A data processing system and method for motion assist monitoring

Abstract

The application relates to the technical field of health monitoring and data processing, and provides a data processing system and method for motion auxiliary monitoring. The data fusion module is used for realizing spatiotemporal alignment of heterogeneous motion data and generating a digital twin representing user motion health, so that the user motion health state can be accurately represented. The risk prediction module is used for constructing a parameter association graph based on a prior knowledge base and historical heterogeneous data, capturing the change mode of the motion risk through a long short-term memory network, mining the coupling relationship between the heterogeneous motion data through a graph convolutional neural network, and finally fusing the two types of information to predict the risk probability, so that the potential injury risk can be more accurately identified. The monitoring management module is used for determining a priority based on the risk probability, a motion target and historical motion performance, generating an intervention strategy through reinforcement learning, and finally simulating and verifying the simulation effect of the intervention strategy through the digital twin, so that the user can be assisted to accurately execute, and the practicality and intelligent level of the intervention strategy are greatly improved.

Description

Technical Field

[0001] This application relates to the field of health monitoring and data processing technology, and in particular to a data processing system and method for motion-assisted monitoring. Background Technology

[0002] With the intelligent upgrade in the sports and health field, sports-assisted monitoring has gradually evolved from the traditional data collection and simple feedback model to multi-dimensional data fusion and risk warning. Currently, sports-assisted devices on the market can collect various types of sports data. Some systems have introduced basic data fusion technology, which can integrate multi-source data into a unified data matrix, providing a foundation for sports status analysis. At the same time, risk prediction functions are also beginning to be applied to sports-assisted monitoring, using preset fixed thresholds or simple statistical models to initially provide warnings of sports injury risks.

[0003] Although existing sports-assisted monitoring systems have acquired a certain level of intelligence, they still have shortcomings in practical applications. Firstly, the multi-source data fusion in existing systems often remains at the level of time alignment and simple stitching, failing to achieve accurate data correlation and failing to explore the deep relationships between different types of features. This results in the fused data not accurately representing the complete state of a user's sports and health. Secondly, existing risk predictions often rely on fixed thresholds or general statistical models, neglecting individual user differences and failing to consider the dynamic changes in sports risk and the coupling relationships between multi-source data. This leads to frequent misjudgments or omissions in risk prediction results, failing to meet the personalized risk warning needs of different users.

[0004] Based on the aforementioned shortcomings of the existing technology, the technical problem to be solved in this application is how to achieve motion-assisted monitoring and data processing through the correlation and fusion of multi-source data. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides a data processing system and method for motion-assisted monitoring.

[0006] In a first aspect, this application provides a data processing system for sports-assisted monitoring, the system comprising: a data fusion module, a risk prediction module, and a monitoring management module;

[0007] The data fusion module is used to receive heterogeneous motion data from user terminals, perform spatiotemporal alignment and fusion processing on the heterogeneous motion data, and generate a digital twin representing the user's exercise health.

[0008] The risk prediction module is used to extract basic and related features from the digital twin, learn historical heterogeneous data based on a predefined prior knowledge base to construct a parameter correlation graph, input the basic and related features into a long short-term memory network to capture the changing patterns of motion risk, input the parameter correlation graph into a graph convolutional neural network to mine the coupling relationship between heterogeneous motion data, and fuse the changing patterns of motion risk with the coupling relationship between heterogeneous motion data to predict the probability of a user's motion injury.

[0009] The monitoring and management module is used to generate intervention strategies based on the user's risk probability of sports injury, combined with the user's sports goals and historical sports performance, through reinforcement learning. The intervention strategies are then synchronized to the digital twin simulation and visualized.

[0010] As an optional implementation, generating a digital twin representing the user's exercise health includes:

[0011] Receive heterogeneous motion data transmitted from user terminals. The heterogeneous motion data includes at least kinematic data, physiological data, and environmental data. Perform integrity verification on the heterogeneous motion data.

[0012] Using the timestamps of motion state changes in kinematic data as the time origin and the coordinate positions of motion state changes as the spatial origin, a spatiotemporal coordinate system is formed. Interpolation algorithms are used to align heterogeneous motion data in time, and coordinate transformation is used to map the heterogeneous motion data to the spatiotemporal coordinate system.

[0013] The spatiotemporally aligned heterogeneous motion data is classified and denoised, and the validity of the heterogeneous motion data is verified after classification and denoising.

[0014] The basic features are extracted from the verified heterogeneous motion data, the correlation between the basic features is explored to generate associated features, and the basic features and associated features are fused through a spatiotemporal attention fusion network to generate a digital twin.

[0015] As an optional implementation, the fusion processing of basic features and related features includes:

[0016] When any basic feature changes by more than a preset threshold, the calculation of the mutual information entropy between the other two basic features is triggered to quantify the coupling strength between the two basic features and the feature change.

[0017] When the basic features change to a value less than or equal to a preset threshold, the correlation coefficient between the basic features is calculated using a sliding time window, and the coupling strength and correlation coefficient are used as the associated features.

[0018] Temporal attention weights are dynamically generated based on the rate of change of features, and spatial attention weights are adaptively generated based on the magnitude of the associated features.

[0019] The spatiotemporal attention fusion network multiplies and fuses temporal attention weights and spatial attention weights to obtain a comprehensive weight. The comprehensive weight is then used to weight and fuse basic features and related features to generate a digital twin that includes both basic features and related features.

[0020] As an optional implementation, the construction of the parameter association map includes:

[0021] Extract basic and related features from the digital twin, and establish mapping relationships between the basic and related features and a predefined prior knowledge base;

[0022] Based on the risk level classification of the prior knowledge base, historical heterogeneous data is classified and learned, and learning weights are assigned to historical heterogeneous data with different risk levels.

[0023] Using basic features as nodes, with node attributes including preset thresholds for basic features and individual user base values, and using associated features as edge weights, the edge weights are corrected by combining feature interaction rules in the prior knowledge base to construct a parameter association graph.

[0024] Based on the passage of time of historical heterogeneous data acquisition and changes in user movement scenarios, the weights of the edges are dynamically adjusted, and the parameter association graph is periodically compared with new heterogeneous movement data to determine whether the edge weights are calibrated based on the prior knowledge base.

[0025] As an optional implementation, the variation pattern of capturing motion risk includes:

[0026] The basic features and associated features are serialized according to time series to form an input sequence that includes information on the temporal changes of features;

[0027] The time when the feature change magnitude in the input sequence is greater than the feature change threshold associated with motion risk in the prior knowledge base is marked as a feature mutation point;

[0028] Long Short-Term Memory (LSTM) networks process input sequences in a hierarchical manner. For labeled feature mutation points, they capture instantaneous changes in features at the first time step, while for input sequences without labeled feature mutation points, they capture trend changes in features at the second time step.

[0029] The risk feature set is generated by integrating instantaneous changes and trend changes of features to characterize the change patterns of motion risk. The risk feature set includes the frequency of feature mutations, cumulative change trends, and duration of risk.

[0030] As an optional implementation, the mining of coupling relationships between heterogeneous motion data includes:

[0031] Using the basic features in the parameter association graph as nodes and the association features as edge weights, we construct the input graph structure of the graph convolutional neural network;

[0032] The input graph structure is input into the graph convolutional neural network. The first convolutional layer extracts the direct correlation strength between nodes, and the second convolutional layer aggregates neighboring nodes to discover the indirect correlation path between nodes.

[0033] Based on the feature interaction rules associated with sports injuries in the prior knowledge base, attention weights are assigned to the associated features related to sports injuries in the parameter association map.

[0034] By integrating direct correlation strength, indirect correlation path, and weighted correlation features, a set of coupling relationships is generated, including the influence weights of direct coupling strength, indirect coupling path, and coupling features, in order to explore the coupling relationships between heterogeneous motion data.

[0035] As an optional implementation, the prediction of the user's risk probability of sports injury includes:

[0036] The risk feature set and the coupling relationship set are mapped to the same feature space. The fusion weights of the risk feature set and the coupling relationship set are dynamically allocated by comparing them with the prior knowledge base based on the frequency of feature mutation and the influence weight of coupling features.

[0037] The risk feature set and the coupling relationship set are weighted and fused by fusion weights to generate a comprehensive risk vector, and the risk level of the current comprehensive risk vector is determined by combining it with the prior knowledge base;

[0038] Based on the frequency of actual sports injuries at the same risk level in historical heterogeneous data, the risk level is adjusted by combining the user's historical sports performance in order to predict the probability of the user developing a sports injury.

[0039] As an optional implementation, synchronizing the intervention strategy to the digital twin simulation and visualizing it includes:

[0040] The intervention strategy is parsed into parameters and instructions for the digital twin, synchronized to the digital twin for simulation, and the user's exercise and health status is monitored after the simulated intervention strategy is executed to obtain the simulation effect;

[0041] The system visualizes the intervention strategy and simulation results, provides the implementation requirements of the intervention strategy, and simultaneously displays a demonstration of the movement process simulated by the digital twin.

[0042] Obtain user feedback on adjustments to the interference strategy, and synchronize the feedback to the digital twin in real time to re-simulate and generate the adjusted intervention strategy.

[0043] As an optional implementation, the generation intervention strategy includes:

[0044] Based on the user's risk probability of sports injury, the priority of generating intervention strategies is determined by combining the user's sports goals and historical sports performance.

[0045] Using risk intervention effectiveness, goal alignment, and user suitability as evaluation indicators, multiple sets of candidate intervention strategies are generated through reinforcement learning;

[0046] By combining user exercise preferences and historical intervention records in the prior knowledge base, candidate intervention strategies are validated to generate intervention strategies, which include adjusting exercise intensity, exercise posture, and rest intervals.

[0047] Secondly, this application provides a data processing method for motion-assisted monitoring, the method comprising: receiving heterogeneous motion data from a user terminal, performing spatiotemporal alignment and fusion processing on the heterogeneous motion data, and generating a digital twin representing the user's exercise health;

[0048] Extract basic and related features from digital twins, and learn from historical heterogeneous data based on a predefined prior knowledge base to construct a parameter association graph;

[0049] Basic and associated features are input into a long short-term memory network to capture the changing patterns of motion risk, and parameter association maps are input into a graph convolutional neural network to explore the coupling relationships between heterogeneous motion data.

[0050] By integrating the changing patterns of sports risks with the coupling relationship between heterogeneous sports data, the probability of a user developing a sports injury can be predicted.

[0051] Based on the user's risk probability of sports injury, and combined with the user's sports goals and historical sports performance, an intervention strategy is generated through reinforcement learning. The intervention strategy is then simulated and visualized in a digital twin.

