010] As a further aspect of the present invention, the initial capability score is input into a state-inversion-based inference calculation model. This inference calculation model, by simulating state evolution under different training environments and unexpected situations, corrects and inversely extrapolates the initial capability score, including: The inference computing model loads a variety of preset virtual training environment configuration files and a database of unexpected events; The inference calculation model uses the initial ability score as the initial state and combines it with the individual basic physical fitness data of the training object to positively deduce the dynamic change process of its ability state in one or more of the virtual training environments. During the forward inference process, the inference calculation model randomly or according to rules injects emergency events from the emergency event library and simulates the training object's reaction based on its current capability state and the resulting change in capability state. Record the final capability state of all deduction paths and compare it with the preset ideal state; Based on the comparison results, the credibility weights and deviation ranges of each score in the initial capability score are calculated in reverse, and the initial capability score is weighted and corrected to obtain the corrected capability score.

[0011] As a further aspect of the present invention, the inference calculation model of the state inversion generates a comprehensive ability profile of the training object under a preset assessment standard, including: The reasoning and calculation model maps the corrected ability score to a preset assessment standard level system to determine the training object's level in four dimensions: operational standard, action efficiency, immediate response, and physiological stability. Based on the four-dimensional levels, the reasoning and calculation model uses a radar chart model to construct the comprehensive capability profile. Each axis of the radar chart represents a capability dimension, the scale of the axis represents the level, and the corrected capability score determines the position of the vertex of the radar chart on the axis. Calculate the area, centroid, and perimeter of the integrated capability profile radar image, and quantify them into a set of profile feature vectors; The profile feature vector is matched with the three standard profiles (excellent, qualified, and needing improvement) stored in the fire training big data knowledge base to determine the overall evaluation category of the training object.

[0012] As a further aspect of the present invention, based on the comprehensive capability profile, the superiority and inferiority patterns of the training subject in different capability dimensions are identified, and the specific behavioral segments or physiological abnormal periods that lead to the occurrence of the inferiority patterns are located, including: The levels of each dimension in the comprehensive capability profile are compared longitudinally with the average level of the training subjects' historical evaluations. Dimensions with improved levels are identified as potential strength patterns, and dimensions with decreased levels or below average levels are identified as potential weakness patterns. For each identified potential weakness pattern, backtrack to the multidimensional assessment feature set that generated the initial capability scores for each dimension; Further, based on the time points corresponding to the multidimensional evaluation feature set, the specific time period in the original real-time training data stream is located; Extract the corresponding continuous behavioral trajectory and physiological state time series from the specific time period, analyze the specific behavioral segments of abnormal actions, operational errors, slow response or physiological index exceeding limits of the training object within the specific time period, and confirm the bottleneck pattern.

[0013] As a further aspect of the present invention, the step of generating personalized reinforcement training program suggestions for the training object based on the identified weakness pattern and its location information includes: Based on the capability dimension corresponding to the aforementioned weakness pattern, several basic training subjects specifically designed to improve the aforementioned capability dimension are matched from a preset training subject library. Based on the specific abnormal behavior fragments in the location information, the matched basic training subjects are adjusted to design targeted repetitive training, decomposition training or anti-interference training content. Based on the overall assessment category and basic physical fitness data of the trainees, the training intensity, duration and repetitions are set for each training item to form an initial reinforcement training module. Integrate all reinforcement training modules targeting different weakness patterns, and sort them according to the principle of starting with the easy and progressing to the difficult, and complementing skills, to form a complete training cycle plan, namely the personalized reinforcement training program recommendation.

[0014] As a further aspect of the present invention, the step of constructing a fire training big data knowledge base and using the fire training big data knowledge base to iteratively optimize the deep evaluation network and the inference calculation model includes: The real-time training data stream, the multi-dimensional evaluation feature set, the initial capability score, the comprehensive capability profile, the identified weakness patterns, and the finally generated personalized reinforcement training scheme suggestions generated from each training evaluation are stored as a complete case in the fire training big data knowledge base. A large amount of case data is periodically extracted from the fire training big data knowledge base to retrain the deep evaluation network and update its network weights, so that its evaluation ability score is more accurate. Meanwhile, by utilizing the enhanced training programs and their corresponding capability improvement results stored in the fire training big data knowledge base that have been verified to be effective through actual training, the deduction rules and parameters of state inversion in the reasoning calculation model are optimized.

[0015] As a further aspect of the present invention, the construction steps of the inference computation model include: The initial capability scores before training, dynamic behavior and physiological data during training, and the actual combat capability results of the final assessment after training of different training subjects in historical fire training are obtained to form the basic data set for model training. The initial ability score of the trainees before training, their basic physical fitness data, and the parameters of the virtual training environment they face are used as input features. The optimization objective is to match the final capability state obtained from the simulation with the actual combat capability results after training. Construct a computational model framework that includes a state transition submodule and a state inversion correction submodule; The computational model framework is trained using the basic dataset. The state transition probability parameters in the state transition submodule and the score correction weight parameters in the state inversion correction submodule are adjusted so that the final capability state inferred by the model under given input features can fit the corresponding actual combat capability result to the greatest extent, thus completing the construction of the inference computational model.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By simulating the evolution of states under different training environments and emergencies, the initial capability scores are corrected and extrapolated. This approach breaks away from the singular evaluation logic of static score calculation, allowing capability scores to adapt to complex and ever-changing scenario conditions. It weakens the limitations of single-assessment data, ensuring that the score data used in the evaluation closely matches the actual capability performance of trainees in variable scenarios, thus establishing a correspondence between the evaluation results and the capability status in real application scenarios. The simulated state evolution processing method can cover emergency scenario variables that conventional assessments cannot address, ensuring that the score correction and extrapolation process aligns with the actual application needs of fire training, eliminating evaluation biases between fixed assessment conditions and real rescue scenarios.

