Training quality prediction method and system based on training damage analysis

Abstract

The invention relates to the technical field of training assistance, in particular to a training quality prediction method and system based on training damage analysis, and the method comprises the steps: multi-modal data collection and accumulative superposition, causal graph construction, attitude risk modeling, data enhancement, dynamic prediction and training plan optimization. According to the method, daily exercise data of a user is continuously collected, a chronic load accumulative calculation model is established, and a time sequence analysis method is adopted to perform long-term tracking and accumulative calculation on core indexes such as exercise amount, heart rate variability and joint activity, so that a complete personal exercise state database is formed; the chronic load accumulation mechanism breaks through the limitation of traditional single training evaluation, comprehensive monitoring and long-term trend analysis of the motion state of the user are achieved, a potential risk mode can be found earlier, a reliable data basis is provided for personalized risk evaluation and scientific training optimization, and the method is suitable for popularization and application. The problems that a traditional motion monitoring system is single in data and one-sided in evaluation are effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of training assistance, and particularly to a training quality prediction method and system based on training injury analysis. Background Art

[0002] In the field of sports training monitoring, traditional evaluation methods mainly rely on the collection and analysis of single training data, and this evaluation method has obvious limitations. Existing technologies usually only collect sports data for specific training periods, lacking continuous tracking of users' daily sports status, resulting in single and discontinuous data samples. Due to the failure to establish a complete personal sports status database, traditional methods are difficult to accurately evaluate the changing trend of users' chronic training loads and cannot timely detect potential sports risks.

[0003] Although current sports monitoring devices on the market can record some sports indicators, they generally have problems of limited data collection dimensions and simple analysis methods. These devices often only focus on the exercise intensity or heart rate changes during the current training and ignore the long-term cumulative effect of sports data. This one-sided data collection method makes the system unable to comprehensively grasp the true sports status of users, resulting in inaccurate risk assessment results.

[0004] In addition, traditional monitoring systems lack an effective data accumulation calculation model and are difficult to scientifically analyze users' long-term sports status. Due to the failure to establish an accumulation mechanism for chronic loads, existing technologies cannot identify the progressive changes in sports status and cannot provide reliable data support for personalized training. These problems seriously restrict the accuracy and practicability of sports monitoring systems, and there is an urgent need for a technical solution that can achieve long-term data tracking and comprehensive analysis. Summary of the Invention

[0005] In order to overcome the problems mentioned in the above background art, the present invention proposes a training quality prediction method and system based on training injury analysis.

[0006] The technical solution of the present invention is: A training quality prediction method based on training injury analysis, comprising the following steps:

[0007] S11: Multimodal data collection and cumulative superposition. Wearable devices and multimodal sensors are used to collect the user's motion data, and the collected data is cumulatively superimposed. Among them, the motion data includes exercise volume, heart rate, muscle load, training volume, training intensity, individual attributes, and environmental factors. The wearable devices and multimodal sensors are used to collect multi-dimensional motion data covering exercise volume, heart rate, muscle load, training volume, training intensity, individual attributes, and environmental factors, and perform cumulative superposition. This comprehensive and integrated data collection method can provide a rich and detailed data basis for the assessment of user training injuries and the prediction of training quality, making subsequent analysis more accurate and reliable, and avoiding one-sided analysis due to single data;

[0008] S12: Causal graph construction. Causal graphs are constructed for daily motion data, mental states, and environmental factors through causal discovery algorithms to identify potential injury paths. Causal graphs are constructed for daily motion data, mental states, and environmental factors through causal discovery algorithms to identify potential injury paths. This method breaks through the limitation of only focusing on the motion data itself in the traditional way, comprehensively considers the impacts of multiple factors such as mental and environmental factors on injuries, and can more deeply and comprehensively explore the potential causes leading to training injuries, providing a scientific basis for formulating targeted preventive measures in the future;