[0052] Compared with existing technologies, the beneficial effects of this application are as follows: The data fusion module achieves spatiotemporal alignment of heterogeneous motion data, avoiding correlation bias caused by differences in data benchmarks, and generates a digital twin representing the user's exercise health. This digital twin can completely and accurately represent the user's exercise health status, providing a more comprehensive and reliable data foundation for subsequent risk prediction and reducing risk prediction bias caused by insufficient data fusion. The risk prediction module constructs a parameter correlation graph based on a priori knowledge base and historical heterogeneous data, ensuring individualized adaptation of the parameter correlation graph. It captures the changing patterns of exercise risk through a long short-term memory network and mines the coupling relationships between heterogeneous motion data through a graph convolutional neural network. Finally, it fuses the two types of information to predict the risk probability, significantly improving the accuracy of risk prediction, enabling more accurate identification of potential injury risks, and effectively reducing the incidence of sports injuries. The monitoring and management module determines the generation priority of intervention strategies based on risk probability, exercise goals, and historical exercise performance, ensuring that the direction of intervention strategies aligns with core needs. It then generates intervention strategies through reinforcement learning, and finally verifies the simulation effect of the intervention strategies through digital twin simulation. The module also visualizes the details of the intervention strategies and exercise demonstrations, assisting users in accurate execution and significantly improving the practicality and intelligence level of the intervention strategies. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0054] Figure 1 A system flowchart of a data processing system for motion-assisted monitoring provided in an embodiment of this application;

[0055] Figure 2 This is a schematic diagram of the structure of an ankle pump motion assist device for a data processing system for motion assist monitoring provided in an embodiment of this application;

[0056] Figure 3 A logic flowchart illustrating the capture of changing patterns of motion risk in a data processing system for motion-assisted monitoring, provided in an embodiment of this application.

[0057] Figure 4 This is a flowchart of a data processing method for motion-assisted monitoring provided in an embodiment of this application. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of this application more apparent and understandable, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0059] Example 1:

[0060] like Figure 1 The diagram shown is a system flowchart of a data processing system for motion-assisted monitoring provided in this application embodiment. The system includes a data fusion module, a risk prediction module, and a monitoring management module.

[0061] In this embodiment, the user terminal uses an ankle pump exercise assist device as an example, combined with other wearable devices and environmental sensors to acquire heterogeneous motion data. A schematic diagram of the ankle pump exercise assist device is shown below. Figure 2 As shown.

[0062] The data fusion module is used to receive heterogeneous motion data from user terminals, perform spatiotemporal alignment and fusion processing on the heterogeneous motion data, and generate a digital twin representing the user's exercise health. The digital twin includes at least the user's real-time motion posture characteristics, physiological state characteristics, and environmental state characteristics.

[0063] Specifically, generating a digital twin representing a user's exercise and health includes:

[0064] Receive heterogeneous motion data transmitted from user terminals. The heterogeneous motion data includes at least kinematic data, physiological data, and environmental data. Perform integrity verification on the heterogeneous motion data.

[0065] Using the timestamps of motion state changes in kinematic data as the time origin and the coordinate positions of motion state changes as the spatial origin, a spatiotemporal coordinate system is formed. Interpolation algorithms are used to perform time alignment on heterogeneous motion data, and coordinate transformation is used to map the heterogeneous motion data to the spatiotemporal coordinate system to complete spatial alignment.

[0066] The spatiotemporally aligned heterogeneous motion data is classified and denoised. The classification and denoising process includes filtering the kinematic data, screening the physiological data using a mode decomposition algorithm, and handling outliers in the environmental data. After classification and denoising, the heterogeneous motion data is validated.

[0067] The basic features, including motion posture features, physiological state features, and environmental state features, are extracted from the verified heterogeneous motion data. The correlation between the basic features is explored to generate associated features. The basic features and associated features are then fused through a spatiotemporal attention fusion network to generate a digital twin.

[0068] Ankle pump exercise assist devices rely on a comprehensive assessment of user exercise health using multiple types of data. These data come from the device's built-in sensors, wearable devices, and environmental sensors. Data loss can occur during transmission due to unstable device connections or temporary sensor malfunctions. Directly using incomplete data will affect the accuracy of feature extraction and digital twin generation. The device receives heterogeneous motion data from the ankle pump exercise assist device itself. Its built-in sensors include an angle sensor at the pedal and a pressure sensor on the sole of the foot to acquire the user's kinematic data, including ankle joint angle and plantar pressure data. A matching smart bracelet is used to acquire the user's heart rate and blood pressure. Oxygen saturation and environmental sensors are used to acquire the temperature and humidity of the exercise environment and to perform integrity verification on heterogeneous exercise data. First, it checks for missing fields, such as ankle angle and heart rate, which are essential data. Then, it examines the time series of heterogeneous exercise data to determine if there are any time gaps, such as a lack of pressure data records for a certain period. If data loss or gaps are found, the device prompts the user to check the device connection status and re-initiates the data transmission request until complete data is obtained. This avoids biases in subsequent analysis due to missing data, ensures that the data fully covers the health status of ankle pump exercise, and provides a reliable basis for subsequent processing.

[0069] Different sensors have different reference points. The device's angle sensor uses the initial position of the pedal as the coordinate reference, the smart bracelet uses its own startup time as the time reference, and the environmental sensor uses the installation position as the spatial reference. The timestamp of the change in motion state in the kinematic data is used as the time origin. For example, when the angle sensor detects that the ankle joint starts to rotate from the neutral position to the plantar flexion direction, and the pressure sensor detects a change in the plantar pressure data, it is determined that the ankle pump movement is started. The time of the ankle pump movement start is set as the time origin to ensure that the time reference is consistent with the actual movement start. The coordinate position of the change in motion state is used as the spatial origin, that is, the projection of the geometric center of the device's pedal onto the ground is used as the spatial origin, forming a spatiotemporal coordinate system, which corresponds to the vertical, forward and backward, and left and right directions of the ankle joint movement, respectively, to ensure that the spatiotemporal coordinates can accurately reflect the movement position of the ankle joint and the corresponding time of the change in motion state.

[0070] To address the differences in sampling frequencies among different sensors, interpolation algorithms are used to time-align heterogeneous motion data. For example, heart rate and temperature / humidity data with low sampling frequencies are upsampled, while ankle angle and pressure data with high sampling frequencies are downsampled, ensuring that the time intervals of all data remain consistent and that complete heterogeneous motion data can be obtained at the same point in time. Coordinate transformation is used to map the heterogeneous motion data to a spatiotemporal coordinate system, achieving a unified spatial reference for various types of data. This eliminates the reference differences between different sensors, enabling the posture, physiological, and environmental data of ankle pump movements to be accurately matched to the same movement time and spatial location, providing a spatiotemporally synchronized data foundation for feature extraction.

[0071] Even after spatiotemporal alignment, the data still suffers from noise interference. Angle sensors experience angle fluctuations due to slight pedal vibrations, leading to deviations between the ankle joint angle and the actual movement angle. Pressure sensors exhibit errors in pressure detection due to foot sweating. Heart rate fluctuates abnormally due to the relative sliding of the smart bracelet and wrist. Environmental data is affected by instantaneous temperature and humidity changes caused by local airflow. To address this, the spatiotemporally aligned heterogeneous motion data undergoes classification and noise reduction. Ankle joint angle data is filtered using a filtering algorithm combined with typical angle change patterns of ankle pump movement to remove random fluctuations caused by pedal vibrations, preserving the actual angle change trend of the ankle joint. Pressure data is smoothed using a sliding window, averaging pressure over multiple sampling periods to eliminate single-detection errors caused by foot sweating and restore the true pressure distribution. Heart rate is modally decomposed using a modal decomposition algorithm, filtering components that match the ankle pump movement rhythm and eliminating abnormal fluctuations caused by smart bracelet sliding. Environmental data undergoes outlier processing, using threshold filtering to remove instantaneous temperature and humidity data exceeding the normal range for indoor exercise, preserving stable trends.

[0072] After classification and noise reduction, the heterogeneous motion data undergoes validity verification. This verification is based on medical standards and physiological knowledge of ankle pump exercises, including referencing the ankle joint angle range for clinical ankle pump exercises, the normal range of heart rate during exercise, the reasonable distribution range of plantar pressure, and the temperature and humidity range for indoor exercise. The system determines whether the noise-reduced data falls within the normal range. If the data exceeds the range, it is considered invalid and removed. This targeted approach eliminates noise interference from various types of data, ensuring that the data accurately reflects the actual situation while removing invalid data, thus providing high-quality data for feature extraction.

[0073] Even after noise reduction, the effective data remains in the form of sensor data and cannot directly characterize the health status of ankle pump movement. It needs to be transformed into basic features that reflect the movement state. At the same time, a single basic feature cannot reflect the correlation between movement, physiology, and environment, so it is necessary to mine correlation features. Basic features including movement posture features, physiological state features, and environmental state features are extracted from the validated heterogeneous movement data. From the ankle joint angle, the plantar flexion angle, dorsiflexion angle, movement cycle, and left and right foot movement symmetry of the ankle joint are extracted to reflect whether the movement posture is standard. From the heart rate data, the average heart rate during exercise, the amplitude of heart rate change, and the heart rate recovery speed after exercise are extracted to reflect whether the physiological load is reasonable. From the pressure data, the peak plantar pressure and pressure distribution uniformity are extracted to reflect the distribution of exercise load. From the environmental data, the average temperature and humidity during exercise are extracted to reflect the influence of the environment on exercise. Thus, basic features are formed.

[0074] Based on the phase division of ankle pump exercise, this study analyzes the correlation between exercise posture and physiological state, exercise posture and environmental state, and physiological state and environmental state at different stages, generating corresponding correlation features. When fusing basic and correlation features through a spatiotemporal attention fusion network, the temporal attention weight is determined according to the rate of change of features during ankle pump exercise, and the spatial attention weight is adaptively generated according to the magnitude of the correlation features. Finally, a digital twin of ankle pump exercise is generated, using a three-dimensional virtual human body model as a carrier to reproduce the user's ankle pump exercise process, while displaying exercise posture, physiological state, environmental state, and correlation features, intuitively showing whether the exercise is standard and whether the physiological load is reasonable. Thus, the raw data is transformed into a digital twin that can directly represent the health of ankle pump exercise. This not only provides accurate feature input for risk prediction, but also intuitively displays exercise details through the digital twin, making it convenient for users or medical staff to observe the exercise status in real time and promptly detect problems such as non-standard exercise or abnormal physiological load.

[0075] Furthermore, the fusion processing of basic features and related features includes:

[0076] When any basic feature changes by more than a preset threshold, the calculation of the mutual information entropy between the other two basic features is triggered to quantify the coupling strength between the two basic features and the feature change.

[0077] When the basic features change to a value less than or equal to a preset threshold, the correlation coefficient between the basic features is calculated using a sliding time window, and the coupling strength and correlation coefficient are used as the associated features.

[0078] Temporal attention weights are dynamically generated based on the rate of change of features, and spatial attention weights are adaptively generated based on the magnitude of the associated features.

[0079] The spatiotemporal attention fusion network multiplies and fuses temporal attention weights and spatial attention weights to obtain a comprehensive weight. The comprehensive weight is then used to weight and fuse basic features and related features to generate a digital twin that includes both basic features and related features.