[0017] By leveraging comprehensive ability profiles to identify the strengths and weaknesses of training subjects, and simultaneously pinpointing specific behavioral segments or physiological abnormalities corresponding to the weaknesses, this approach overcomes the fuzzy judgment logic at the ability dimension level, directly identifying the source of ability deficiencies and preventing broad ability assessments from influencing subsequent training plans. Personalized reinforcement training programs generated based on this positioning information allow the setting of key skill items, training scenarios, and target ranges for physiological indicators to directly correspond to the actual problems of the training subjects. This ensures a precise match between the program content and the training subjects' ability weaknesses, aligning the planning logic of the training program with the actual needs of weakness correction. Attached Figure Description

[0018] Figure 1 The flowchart is a big data-based intelligent assessment and management method for fire training and assessment as described in this invention. Figure 2 A flowchart for extracting multidimensional evaluation feature sets; Figure 3 A dynamic change graph of capability status under different virtual training environments; Figure 4 Radar diagram illustrating the comprehensive capabilities of fire training and assessment. Figure 5 This is a diagram for in-depth evaluation of the network iteration effect analysis. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] See Figure 1The system obtains initial capability scores for trainees across various assessment dimensions in fire training assessments. These initial scores are then input into a state-inversion-based inference model, which simulates state evolution under different training environments and emergencies to correct and inversely extrapolate the initial capability scores. The state-inversion inference model generates a comprehensive capability profile of the trainees under preset assessment standards, containing scores for multiple interrelated capability dimensions. Based on this profile, the system identifies the trainees' strengths and weaknesses across different capability dimensions, pinpointing specific behavioral segments or physiological abnormalities that lead to these weaknesses. Based on the identified weakness patterns and their location, personalized reinforcement training program suggestions are generated for the trainees, including key skills requiring focused training, suggested training scenarios, and expected physiological target ranges. Finally, all historical assessment data of trainees is integrated to construct a fire training big data knowledge base, which is then used to iteratively optimize the deep assessment network and inference model.

[0022] In one embodiment of the present invention, the method continuously collects real-time training data streams from fire training subjects. These real-time training data streams include multi-sensor motion data, physiological index data, task scenario video streams, and operation instruction logs. The method performs multi-source heterogeneous data synchronization and fusion on the real-time training data streams to construct a continuous behavioral trajectory of the training subject in three-dimensional space and its accompanying physiological state time series. Specifically, this includes multi-sensor motion data, physiological index data, task scenario video streams, and operation instruction logs from different sensors, all uniformly timestamped to achieve millisecond-level time alignment. From the task scenario video stream, the method extracts the three-dimensional coordinates of key points of the human skeleton of the training subject in real time using visual recognition technology to form a visual behavioral trajectory sequence. The method performs attitude calculation on the inertial measurement unit data from the multi-sensor motion data and performs spatial registration and Kalman filtering fusion with the visual behavioral trajectory sequence to generate a continuous behavioral trajectory. Simultaneously, the method aligns the heart rate, blood oxygen, and body temperature data from the physiological index data with each time point of the continuous behavioral trajectory according to the timestamp, forming a synchronized physiological state time series. Based on the continuous behavioral trajectory and physiological state time series, a multi-dimensional evaluation feature set reflecting the training subject's operational standardization, action efficiency, responsiveness, and physiological stability is extracted. A deep evaluation network with multiple hidden layers is constructed, and a multidimensional evaluation feature set is input into the deep evaluation network. The deep evaluation network performs layer-by-layer nonlinear transformation and feature abstraction on the multidimensional evaluation feature set through its multiple hidden layers, and calculates the initial ability scores of the training object in each evaluation dimension in the fire training assessment.

[0023] In practical implementation, a firefighter's assessment task of laying fire hoses and spraying water to extinguish fires in a simulated dark and smoky environment was used as an example scenario. Trainees wore motion capture suits with integrated inertial measurement units, heart rate monitors, and pulse oximeter clips. The training helmets were equipped with head-mounted cameras. A network of fixed cameras in the simulated training area captured video streams of the task scene from multiple angles. Pressure sensors and switch sensors were installed on the fire hose interface valves and simulated water guns operated by the trainees to generate operation command logs. In this implementation, the continuously collected real-time training data stream included multi-sensor motion data output from the motion capture suit, physiological indicator data output from the heart rate monitor and pulse oximeter clips, the task scene video stream output from the head-mounted camera and fixed camera network, and operation command logs recorded by the pressure and switch sensors. In some embodiments, the process of synchronizing and fusing multi-source heterogeneous data from the real-time training data stream first requires providing a unified time reference for all data sources. In practice, a central synchronization server sends precise clock synchronization signals to all sensors and cameras. Acceleration and angular velocity readings from multi-sensor motion data, heart rate and blood oxygen saturation readings from physiological indicators, each frame of the task scene video stream, and each switch event from the operation command log are all marked with millisecond-level timestamps from the central synchronization server, achieving millisecond-level time alignment. When extracting the visual behavior trajectory sequence from the task scene video stream, computer vision algorithms identify the human joints of the training object in each frame of video image and, combined with multi-view video, calculate the three-dimensional coordinates of the human skeletal key points in the world coordinate system using triangulation principles. These are then arranged in timestamp order to form the visual behavior trajectory sequence. When fusing to generate continuous behavior trajectories, inertial measurement unit (IMU) data from the multi-sensor motion data is used to obtain the pose quaternions of the training object's limbs through a pose calculation algorithm. The local coordinate system defined by the IMU data is spatially registered with the world coordinate system defined by the visual behavior trajectory sequence. A Kalman filter is then used to fuse and filter the visual behavior trajectory sequence with the pose-calculated IMU trajectory, thereby generating a smooth and accurate continuous behavior trajectory of the training object in three-dimensional space.

[0024] It is understandable that the construction of the physiological state time series and the generation of the continuous behavioral trajectory are carried out simultaneously. In practice, the heart rate, blood oxygen, and body temperature data in the physiological indicators also carry millisecond-level timestamps. The data processing system aligns each physiological data point to the corresponding time point of the continuous behavioral trajectory according to the timestamp, thus forming a physiological state time series that strictly corresponds to each moment of the continuous behavioral trajectory. For example, when the continuous behavioral trajectory records the moment when the training subject runs to a certain position, the physiological state time series simultaneously records that the training subject's heart rate at that moment is 150 beats per minute. In practice, the process of extracting a multidimensional evaluation feature set based on the continuous behavioral trajectory and the physiological state time series involves multiple quantitative calculations. From the continuous behavioral trajectory, we can calculate the angle change curves of the shoulder and elbow joints during the water hose throwing action, the motion speed curve of the hip joint during movement, and the actual movement path length from the starting point to the water distributor installation point when the training subject performs the standard operation of "quickly laying water hose and connecting the water distributor". The quantitative deviation characteristics of operational standardization are obtained by calculating the mean square error after dynamically time-warping and matching the calculated actual joint angle curve with the preset standard action template curve. The speed and path optimization characteristics of action efficiency are calculated by analyzing the total time taken from the start of the task to the completion of the key sub-task of "stable hit of the fire point by the water jet", and the ratio of the total length of the movement path to the preset optimal path length. In specific implementation, the deep evaluation network is a pre-constructed deep neural network model. The multidimensional evaluation feature set is input into the deep evaluation network as a feature vector. The first layer of the deep evaluation network standardizes the input multidimensional evaluation feature set and initially projects the high-dimensional features into a lower-dimensional space. The middle hidden layers of the deep evaluation network consist of fully connected layers and nonlinear activation functions. These hidden layers perform layer-by-layer nonlinear transformation and feature abstraction on the initially processed features, learning the complex high-order correlation patterns between the quantitative deviation characteristics of operational standardization, the speed and path optimization characteristics of action efficiency, the delay and accuracy characteristics of immediate response, and the fluctuation characteristics of physiological stability. The final layer of the deep evaluation network is a multi-task output layer. It maps the high-level features, abstracted from the preceding hidden layers, into four specific scores: operational standardization score, action efficiency score, immediate response score, and physiological stability score. These four scores together constitute the initial capability score of the trainee in this fire training assessment. Optionally, a quantitative matching degree calculation formula can be introduced when calculating the quantitative deviation features of operational standardization. For a standard action template T containing N keyframes and an actual action sequence A, after dynamic time warping alignment, the angle deviation of the i-th joint at the j-th alignment point can participate in the overall standardization score calculation. The contribution calculation method of joint angle deviation is as follows: in: This represents the actual angle of the i-th joint at the j-th alignment point. Compared to standard angle The absolute deviation. The deviations of all joints at all alignment points are aggregated and used to generate a portion of the quantized deviation feature vector for operational normalization.

