A method and system for monitoring and intervening in the physical fitness of adolescents based on multi-modal fusion

Abstract

The present application belongs to the technical field of medical health, and particularly relates to a youth physique dynamic monitoring and intervention method and system based on multi-modal fusion. Physique data is collected and processed through a stereoscopic vision sensor array and a deep learning algorithm, multi-modal psychological data is collected through an AI drawing interactive process, and a personalized intervention scheme is generated through cross-modal semantic fusion and intelligent decision-making, thereby solving the problems of single data dimension, imperfect evaluation system, homogeneous intervention and scattered data of a traditional monitoring system, and realizing full-process intelligent and personalized dynamic management of youth physique health.

Description

Technical Field

[0001] This invention belongs to the field of medical and health technology, specifically relating to a method and system for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion. Background Technology

[0002] Adolescent physical health management is an important branch of health management. Currently, wearable device-based physiological parameter monitoring systems and psychological questionnaire systems are widely used. However, wearable device-based physiological parameter monitoring systems primarily focus on basic physiological indicators such as heart rate and step count, failing to acquire in-depth physical health data such as movement posture quality and movement standardization. This results in insufficient accuracy and specificity in monitoring data related to the physical characteristics of adolescents. Relying on periodic physical tests throughout the school year leads to low data collection frequency and coarse data granularity, making it impossible to construct a continuous health status trajectory. Intervention programs generated based on sparse datasets lack individual variability, making it difficult to achieve precise personalized health management, resulting in poor user compliance and intervention effectiveness.

[0003] Traditional psychological assessments rely on paper or electronic questionnaires, and data collection is greatly affected by adolescents' expressive abilities and subjective will, making it difficult to guarantee authenticity. In addition, the questionnaire collection cycle is long, making it impossible to capture dynamic changes in psychological state, resulting in delayed psychological assessment results and making it difficult to achieve early intervention.

[0004] Furthermore, the aforementioned physical and psychological data are scattered across different systems, lacking an effective multimodal data fusion mechanism, making it impossible to form a unified health profile of students. Summary of the Invention

[0005] To address the problems in the background technology, this invention provides a method and system for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion.

[0006] The technical solution of the present invention is as follows: This invention provides a method for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion, comprising: S1: Acquire the video stream of youth sports at the target sports venue, extract the three-dimensional coordinate values ​​of multiple target skeletal points in the video stream in real time, and construct the temporal trajectory data of the target skeletal points; extract the temporal features of the temporal trajectory data of the target skeletal points and map them to preset sports quality indicators, calculate the scores of each sports quality indicator; according to the individual ID, integrate the scores of each sports quality indicator with multidimensional physical fitness data into a multidimensional physical fitness feature vector; S2: Acquire the speech data, temporal intensity data, and image data of the painting process frames of teenagers during the AI ​​painting interaction process, extract features from each, and obtain the corresponding speech features, intensity features, and painting image features; after aligning by timestamp, perform weighted fusion to obtain a multi-dimensional psychological feature vector; S3: Multidimensional physical feature vectors and multidimensional psychological feature vectors are processed through multi-head attention to generate multidimensional health status embedding vectors. Then, risk prediction is performed using a gated recurrent unit to obtain the risk prediction result. Using risk prediction results as the state space and intervention strategies as the action space, the intervention strategies are optimized through reinforcement learning to generate personalized plans that include exercise prescriptions, psychological adjustment, and nutritional advice.

[0007] In addition, it also includes updating personalized plans using federated learning, specifically: Personalized plans are transformed into gamified tasks, class health heatmaps, and family suggestions, which are then sent to students, teachers, and parents for execution. Feedback data is collected from each platform, and federated learning is used to aggregate the feedback data to update the personalized plans.

[0008] Furthermore, S3 describes using a gated loop unit to perform risk prediction and obtain the risk prediction result, specifically as follows: The multidimensional health status embedding vectors are arranged into a time series in chronological order. After being processed by a multi-layer gated recurrent unit network, they are mapped to probability values ​​through a sigmoid activation function, corresponding to the predicted probabilities of various health risks.