[0080] In ankle pump exercises, the range of variation in baseline characteristics is directly related to exercise risk. For example, abnormal fluctuations in ankle joint angle are associated with potential joint damage, while sudden changes in heart rate indicate that the body's load exceeds its tolerance range. If a uniform correlation method is used for all fluctuations, it is impossible to distinguish between normal exercise fluctuations and fluctuations with potential risks, which will lead to the correlation characteristics failing to accurately reflect the risk relationship. The preset thresholds for baseline characteristics need to be determined in combination with the clinical guidelines for ankle pump exercises and the individual circumstances of users. First, the safe fluctuation range of baseline characteristics in ankle pump exercises of healthy individuals should be retrieved from the prior knowledge base, and then the safe fluctuation range should be dynamically adjusted based on whether the user has undergone surgery. For example, if a user has limited ankle joint mobility after surgery, the angle change threshold needs to be lowered.

[0081] Using an ankle pump exercise cycle as a sliding comparison window, which encompasses the entire process from plantar flexion to dorsiflexion, the differences between the basic features in the current window and the previous window are continuously calculated. This represents the feature changes of the basic features within the ankle pump exercise cycle. When any basic feature shows a feature change greater than a preset threshold (i.e., the difference between the basic features exceeds the preset threshold), the mutual information entropy between the other two basic features is calculated to quantify the coupling strength between the two basic features and their feature changes, thus quantifying the degree of correlation among the three basic features. When a basic feature shows a feature change less than or equal to a preset threshold (i.e., the difference between the basic features is less than or equal to the preset threshold), multiple consecutive ankle pump exercise cycles are used as sliding time windows. The correlation coefficient between the basic features within the window is calculated using the Pearson correlation coefficient, and the coupling strength and correlation coefficient are used as associated features. This allows for the differentiation of movement fluctuations by using preset thresholds and the matching of corresponding associated features. This ensures that associated features reflect both the feature interactions during normal movement and the risk associations under abnormal fluctuations, avoiding the omission of risk information caused by general calculations and providing a basis for subsequent weight allocation.

[0082] In ankle pump exercises, the rate of change of features determines the urgency of the risk. For example, the risk of injury increases significantly when the ankle joint angle changes rapidly. The magnitude of the associated feature values ​​determines the importance of the risk association; for example, high-value associated features often correspond to a closer risk relationship. If a fixed-weight fusion is used, the information of critical risk times and important associated features will be diluted, making it impossible to accurately focus on risk nodes. A time attention weight is dynamically generated based on the rate of change of features, where the rate of change of features refers to the ratio of the difference between adjacent time-based basic features to the time interval. The greater the rate of change of features, the greater the time attention weight, and vice versa. Then, the magnitude of associated feature values ​​is read, and the values ​​are uniformly mapped to a fixed interval through feature normalization. Simultaneously, the greater the magnitude of associated feature values, the greater the spatial attention weight, and vice versa. Dynamic spatiotemporal attention weights are generated by the rate of change and the magnitude of associated feature values, enabling the spatiotemporal attention weights to accurately adapt to risk changes. The weights of critical risk times and important associated features are significantly increased, while the weights of regular times and ordinary associated features are reasonably allocated, preventing risk information from being masked by regular information and improving the ability of subsequent fusion processes to capture risk information.

[0083] Basic features reflect the single-dimensional state of ankle pump movement, while associated features reflect the relationships between dimensions. Simply superimposing these two types of features leads to fragmented feature information, failing to intuitively represent the risk-related scenarios. The spatiotemporal attention fusion network comprises a four-layer functional structure. First, the weight fusion layer reads the temporal attention weights and spatial attention weights, generating a comprehensive weight through multiplication. This comprehensive weight simultaneously reflects the urgency of risk in the temporal dimension and the importance of association in the spatial dimension. Second, the feature weighting layer inputs the basic and associated features into the feature weighting layer, performing weighted fusion processing on the basic and associated features according to the comprehensive weight. High-risk features with large comprehensive weights are marked and their proportion in the fusion result is amplified, while ordinary features with small weights are given ordinary weights to avoid redundant information interference.

[0084] Next is the residual connection layer. To prevent the loss of original information of basic features during the weighting process, the unweighted original data of basic features is superimposed with the weighted feature data, ensuring that the fused data retains both the original motion state and reflects the risk weights. Finally, there is the twin generation layer, which inputs the fused feature data into the 3D model to generate a digital twin synchronized with the user's ankle pump movement. The digital twin is based on the human body model, highlighting the ankle joint area, reproducing the ankle joint's movement trajectory in real time, superimposing basic features, connecting high-risk associated features through special markers, and setting a risk node backtracking function for viewing details. This achieves the organic integration of the original motion state and risk association. The generated digital twin can not only completely reproduce the ankle pump movement process, but also intuitively highlight the basic and associated features of movement risks, making it easy for users to observe the motion state in real time and avoiding risk misjudgment caused by feature fragmentation.

[0085] The risk prediction module extracts basic and related features from the digital twin, learns from historical heterogeneous data based on a predefined prior knowledge base to construct a parameter correlation graph, inputs the basic and related features into a long short-term memory network to capture the changing patterns of motion risk, inputs the parameter correlation graph into a graph convolutional neural network to mine the coupling relationship between heterogeneous motion data, and integrates the changing patterns of motion risk with the coupling relationship between heterogeneous motion data to predict the probability of a user's motion injury risk.

[0086] Specifically, constructing the parameter correlation graph includes:

[0087] Extract basic and related features from the digital twin, and establish mapping relationships between the basic and related features and a predefined prior knowledge base;

[0088] Based on the risk level classification of the prior knowledge base, historical heterogeneous data is classified and learned, and learning weights are assigned to historical heterogeneous data with different risk levels.

[0089] Using basic features as nodes, with node attributes including preset thresholds for basic features and individual user base values, and using associated features as edge weights, the edge weights are corrected by combining feature interaction rules in the prior knowledge base to construct a parameter association graph.

[0090] Based on the passage of time of historical heterogeneous data acquisition and changes in user movement scenarios, the weights of edges are dynamically adjusted, and the parameter association graph is periodically compared with new heterogeneous movement data to determine whether the edge weights are calibrated based on the prior knowledge base.

[0091] It should be explained that the prior knowledge base is a structured collection of knowledge that supports the monitoring and management system in realizing the functions of data fusion, risk prediction, and monitoring and management throughout the entire process. Its core is to integrate sports medicine standards, clinical rehabilitation experience, research results in the field of sports health, and the patterns of users' personalized sports data, providing a unified knowledge benchmark and decision-making basis for each module of the system. This prior knowledge base is not statically stored, but can be iteratively optimized as users' heterogeneous sports data accumulates, sports scenarios change, and medical knowledge is updated. Ultimately, it provides accurate support for users' sports health status assessment, risk prediction, and intervention strategy generation, ensuring that the system functions meet the requirements of medical safety, individual adaptability, and sports science.

[0092] The prior knowledge base includes preset thresholds for basic features, basic standards in sports medicine, feature interaction rules, individual user benchmarks, and simulation effects. The initial construction of the prior knowledge base is based on general medical knowledge, supported by clinical data, and supplemented by heterogeneous user exercise data. It is achieved through knowledge acquisition, structured processing, correlation verification, and initial adaptation to ensure that the prior knowledge base has basic usability and medical compliance. Initial knowledge is obtained from authoritative medical resources, clinical practice data, and initial user data to form an original knowledge set. This includes obtaining general standards and norms in the field of sports medicine through literature retrieval, obtaining clinical data through cooperation with sports and health institutions, and recruiting different types of users to obtain small-scale exercise data in the early stages of system launch. It must be stated that the data is only used for experiments.

[0093] The original knowledge set is then transformed into structured data conforming to the prior knowledge base data structure, eliminating differences in knowledge format and logical conflicts. The original knowledge set is classified according to the content included in the prior knowledge base, and a corresponding field structure is matched for each category of knowledge. The expression format and units of knowledge are unified to eliminate ambiguity and conflicts. At the same time, a unique identifier is added to each piece of structured knowledge, and knowledge association is performed. Next, the accuracy and relevance of the prior knowledge base are ensured through two methods: medical expert review and clinical data backtesting verification, to prevent erroneous knowledge from entering the monitoring and management system. Finally, based on the user's heterogeneous motion data and typical motion scenarios, the verified structured knowledge is initially adapted to form the final prior knowledge base. Typical motion scenarios include early postoperative rehabilitation indoors, indoor health fitness, and professional training in gyms.

[0094] Parameter association maps need to rely on the correspondence between features and movement risks to have risk characterization capabilities. If the map is constructed directly using feature dimensions and feature interaction rules that are not associated with a prior knowledge base, it will be impossible to determine the risk meaning of the features through the parameter association map. Basic and associated features related to ankle pump movement are extracted from the digital twin. Basic features include ankle plantar flexion angle, ankle dorsiflexion angle, and movement cycle, which reflect movement posture; average heart rate and blood oxygen saturation, which reflect physiological state; and temperature and humidity, which reflect environmental state. Associated features include coupling strength and correlation coefficient. The predefined prior knowledge base includes feature interaction rules and feature dimension mapping tables. Feature interaction rules include that an ankle dorsiflexion angle exceeding a preset threshold can easily cause ligament damage and that a persistently high heart rate indicates myocardial overload. The feature dimension mapping table classifies basic features and maps them to biomechanical parameters, physiological indicators, and environmental influencing factors in the prior knowledge base, and maps associated features to feature interaction rules.

[0095] Then, mapping relationships are established between basic features and associated features and predefined prior knowledge bases one by one. For example, the plantar flexion angle of the ankle joint is mapped to the joint range of motion in the biomechanical parameters in the prior knowledge base, and the average heart rate during exercise is mapped to the circulatory system load in the physiological indicators in the prior knowledge base. At the same time, associated features are matched with specific rules in the feature interaction rules. For example, the coupling strength between a sudden change in ankle joint angle and an increase in heart rate is mapped to the feature interaction rule for physiological stress caused by abnormal joint movement. Finally, a mapping relationship between features and prior knowledge base is formed. This ensures that each node and edge of the subsequent parameter association graph can correspond to a clear feature interaction rule, avoids invalid graph structures with no risk meaning for features, and provides a unified knowledge benchmark for subsequent processing, reducing processing deviations caused by feature dimension confusion.

[0096] The reference value of historical heterogeneous data with different risk levels for constructing parameter association maps varies significantly. For example, high-risk data, such as ankle discomfort and a sudden increase in heart rate during a particular exercise, is more indicative of an individual user's risk association patterns than low-risk data with no abnormalities over a long period. If all data are learned with the same weight, the influence of high-value data will be diluted, resulting in the parameter association map failing to accurately match the individual user's baseline values. Therefore, it is necessary to combine the risk levels of the prior knowledge base to classify and learn historical heterogeneous data and assign differentiated weights. First, the risk level classification criteria are retrieved from the prior knowledge base. This classification criteria divide ankle pump exercise risk into three levels: low, medium, and high. Low risk refers to no basic features exceeding the safe fluctuation range and no related features triggering feature interaction rules. Medium risk refers to a single basic feature exceeding the safe fluctuation range but related features not triggering feature interaction rules or related features only triggering low-risk feature interaction rules. High risk refers to multiple basic features exceeding the safe fluctuation range or related features triggering high-risk feature interaction rules.