[0025] In one embodiment of the present invention, see [reference] Figure 2 This method calculates the joint angles, movement speed, and path length of the training subject when performing standard fire-fighting operations from continuous behavioral trajectories, and performs dynamic time warping matching with preset standard action templates to extract quantitative deviation features of operational standardization. It analyzes the time taken from the start of the task to the completion of key sub-tasks, the total length of the movement path, and the proportion of ineffective movements in the continuous behavioral trajectory to calculate the speed and path optimization features of action efficiency. It detects the time delay between the appearance of a sudden situation signal and the training subject's first effective response action in the task scene video stream and operation instruction log, as well as the accuracy of the response action, to extract the delay and accuracy features of immediate response. It statistically analyzes the fluctuation range of the physiological state time series throughout the training process and the duration exceeding a set threshold, calculating its heart rate variability index to obtain the fluctuation features of physiological stability. The first layer of the deep evaluation network standardizes and performs preliminary feature dimensionality reduction on the input multidimensional evaluation feature set. Multiple hidden layers in the middle of the deep evaluation network learn complex high-order correlations between the quantitative deviation features of operational standardization, the speed and path optimization features of action efficiency, the delay and accuracy features of immediate response, and the fluctuation features of physiological stability through nonlinear activation functions. The last layer of the deep evaluation network is a multi-task output layer, which maps the deeply abstracted features to operational standardization scores, action efficiency scores, responsiveness scores, and physiological stability scores. The operational standardization scores, action efficiency scores, responsiveness scores, and physiological stability scores together constitute the initial ability score.

[0026] In practice, extracting quantitative deviation features of operational standardization from continuous behavioral trajectories involves the calculation and matching of joint angles. The data processing system calculates the continuous change curves of the shoulder joint abduction angle in the coronal plane and the elbow joint flexion and extension angle over time when the training object performs the standard firefighting operation of "standing upright and holding a gun to spray water". At the same time, it calculates the velocity vector of the training object's torso in the horizontal plane and the actual movement path length from the water source to the final water spraying position. The preset standard action template stores the reference curves and values ​​of joint angles, movement speed and path length under ideal conditions. The actual calculated joint angle curves are aligned and matched with the reference curves in the standard action template through a dynamic time warping algorithm. The cumulative distance deviation between the two curves on the warped path is calculated. This cumulative distance deviation value is extracted as a core quantitative deviation feature of operational standardization. In some embodiments, the speed and path optimization characteristics of action efficiency are obtained through spatiotemporal analysis of continuous behavior trajectories. The system analyzes the time taken from the start of the task (receiving the dispatch instruction) to the completion of the key sub-task (successfully breaching the door obstacle) in the continuous behavior trajectory. It records the total length of the training object's movement path throughout the entire task and calculates the proportion of ineffective movement segments (such as back-and-forth movement or movement outside the task area) to the total path length. Based on the time taken, total length, and proportion of ineffective movement, the speed and path optimization characteristics reflecting action efficiency are calculated. An optional path optimization index calculation formula is provided. in: This represents the path optimization index. This indicates the total length of the actual movement path. This indicates the preset optimal path length. This indicates the actual time elapsed from the start of the task to the completion of the critical subtask. This indicates the preset standard time. This indicates the percentage of invalid moves.

[0027] In practical implementation, the latency and accuracy characteristics of the immediate response are derived from the joint analysis of the task scenario video stream and the operation instruction log. The system detects the timestamps of the flash and explosion sound effects simulating "fire flashover" in the task scenario video stream, and at the same time analyzes the timestamps generated by the training subject's operation of the "emergency evacuation signal button" or "switch water spray mode" command in the operation instruction log. It calculates the time delay from the appearance of the emergency signal to the training subject's first effective response action judged by the system. The accuracy of the response action is evaluated by comparing the degree of conformity between the actual operation sequence executed by the training subject and the standard response operation sequence in the contingency plan for the emergency. The delay time and the degree of conformity of the operation sequence together constitute the latency and accuracy characteristics of the immediate response. In some embodiments, the fluctuation characteristics of physiological stability are statistically derived from physiological state time series. The system statistically analyzes the maximum, minimum, and average values ​​of heart rate and respiratory rate throughout the entire training process, calculates their fluctuation range, records the duration of heart rate exceeding the safety threshold of 180 beats per minute or blood oxygen saturation falling below the 90% threshold, and calculates time-domain and frequency-domain indices of heart rate variability based on the heart rate interval sequence. These fluctuation ranges, over-limit durations, and heart rate variability indices collectively constitute the fluctuation characteristics of physiological stability. It can be understood that the deep evaluation network performs layer-by-layer nonlinear transformation and feature abstraction on the multidimensional evaluation feature set through its multiple hidden layers. The first layer of the deep evaluation network receives a multidimensional evaluation feature set input vector consisting of quantitative deviation features of operational standardization, speed and path optimization features of action efficiency, delay and accuracy features of immediate response, and fluctuation characteristics of physiological stability. The first layer standardizes each dimension of the input vector, scaling it to a distribution with zero mean and unit variance, and achieves preliminary feature dimensionality reduction through linear transformation and activation functions. The deep evaluation network consists of multiple hidden layers in the middle, which are composed of a series of fully connected layers and nonlinear activation functions. These hidden layers perform layer-by-layer nonlinear transformation and feature abstraction on the features processed by the first layer, and learn complex high-order relationships between quantitative deviation features of operational standardization, speed and path optimization features of action efficiency, delay and accuracy features of immediate response, and fluctuation features of physiological stability. For example, a neuron in the middle layer may learn a complex interaction pattern such as "under high heart rate variability, the nonlinear influence of joint angle deviation on the final operational standardization score is weakened".