[0009] S13: Pose risk modeling. LSTM is used to analyze time series pose data, detect abnormal patterns, and generate intervention suggestions through counterfactual reasoning. LSTM is used to analyze time series pose data, detect abnormal patterns, and generate intervention suggestions through counterfactual reasoning. LSTM's powerful processing ability for time series data enables it to accurately capture the time-dependent relationships and changing trends in pose data and timely detect abnormal pose patterns. The intervention suggestions generated by counterfactual reasoning are forward-looking and targeted, and can help users adjust their poses in a timely manner to reduce the risk of injury;

[0010] S14: Data augmentation. Diffusion models are used to generate synthetic data of high-risk actions to expand the training set, and adversarial training is used to simulate pose offsets under fatigue conditions. Diffusion models are used to generate synthetic data of high-risk actions to expand the training set, and adversarial training is used to simulate pose offsets under fatigue conditions. Data augmentation not only solves the problem of scarce data of high-risk actions that may exist in the actual data collection process, enriches the training data, but also enables the model to learn more complex and realistic features by simulating pose offsets under fatigue conditions, improving the generalization ability and prediction accuracy of the model;

[0011] S15: Dynamic prediction: Establish a dynamic prediction model, train the dynamic prediction model using the expanded training set, and then use the dynamic prediction model to predict the probability of injury in the next 7 days. Establish a dynamic prediction model, train the model using the expanded training set, and predict the probability of injury in the next 7 days. The dynamic prediction model can update the prediction results in real time based on the user's latest training data and status, providing a more timely and accurate injury risk assessment. The 7-day prediction cycle provides users and coaches with sufficient time to adjust training plans and take preventive measures to effectively avoid or mitigate the occurrence of injuries.

[0012] S16: Training plan optimization: load adjustment based on ACWR thresholds and corrective exercises for abnormal posture. Load adjustment based on ACWR thresholds and corrective exercises for abnormal posture. Adjusting load based on ACWR thresholds ensures that training loads are within a reasonable range, avoiding injuries caused by overtraining while ensuring training effectiveness. Corrective exercises for abnormal posture can specifically improve the user's movement patterns, reduce the risk of injury caused by abnormal posture, improve the scientific nature and safety of training, and ultimately enhance overall training quality.

[0013] Preferably, when accumulating and superimposing the collected data, the following steps are specifically included:

[0014] S21: Data processing: First, we use sliding window technology to align the data of exercise volume, heart rate, joint angles, gait, muscle oxygen content, and blood lactate. Then, we use Z-score standardization to remove outliers. Using sliding window technology to align multi-dimensional data such as exercise volume, heart rate, joint angles, gait, muscle oxygen content, and blood lactate, we can synchronize and align data of different types and different collection frequencies in the time dimension. This provides a unified time reference for subsequent data analysis and modeling, avoids analysis bias caused by data time misalignment, and ensures the comparability and consistency of data in the time series.

[0015] S22: Feature engineering, structured conversion of data, organizing the original data into a tabular form, and then extracting features from the processed data, where the extracted features include dynamic stability index and energy metabolism characteristics; organizing the original unstructured or semi-structured data into a tabular form, making the data more standardized, easy to understand and process. The structured data format facilitates subsequent feature extraction, model training and algorithm application, improving the efficiency of data processing and analysis; extracting key features such as dynamic stability index and energy metabolism characteristics can mine representative and predictive information from the original data. The dynamic stability index can reflect the user's body balance and coordination during exercise, and the energy metabolism characteristics are closely related to the user's exercise intensity and energy consumption. These features provide an important basis for subsequent injury prediction and training quality assessment, and help to more deeply understand the user's exercise status and potential risks;

[0016] S23: Load time series analysis, using exponential weighting to calculate chronic load and increase the weight of recent data.