[0097] The system retrieves historical heterogeneous data, specifically users' historical heterogeneous motion data, and labels each piece of historical heterogeneous data with a risk level according to the risk level standards of the prior knowledge base. For example, if a piece of historical heterogeneous data shows that the ankle dorsiflexion angle exceeds a preset threshold and the heart rate suddenly increases, and the associated features trigger the feature interaction rules of abnormal joint movement leading to physiological stress, it is labeled as high risk. If a piece of data only shows that the ambient temperature is slightly higher than the normal range, but there are no other basic feature changes exceeding the preset threshold, and the associated features do not trigger the feature interaction rules, it is labeled as low risk. Learning weights are assigned to historical heterogeneous data of different risk levels, with high-risk data having the highest weight, medium-risk data having the second highest weight, and low-risk data having the lowest weight. At the same time, the adaptability of the exercise scenario to the user's physical condition is considered to ensure that the learning results are more in line with the current actual application needs. This highlights the learning priority of high-risk and historical data in the same scenario, enabling the parameter association graph to more accurately capture the risk association patterns of individual users, avoiding insufficient adaptability of the parameter association graph due to interference from general low-value historical heterogeneous data, and improving the accuracy of the parameter association graph in representing individual user risks.

[0098] Simply constructing a parameter association graph using basic features as nodes ignores individual user differences. For example, the same ankle angle change has different risk implications for users in the early postoperative recovery period and those in the later recovery period. It is necessary to reflect individual user baselines and adjust edge weights by incorporating feature interaction rules from the prior knowledge base. Using basic features that have been mapped to the prior knowledge base as nodes in the parameter association graph, each node includes node attributes such as a preset threshold for the basic feature and a user's individual baseline value. The preset threshold for the basic feature includes the safe fluctuation range of the corresponding dimension retrieved from the prior knowledge base, such as the general safe threshold for ankle dorsiflexion angle and the safe range for average heart rate during exercise. The user's individual baseline value is obtained by calculating the average value of the basic feature in high-weight historical heterogeneous data. For example, the average plantar flexion angle and average heart rate of a user during multiple low-risk ankle pump exercises are calculated as the user's individual baseline value.

[0099] The numerical values ​​of associated features are used as the weights of the edges in the parameter association graph. For example, the coupling strength between a sudden change in ankle angle and an increase in heart rate, as well as the correlation coefficient between ambient temperature and exercise cycle, are directly used as the weights of the edges between corresponding nodes. Then, feature interaction rules from the prior knowledge base are invoked. These rules include, for example, that when the ankle angle exceeds a certain percentage of the user's baseline value, its correlation strength with heart rate needs to be appropriately amplified, and when the ambient temperature exceeds the normal suitable range, its correlation strength with exercise cycle needs to be corrected according to specific logic. The weights of the edges are adjusted according to these feature interaction rules. For example, if the weight of a certain edge conforms to the rule that the ankle angle exceeds a certain percentage of the individual's baseline value, the weight is adjusted to a value that better reflects the risk association, ultimately forming the parameter association graph. This ensures that the nodes of the parameter association graph not only include general medical safety standards but also reflect the user's individual exercise ability benchmark. The edge weights not only reflect the natural correlation strength of features but also meet the adjustment requirements of medical rules for risk association, ensuring that the parameter association graph can accurately reflect the user's individual risk association status.

[0100] Users' athletic ability changes over time. For example, during postoperative rehabilitation, ankle joint mobility gradually increases, and the corresponding risk association patterns also change. The correlation strength between ankle joint angle and heart rate in the early postoperative recovery period will inevitably differ from that in the later recovery period. At the same time, users' exercise scenarios will change, such as from indoor rehabilitation training to home self-training, and the impact of environmental factors on exercise will also change. To address the changes in users' athletic ability over time, as the sampling time of historical heterogeneous data is further away from the current time, the influence of this historical heterogeneous data on the weight of edges in the parameter association graph gradually decreases, thereby ensuring that the parameter association graph can prioritize reflecting recent user risk association patterns.

[0101] When a user's exercise scenario changes, different adjustment logic is set for the weights of different edges under different exercise scenarios. For example, in outdoor scenarios, the weights of edges related to ambient temperature and ankle angle, as well as ambient humidity and exercise cycle, need to be adjusted according to the characteristics of the scenario. The weights of the corresponding edges are temporarily adjusted to ensure that the parameter association graph adapts to the risk influencing factors of the current scenario. Then, the current parameter association graph is compared with recently acquired heterogeneous exercise data periodically. The deviation between the weight of the edge in the parameter association graph and the corresponding feature association strength in the new heterogeneous exercise data is calculated. If the deviation of the weight of a certain edge from the feature association strength in the new heterogeneous exercise data exceeds the allowable range, the weight of the edge is corrected based on the average level of feature association strength in the new heterogeneous exercise data. After correction, the parameter association graph is updated and stored. This makes the parameter association graph time-adaptive and scenario-adaptive. It can update the risk association rules with changes in the user's exercise ability and adjust the risk influence weights with changes in exercise scenarios. At the same time, by periodically calibrating and correcting the initial deviation, the parameter association graph avoids the risk prediction error from increasing due to staticization, ensuring that the parameter association graph can always accurately reflect the user's current risk association status.

[0102] Furthermore, such as Figure 3 As shown, the changing patterns of motion risk capture include:

[0103] The basic features and associated features are serialized according to time series to form an input sequence that includes information on the temporal changes of features;

[0104] The time when the feature change magnitude in the input sequence is greater than the feature change threshold associated with motion risk in the prior knowledge base is marked as a feature mutation point;

[0105] Long Short-Term Memory (LSTM) networks process input sequences in a hierarchical manner. For labeled feature mutation points, they capture instantaneous changes in features using a short time step, while for input sequences without labeled feature mutation points, they capture trend changes in features using a long time step.

[0106] The risk feature set is generated by integrating instantaneous changes and trend changes of features to characterize the change patterns of motion risk. The risk feature set includes the frequency of feature mutations, cumulative change trends, and duration of risk.

[0107] The core advantage of Long Short-Term Memory (LSTM) networks is their ability to process time-series data and capture the patterns of data change over time. However, the basic and associated features extracted from digital twins are discrete, instantaneous data, such as ankle angle at a certain time, heart rate during exercise, and feature coupling strength over a certain period. These discrete data cannot directly reflect the continuous changes of features as the exercise progresses. If directly input into LTM networks, they cannot capture the dynamic changes in exercise risk. Therefore, features must be integrated into a continuous input sequence according to the time series. The basic and associated features of ankle pump movement are extracted from digital twins. Basic features include ankle plantar flexion angle, ankle dorsiflexion angle, average heart rate during exercise, ambient temperature, and ambient humidity. Associated features include the coupling strength between ankle angle and heart rate, as well as the correlation coefficient between ambient temperature and exercise cycle. These feature data are sorted according to the order of sampling time.

[0108] Using the ankle pump exercise cycle as a temporal unit, which is the time it takes for a user to complete one full plantar flexion and dorsiflexion movement, the feature data within each ankle pump exercise cycle are integrated. This includes data such as the maximum plantar flexion angle, maximum dorsiflexion angle, average heart rate, and average ambient temperature, as well as the average value of feature coupling strength and correlation coefficient within that cycle. This integrated data is then used as a temporal unit. The temporal units from multiple consecutive exercise cycles are then concatenated chronologically to form an input sequence that includes the temporal changes of features. This input sequence fully reflects the changes in each feature across multiple exercise cycles. For example, from the initial cycle to subsequent cycles, the ankle angle gradually increases from the normal range and exceeds a preset threshold, and the heart rate gradually increases with the angle. This transforms discrete instantaneous data into a continuous temporal sequence, fully preserving the trend and continuity of feature changes throughout the exercise process. This aligns with the characteristics of Long Short-Term Memory (LSTM) networks in processing temporal data and avoids the inability of LTM networks to capture the dynamic evolution of risks due to fragmented input data.

[0109] The occurrence of exercise risks is often directly related to feature mutations. For example, a sudden increase in ankle joint angle may be caused by incorrect user posture, while a sudden increase in heart rate indicates that the body load has suddenly exceeded tolerance. These mutation points are risk warning nodes. If the Long Short-Term Memory (LSTM) network cannot identify these mutation points when processing the input sequence, it will treat mutated data the same as normal data, causing the LTM network to fail to focus on key risk information. The feature change thresholds associated with ankle pump exercise risk are retrieved from the prior knowledge base. These feature change thresholds are based on the abnormal feature change standards set by medical research and clinical experience. For example, a single change in ankle joint angle exceeding the safe fluctuation range is judged as a mutation, a single change in heart rate exceeding the normal physiological range is judged as a mutation, and a single change in feature coupling strength exceeding the normal association level is judged as a mutation.

[0110] The algorithm iterates through each temporal unit in the input sequence, calculating the magnitude of change in the corresponding feature between the current temporal unit and the previous temporal unit. For example, it calculates the difference in ankle plantar flexion angle, average heart rate, and coupling strength between adjacent cycles. The algorithm assesses the magnitude of change for each feature. If the magnitude of change for any feature in a temporal unit exceeds the feature change threshold associated with exercise risk in the prior knowledge base, the time corresponding to that temporal unit is marked as a feature mutation point, and the mutation type is labeled, such as ankle angle mutation, heart rate mutation, and coupling strength mutation. If the magnitude of change for all features does not exceed the feature change threshold, the temporal unit is marked as a normal temporal unit, ultimately forming an input sequence marked with feature mutation points. This accurately locates key risk nodes in the input sequence, preventing the dilution of mutation information due to indiscriminate attention when processing long short-term memory (LSM) sequences. It also provides clear priority processing targets for subsequent hierarchical processing by the LSM, ensuring that the LSM prioritizes capturing risk information corresponding to mutation points and improving the targeting of risk change pattern capture.

[0111] The risk information types corresponding to feature mutation points and normal time units are different. Feature mutation points reflect sudden risks, and it is necessary to accurately capture the details of changes in order to judge the severity of the risks. For example, the specific situation of a sudden increase in ankle joint angle within a cycle and the speed of sudden increase in heart rate. Normal time units reflect cumulative risks, and it is necessary to grasp the overall trend of feature changes in order to identify the risk accumulation process. For example, the overall pattern of gradual increase in heart rate over multiple cycles. If the Long Short-Term Memory Network uses a single time step to process the entire input sequence, it will lead to insufficient capture of mutation point details or incomplete capture of the trend of normal time units. Therefore, it is necessary to process the input sequence in layers. The Long Short-Term Memory Network has a built-in dual time step processing branch to adapt to the processing needs of feature mutation points and normal time units respectively.