[0028] In practical implementation, the final layer of the deep evaluation network is a multi-task output layer. This layer maps the high-dimensional feature vectors, after deep abstraction through multiple hidden layers, to four output scores: operational standardization, action efficiency, responsiveness, and physiological stability. These scores collectively constitute the initial capability scores for each evaluation dimension of the training subject in this fire training assessment. Optionally, the multi-task output layer can adopt a structure that shares underlying features but has independent weighted branches. Each branch is responsible for regressing the capability score for a specific dimension. The loss function of the multi-task output layer is a weighted sum of the regression loss terms for each dimension's score. The backpropagation algorithm optimizes all parameters of the deep evaluation network, enabling the network to simultaneously and accurately calculate the operational standardization, action efficiency, responsiveness, and physiological stability scores from the multi-dimensional evaluation feature set.

[0029] In one embodiment of the present invention, the inference computing model loads a preset set of multiple virtual training environment configuration files and a database of contingency events. The inference computing model uses an initial ability score as the initial state and, combined with the trainee's basic physical fitness data, forward-deduces the dynamic changes in their ability state within one or more virtual training environments. During the forward deduction process, the inference computing model randomly or systematically injects contingency events from the contingency event database and simulates the trainee's reactions based on their current ability state and the resulting changes in their ability state. The final ability state of all deduction paths is recorded and compared with a preset ideal state. Based on the comparison results, the reliability weights and deviation ranges of each score in the initial ability score are calculated in reverse, and the initial ability score is weighted and corrected to obtain the corrected ability score. The construction steps of the inference computing model include obtaining the initial ability scores before training, dynamic behavior and physiological data during training, and the final combat capability results of different trainees in historical fire training, constituting the basic data set for model training. The trainee's initial ability score before training, personal basic physical fitness data, and the parameters of the virtual training environment are used as input features. The optimization objective is to match the final capability state obtained from simulation with the actual combat capability results after training. A computational model framework is constructed, including a state transition submodule and a state inversion correction submodule. The computational model framework is trained using a basic dataset, and the state transition probability parameters in the state transition submodule and the score correction weight parameters in the state inversion correction submodule are adjusted so that the final capability state obtained by the model under given input features can fit the corresponding actual combat capability results to the greatest extent, thus completing the construction of the inference computational model.

[0030] In practical implementation, the inference computing model loads various pre-set virtual training environment configuration files and an emergency event library. The virtual training environment configuration files define parameters for different training scenarios. For example, a virtual training environment configuration file for a "multi-story building fire interior attack and search and rescue" task might include the three-dimensional structure of the building's interior, simulated smoke concentration gradient distribution, environmental temperature field change curves, and combustible load parameters for different rooms. The emergency event library stores various pre-set emergency event logics and triggering conditions, such as "room flashover," "partial building collapse," "communication interruption," or "sudden change in the condition of the injured." In an assessment of firefighter Zhang San, the inference computing model uses Zhang San's initial ability scores obtained in the assessment: 85 for operational standardization, 78 for action efficiency, 70 for immediate response, and 65 for physiological stability as the initial state. Simultaneously, it combines Zhang San's basic physical fitness data, such as maximum oxygen uptake and muscle endurance level, to forward deduce the dynamic changes in Zhang San's ability state in a virtual training environment called "high-rise residential fire." In some embodiments, during the forward simulation, the inference computing model injects emergency events from the emergency event library according to preset probability rules. For example, at the 5th minute of the simulation, the system triggers the "sudden change in wind direction at the fire site causing backflow of dense smoke" event based on the event probability. The inference computing model simulates Zhang San's reaction based on his current ability state. For example, a low current immediate response score may lead to a delayed reaction, and a low current physiological stability score may lead to a surge in heart rate, exacerbating judgment errors. The model calculates the changes in ability state caused by these reactions. For example, the immediate response score further decreases due to the delay, and the physiological stability score decreases due to stress. The inference computing model records the final ability state of all simulation paths. For example, after 1000 Monte Carlo simulations, 1000 sets of final operational standardization scores, action efficiency scores, immediate response scores, and physiological stability scores are obtained, and the median or mean of these simulation results is compared with the preset ideal state.

[0031] In practice, based on the comparison results, the inference calculation model reverse-calculates the credibility weights and deviation ranges of each score in the initial ability score, and then performs a weighted correction on the initial ability score to obtain the corrected ability score. A method for calculating the corrected ability score is provided. Optional formulas: in: This represents the corrected ability score for the k-th ability dimension. This represents the initial ability score for the k-th ability dimension. This represents the expected value of the final state score of the k-th capability dimension across all deduction paths. The credibility weight coefficient is calculated based on the deviation between the initial capability score and the distribution of the inference results. This is a normalization factor related to environmental difficulty and the frequency of emergencies. In this way, dimensions with high consistency with multiple simulation results in the initial capability score receive higher weights, while dimensions that deviate significantly from the ideal state or have high dispersion in different simulation paths are corrected based on the simulation expectation, resulting in a more robust corrected capability score. It can be understood that the construction of the inference calculation model relies on historical data. The construction steps include obtaining the initial capability scores of different training subjects before training, dynamic behavior and physiological data during training, and the final combat capability results after training, forming the basic data set for model training. In specific implementation, the initial capability scores of the training subjects before training, their basic physical fitness data, and the parameters of the virtual training environment they face are used as input features. The matching degree between the final capability state obtained from the simulation and the actual combat capability results after training is used as the optimization objective. A computational model framework including a state transition submodule and a state inversion correction submodule is constructed. The state transition submodule is responsible for simulating the probability of transitioning from the current capability state to the next capability state under given environment and unforeseen circumstances. The state inversion correction submodule is responsible for calculating the correction weights for the initial input based on the difference between the inference result and the ideal state. The computational model framework is trained using a basic dataset, and gradient descent or other optimization algorithms are used to adjust the state transition probability parameters in the state transition submodule and the score correction weight parameters in the state inversion correction submodule. This ensures that the final capability state inferred by the model under given input features best fits the corresponding actual combat capability result, thus completing the construction of the inference computational model. In some embodiments, cross-validation is used during training to evaluate the fitting effect and prevent overfitting, ensuring that the trained inference computational model can effectively correct and invert the initial capability score of new training objects.