[0009] Furthermore, the multidimensional physical feature vector and multidimensional psychological feature vector mentioned in S3 are processed through multi-head attention to generate a multidimensional health state embedding vector, specifically: Based on the standardized multidimensional physical and psychological feature vectors, physical feature node sets, psychological feature node sets, and directed edges are constructed respectively; the resulting heterogeneous physical and psychological graph is processed by a heterogeneous graph attention network to generate a graph global fusion feature vector. After concatenating the graph global fusion feature vector, the standardized multidimensional physical feature vector, and the multidimensional psychological feature vector into a cross-modal feature matrix, it is integrated into an attention output matrix through multi-head attention processing. Global average pooling along the temporal dimension is performed on the attention output matrix to obtain a multidimensional health state embedding vector.

[0010] Furthermore, the temporal features described in S1 include: target bone point displacement velocity, joint angle change rate, and center of gravity displacement standard deviation.

[0011] Furthermore, the speech features mentioned in S2 are obtained by extracting Mel frequency cepstral coefficients from the speech data and then processing them using the BERT model.

[0012] Furthermore, the force characteristics described in S2 include peak force, fluctuation range, average force, and force changes before and after the stroke pauses.

[0013] Furthermore, the painting image features mentioned in S2 are obtained by extracting the HSV histogram from the painting process frame image data and then processing it using the Vision Transformer model.

[0014] Furthermore, after obtaining the multidimensional psychological feature vector as described in S2, the process also includes constructing a dynamic psychological profile, specifically: Based on multidimensional psychological feature vectors, stress-related features at different time points are extracted, stress scores at each time point are calculated, and the ratio of the change in stress scores between adjacent time points to the time interval is calculated as the gradient of stress level change. Based on multidimensional psychological feature vectors, the categories of adolescent emotional states are determined, and the ratio of the number of times a certain emotional state changes to another emotional state to the total number of times the emotional state occurs is used as the probability of emotional state transition. Based on multidimensional psychological feature vectors, creativity features are quantified, and the creativity index is obtained by weighted summation according to preset weights. Dynamic psychological profiles are constructed using stress level change gradient, emotional state shift probability, and creativity index as nodes.

[0015] This invention also provides a multimodal fusion-based dynamic monitoring and intervention system for adolescent physical fitness, comprising: Physical fitness data processing module: used to acquire video streams of adolescent sports activities at target sports venues, extract the three-dimensional coordinates of multiple target skeletal points in the video streams in real time, construct temporal trajectory data of target skeletal points; extract the temporal features of the temporal trajectory data of target skeletal points and map them to preset sports quality indicators, calculate the scores of each sports quality indicator; and integrate the scores of each sports quality indicator with multidimensional physical fitness basic data into a multidimensional physical fitness feature vector according to individual ID. The psychological data processing module is used to acquire speech data, temporal intensity data, and image data of the painting process frames of teenagers during the AI ​​painting interaction process, extract features from each, and obtain corresponding speech features, intensity features, and painting image features; after being aligned by timestamp, they are weighted and fused to obtain a multi-dimensional psychological feature vector; Risk prediction and intervention module: Multidimensional physical feature vectors and multidimensional psychological feature vectors are processed through multi-head attention to generate multidimensional health status embedding vectors. Then, risk prediction is performed using a gated recurrent unit to obtain the risk prediction results. Using risk prediction results as the state space and intervention strategies as the action space, the intervention strategies are optimized through reinforcement learning to generate personalized plans that include exercise prescriptions, psychological adjustment, and nutritional advice.

[0016] Beneficial effects This invention collects and processes physical fitness data through a stereoscopic vision sensor array and deep learning algorithms, collects multimodal psychological data through AI drawing interaction, and generates personalized intervention plans through cross-modal semantic fusion and intelligent decision-making. It solves the problems of single data dimension, imperfect evaluation system, homogeneous intervention and data dispersion in traditional monitoring systems, and realizes intelligent and personalized dynamic management of adolescent physical health throughout the entire process. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a method for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion, as described in Example 1.

[0018] Figure 2 This is a schematic diagram of the adaptive intervention process in step S3 of Example 1. Detailed Implementation

[0019] The following examples are intended to illustrate the present invention, and not to further limit the invention.