[0112] For time-series units marked as feature mutation points, a first time-step branch is activated, using a short time step. This branch processes feature data for one motion cycle at a time, focusing on capturing the details of instantaneous feature changes before and after the mutation point. For example, it simultaneously processes feature data from the motion cycle before the mutation point, the mutation point itself, and the motion cycle after the mutation point, analyzing the magnitude and rate of feature changes before and after the mutation, and finally outputting a vector including the instantaneous feature changes. For time-series units marked as normal time-series units, a second time-step branch is activated, using a long time step. The branches use multiple consecutive temporal units as time steps. For example, each branch processes feature data from multiple consecutive motion cycles. The feature data within the window is integrated through a sliding window, and the overall trend of feature changes within the window is calculated. For example, the overall fluctuation direction of ankle joint angle and the overall rise and fall trend of heart rate within multiple cycles are calculated. Finally, the output is a vector that includes the feature trend changes. This enables the Long Short-Term Memory Network to accurately capture the instantaneous feature changes corresponding to sudden risks and to fully grasp the trend feature changes corresponding to cumulative risks. It avoids the one-sided capture of change patterns caused by a single processing method and ensures that the Long Short-Term Memory Network can comprehensively acquire risk-related feature change information.

[0113] The vectors of instantaneous changes and vectors of trend changes in features output by the hierarchical processing of the Long Short-Term Memory (LSTM) network are independent information carriers, corresponding to sudden risks and cumulative risks, respectively. If directly used for risk prediction, it would lead to the separation of the two types of risk information. For example, it is impossible to determine whether there is a cumulative increase in heart rate based solely on instantaneous angle changes, and it is impossible to identify sudden postural abnormalities based solely on heart rate trends. By integrating the two types of feature vectors output from the LSM network, the frequency of feature mutations is extracted from the vectors of instantaneous changes in features. That is, the total number of all feature mutation points in the input sequence is counted, while the proportion of different types of feature mutation points is distinguished, such as the proportion of ankle angle mutations and heart rate mutations, to reflect the frequency of sudden risks and the main sources of risk during exercise.

[0114] The cumulative trend is extracted from the vector of feature trend changes. This involves analyzing the overall direction of change of each basic feature throughout the input sequence. For example, the ankle angle gradually shifts from the normal range towards a preset threshold, and the heart rate shows a continuous or fluctuating upward trend. The stability of the trend is also marked, such as whether the trend is continuous without reverse fluctuations and whether there are intermittent pauses. The duration of risk is also extracted from the vector of feature trend changes. This involves counting the number of time-series units where a certain type of risk-related trend persists, reflecting the degree of continuous impact of cumulative risk. Risk-related trends include heart rate exceeding the normal range and ankle angle deviating from the preset threshold range. The three types of information—feature mutation frequency, cumulative trend, and duration of risk—are integrated into a structured risk feature set. Each information dimension is accompanied by a corresponding risk association description. For example, a high feature mutation frequency indicates a sudden change in ankle angle, suggesting frequent abnormal postures, and a continuous increase in heart rate indicates a gradual increase in physical load. This achieves an organic integration of sudden and cumulative risks, avoiding the one-sided risk judgment caused by feature fragmentation, and ensuring that the risk feature set can comprehensively and accurately represent the dynamic change pattern of exercise risk.

[0115] Furthermore, exploring the coupling relationships between heterogeneous motion data includes:

[0116] Using the basic features in the parameter association graph as nodes and the association features as edge weights, we construct the input graph structure of the graph convolutional neural network;

[0117] The input graph structure is input into the graph convolutional neural network. The first convolutional layer extracts the direct correlation strength between nodes, and the second convolutional layer aggregates neighboring nodes to discover the indirect correlation path between nodes.

[0118] Based on the feature interaction rules associated with sports injuries in the prior knowledge base, high attention weights are assigned to the associated features related to sports injuries in the parameter association map.

[0119] By integrating direct correlation strength, indirect correlation path, and weighted correlation features, a set of coupling relationships is generated, including the influence weights of direct coupling strength, indirect coupling path, and coupling features, in order to explore the coupling relationships between heterogeneous motion data.

[0120] Graph convolutional neural networks are used to process data with associated structures. The parameter association graph has integrated the association information of heterogeneous motion data. If the input graph structure is reconstructed without the parameter association graph, it will not only waste the existing accurate association information, but also cause the subsequent coupling relationship mining to deviate from the individual risk patterns of users because the new structure is disconnected from the logic of the parameter association graph. The core elements directly related to ankle pump exercise risk are extracted from the parameter association graph. The nodes are selected from five basic features: ankle plantar flexion angle, ankle dorsiflexion angle, average heart rate during exercise, plantar pressure, and ambient temperature during exercise. The weights of the edges correspond to the association features between nodes. The relationship between the above nodes and edges is transformed into a format that can be recognized by the graph convolutional neural network. That is, a feature vector is constructed for each node. The feature vector includes the real-time monitoring value of the basic feature, the preset threshold, and the individual user base value, ensuring that the node not only reflects the current state, but also reflects the safety boundary and individual differences.

[0121] The edge weights are directly used as the weight parameters for connections between nodes, while strictly preserving the connection logic between nodes. For example, the ankle dorsiflexion angle node needs to be connected to both the average heart rate and plantar pressure nodes, corresponding to the medical pattern that abnormal joint mobility easily leads to heart rate stress and uneven pressure distribution in clinical practice. This results in a complete input graph structure containing node feature vectors, edge weights, and connection relationships, ensuring that the structure meets both the computational requirements of graph neural convolutional networks and the medical risk association logic of ankle pump movement. This maximizes the individualized and medical advantages of parameter association graphs, avoids the inefficiency and information loss caused by repeated structure construction, and ensures a high degree of compatibility between the input graph structure and the processing logic of graph convolutional neural networks, laying the foundation for subsequent processing and reducing computational deviations caused by structural mismatch.

[0122] The coupling relationships in heterogeneous motion data exhibit hierarchical differences. Direct associations reflect the immediate mutual influence between features; for example, a sudden change in ankle joint angle directly leads to a change in plantar pressure distribution. Indirect associations reflect the chain effect of multiple features; for example, an increase in ambient temperature during exercise prolongs the exercise cycle, thereby increasing the duration of continuous stress on the ankle joint. Processing with only a single-layer convolution will miss indirect association paths, which are often key triggers for cumulative injuries. The constructed input graph structure is input into a graph convolutional neural network. The graph convolutional neural network includes two functionally independent convolutional branches. The first convolutional branch focuses on... Extracting the direct correlation strength, the convolution kernel of this branch only operates on the direct connection relationship between nodes. For example, for the ankle plantar flexion angle node, it only calculates the feature interaction between it and the directly connected plantar pressure and average heart rate nodes. That is, it integrates the node's own real-time measurement value, preset threshold, and user's individual base value with the corresponding information of connected nodes through convolution operation, and outputs the direct correlation strength between the two types of features. The magnitude of the direct correlation strength directly reflects the degree of mutual influence between the two. For example, a high direct correlation strength between ankle plantar flexion angle and plantar pressure indicates that the user's ankle angle change has a more significant impact on plantar pressure.

[0123] The second convolutional branch focuses on uncovering indirect correlation paths. Building upon the direct correlation strength output from the first layer, this branch expands the computational scope to the neighboring nodes of a node. For example, the direct connected node of the ambient temperature node is the motion cycle, and the direct connected node of the motion cycle is the ankle dorsiflexion angle. Therefore, the ankle dorsiflexion angle becomes a neighboring node of the ambient temperature node. By aggregating the feature information of neighboring nodes through the convolutional kernel, it can not only identify indirect correlation paths such as ambient temperature, motion cycle, and ankle dorsiflexion angle, but also calculate the overall correlation strength of these indirect correlation paths. The strength value reflects the cumulative effect of the chain reaction. For example, a high correlation strength in an indirect correlation path indicates that the indirect influence of ambient temperature on the ankle joint angle is more pronounced by extending the movement cycle. After the two convolutional operations are completed, the direct correlation strength and the indirect correlation paths and their corresponding strength values ​​are stored separately, and the feature type corresponding to each correlation is labeled. This comprehensively captures the two types of coupling relationships in heterogeneous motion data, avoiding the omission of immediate risks due to neglecting direct correlations and preventing the misjudgment of cumulative risks due to neglecting indirect correlations. This ensures that the coupling relationship mining can fully reflect the immediate and cumulative risk triggers, providing a more comprehensive correlation basis for subsequent risk probability prediction.

[0124] The association information extracted by hierarchical convolution contains a large amount of content that is unrelated to or weakly associated with injury risk. For example, the direct association between ambient temperature and average heart rate during exercise has minimal impact on injury within the normal temperature range, while the direct association between ankle dorsiflexion angle and plantar pressure, as well as the indirect association between ambient temperature, exercise cycle, and ankle load, are directly related to ligament injuries and joint strain. If all associations are assigned the same weight, key risk associations will be diluted by non-risk associations, and the graph convolutional neural network will be unable to focus on core risk points. Feature interaction rules associated with sports injuries are retrieved from a prior knowledge base. These feature interaction rules integrate clinical rehabilitation experience and sports medicine research results. Each feature interaction rule clearly defines the corresponding association type and risk level, including direct and indirect associations.

[0125] The association information output by each layer of convolution is compared with the feature interaction rules. For the strength of direct association, if a direct association conforms to the feature interaction rule of a certain injury association, and the current association strength has exceeded the risk threshold set by the feature interaction rule, then the weight of the direct association is increased according to the rule requirements, and it is marked as a high-risk direct association. For example, if a feature interaction rule stipulates that the weight is increased to 1.2 times when the ankle dorsiflexion angle exceeds the user's individual baseline value, then the corresponding direct association is adjusted according to this ratio. If the direct association does not conform to any feature interaction rule, or the strength does not reach the risk threshold, then the original weight is maintained and it is marked as a regular direct association.

[0126] For indirect association paths, if any link in a path conforms to the characteristic interaction rules of damage association, and the overall strength of the indirect association path reaches the risk threshold, then the association strength weight of the indirect association path is amplified and marked as a high-risk indirect association path. At the same time, key risk links in the indirect association path, such as prolonged movement cycle, are also marked. If no link in the indirect association path conforms to the characteristic interaction rules or the strength does not meet the threshold, then the original weight is maintained and it is marked as a regular indirect association path. This accurately filters and strengthens coupling relationships directly related to damage, effectively filters out interference from non-risk associations, and enables the subsequent fusion of coupling relationship sets to focus on core risk points, enhance the supporting value of coupling relationships for risk prediction, and avoid risk judgment bias caused by messy association information.