[0032] See Figure 3This is a dynamic chart showing the change in firefighters' overall ability scores under different virtual training environments. It illustrates the decay trend of firefighters' comprehensive ability scores over time in four virtual environments during fire training assessments. This is a core visualization result from the state inversion reasoning calculation – forward deduction phase. Initially, the comprehensive ability score is 85 points in all environments, representing the firefighter's initial ability level. The score decays the slowest in the conventional environment, remaining at 76 points after 9 minutes, indicating that conventional training causes the least ability loss. The decay rate is moderate in the dense smoke environment, dropping to 62 points after 9 minutes; smoke interference significantly affects operational performance. The decay is even faster in the high-temperature environment, dropping to 58 points after 9 minutes; high temperatures exacerbate physiological stress and judgment errors. The decay is fastest in the complex structure environment, with the score dropping to only 57 points after 9 minutes; complex building structures significantly increase task difficulty and ability consumption. The harsher the environment, the faster the ability score decreases, reflecting the significant negative impact of environmental stress on firefighters' overall performance.

[0033] In one embodiment of the present invention, the inference calculation model maps the corrected ability score to a preset assessment standard level system to determine the training object's level in four dimensions: operational standardization, action efficiency, immediate response, and physiological stability. Based on the levels of the four dimensions, the inference calculation model uses a radar chart model to construct a comprehensive ability profile. Each axis of the radar chart represents an ability dimension, the scale of the axis represents the level, and the corrected ability score determines the vertex position of the radar chart on the axis. The area, centroid, and perimeter of the comprehensive ability profile radar chart are calculated and quantified into a set of profile feature vectors. The profile feature vectors are matched with the three standard profiles of excellent, qualified, and needing improvement stored in the fire training big data knowledge base to determine the overall assessment category of the training object. The levels of each dimension in the comprehensive ability profile are compared longitudinally with the average level of the training object's historical assessments to identify dimensions with improved levels as potential strength patterns and dimensions with decreased levels or below average levels as potential weakness patterns. For each identified potential weakness pattern, the model backtracks to the multi-dimensional assessment feature set corresponding to the initial ability score of the generated dimension. Further, based on the time points corresponding to the multidimensional evaluation feature set, specific time periods are located in the original real-time training data stream. From these specific time periods, corresponding continuous behavioral trajectories and physiological state time series are extracted. Analyzing specific behavioral segments of abnormal actions, operational errors, slow responses, or exceeding physiological limits of the training subjects within these specific time periods helps identify the bottleneck pattern.

[0034] In practice, the inference calculation model maps the corrected ability scores to a preset assessment standard level system. The assessment standard level system maps the percentage score of each ability dimension to discrete levels. For example, a preset mapping rule is that a score between 90 and 100 corresponds to the "excellent" level, between 80 and 89 corresponds to the "good" level, between 70 and 79 corresponds to the "qualified" level, and below 70 corresponds to the "needs improvement" level. Suppose that in an assessment, the corrected ability scores of the training subject Li Si are 85 points for operational standardization, 78 points for action efficiency, 70 points for responsiveness, and 65 points for physiological stability. The inference calculation model maps the 85 points for operational standardization to the "good" level, the 78 points for action efficiency to the "qualified" level, the 70 points for responsiveness to the "qualified" level, and the 65 points for physiological stability to the "needs improvement" level. In some embodiments, based on four-dimensional levels, the inference computation model uses a radar chart model to construct a comprehensive capability profile. The four axes of the radar chart represent four capability dimensions: operational standardization, action efficiency, immediate response, and physiological stability. The scale of each axis increases from the center outwards, representing a level from "needs improvement" to "excellent". The corrected capability score determines the vertex position of the radar chart on the corresponding axis. For example, the physiological stability axis corresponds to a score of 65, and its vertex falls in the "needs improvement" level range closer to the center. The operational standardization axis corresponds to a score of 85, and its vertex falls in the "good" level range further away from the center. Connecting the four vertices forms a closed polygon, which is the comprehensive capability profile of the training object. In practical implementation, the comprehensive capability profile needs to be quantified into a set of profile feature vectors for subsequent analysis. The area, centroid, and perimeter of the comprehensive capability profile radar image are calculated. The area of ​​the radar image is obtained by calculating the area of ​​the closed polygon, the centroid is obtained by calculating the arithmetic mean coordinates of the vertices of the polygon, and the perimeter is obtained by calculating the sum of the lengths of the sides of the polygon. In order to make a mathematical comparison with consistent dimensions, the original calculated values ​​of area, centroid coordinates, and perimeter are normalized. The area is divided by the maximum possible area of ​​the radar image, the centroid coordinates are divided by the radius of the radar image for normalization, and the perimeter is divided by the maximum possible perimeter of the radar image. The processed area, the X component of the normalized centroid coordinates, the Y component of the normalized centroid coordinates, and the normalized perimeter together constitute a set of dimensionless profile feature vectors. The inference calculation model performs similarity matching between the profile feature vector and the feature vectors of three standard profiles (excellent, qualified, and requiring improvement) stored in the fire training big data knowledge base. The feature vectors of the standard profiles also consist of normalized area, centroid coordinates, and perimeter. One method for calculating similarity is to use weighted Euclidean distance. This calculation method is as follows: in: This indicates that the feature vector of the comprehensive capability profile in this assessment is related to the first [item] in the knowledge base. The weighted Euclidean distance between the feature vectors of the standard profiles; the smaller the distance value, the higher the similarity. This represents the first eigenvector of the current profile. A normalized component, Indicates the first The first standard profile feature vector A normalized component, These are the preset weighting coefficients for the corresponding components, and satisfy the following conditions: All components , All are dimensionless normalized values, weights Since it is a dimensionless coefficient, the distance is... Since it is a dimensionless scalar, the dimensions of both sides of the formula are kept consistent. After calculating the distance between the current profile and the three standard profiles of "needs improvement", "qualified" and "excellent", the category corresponding to the standard profile with the smallest distance is determined as the overall evaluation category of the training object.

[0035] In practice, the comprehensive ability profile is used to identify the strengths and weaknesses of training subjects across different ability dimensions. Specifically, the levels of each dimension in the comprehensive ability profile are compared longitudinally with the average level of the training subject's historical assessments. Dimensions showing improvement are identified as potential strengths, while dimensions showing decline or falling below the historical average are identified as potential weaknesses. See Table 1 for an example of this comparison.