[0020] Example 1 like Figure 1 This embodiment provides a method for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion, including: S1: Acquire the video stream of youth sports at the target sports venue, extract the three-dimensional coordinate values ​​of multiple target skeletal points in the video stream in real time, and construct the temporal trajectory data of the target skeletal points; extract the temporal features of the temporal trajectory data of the target skeletal points and map them to preset sports quality indicators, calculate the scores of each sports quality indicator; according to the individual ID, integrate the scores of each sports quality indicator with multidimensional physical fitness basic data into a multidimensional physical fitness feature vector.

[0021] Considering the smaller bone structure and greater range of motion of adolescents, this invention uses a stereoscopic vision sensor array deployed at the target sports venue to collect video streams of adolescents' movements, ensuring the continuity and accuracy of motion posture capture.

[0022] Furthermore, to quantify the quality of adolescent athletic performance, the temporal characteristics include: target skeletal point displacement velocity, joint angle change rate, and standard deviation of center of gravity displacement.

[0023] In addition, to facilitate the establishment of a personalized longitudinal physical fitness database, the scores of various exercise quality indicators are integrated with multidimensional physical fitness baseline data into a multidimensional physical fitness feature vector according to individual ID, specifically as follows: Using the individual ID of the adolescent as an index, the scores of various sports quality indicators are integrated and standardized with multidimensional physical fitness data. The standardized data is then processed by a fully connected layer to obtain a multidimensional physical fitness feature vector identified by the individual ID.

[0024] In practice, the process of non-visual physical fitness monitoring and dynamic database construction is as follows: S11: Multiple high-definition stereo vision sensors are deployed in sports venues such as school playgrounds and gymnasiums. The sensors are connected to edge computing nodes via a 5G private network. When teenagers are running, jumping rope, playing basketball, or engaging in other sports, the edge computing nodes automatically trigger the sensors to collect video streams. The human joint point algorithm extracts the three-dimensional coordinate values ​​of 21 target skeletal points (head, shoulder, elbow, wrist, hip, knee, ankle, etc.) in real time, generating 30 sets of coordinate data per second. This constructs the temporal trajectory data of the target skeletal points, forming a spatiotemporal model of the movement posture.

[0025] S12: Based on the temporal convolutional network (TCN), analyze the temporal trajectory data of the target skeleton points. Extract spatiotemporal features such as target skeleton point displacement velocity, joint angle change rate, and center of gravity displacement standard deviation through three layers of convolutional kernels (sizes 3×3, 5×5, and 7×7). Map these features to multiple motion quality indicators such as motion standard, joint range of motion, and motion rhythm stability. Calculate the score for each motion quality indicator using the "spatiotemporal feature value - standard threshold" comparison method.

[0026] S13: Using the individual ID of the adolescent as an index, integrate and standardize the scores of various sports quality indicators with multidimensional physical fitness data. The standardized data is then processed by a fully connected layer to obtain a multidimensional physical fitness feature vector identified by the individual ID.

[0027] The aforementioned multidimensional physical fitness data are multidimensional structured data obtained according to body composition and morphology (such as weight and body fat percentage), metabolism and endocrine (such as basal metabolic rate and historical curves), cardiovascular and respiratory function (such as maximum oxygen uptake and its trend), athletic quality and skills (athletic skill development map), flexibility, stability and recovery (dynamic and static flexibility), and neurosensory and adaptive classification (hand-eye coordination).

[0028] The key focus of this invention is monitoring the changing patterns of exercise quality indicators and structured data. This reveals whether training is effective and what state the body is in (improvement, plateau, or excessive fatigue), thereby providing each individual with a unique, data-driven training, nutrition, and recovery plan. This operation uses time-series trend analysis to depict an individual's physical health status in a long-term, multi-dimensional, and dynamic manner.

[0029] This invention employs a stereo vision sensor array and deep learning algorithms to achieve non-contact, high-frequency physical fitness data collection, moving beyond traditional periodic physical tests and constructing a continuous health status trajectory, providing data support for early risk warning.