[0017] Preferably, when the exponential weighting method is used to calculate the chronic load, the calculation frequency is once a day, and the calculation principle formula is:

[0018] CL-α·AL current +(1-α)·CL previous ;

[0019] Among them, CL is the chronic load, that is, the cumulative superposition data result, α is the attenuation factor, and the value range of α is 0.15-0.25, AL current is the current acute load, CL prevlous is the chronic load value calculated in the previous round, that is, the chronic load value calculated the day before, where the attenuation factor α increases with the chronic load value CL calculated in the previous round. previows increases with the increase of

[0020] Using an exponential weighting method to calculate chronic load and increasing the weight of recent data can better reflect the user's recent and long-term training load. Traditional load calculation methods may assign the same weight to data at different time points, failing to accurately reflect the impact of recent training on the user's physical condition. By giving more weight to recent data, the exponential weighting method can promptly capture the changing trends of the user's training load, more accurately assess the user's fatigue level and injury risk, and provide a scientific basis for developing reasonable training plans.

[0021] Preferably, when daily exercise data is combined with psychological state and environmental factors through a causal discovery algorithm to construct a causal graph and identify potential injury paths, the following steps are specifically included:

[0022] S31: Data preparation and variable definition. The core variables defined include motion data, mental state, environmental factors, and injury endpoints, and they are classified using the 5M1E classification method. Classifying core variables such as motion data, mental state, environmental factors, and injury endpoints using the 5M1E classification method makes the variable definition more comprehensive and systematic. The 5M1E (man, machine, material, method, environment, measurement) classification method covers multiple key dimensions affecting training injuries, ensuring that various possible influencing factors can be considered when constructing the causal diagram, avoiding omission of important variables, and providing a solid foundation for subsequent causal relationship analysis;

[0023] S32: Initialize the complete graph by initializing all variables as a fully connected undirected graph. Initializing all variables as a fully connected undirected graph provides a comprehensive initial framework for subsequent conditional independence tests and causal relationship identification. This method ensures that no possible causal relationships are omitted during the analysis, providing sufficient exploration space for the subsequent application of the iterative edge deletion and edge orientation rules;

[0024] S33: Conditional independence test. Use the partial correlation coefficient to judge variable independence and perform iterative edge deletion. Among them, the rule of iterative edge deletion is: if variable X and Y are independent given the conditional set Z, then delete the edge of X - Y. Use the partial correlation coefficient to judge variable independence and gradually remove unreasonable edges through the iterative edge deletion rule. This method is based on statistical principles and can accurately identify the conditional independence between variables, thereby removing spurious causal relationships and retaining the edges with true causal associations. Through the iterative edge deletion process, the structure of the causal diagram can be gradually optimized to make it more in line with the actual situation;

[0025] S34: Edge orientation rules, including collider structure identification and V - structure verification. Among them, the principle of collider structure identification is: if there is a structure X→Z←Y and X and Y are independent, then orient it as X→Z←Y. The application of edge orientation rules such as collider structure identification and V - structure verification can accurately determine the causal direction between variables. The principle of collider structure identification is based on the independence relationship between variables, which can effectively identify the collider structure and determine its causal direction; V - structure verification further ensures the accuracy of the causal direction. The application of these rules makes the causal diagram more scientific and reasonable, and can accurately reflect the causal relationship between variables;

[0026] S35: Output the undirected graph skeleton and mark the retained causal relationship edges. Output the undirected graph skeleton and mark the retained causal relationship edges, then perform hierarchical partitioning and draw the causal diagram using the Mermaid syntax, making the causal relationship more intuitive and easy to understand. The Mermaid syntax is a simple and easy - to - use chart drawing tool that can quickly generate clear and beautiful causal diagrams, facilitating viewing and analysis by researchers and decision - makers;

[0027] S36: Causal graph construction and path identification, hierarchical division is carried out and the causal graph is drawn using Mermaid syntax, and then critical path analysis is performed; through critical path analysis, the causal path with the greatest impact on training injuries can be identified, providing an important basis for formulating targeted preventive measures. Critical path analysis can focus on the most important causal relationships, avoiding getting lost among numerous causal relationships, and improving the efficiency and pertinence of the analysis;

[0028] S37: Expert verification and correction, combining medical literature to confirm the rationality of the path, and verifying the causal strength through intervention experiments; the process of expert verification and correction is a process of continuous improvement and optimization. By continuously combining new medical research results and intervention experiment data, the causal graph can be updated and improved in a timely manner to make it more in line with the actual situation. This helps to improve the accuracy of training injury prediction and the scientific nature of training plan formulation, providing better protection for athletes' training and health.