[0127] Direct correlation strength, indirect correlation path, and weighted correlation features reflect coupling relationships from different dimensions. Direct correlation strength reflects the immediate impact, indirect correlation path reflects the chain-like impact logic, and weighted correlation features reflect risk priority. Using only one type of information will lead to an incomplete representation of the coupling relationship. For example, direct correlation alone cannot explain how environmental temperature indirectly affects joint damage, and weighted correlation features alone cannot reflect the background information of conventional correlations. Therefore, it is necessary to organically integrate the three types of information to form a set that can comprehensively and hierarchically reflect the coupling relationship, providing a complete correlation basis for risk probability prediction. Integrating direct correlation strength, indirect correlation path, and weighted correlation features first retains the information on direct correlation strength, namely the original strength value, adjusted weighted value, and corresponding feature type of each direct correlation, as well as the marker of whether it belongs to a high-risk direct correlation, to ensure that the impact degree and risk attributes of various direct correlations can be intuitively judged. Secondly, the content of indirect correlation path is integrated, namely the correlation chain of each indirect correlation path, the original strength and weighted strength of the indirect correlation path as a whole, and the marking of key risk links in the indirect correlation path, to ensure that the logic of chain-like impact and core risk points can be clearly traced.

[0128] Finally, direct and indirect correlations are established. For example, the direct correlation between ambient temperature and exercise cycle is linked to the indirect path between ambient temperature, exercise cycle, and ankle load, illustrating that the direct correlation is the foundation of the indirect correlation path. The weighted differences between the two are also marked to reflect the cumulative risk effect of the indirect path. The resulting set of coupling relationships, in a structured form, includes three core categories: direct coupling strength, indirect coupling path, and the influence weight of coupling characteristics. These three categories are interconnected and complementary, clearly defining the strength and risk level of each type of correlation while also reflecting the logical relationships between them. This achieves seamless integration of the three types of correlation information, avoiding the limitations of a single information dimension. The set of coupling relationships can fully cover both immediate and chain-like correlations, and can focus on core risks through risk priority marking, providing comprehensive and accurate correlation basis for subsequent risk probability prediction and ensuring that the prediction process fully considers the impact of various risk correlations.

[0129] Specifically, predicting the probability of a user developing a sports injury includes:

[0130] The risk feature set and the coupling relationship set are mapped to the same feature space. The fusion weights of the risk feature set and the coupling relationship set are dynamically allocated by comparing them with the prior knowledge base based on the frequency of feature mutation and the influence weight of coupling features.

[0131] The risk feature set and the coupling relationship set are weighted and fused by fusion weights to generate a comprehensive risk vector, and the risk level of the current comprehensive risk vector is determined by combining it with the prior knowledge base;

[0132] Based on the frequency of actual sports injuries at the same risk level in historical heterogeneous data, the risk level is adjusted by combining the user's historical sports performance in order to predict the probability of the user developing a sports injury.

[0133] The risk feature set focuses on time-series risk changes, including the frequency of feature mutations, cumulative change trends, and duration of risk. This type of information is essentially a statistical summary of time-series data. The coupling relationship set focuses on the risk associated with features, including the strength of direct coupling, indirect coupling paths, and the influence weight of coupled features. This type of information is essentially a quantification of the degree of correlation between features. The feature dimensions and representation logic of the two sets are completely different. Direct fusion will lead to distortion of the fusion result due to dimension incompatibility. The risk feature set and the coupling relationship set are mapped to the same feature space. Here, the two sets are mapped to the feature space of risk contribution dimension, where risk contribution is the actual degree to which a feature promotes motion risk. The magnitude of the risk contribution directly reflects the risk influence weight of that feature.

[0134] The mapping for the risk feature set corresponds to a higher risk contribution rate for higher frequency of feature mutations, a more risk-oriented cumulative change trend, and a longer duration of risk. This means converting feature mutation frequency into mutation risk contribution (frequent mutations correspond to high contribution), cumulative change trend into trend risk contribution (a trend oriented towards risk corresponds to high contribution), and risk duration into duration risk contribution (longer duration corresponds to high contribution). Similarly, the mapping for the coupling relationship set corresponds to a higher risk contribution rate for higher direct coupling strength, a more direct indirect coupling path towards the damage point, and a higher influence weight of the coupling features. The direct coupling strength is transformed into direct association contribution, the indirect coupling path strength into indirect association contribution, and the influence weight of coupled features into weight contribution. After the mapping is completed, all features of the two sets are based on risk contribution as a unified dimension, ensuring that the weights can be calculated based on a unified dimension during subsequent fusion, and that the original risk meaning corresponding to each contribution can be clearly traced. This solves the problem of dimension incompatibility between the two sets, making time-series risk changes and feature-related risks comparable and fusionable under the same risk contribution dimension, avoiding fusion distortion caused by dimension differences, and ensuring that the weight allocation can accurately reflect the actual risk impact of various features.

[0135] In different ankle pump exercise scenarios, the importance of temporal risk changes and feature-related risks for injury prediction varies significantly. For example, when a user experiences frequent sudden changes in ankle angle, i.e., the frequency of feature changes is high, temporal risk changes are the main driving factor of the current risk, and risk should be judged based on these features first. When a user does not have obvious temporal changes, but the coupling strength between ankle angle and plantar pressure is extremely high, i.e., feature-related risks are prominent, feature-related risks become the main driving factor, and risk should be judged based on associated features first. If a fixed weight fusion is used, it will not be able to adapt to the differences in scenarios, causing the prediction results to be biased towards secondary risk factors. Therefore, it is necessary to dynamically allocate fusion weights based on comparison with the prior knowledge base to ensure that the fusion weights are highly matched with the current dominant risk factors.

[0136] The system retrieves thresholds corresponding to the frequency of feature mutations and the influence weights of coupled features from a prior knowledge base. These two indicators are then compared to their respective thresholds to determine the dominant risk factor. Based on this determination, a fusion weight is allocated: the fusion weight of the dominant factor increases, while the fusion weight of non-dominant factors decreases. For example, when the contribution of mutation risk in the risk feature set exceeds the corresponding threshold (i.e., mutations are frequent and mostly joint-related), the fusion weight of the risk feature set becomes dominant. Similarly, when the contribution of weight in the coupled relationship set exceeds the corresponding threshold (i.e., there are many high-risk associations with high strength), the fusion weight of the coupled relationship set becomes dominant. When the risk contribution of neither set exceeds the corresponding threshold, the fusion weight is evenly distributed. When the risk contribution of both sets exceeds the corresponding threshold, the fusion weight tilts towards the set with the higher risk contribution. This achieves dynamic adaptation of the fusion weight to specific scenarios, avoiding the one-size-fits-all problem caused by fixed weights. It ensures that the fusion weight prioritizes the most critical risk factors and improves the representation accuracy of the comprehensive risk vector.

[0137] After dynamic weight allocation, the risk contribution of the two sets has a clear priority difference. If the risk contribution of the two sets is simply superimposed, the guiding role of the fusion weight will not be reflected, and the comprehensive risk information will not be able to focus on the core risks. At the same time, the severity of risks cannot be intuitively judged by risk contribution alone, so it needs to be transformed into a clear risk level. The risk contribution of the two sets after mapping is weighted and fused according to the allocated fusion weight. That is, the contribution of mutation risk, trend risk, and duration risk in the risk feature set is multiplied by the fusion weight of the risk feature set to obtain the weighted time-series risk contribution. The contribution of direct association, indirect association, and weight in the coupling relationship set is multiplied by the fusion weight of the coupling relationship set to obtain the weighted association risk contribution.

[0138] All weighted risk contributions are integrated into a single comprehensive risk vector. Each dimension of the vector corresponds to a type of weighted risk contribution, and the source of risk for each dimension is labeled to ensure clear traceability of the composition of the comprehensive risk vector. The correspondence rules between the comprehensive risk vector and risk level are retrieved from a prior knowledge base. These rules are based on a large amount of clinical ankle pump injury data. If the overall value of the comprehensive risk vector is in the low range and no dimension's weighted contribution reaches the high-risk standard, it is judged as a low-risk level. If the overall value of the comprehensive risk vector is in the medium range, or only a few dimensions' weighted contributions reach the high-risk standard, it is judged as a medium-risk level. If the overall value of the comprehensive risk vector is in the high range, or a large number of dimensions' weighted contributions reach the high-risk standard, it is judged as a high-risk level. The overall value of the comprehensive risk vector in the low, medium, and high ranges is obtained through threshold comparison.

[0139] By analyzing the numerical range and number of high-risk dimensions of the current comprehensive risk vector against this rule, the corresponding risk levels are determined, including low, medium, and high risk levels. Key criteria for risk level determination are also recorded. For example, a medium risk level indicates that the comprehensive risk vector is in the middle range, and the dimension of mutation risk contribution reaches the high-risk standard, providing a clear level background for subsequent probability correction. This integrates the risk information from the two sets into a unified comprehensive risk vector according to the fusion weight, avoiding information fragmentation. Furthermore, by classifying the risk levels, the quantitative vector is transformed into intuitively judgeable risk levels, improving the intuitiveness of risk judgment and providing clear level anchors for subsequent probability correction, avoiding data matching bias caused by unclear levels.

[0140] Risk levels are determined based on general medical rules, which do not fully consider individual user differences. For example, some users have never experienced an injury at the medium-risk level in their history, while another user frequently experiences joint discomfort at the same level. Directly assigning the initial level to the general injury probability would ignore the individual user's exercise risk patterns, leading to a significant deviation between the predicted probability and the actual situation. To correct this, the frequency of actual sports injuries at the same risk level is retrieved from the user's historical heterogeneous data, and the risk level is adjusted in conjunction with the user's historical exercise performance. First, ankle pump exercise data at the same risk level in the historical heterogeneous data is recorded, including characteristic changes during exercise, coupling relationships, and exercise scenarios. The number of actual injuries in the same risk level exercises in the historical heterogeneous data, as well as the characteristic combinations at the time of injury, are also counted. Finally, the user's historical exercise performance at the same risk level is recorded, including the quality of exercise completion and physical adaptation.

[0141] First, the frequency of actual sports injuries at the same risk level in historical heterogeneous data is calculated. This is the ratio of the number of actual injuries in sports at the same risk level to the total number of sports at the same risk level. This frequency reflects the individual injury probability benchmark for the user at that risk level. Second, the similarity between the current sports scenario and feature combination and the historical injury scenario and feature combination is compared. If the similarity to historical injury scenarios is high, the historical injury frequency is appropriately increased as the corrected probability. If it is more similar to historical injury-free scenarios, the historical injury frequency is appropriately decreased. Finally, combined with historical sports performance at the same risk level, if the user's historical sports performance at the same risk level is up to standard and there is no injury, the corrected probability is further confirmed. If the historical sports performance is not up to standard and there is injury, the corrected probability is re-verified to ensure that the correction result fits the user's individual ability and risk pattern. Finally, the corrected risk probability of the user's current sports injury is output. In this way, the risk level is corrected by the user's individual historical data, avoiding prediction bias caused by the general probability ignoring individual differences. This ensures that the final output risk probability can accurately reflect the user's own sports risk pattern, improving the individualization and accuracy of risk prediction.

[0142] The monitoring and management module is used to generate intervention strategies based on the user's risk probability of sports injury, combined with the user's sports goals and historical sports performance, through reinforcement learning. The intervention strategies are then synchronized to the digital twin simulation and visualized.

[0143] Furthermore, the generated intervention strategies include:

[0144] Based on the user's risk probability of sports injury, the priority of generating intervention strategies is determined by combining the user's sports goals and historical sports performance.