[0030] S2: Acquire the speech data, temporal intensity data, and image data of the painting process frames of teenagers during the AI ​​painting interaction process, extract features from each, and obtain the corresponding speech features, intensity features, and painting image features; after aligning by timestamp, perform weighted fusion to obtain a multi-dimensional psychological feature vector.

[0031] This stage achieves psychological state assessment through multimodal data collection during human-computer interaction. The speech features are obtained by extracting Mel-frequency cepstral coefficients from the speech data and then processing them using the BERT model.

[0032] The force characteristics include peak force, fluctuation range, average force, and force changes before and after the stroke pauses.

[0033] The painting image features are obtained by extracting HSV histograms from the painting process frame image data and then processing them using the VisionTransformer model.

[0034] In practice, the generative AI psychological analysis and psychological profile construction process is as follows: S21: Acquire multimodal interaction data collected by the student AI interactive terminal (with microphone, pressure-sensitive touch screen, and camera). Among them, Mel-frequency cepstral coefficients (MFCC) are extracted from the speech data, and environmental noise is filtered out by wavelet transform denoising algorithm; HSV histograms (hue, saturation, and brightness distribution) are extracted from the frame image data of the painting process; outliers are removed from the temporal intensity data by sliding window smoothing algorithm, and all data are aligned by timestamp.

[0035] S22: Employs the BERT model to perform sentiment analysis and semantic decoding on speech data, adapting to the colloquial expressions of teenagers, and identifying emotional tendencies and semantic coherence; analyzes the changing trends of stress time-series data, including peak intensity, fluctuation amplitude, average intensity, and intensity changes before and after pen stroke pauses; and uses the Vision Transformer model to analyze the thematic consistency, compositional complexity, and color emotional tendencies of paintings.

[0036] S23: Align speech features, intensity features, and painting image features according to timestamps, remove temporal deviations, and generate a multi-dimensional core psychological feature vector through a weighted fusion algorithm.

[0037] S24: After obtaining the multi-dimensional psychological feature vector, the process also includes constructing a dynamic psychological profile, specifically: Based on multidimensional psychological feature vectors, with each 100ms as a time node, pressure-related features such as the amplitude of force fluctuation and the number of pauses in the pen stroke are extracted at different time nodes. The pressure score of each time node is calculated, and the ratio of the change in pressure score between adjacent time nodes to the time interval is calculated as the gradient of pressure level change. Based on multidimensional psychological feature vectors, the emotional state categories of adolescents are determined as positive, neutral, and negative. The ratio of the number of times a certain emotional state changes to another emotional state to the total number of times the emotional state occurs is used as the probability of emotional state transition. Based on multidimensional psychological feature vectors, creativity features are quantified: the complexity of painting image features and color diversity are summed according to preset weights to obtain the creativity index. Using the gradient of stress level changes, the probability of emotional state transitions, and the creativity index as nodes, the edges between nodes are defined as "collection timestamp associations," recording the numerical relationships of the three major nodes under the same time dimension to construct a dynamic psychological profile.

[0038] S25: After each new AI painting interaction is completed, the node data, time sequence trend and correlation in the portrait are automatically updated. Historical versions are retained to support retrospective comparison. At the same time, abnormal warnings are triggered by combining age group psychological thresholds to form a complete dynamic psychological portrait.

[0039] Based on the cognitive characteristics and behavioral habits of adolescents, this invention objectively collects multimodal psychological correlation data through AI drawing interaction scenarios. Through feature extraction, semantic decoding and temporal modeling, it constructs a dynamically updated psychological profile of adolescents, providing accurate and continuous psychological feature support for subsequent multimodal fusion, while realizing real-time tracking and abnormal warning of psychological state.

[0040] S3: Multidimensional physical feature vectors and multidimensional psychological feature vectors are processed through multi-head attention to generate multidimensional health status embedding vectors. Then, risk prediction is performed using a gated recurrent unit to obtain the risk prediction result. Using risk prediction results as the state space and intervention strategies as the action space, the intervention strategies are optimized through reinforcement learning to generate personalized plans that include exercise prescriptions, psychological adjustment, and nutritional advice.