[0036] Table 1: Comparison of Training Subjects' Comprehensive Ability Profile with Historical Data Understandably, based on the comparison results shown in Table 1, the evaluation levels of trainee Li Si in the three dimensions of operational standardization, action efficiency, and responsiveness are on par with the historical average, with no advantageous patterns identified. However, the physiological stability dimension has decreased from the historical average of "qualified" to the "needs improvement" level in this evaluation, thus physiological stability is identified as a potential weakness. For each identified potential weakness, the system needs to backtrack to the multidimensional evaluation feature set that generated the initial ability score for that dimension. For the "physiological stability" weakness, the system backtracks to the fluctuation characteristics of physiological stability used when generating the physiological stability score. These fluctuation characteristics include specific values ​​such as heart rate fluctuation range, duration of blood oxygen saturation exceeding the threshold, and heart rate variability index. Further, based on the time points corresponding to the multidimensional evaluation feature set, the system locates the specific time period in the original real-time training data stream. For example, the system finds that the feature data causing excessively low heart rate variability index and excessively long duration of heart rate exceeding the threshold mainly correspond to the time period from the 10th to the 15th minute in the real-time training data stream. By extracting corresponding continuous behavioral trajectories and physiological state time series from specific time periods, the system analyzes specific behavioral segments of abnormal actions, operational errors, slow responses, or exceeding physiological limits of the training subjects within those time periods to identify the bottleneck pattern. For example, the system extracts continuous behavioral trajectories from the 10th to the 15th minute, showing that the training subject, Li Si, is performing a "weighted climbing" task. The synchronized physiological state time series shows that his heart rate rapidly exceeds the safety threshold at 11 minutes and 30 seconds and remains there for about 3 minutes. Combined with the task scene video stream, it can be confirmed that Li Si exhibited behavioral segments of disordered breathing rhythm and sluggish gait during this time period, thus confirming the "physiological stability" bottleneck pattern and its specific reasons for its occurrence in the high-intensity weighted climbing segment.

[0037] See Figure 4 This is a radar chart illustrating the comprehensive capabilities of firefighters in training and assessment. It compares the current assessment of firefighters with their historical average performance across four core dimensions, representing a visual outcome of the comprehensive capability profile generation and weakness identification phase. Operational standardization and action efficiency are both above the historical average, indicating that firefighters demonstrate stable performance in basic operational proficiency and execution efficiency, which are their core strengths. Response timeliness and physiological stability are both below the historical average, representing weaknesses that require immediate improvement. Insufficient response timeliness may lead to delayed reactions and errors in handling emergencies. Insufficient physiological stability may result in heart rate spikes and physical exhaustion during prolonged or high-intensity training. The current capability profile exhibits a "strong operational capability, weak response capability" characteristic, indicating a biased capability structure that necessitates a targeted reinforcement training program.

[0038] In one embodiment of the present invention, based on the capability dimension corresponding to the weakness pattern, several basic training subjects specifically designed to improve that capability dimension are matched from a preset training subject library. Combined with specific abnormal behavior fragments in the location information, the matched basic training subjects are adjusted to design targeted repetitive training, decomposed training, or anti-interference training content. Based on the overall assessment category and basic physical fitness data of the training subject, the training intensity, duration, and repetition count are set for each designed training content, forming a preliminary reinforcement training module. All reinforcement training modules targeting different weakness patterns are integrated and sorted according to the principle of starting with the easy and progressing to the difficult, and complementing skills, to form a complete training cycle plan, i.e., a personalized reinforcement training scheme recommendation. The real-time training data stream, multi-dimensional assessment feature set, initial capability score, comprehensive capability profile, identified weakness pattern, and the finally generated personalized reinforcement training scheme recommendation generated from each training assessment are stored as a complete case in the fire training big data knowledge base. Massive amounts of case data are periodically extracted from the fire training big data knowledge base to retrain the deep assessment network, updating its network weights to make its assessment capability scores more accurate. At the same time, by utilizing the enhanced training programs and their corresponding capability improvement results stored in the fire training big data knowledge base that have been verified by actual training, the deduction rules and parameters of state inversion in the reasoning calculation model are optimized.

[0039] In practice, basic training subjects are matched from a pre-set training subject library based on the capability dimensions corresponding to the weakness patterns. For example, if the weakness patterns of the training subject Wang Wu are identified as "immediacy of response" and "standardization of operation", the system will match several basic training subjects specifically for improving "immediacy of response" from the training subject library, such as "sudden fire signal recognition and response training" and "multi-task parallel handling of interference training". At the same time, basic training subjects specifically for improving "standardization of operation" will be matched, such as "standardized process training for laying water hoses in confined spaces" and "decomposition training of demolition equipment operation actions". By combining specific abnormal behavior fragments from the location information, the matched basic training subjects are adjusted to design targeted training content. For example, the location information shows that Wang Wu had a 3-second response delay and an incorrect "forward water spraying" operation in the case of "sudden deflagration". For the basic subject of "sudden fire signal recognition and response training", the system designs repetitive training content that adds the sudden signal of "deflagration flash and loud noise" and forces the execution of "emergency evacuation" before "cutting off the fire with the water distributor". For "standardized process training for laying water hose in confined space", based on the "water hose kinking" behavior fragment that appears in the operation, the system designs decomposed training content that focuses on strengthening "water hose release techniques" and "tidying up while moving".

[0040] In some embodiments, training parameters are set for each training content designed based on the overall assessment category and basic physical fitness data of the trainee. The overall assessment category of trainee Wang Wu is "needs improvement". His personal basic physical fitness data shows that his cardiopulmonary function rating is C and his upper limb strength rating is B. The system sets the training intensity for the targeted "sudden fire signal recognition and response" repetitive training content to "moderate to high", the duration to be 15 minutes per training session, and the repetition frequency to 3 times per week. The training intensity for the "water hose release technique" decomposition training content is set to "moderate", the duration to be 20 minutes per training session, and the repetition frequency to 2 times per week. These training intensities, durations, and repetition frequencies together constitute the initial reinforcement training module. Integrate all reinforcement training modules targeting different weakness patterns and sort them according to the principle of starting with the easy and progressing to the difficult, and complementing skills to form a complete training cycle plan. For example, combine the decomposition training of "water hose release technique" with "multi-task parallel handling of interference training" in the same training day, first conduct relatively simple action decomposition training and then conduct complex comprehensive anti-interference training, and arrange the different training modules within a week to form a personalized reinforcement training plan suggestion for four weeks, with the focus of training content on each week gradually increasing.