[0145] Using risk intervention effectiveness, goal alignment, and user suitability as evaluation indicators, multiple sets of candidate intervention strategies are generated through reinforcement learning;

[0146] By combining user exercise preferences and historical intervention records in the prior knowledge base, candidate intervention strategies are validated to generate intervention strategies, which include adjusting exercise intensity, exercise posture, and rest intervals.

[0147] The priorities differ across different sports scenarios. For example, when the risk of injury is high during ankle pump exercises, risk reduction should be prioritized. However, when the risk is low but there are specific time requirements for the exercise goals, the progress of the goals should be prioritized. If the priorities are not clearly defined, the generated intervention strategies may prioritize safety over goals or goals over suitability. The system obtains the probability of a user's sports injury risk and retrieves the user's exercise goals, such as the need for ankle joint mobility to recover to the level of independent walking within a specific time period after surgery and the need to complete a basic number of standard ankle pump exercises daily. It also extracts the user's historical exercise performance, including the completion status of exercise plans under the same exercise goals, whether there have been any injuries or discomforts, and feedback on the implementation of past intervention strategies.

[0148] The priority of intervention strategy generation is determined by comprehensively considering the user's risk probability of sports injury, sports goals, and historical sports performance. If the risk probability is high, safety assurance is the highest priority, and the intervention strategy must focus on reducing the risk of injury, allowing for adjustments to the goal progress within a reasonable range to avoid increasing risk due to pursuing the goal. If the risk probability is low and the sports goal has a clear time limit, goal advancement is the highest priority, and the intervention strategy must adapt to the goal requirements as much as possible within the safety boundary to ensure that the sports progress meets the expected plan. If historical sports performance shows that a certain type of intervention strategy has poor user adaptability, such as muscle fatigue easily occurring after adjusting the exercise intensity, individual adaptability is the highest priority, and the intervention strategy must avoid such adjustment methods, prioritizing intervention strategies that have had good user performance in the past, such as minor adjustments to exercise posture. This results in a clear generation priority ranking, providing a clear direction for the subsequent generation of candidate intervention strategies. This avoids confusion in the direction of intervention strategies due to conflicting needs, ensures that candidate intervention strategies focus on the current core needs, reduces the generation of ineffective intervention strategies that do not meet the needs, improves the efficiency and targeting of subsequent candidate intervention strategy selection, and provides a clear optimization direction for reinforcement learning.

[0149] Single intervention strategies are difficult to adapt to complex exercise scenarios. For example, users may have multiple needs, such as reducing risk, advancing rehabilitation goals, and preferring specific exercise methods. Furthermore, the actual effects of different intervention strategies are uncertain. If only one set of intervention strategies is generated, there will be no alternative if the strategy fails to meet the user's actual needs or performs poorly. Risk intervention effect, goal fit, and user adaptability are set as evaluation indicators for reinforcement learning. Risk intervention effect is based on medical conclusions in the prior knowledge base regarding the association between ankle pump exercise strategies and risk. For example, adjusting ankle joint movement posture can reduce joint load risk, and extending rest between sets can alleviate muscle fatigue risk. This is used to evaluate the expected risk reduction effect after the intervention strategy is implemented. Goal fit is assessed by combining the user's exercise goals and evaluating whether the intervention strategy can still ensure goal advancement after adjustment. For example, whether adjusting the number of exercises can still gradually improve ankle joint mobility, and whether adjusting the exercise time can still complete the daily basic exercise volume. User adaptability is based on the user's historical exercise preferences, such as whether they reject specific movement adjustments and whether they are accustomed to interval training. This is used to evaluate the degree of matching between the intervention strategy and the user's daily exercise habits.

[0150] Reinforcement learning assigns corresponding weights to three types of evaluation indicators based on a defined generation priority. For example, when safety is prioritized, the weight of risk intervention effect is the highest, and when the goal is prioritized, the weight of goal fit is the highest. Through multiple rounds of iterative learning, multiple sets of candidate intervention strategies are generated. For high-risk scenarios, candidate intervention strategies with risk reduction as the core are generated, such as adjusting the ankle joint range of motion to a safe fluctuation range, extending the rest time after a single exercise session, and replacing high-load exercises with low-load exercises. For low-risk scenarios with urgent goals, candidate intervention strategies with goal advancement as the core are generated, such as increasing the number of daily exercise sets within a safe range, optimizing the exercise rhythm to improve rehabilitation efficiency, and adjusting exercise time according to user preferences. This generates multiple sets of candidate intervention strategies covering different adjustment directions, avoiding the limitations of a single intervention strategy. At the same time, the three types of evaluation indicators ensure that each set of intervention strategies takes into account safety, goals, and user suitability, reducing one-sided strategies that only meet a single need. This provides rich and high-quality candidate intervention strategies for subsequent verification and screening, improving the rationality and feasibility of the final intervention strategy.

[0151] Reinforcement learning-generated candidate intervention strategies may be theoretically feasible but practically unsuitable. For example, a candidate intervention strategy may meet the evaluation criteria, but the user may have experienced joint discomfort when implementing such a strategy in the past. Or, a candidate intervention strategy may conflict with medical guidelines in the prior knowledge base, such as the adjustment method not conforming to the rehabilitation principles of postoperative ankle pump exercises. Directly applying such candidate intervention strategies may lead to user discomfort or affect the rehabilitation effect. Therefore, it is necessary to verify the candidate intervention strategies in conjunction with medical guidelines and user history, eliminate unsuitable strategies, and optimize details to ensure that the final intervention strategy is both safe and compliant, and tailored to the individual user situation. Medical guidelines, user exercise preferences, and historical intervention records are extracted from the prior knowledge base. The medical guidelines include that the adjustment of ankle joint range of motion must meet the requirements of the postoperative rehabilitation stage to avoid overstretching and causing injury, and that the adjustment of rest intervals must balance blood circulation and muscle memory formation, and should not be too short or too long. Exercise preference records include the user's aversion to movements that involve bending the knee joint significantly and the habit of exercising at fixed times. Historical intervention records include the implementation status of past intervention strategies, the physical feedback after implementation, and the trend of risk changes.

[0152] Each candidate intervention strategy was validated against medical guidelines, user exercise preferences, and historical intervention records. From the validated strategies, the one with the best performance across three evaluation metrics and the most suitable user profile after detailed optimization was selected as the primary intervention strategy. Two backup strategies were also generated to address potential risk rebound or user discomfort during the implementation of the primary intervention strategy, ensuring flexibility in its execution. This process eliminated intervention strategies that did not conform to medical guidelines or were unsuitable for the user's historical circumstances, preventing a disconnect between theoretical strategies and practical application. Furthermore, detailed optimization improved user acceptance and feasibility, ensuring the final intervention strategy was both safe and effective, while also aligning with individual user habits, thus enhancing adherence and overall effectiveness.

[0153] Specifically, synchronizing intervention strategies to the digital twin simulation and visualizing them includes:

[0154] The intervention strategy is parsed into parameters and instructions that the digital twin can recognize, synchronized to the digital twin for simulation, and the user's exercise and health status is monitored after the simulated intervention strategy is executed to obtain the simulation effect;

[0155] The system visualizes the intervention strategy and its simulated effects, provides the requirements for implementing the intervention strategy, and simultaneously displays a demonstration of the movement process simulated by the digital twin to help users understand the details of the intervention strategy's implementation.

[0156] Obtain user feedback on adjustments to the interference strategy, and synchronize the feedback to the digital twin in real time to re-simulate and generate the adjusted intervention strategy.

[0157] Intervention strategies are typically described in natural language, such as adjusting ankle range of motion and extending rest intervals. Digital twins cannot directly recognize these descriptions and must convert them into executable parameter commands. Furthermore, the effectiveness of intervention strategies is uncertain; direct application to actual exercise can lead to risks or target deviations. The natural language description of the final intervention strategy is converted into parameter commands recognizable by the digital twin. Several pre-strategies include adjusting exercise posture, which involves resolving ankle pump movements into posture parameters such as the range of ankle plantar flexion and dorsiflexion angles, the limiting angle of knee flexion, and the specific position of foot support. Other pre-strategies include adjusting exercise intensity, i.e., reducing exercise intensity, which is resolved into intensity parameters such as the number of repetitions per exercise, the interval between sets, and the total number of sets per day. Finally, some pre-strategies include adjusting rest intervals, i.e. extending rest intervals, which is resolved into time parameters such as rest time between sets and the intervals between different exercise times per day.

[0158] The parsed parameters and instructions are synchronized to the digital twin for simulation. Intervention strategies are executed based on the user's current exercise and health status represented by the digital twin. This exercise and health status includes the current ankle range of motion, muscle fatigue level, and heart rate. During the simulation, changes in the digital twin's exercise and health status are monitored in real time. For example, after adjusting the angle range, it is observed whether the ankle load decreases, whether the heart rate remains within a safe range, and whether the rehabilitation progress is still progressing. If the simulation results show that the risk is still higher than the safe range after the intervention strategy is implemented, or the target progress deviates too much, the parsing and simulation of backup strategies are automatically triggered until an intervention strategy with satisfactory simulation results is selected. Status changes during the simulation are recorded, including risk and target progress trends, generating a simulation effect report. This resolves the information barrier between the intervention strategy and the digital twin, ensuring the digital twin can accurately execute the strategy. Simulation also verifies the simulation effect of the intervention strategy in advance, avoiding risks or target deviations caused by direct application to actual exercise. It also allows for the early selection of the optimal intervention strategy, improving the safety and effectiveness of intervention strategy execution.

[0159] If the parameters and simulation effects of the intervention strategy are presented in plain text, users, especially non-professional postoperative patients, will find it difficult to understand intuitively. For example, adjusting the ankle dorsiflexion angle to a safe range cannot be accurately conveyed through text, which will lead to deviations in actual execution. First, the user's current risk level, the name of the recommended intervention strategy, and the simulation effect should be displayed intuitively. The simulation effect can be distinguished by color to help users quickly obtain the core conclusions. Then, after clicking on the intervention strategy, the range of ankle plantar flexion and dorsiflexion angles, the restrictions on knee flexion, and the placement of the foot on the device pedal can be displayed directly. The specific rest requirements for extending the rest between sets, the number of sets and the number of repetitions per day, and the contents of the simulation effect report can be displayed simultaneously, including trend graphs showing changes in risk and goals.

[0160] It also provides a digital twin simulation of the movement process, intuitively presenting the standard ankle pump posture corresponding to the intervention strategy in a three-dimensional virtual model, such as the limb position and joint angle during box-assisted movement, marking key force exertion points and precautions. The demonstration process supports pause and replay to facilitate user comparison and learning; thus, it transforms complex intervention strategies and simulation effects into intuitive and easy-to-understand visualizations, solving the problem of user comprehension difficulties, avoiding execution deviations caused by unclear information transmission, and providing standard movement references through movement demonstrations to help users master the correct movement methods and improve the accuracy and effectiveness of intervention strategy execution.