[0041] First, in the process of multimodal data fusion and health risk prediction, the multidimensional physical feature vector and multidimensional psychological feature vector are processed through multi-head attention to generate a multidimensional health state embedding vector, specifically: Based on the standardized multidimensional physical and psychological feature vectors, physical feature node sets, psychological feature node sets, and directed edges are constructed respectively; the resulting heterogeneous physical and psychological graph is processed by a heterogeneous graph attention network to generate a graph global fusion feature vector.

[0042] Furthermore, based on the standardized multidimensional physical feature vector and multidimensional psychological feature vector, a set of physical feature nodes, a set of psychological feature nodes, and directed edges are constructed respectively, specifically as follows: Based on the standardized multidimensional physical and psychological feature vectors, core feature dimensions are extracted to construct corresponding physical and psychological feature node sets. Each node is configured with an individual ID, node type, collection timestamp, and initial feature vector. For each pair of semantically related physical and psychological feature nodes, a semantically related directed edge is generated, and the weight of the semantically related directed edge is determined by the Pearson correlation coefficient between the physical and psychological feature nodes. For each pair of temporally related physical and psychological feature nodes, a temporally related directed edge is generated, and the weight of the temporally related directed edge is determined by the exponential decay function value between the physical and psychological feature nodes.

[0043] After concatenating the graph global fusion feature vector, the standardized multidimensional physical feature vector, and the multidimensional psychological feature vector into a cross-modal feature matrix, it is integrated into an attention output matrix through multi-head attention processing. Global average pooling along the temporal dimension is performed on the attention output matrix to obtain a multidimensional health state embedding vector.

[0044] This invention first constructs a physical-psychological heterogeneous graph and completes graph feature learning, then combines multi-head attention for cross-modal deep fusion, and finally performs pooling compression to generate a multi-dimensional health state embedding vector that encodes cross-modal associations, temporal patterns, and intrinsic relationships of features. The multi-dimensional health state embedding vector retains the cross-modal feature associations mined by multi-head attention and integrates the intrinsic relationships of physical-psychological features and temporal evolution patterns encoded by the graph structure, which can effectively improve the accuracy of risk prediction.

[0045] Furthermore, the process of using a gated loop unit to perform risk prediction and obtain the risk prediction result is as follows: The multidimensional health status embedding vectors are arranged into a time sequence in chronological order. After being processed by a multi-layer gated recurrent unit network, they are mapped to probability values ​​through a sigmoid activation function, corresponding to the predicted probabilities of various health risks, such as the predicted probabilities of obesity risk coefficient, sports injury probability, and psychological stress index in the next 30 days.

[0046] This invention uses multi-head attention processing to explore cross-modal correlations between physical and psychological characteristics, and then uses gated recurrent units to accurately capture the long-term evolution of health status, thereby enabling the prediction of future health risks and providing more comprehensive and accurate assessment results.

[0047] Then, as Figure 2 During the adaptive intervention process, deep reinforcement learning is used. The risk prediction results are used as the state space, and the intervention measures (exercise prescriptions, psychological adjustment, and nutritional advice in the intervention action library) are used as the action space. The intervention strategy is optimized through the proximal policy optimization algorithm (PPO) to generate a personalized plan that includes exercise prescriptions, psychological adjustment, and nutritional advice.

[0048] In addition, it also includes adapting personalized solutions to specific scenarios and updating personalized solutions using federated learning, specifically: Personalized plans are transformed into gamified tasks, class health heatmaps, and family suggestions, which are then sent to students, teachers, and parents for execution. Feedback data is collected from each platform, and federated learning is used to aggregate the feedback data to update the personalized plans.

[0049] In practice, exercise prescriptions are transformed into gamified tasks, and psychological adjustment plans are transformed into AI drawing interactive tasks, which are sent to students to increase their participation. Personalized plans are integrated with class health data to generate a class health heat map, which is sent to teachers to help them understand the overall health status of the class. Personalized plans are also transformed into family-adapted suggestions, such as parent-child exercise and family nutrition planning, which are sent to parents to promote collaborative management between home and school.