[0041] It is understandable that constructing a fire training big data knowledge base and using it to iteratively optimize the deep assessment network and inference calculation model is an ongoing process. In specific implementation, each training assessment generates a real-time training data stream, multi-dimensional assessment feature set, initial capability score, comprehensive capability profile, identified weakness patterns, and finally generated personalized reinforcement training scheme suggestions, all of which are stored as a complete case in the fire training big data knowledge base. The case storage format includes structured assessment result data and unstructured raw sensor data indexes. Massive amounts of case data are periodically extracted from the fire training big data knowledge base to retrain the deep assessment network and update its network weights. For example, after accumulating 500 new cases, the system uses the multi-dimensional assessment feature set from all historical cases as input and the capability score, corrected by manual calibration or back-calculation from actual combat results, as the target output to fully retrain the deep assessment network, making its assessment capability scores more accurate. Simultaneously, the inference rules and parameters of the state inversion in the inference calculation model are optimized by utilizing the effective reinforcement training programs and their corresponding capability improvement results stored in the fire training big data knowledge base. For example, the knowledge base records 100 cases where the physiological stability scores of trainees improved after adopting the "intermittent stair climbing training for tachycardia" program. When the inference calculation model infers the state of trainees with similar weakness patterns, it will adjust the state transition probability parameters related to "physiological stability recovery after high-intensity cardiopulmonary load" in the state transition submodule to make its inference more consistent with the actual training effect. Optionally, the optimization process of the knowledge base can adopt an incremental learning strategy formula, and the score correction weight parameters of the state inversion correction submodule in the inference calculation model can be adjusted. The parameters can be dynamically adjusted based on the effectiveness of historical solutions. Update method: in: This indicates the updated parameters. This indicates the parameters before the update. Indicates the learning rate. This represents the actual measured improvement in ability after adopting the historical reinforcement training program. This represents the improvement in the predictive ability of the inference and computation model. Through this iteration, the model's inference and prediction continuously approach the actual data.

[0042] See Figure 5This is a diagram analyzing the iterative effects of the deep evaluation network in the intelligent assessment system for fire training and assessment. It illustrates the iterative optimization effect of the deep evaluation network as training cases accumulate. The assessment accuracy shows a continuous upward trend with the increase in the number of accumulated cases, rising from an initial 0.72 to 0.94, indicating that the model's assessment accuracy for fire training capabilities significantly improves after training with more data. The growth rate is stable, with an average increase in accuracy of approximately 0.04-0.05 for every 100 additional cases. The assessment error shows a continuous downward trend with the increase in the number of accumulated cases, decreasing from an initial 0.26 to 0.06, indicating that the deviation between the model's predictions and the actual values ​​is constantly narrowing. The magnitude of the error decrease highly corresponds to the magnitude of the accuracy increase, reflecting the consistency of model optimization. When the number of cases reaches 500, the accuracy approaches 0.95, and the error drops to 0.06, indicating that the model has entered a relatively high accuracy range, and the marginal benefit of further adding cases may gradually decrease.

[0043] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A big data-based intelligent assessment and management method for fire training and assessment, characterized in that: include: Obtain the initial competency scores of the trainees in each assessment dimension of the fire training assessment; The initial capability score is input into the inference calculation model based on state inversion. The inference calculation model corrects and inverses the initial capability score by simulating the state evolution under different training environments and sudden situations. The inference and calculation model based on state inversion generates a comprehensive ability profile of the training object under a preset assessment standard. The comprehensive ability profile includes scores of multiple interrelated ability dimensions. Based on the comprehensive capability profile, the strengths and weaknesses of the training subjects in different capability dimensions are identified, and the specific behavioral segments or physiological abnormal periods that cause the weakness patterns to appear are located. Based on the identified weakness patterns and their location information, a personalized reinforcement training plan suggestion is generated for the training object. The personalized reinforcement training plan suggestion includes the skills that need to be focused on training, the suggested training scenarios, and the expected physiological indicator target range. By integrating all assessment data of historical training subjects, a fire training big data knowledge base is constructed, and the deep assessment network and the inference calculation model are iteratively optimized using the fire training big data knowledge base.

2. The intelligent assessment and management method for fire training and assessment based on big data as described in claim 1, characterized in that, The acquisition of the initial competency scores of the training subjects in each assessment dimension of the fire training assessment includes: Continuously collect real-time training data streams from fire training subjects, including multi-sensor motion data, physiological index data, task scene video streams, and operation instruction logs; The real-time training data stream is synchronized and fused with multi-source heterogeneous data to construct the continuous behavioral trajectory of the training object in three-dimensional space and its accompanying physiological state time series. Based on the continuous behavioral trajectory and the physiological state time series, a multi-dimensional evaluation feature set reflecting the training object's operational standardization, action efficiency, responsiveness and physiological stability is extracted; Construct a deep evaluation network containing multiple hidden layers, and input the multidimensional evaluation feature set into the deep evaluation network; The deep evaluation network performs layer-by-layer nonlinear transformation and feature abstraction on the multidimensional evaluation feature set through its multiple hidden layers, and calculates the initial ability scores of the training object in each evaluation dimension in the fire training assessment. The real-time training data stream is synchronized and fused with multi-source heterogeneous data to construct the continuous behavioral trajectory of the training object in three-dimensional space and its accompanying physiological state time series, specifically including: The motion data from the multi-sensor sensors, the physiological index data, the task scene video stream, and the operation instruction log from different sensors are all timestamped to achieve millisecond-level time alignment. From the video stream of the task scene, the three-dimensional coordinates of the key points of the human skeleton of the training object are extracted in real time using visual recognition technology to form a visual behavior trajectory sequence. The attitude calculation is performed on the inertial measurement unit data in the multi-sensor motion data, and spatial registration and Kalman filtering are performed on the visual behavior trajectory sequence to generate the continuous behavior trajectory. Synchronously, the heart rate, blood oxygen, and body temperature data in the physiological indicators are aligned with each point in time of the continuous behavioral trajectory according to the timestamp, forming a synchronized time series of the physiological state.

3. The intelligent assessment and management method for fire training and assessment based on big data as described in claim 2, characterized in that, Based on the continuous behavioral trajectory and the physiological state time series, a multi-dimensional evaluation feature set reflecting the training subject's operational standardization, action efficiency, responsiveness, and physiological stability is extracted, including: From the continuous behavioral trajectory, calculate the joint angles, movement speeds, and path lengths of the training subject when performing standard fire-fighting operations, and perform dynamic time warping matching with the preset standard action template to extract the quantitative deviation features of the operational standardization. By analyzing the continuous behavioral trajectory, including the time taken from the start of the task to the completion of the key sub-task, the total length of the movement path, and the proportion of invalid movements, the speed and path optimization characteristics of the action efficiency are calculated. The time delay between the appearance of the emergency signal and the first effective response action of the training object is detected in the task scene video stream and the operation instruction log, as well as the accuracy of the response action, and the delay and accuracy features of the immediate response are extracted. By statistically analyzing the fluctuation range of the physiological state time series throughout the entire training process and the duration of exceeding a set threshold, the heart rate variability index is calculated to obtain the fluctuation characteristics of the physiological stability.