[0161] Digital twin simulations are based on a user's current health status, but cannot fully capture the user's actual experience. For example, the simulation may show that adjusting the angle is feasible, but the user may not be able to reach that angle due to joint pain, or the user may have temporary needs and be unable to complete the specified number of sets as required by the intervention strategy. If user feedback is ignored and the intervention strategy is enforced, it will lead to user discomfort or low execution rate. A user feedback entry point is set up in the visualization interface, allowing users to submit adjustment feedback, including feedback on movement adaptability, time and intensity, and preference. Movement adaptability feedback includes joint pain when performing a certain posture; time and intensity feedback includes being unable to complete the specified number of sets due to time constraints on the day, and the current exercise intensity being too high or too low; preference feedback includes wanting to adjust the exercise time and rejecting a certain type of rest method, etc.

[0162] After obtaining adjustment feedback, the feedback is transformed into specific adjustment requirements, which are then parsed into parameter instructions and transmitted to the digital twin for re-simulation. During the re-simulation, the focus is on verifying the simulation effect of the adjusted intervention strategy. If the re-simulation effect is satisfactory, the adjusted intervention strategy is generated, and the content of the visualization interface is updated. If the simulation effect is unsatisfactory, further adjustment plans are discussed with the user until an intervention strategy acceptable to the user and meeting the simulation effect is generated. This forms a closed loop from intervention strategy generation, simulation, feedback to optimization, fully respecting the user's actual feelings and temporary needs, avoiding one-way output of the intervention strategy, improving the user's acceptance and willingness to implement the intervention strategy, and ensuring that the adjusted intervention strategy still takes into account safety and objectives through re-simulation, avoiding effect deviations caused by user feedback adjustments, and ensuring the effectiveness of the intervention.

[0163] Example 2:

[0164] like Figure 4 The diagram shown is a flowchart of a data processing method for motion-assisted monitoring provided in this application embodiment. The method includes:

[0165] It receives heterogeneous motion data from user terminals, performs spatiotemporal alignment and fusion processing on the heterogeneous motion data, and generates a digital twin representing the user's exercise health.

[0166] Extract basic and related features from digital twins, and learn from historical heterogeneous data based on a predefined prior knowledge base to construct a parameter association graph;

[0167] Basic and associated features are input into a long short-term memory network to capture the changing patterns of motion risk, and parameter association maps are input into a graph convolutional neural network to explore the coupling relationships between heterogeneous motion data.

[0168] By integrating the changing patterns of sports risks with the coupling relationship between heterogeneous sports data, the probability of a user developing a sports injury can be predicted.

[0169] Based on the user's risk probability of sports injury, and combined with the user's sports goals and historical sports performance, an intervention strategy is generated through reinforcement learning. The intervention strategy is then simulated and visualized in a digital twin.

[0170] Since the principle of the method in this application embodiment is similar to that of the system described in this application embodiment, the implementation of the method is the same as that of the system, and the repeated parts will not be described again.

Claims

1. A data processing system for motion assist monitoring, characterized by, The application relates to a health monitoring system and method. The system comprises a data fusion module, a risk prediction module and a monitoring management module. The data fusion module is used for receiving heterogeneous motion data from a user terminal, performing spatio-temporal alignment and fusion processing on the heterogeneous motion data, and generating a digital twin representing user motion health. The generation of the digital twin representing user motion health comprises: receiving heterogeneous motion data transmitted from the user terminal, the heterogeneous motion data at least comprising kinematic data, physiological data and environmental data, and performing integrity verification on the heterogeneous motion data; taking the time stamp of the motion state change in the kinematic data as a time origin and the coordinate position of the motion state change as a space origin to form a spatio-temporal coordinate system, performing time alignment on the heterogeneous motion data by means of an interpolation algorithm, and mapping the heterogeneous motion data to the spatio-temporal coordinate system by means of coordinate conversion; performing classification and noise reduction processing on the spatio-temporally aligned heterogeneous motion data, performing effectiveness verification on the heterogeneous motion data after the classification and noise reduction processing; extracting basic features from the verified heterogeneous motion data, mining the correlation between the basic features to generate associated features, and performing fusion processing on the basic features and the associated features by means of a spatio-temporal attention fusion network to generate a digital twin; the fusion processing on the basic features and the associated features comprises: when any basic feature has a feature change greater than a preset threshold, the mutual information entropy between the other two basic features is calculated to quantify the coupling strength of the two basic features and the feature change; when the basic feature has a feature change less than or equal to the preset threshold, the correlation coefficient between the basic features is calculated by means of a sliding time window, and the coupling strength and the correlation coefficient are taken as the associated features; dynamically generating time attention weights based on the change rate of the feature change, and adaptively generating space attention weights according to the numerical size of the associated features; the spatio-temporal attention fusion network performs product fusion on the time attention weights and the space attention weights to obtain comprehensive weights, and performs weighted fusion on the basic features and the associated features by means of the comprehensive weights to generate a digital twin comprising the basic features and the associated features; the risk prediction module is used for extracting the basic features and the associated features from the digital twin, learning historical heterogeneous data based on a predefined prior knowledge base to construct a parameter correlation graph, inputting the basic features and the associated features into a long short-term memory network to capture the change mode of the motion risk, inputting the parameter correlation graph into a graph convolutional neural network to mine the coupling relationship between the heterogeneous motion data, and fusing the change mode of the motion risk and the coupling relationship between the heterogeneous motion data to predict the risk probability of the user suffering from motion injury; the monitoring management module is used for generating an intervention strategy by means of reinforcement learning based on the risk probability of the user suffering from motion injury and in combination with the motion target and the historical motion performance of the user, and synchronously simulating and visually displaying the intervention strategy to the digital twin.

2. A data processing system for motion assist monitoring as defined in claim 1, characterized in that, the construction of the parameter correlation graph comprises: mapping the basic features and the associated features to the predefined prior knowledge base respectively; The risk level is divided based on the prior knowledge base, and historical heterogeneous data is classified and learned to assign learning weights to historical heterogeneous data of different risk levels. The basic features are taken as nodes, the node attributes include preset thresholds of the basic features and individual base values of users, the associated features are taken as weights of edges, and the weights of edges are corrected in combination with the feature interaction rules in the prior knowledge base to construct a parameter association graph. The weights of edges are dynamically adjusted according to the change of the time of historical heterogeneous data and the change of the user motion scene, and the parameter association graph is compared with new heterogeneous motion data at regular intervals to determine whether the weights of edges are calibrated based on the prior knowledge base.

3. A data processing system for motion assist monitoring as defined in claim 2, characterized in that, The change pattern of the motion risk includes: The basic features and the associated features are sequentially processed in time sequence to form an input sequence including time sequence change information of the features; The time points at which the change amplitudes of the features in the input sequence are greater than the feature change thresholds of the associated motion risks in the prior knowledge base are marked as feature mutation points; The long short-term memory network processes the input sequence in layers, captures the instantaneous change of the features through a first time step for the marked feature mutation points, and captures the trend change of the features through a second time step for the input sequence of the unmarked feature mutation points; The instantaneous change of the features and the trend change of the features are fused to generate a risk feature set representing the change pattern of the motion risk, and the risk feature set includes a feature mutation frequency, a cumulative change trend and a risk duration.

4. A data processing system for motion assist monitoring as defined in claim 3, characterized in that, The coupling relationship between the heterogeneous motion data includes: The basic features in the parameter association graph are taken as nodes, and the associated features are taken as weights of edges to construct an input graph structure of a graph convolutional neural network; The input graph structure is input into the graph convolutional neural network, the direct association strength between nodes is extracted through a first layer of convolution, and the indirect association path between nodes is mined through a second layer of convolution by aggregating neighbor nodes; Based on the feature interaction rules of the associated motion injuries in the prior knowledge base, attention weights are given to the associated features related to the motion injuries in the parameter association graph; The direct association strength, the indirect association path and the weighted associated features are fused to generate a coupling relationship set including a direct coupling strength, an indirect coupling path and an influence weight of a coupling feature, so as to mine the coupling relationship between the heterogeneous motion data.

5. A data processing system for motion assist monitoring as defined in claim 4, characterized in that, The risk probability of the user suffering from the motion injury includes: The risk feature set and the coupling relationship set are mapped to the same feature space, the fusion weights of the risk feature set and the coupling relationship set are dynamically assigned based on the comparison between the feature mutation frequency and the influence weight of the coupling feature and the prior knowledge base; The risk feature set and the coupling relationship set are weighted and fused through the fusion weights to generate a comprehensive risk vector, and the risk level of the current comprehensive risk vector is determined in combination with the prior knowledge base; The risk level is corrected based on the frequency of the actual motion injury in the historical heterogeneous data under the same risk level and the historical motion performance of the user to predict the risk probability of the user suffering from the motion injury.

6. A data processing system for motion assist monitoring as defined in claim 5, characterized by The intervention strategy is synchronized to the digital twin simulation and visually displayed. The intervention strategy is parsed as a parameter instruction of the digital twin, is synchronized to the digital twin for simulation, and the motion health status of the user after the simulated intervention strategy is executed is monitored to obtain a simulation effect; The intervention strategy and the simulation effect are visually displayed, execution requirements of the intervention strategy are provided, and a motion process demonstration simulated by the digital twin is synchronously displayed; Adjustment feedback of the user on the intervention strategy is obtained, and the adjustment feedback is synchronized to the digital twin in real time, and an adjusted intervention strategy is simulated and generated.

7. A data processing system for motion assist monitoring as defined in claim 6, characterized by The intervention strategy is generated by: Based on the risk probability of the user suffering from a motion injury, the generation priority of the intervention strategy is determined in combination with the motion target and the historical motion performance of the user; Taking the risk intervention effect, the target matching degree and the user adaptability as evaluation indexes, a plurality of candidate intervention strategies are generated through reinforcement learning; In combination with the user motion preference and the historical intervention record in the prior knowledge base, the candidate intervention strategies are verified to generate the intervention strategy, and the intervention strategy includes adjusting the motion intensity, the motion posture and the rest interval.

8. A data processing method for motion assist monitoring, implemented based on the data processing system for motion assist monitoring according to any one of claims 1-7, characterized in that, It includes: Receiving heterogeneous motion data from a user terminal, performing spatio-temporal alignment and fusion processing on the heterogeneous motion data, and generating a digital twin representing the motion health of the user; Extracting basic features and associated features from the digital twin, learning historical heterogeneous data based on a predefined prior knowledge base to construct a parameter association graph; The basic features and the associated features are input into a long short-term memory network to capture the change pattern of the motion risk, and the parameter association graph is input into a graph convolutional neural network to mine the coupling relationship between the heterogeneous motion data; The change pattern of the motion risk and the coupling relationship between the heterogeneous motion data are fused to predict the risk probability of the user suffering from a motion injury; Based on the risk probability of the user suffering from a motion injury, the intervention strategy is generated through reinforcement learning in combination with the motion target and the historical motion performance of the user, and the intervention strategy is synchronized to the digital twin for simulation and visual display.

Images