[0050] Collect feedback data from various sources, including task completion rates and changes in physical fitness indicators after exercise for students; suggestions for classroom intervention adjustments and key individual cases for teachers; and family activity participation and dietary adjustment feedback from parents. While ensuring data privacy, aggregate feedback data from all sources using federated learning to update personalized plans and improve monitoring accuracy and intervention effectiveness.

[0051] This invention collects and processes physical fitness data through a stereo vision sensor array and deep learning algorithms, collects multimodal psychological data through AI drawing interaction, generates personalized intervention plans through cross-modal semantic fusion and intelligent decision-making, and achieves closed-loop optimization through multi-terminal collaboration and federated learning. It solves the problems of single data dimension, imperfect evaluation system, homogeneous intervention and data dispersion in traditional monitoring systems, and realizes intelligent, personalized and dynamic management of adolescent physical health throughout the whole process.

[0052] This invention also provides a multimodal fusion-based dynamic monitoring and intervention system for adolescent physical fitness, comprising: Physical fitness data processing module: used to acquire video streams of adolescent sports activities at target sports venues, extract the three-dimensional coordinates of multiple target skeletal points in the video streams in real time, construct temporal trajectory data of target skeletal points; extract the temporal features of the temporal trajectory data of target skeletal points and map them to preset sports quality indicators, calculate the scores of each sports quality indicator; and integrate the scores of each sports quality indicator with multidimensional physical fitness basic data into a multidimensional physical fitness feature vector according to individual ID. The psychological data processing module is used to acquire speech data, temporal intensity data, and image data of the painting process frames of teenagers during the AI ​​painting interaction process, extract features from each, and obtain corresponding speech features, intensity features, and painting image features; after being aligned by timestamp, they are weighted and fused to obtain a multi-dimensional psychological feature vector; Risk prediction and intervention module: Multidimensional physical feature vectors and multidimensional psychological feature vectors are processed through multi-head attention to generate multidimensional health status embedding vectors. Then, risk prediction is performed using a gated recurrent unit to obtain the risk prediction results. Using risk prediction results as the state space and intervention strategies as the action space, the intervention strategies are optimized through reinforcement learning to generate personalized plans that include exercise prescriptions, psychological adjustment, and nutritional advice.

[0053] This invention utilizes a physical fitness data processing module to collect and monitor multidimensional physical fitness data, a psychological data processing module to perform generative AI psychological analysis, and a risk prediction and intervention module to perform cross-modal semantic fusion and intelligent decision-making to generate personalized intervention plans, thereby significantly improving the intelligence and precision of adolescent physical health management.

[0054] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion, characterized in that, include: S1: Acquire video streams of youth sports activities at the target sports venue, extract the three-dimensional coordinates of multiple target skeletal points in the video stream in real time, and construct temporal trajectory data of the target skeletal points; Extract the temporal features of the target skeletal point temporal trajectory data and map them to preset motion quality indicators to calculate the scores of each motion quality indicator; according to the individual ID, integrate the scores of each motion quality indicator with multidimensional physical fitness data into a multidimensional physical fitness feature vector; S2: Acquire the speech data, temporal intensity data, and painting process frame image data of teenagers during the AI ​​painting interaction process, and extract features respectively to obtain the corresponding speech features, intensity features, and painting image features; After aligning by timestamp and weighting and fusing, a multidimensional psychological feature vector is obtained; S3: Multidimensional physical feature vectors and multidimensional psychological feature vectors are processed through multi-head attention to generate multidimensional health status embedding vectors. Then, risk prediction is performed using a gated recurrent unit to obtain the risk prediction result. Using risk prediction results as the state space and intervention strategies as the action space, the intervention strategies are optimized through reinforcement learning to generate personalized plans that include exercise prescriptions, psychological adjustment, and nutritional advice.

2. The method for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion according to claim 1, characterized in that, This also includes updating personalized plans using federated learning, specifically: Personalized plans are transformed into gamified tasks, class health heatmaps, and family suggestions, which are then sent to students, teachers, and parents for execution. Feedback data is collected from each platform, and federated learning is used to aggregate the feedback data to update the personalized plans.