[0029] Preferably, when constructing the causal graph and identifying the path, the causal graph is divided into the following levels:

[0030] A11: Cause layer, including environmental factors and psychological states. Among them, environmental factors include slope, temperature, and road surface type, and psychological states include stress index and sleep quality score;

[0031] A12: Intermediate layer, including motion data and physiological indicators. Motion data includes gait asymmetry and step frequency, and physiological indicators include knee joint pressure and creatine concentration;

[0032] A13: Result layer, which is the injury endpoint.

[0033] Preferably, when using LSTM to analyze time series pose data, detect abnormal patterns, and generate intervention suggestions through counterfactual reasoning, it specifically includes:

[0034] S41: LSTM modeling, capturing long-term pose patterns through LSTM units. The input layer receives the joint angle sequence, the hidden layer dynamically updates the state of the memory unit through the gating mechanism, and defines the LSTM output distribution of the normal pose, calculates the prediction error of the test sequence, and when the error exceeds the threshold, it is determined as an abnormal pattern; LSTM (Long Short-Term Memory Network), with its unique gating mechanism, can effectively capture the long-term dependencies in time series pose data. When processing the joint angle sequence, the input layer receives the original data, and the hidden layer dynamically updates the state of the memory unit through the gating mechanism, enabling the model to learn the complex patterns of pose changes over time. Compared with traditional time series analysis methods, LSTM can better process data with a long time span, avoid information loss, and thus more accurately detect abnormal patterns in poses;

[0035] S42: Critical Risk Indicator Calculation. Based on the identified key indicators for calculating training injuries, including combining kinematic equations and ground reaction force data to calculate the peak knee joint pressure. The principle formula is:

[0036] P knee -F GRF ·cos(θ knee );

[0037] where F GRF is the ground reaction force, θ knee is the valgus angle of the knee joint, and P knee is the peak knee joint pressure;

[0038] Based on the identified key indicators for calculating training injuries, combine kinematic equations and ground reaction force data to calculate key indicators such as the peak knee joint pressure. These indicators reflect the potential risks that abnormal postures may pose to the body from multiple dimensions, providing a quantitative basis for evaluating the risk of training injuries. For example, the peak knee joint pressure is an important indicator for evaluating knee joint load and injury risk. By accurately calculating this indicator, it is possible to timely detect situations where the knee joint is under excessive pressure, providing support for taking preventive measures;

[0039] S43: Application of Causal Diagram. Screen causal paths through conditional independence tests and retain significant paths; screen causal paths through conditional independence tests and retain significant paths to construct a causal diagram. This process helps to deeply understand the causal relationship between abnormal postures and potential injuries, revealing which factors are the key factors leading to abnormal postures and injuries. Through the causal diagram, the causal flow between various variables can be clearly seen, providing a theoretical basis for formulating targeted intervention measures;

[0040] S44: Counterfactual Intervention Simulation. Set counterfactual conditions and simulate the distribution of postural data after intervention through a generative model; set counterfactual conditions and simulate the distribution of postural data after intervention through a generative model to conduct counterfactual intervention simulation. This process enables the evaluation of the effects of different intervention measures without actually conducting interventions, providing a reference for selecting the optimal intervention plan. Through counterfactual reasoning, it is possible to predict the possible changes in postural data after taking a certain intervention measure, thereby pre-judging the effectiveness of the intervention measure in advance;

[0041] S45: Dynamic suggestion generation combines historical user data with real-time posture feedback to recommend action correction solutions. This approach fully accounts for individual differences and provides personalized intervention suggestions based on the user's actual situation. Users can adjust their posture based on real-time feedback, improving training effectiveness and safety. Personalized suggestions also increase user acceptance and engagement, facilitating better implementation of intervention measures.