4. The intelligent assessment and management method for fire training and assessment based on big data as described in claim 3, characterized in that, The deep evaluation network performs layer-by-layer nonlinear transformation and feature abstraction on the multidimensional evaluation feature set through its multiple hidden layers, calculating the initial capability scores of the training object in each evaluation dimension in the fire training assessment, including: The first layer of the deep evaluation network standardizes and performs preliminary feature dimensionality reduction on the input multidimensional evaluation feature set; The deep evaluation network uses multiple hidden layers in the middle to learn complex high-order relationships between the quantitative deviation characteristics of operational norms, the speed and path optimization characteristics of action efficiency, the delay and accuracy characteristics of strain immediacy, and the fluctuation characteristics of physiological stability through nonlinear activation functions. The last layer of the deep evaluation network is a multi-task output layer, which maps the deeply abstracted features to operational standardization scores, action efficiency scores, responsiveness scores, and physiological stability scores. The operational standardization scores, action efficiency scores, responsiveness scores, and physiological stability scores together constitute the initial capability score.

5. The intelligent assessment and management method for fire training and assessment based on big data as described in claim 4, characterized in that, The initial capability score is input into a state-inversion-based inference model. This inference model, by simulating state evolution under different training environments and unexpected situations, corrects and inverses the initial capability score, including: The inference computing model loads a variety of preset virtual training environment configuration files and a database of unexpected events; The inference calculation model uses the initial ability score as the initial state and combines it with the individual basic physical fitness data of the training object to positively deduce the dynamic change process of its ability state in one or more of the virtual training environments. During the forward inference process, the inference calculation model randomly or according to rules injects emergency events from the emergency event library and simulates the training object's reaction based on its current capability state and the resulting change in capability state. Record the final capability state of all deduction paths and compare it with the preset ideal state; Based on the comparison results, the credibility weights and deviation ranges of each score in the initial capability score are calculated in reverse, and the initial capability score is weighted and corrected to obtain the corrected capability score.

6. The intelligent assessment and management method for fire training and assessment based on big data as described in claim 5, characterized in that, The inference and computation model based on state inversion generates a comprehensive ability profile of the training subject under a preset assessment standard, including: The reasoning and calculation model maps the corrected ability score to a preset assessment standard level system to determine the training object's level in four dimensions: operational standard, action efficiency, immediate response, and physiological stability. Based on the four-dimensional levels, the reasoning and calculation model uses a radar chart model to construct the comprehensive capability profile. Each axis of the radar chart represents a capability dimension, the scale of the axis represents the level, and the corrected capability score determines the position of the vertex of the radar chart on the axis. Calculate the area, centroid, and perimeter of the integrated capability profile radar image, and quantify them into a set of profile feature vectors; The profile feature vector is matched with the three standard profiles (excellent, qualified, and needing improvement) stored in the fire training big data knowledge base to determine the overall evaluation category of the training object.

7. The intelligent assessment and management method for fire training and assessment based on big data as described in claim 6, characterized in that, Based on the comprehensive capability profile, the training subject's strengths and weaknesses across different capability dimensions are identified, and the specific behavioral segments or physiological abnormalities leading to the weakness patterns are pinpointed, including: The levels of each dimension in the comprehensive capability profile are compared longitudinally with the average level of the training subjects' historical evaluations. Dimensions with improved levels are identified as potential strength patterns, and dimensions with decreased levels or below average levels are identified as potential weakness patterns. For each identified potential weakness pattern, backtrack to the multidimensional assessment feature set that generated the initial capability scores for each dimension; Further, based on the time points corresponding to the multidimensional evaluation feature set, the specific time period in the original real-time training data stream is located; Extract the corresponding continuous behavioral trajectory and physiological state time series from the specific time period, analyze the specific behavioral segments of abnormal actions, operational errors, slow response or physiological index exceeding limits of the training object within the specific time period, and confirm the bottleneck pattern.

8. The intelligent assessment and management method for fire training and assessment based on big data as described in claim 7, characterized in that, The step of generating personalized reinforcement training program suggestions for the training subjects based on the identified weakness patterns and their location information includes: Based on the capability dimension corresponding to the aforementioned weakness pattern, several basic training subjects specifically designed to improve the aforementioned capability dimension are matched from a preset training subject library. Based on the specific abnormal behavior fragments in the location information, the matched basic training subjects are adjusted to design targeted repetitive training, decomposition training or anti-interference training content. Based on the overall assessment category and basic physical fitness data of the trainees, the training intensity, duration and repetitions are set for each training item to form an initial reinforcement training module. Integrate all reinforcement training modules targeting different weakness patterns, and sort them according to the principle of starting with the easy and progressing to the difficult, and complementing skills, to form a complete training cycle plan, namely the personalized reinforcement training program recommendation.

9. The intelligent assessment and management method for fire training and assessment based on big data as described in claim 8, characterized in that, The construction of a fire training big data knowledge base, and the iterative optimization of the deep evaluation network and the inference calculation model using the fire training big data knowledge base, includes: The real-time training data stream generated from each training evaluation, the multi-dimensional evaluation feature set, the initial capability score, the comprehensive capability profile, the identified weakness patterns, and the finally generated personalized reinforcement training scheme suggestions are stored as a complete case in the fire training big data knowledge base. A large amount of case data is periodically extracted from the fire training big data knowledge base to retrain the deep evaluation network and update its network weights, so that its evaluation ability score is more accurate. Meanwhile, by utilizing the enhanced training programs and their corresponding capability improvement results stored in the fire training big data knowledge base that have been verified to be effective through actual training, the deduction rules and parameters of state inversion in the reasoning calculation model are optimized.

10. The intelligent assessment and management method for fire training and assessment based on big data as described in claim 9, characterized in that, The steps for constructing the inference computation model include: The initial capability scores of different trainees before training, dynamic behavior and physiological data during training, and the actual combat capability results of the final assessment after training are obtained from historical fire training to form the basic data set for model training. The initial ability score of the trainees before training, their basic physical fitness data, and the parameters of the virtual training environment they face are used as input features. The optimization objective is to match the final capability state obtained from the simulation with the actual combat capability results after training. Construct a computational model framework that includes a state transition submodule and a state inversion correction submodule; The computational model framework is trained using the basic dataset. The state transition probability parameters in the state transition submodule and the score correction weight parameters in the state inversion correction submodule are adjusted so that the final capability state inferred by the model under given input features can fit the corresponding actual combat capability result to the greatest extent, thus completing the construction of the inference computational model.

Images