3. The method for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion according to claim 1, characterized in that, S3 describes the use of a gated loop unit for risk prediction, yielding the risk prediction result as follows: The multidimensional health status embedding vectors are arranged into a time series in chronological order. After being processed by a multi-layer gated recurrent unit network, they are mapped to probability values ​​through a sigmoid activation function, corresponding to the predicted probabilities of various health risks.

4. The method for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion according to claim 1, characterized in that, The multidimensional physical feature vector and multidimensional psychological feature vector mentioned in S3 are processed through multi-head attention to generate a multidimensional health state embedding vector, specifically: Based on the standardized multidimensional physical and psychological feature vectors, physical feature node sets, psychological feature node sets, and directed edges are constructed respectively; the resulting heterogeneous physical and psychological graph is processed by a heterogeneous graph attention network to generate a graph global fusion feature vector. After concatenating the graph global fusion feature vector, the standardized multidimensional physical feature vector, and the multidimensional psychological feature vector into a cross-modal feature matrix, it is integrated into an attention output matrix through multi-head attention processing. Global average pooling along the temporal dimension is performed on the attention output matrix to obtain a multidimensional health state embedding vector.

5. The method for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion according to claim 1, characterized in that, The temporal features described in S1 include: target bone point displacement velocity, joint angle change rate, and center of gravity displacement standard deviation.

6. The method for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion according to claim 1, characterized in that, The speech features mentioned in S2 are obtained by extracting Mel frequency cepstral coefficients from speech data and then processing them using the BERT model.

7. The method for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion according to claim 1, characterized in that, The force characteristics described in S2 include peak force, fluctuation range, average force, and force changes before and after the stroke pauses.

8. The method for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion according to claim 1, characterized in that, The painting image features mentioned in S2 are obtained by extracting the HSV histogram from the painting process frame image data and then processing it using the VisionTransformer model.

9. The method for dynamic monitoring and intervention of adolescent physical fitness based on multimodal fusion according to claim 1, characterized in that, After obtaining the multidimensional psychological feature vector as described in S2, the process also includes constructing a dynamic psychological profile, specifically: Based on multidimensional psychological feature vectors, stress-related features at different time points are extracted, stress scores at each time point are calculated, and the ratio of the change in stress scores between adjacent time points to the time interval is calculated as the gradient of stress level change. Based on multidimensional psychological feature vectors, the categories of adolescent emotional states are determined, and the ratio of the number of times a certain emotional state changes to another emotional state to the total number of times the emotional state occurs is used as the probability of emotional state transition. Based on multidimensional psychological feature vectors, creativity features are quantified, and the creativity index is obtained by weighted summation according to preset weights. Dynamic psychological profiles are constructed using stress level change gradient, emotional state shift probability, and creativity index as nodes.

10. A dynamic monitoring and intervention system for adolescent physical fitness based on multimodal fusion, characterized in that, include: Physical fitness data processing module: used to acquire video streams of adolescent sports activities in target sports venues, extract the three-dimensional coordinate values ​​of multiple target skeletal points in the video streams in real time, and construct temporal trajectory data of the target skeletal points; Extract the temporal features of the target skeletal point temporal trajectory data and map them to preset motion quality indicators to calculate the scores of each motion quality indicator; according to the individual ID, integrate the scores of each motion quality indicator with multidimensional physical fitness data into a multidimensional physical fitness feature vector; The psychological data processing module is used to acquire the speech data, temporal intensity data, and painting process frame image data of teenagers during the AI ​​painting interaction process, and to extract features from them to obtain the corresponding speech features, intensity features, and painting image features. After aligning by timestamp and weighting and fusing, a multidimensional psychological feature vector is obtained; Risk prediction and intervention module: Multidimensional physical feature vectors and multidimensional psychological feature vectors are processed through multi-head attention to generate multidimensional health status embedding vectors. Then, risk prediction is performed using a gated recurrent unit to obtain the risk prediction results. Using risk prediction results as the state space and intervention strategies as the action space, the intervention strategies are optimized through reinforcement learning to generate personalized plans that include exercise prescriptions, psychological adjustment, and nutritional advice.

Images