[0042] Preferably, when using the diffusion model to generate synthetic data of high-risk actions and expand the training set, the following steps are specifically included:

[0043] S51: Data preparation: Collect real-world injury scenario data and corresponding conditional labels, and extract the spatiotemporal characteristics of the injury motion using sensors and biomechanical models as conditional inputs. Collecting real-world injury scenario data and its corresponding conditional labels provides a realistic and instructive reference for generating synthetic data. Real-world data can reflect the characteristics and patterns of high-risk movements in actual sports, while conditional labels clarify the impact of different factors on the movements, making the generated synthetic data more realistic and enriching the diversity of the training set.

[0044] S52: Conditional Diffusion, using the Transformer as a denoising network, integrates conditional information during the back-diffusion process and gradually adds Gaussian noise to the original action data to generate a noisy feature sequence. Synthetic actions are then generated through conditional guidance. Using the Transformer as a denoising network, conditional information can be efficiently integrated during the back-diffusion process. This makes it easy to generate diverse action samples based on different conditional parameters when generating synthetic data, meeting different training requirements.

[0045] S53: Controllable Generation and Verification: This approach generates diverse high-risk action samples by adjusting conditional parameters. Biomechanical simulation is then used to verify the physical plausibility of the synthesized data, ensuring that joint forces conform to medical standards. This rich synthetic data provides the model with more training samples, helping to reduce overfitting to limited real-world data. After training on a large amount of synthetic data, the model can learn more generalized features and patterns, improving its recognition accuracy for real-world high-risk actions.

[0046] Preferably, when simulating posture deviation in a fatigue state through confrontation training, the method specifically includes:

[0047] S61: Fatigue data collection and preprocessing: collect motion data under real fatigue conditions and label them with fatigue indexes. This provides a reliable reference for subsequent posture deviation simulations.

[0048] S62: Model architecture design, including a generator and a discriminator. The input of the generator is random noise and fatigue labels, and the output is simulated fatigue pose offset data. The input of the discriminator is real and generated pose data, and the output is the Wasserstein distance score. A gradient penalty term is used to constrain the Lipschitz continuity of the discriminator. The generator takes random noise and fatigue labels as inputs and can learn the internal relationship between the fatigue state and the pose offset, thereby generating more realistic simulated data. This simulation method based on real data improves the credibility and effectiveness of the generated data. After adversarial training, the generator can output high-quality simulated fatigue pose offset data. These data not only resemble real fatigue poses in morphology but also have high similarity in motion characteristics and mechanical properties, providing rich data resources for subsequent research and applications;

[0049] S63: Adversarial training, using the RMSProp optimizer to alternately optimize the generator and the discriminator. Using the RMSProp optimizer to alternately optimize the generator and the discriminator can dynamically adjust the learning rate according to different optimization objectives, improving the training efficiency and convergence speed. This optimization method enables the generator and the discriminator to continuously evolve during the process of competing with each other, ultimately generating more realistic simulated data;

[0050] S64: Class division, counting the frequency of damage types and identifying the minority classes that need to be enhanced. By counting the frequency of damage types, the minority class samples that need to be enhanced are identified. In actual data, certain types of fatigue pose offsets may occur less frequently, resulting in weak recognition ability of the model for these classes. Through class division and minority class identification, these classes can be enhanced targetedly to improve the overall performance of the model;

[0051] S65: Oversampling process, for each minority class sample, calculating its k-nearest neighbors, selecting neighboring samples based on the Euclidean distance, and randomly interpolating between the minority class sample and its neighbors to generate new samples. Using the oversampling process, for each minority class sample, calculating its k-nearest neighbors and selecting neighboring samples based on the Euclidean distance for random interpolation to generate new samples. This method can increase the number of minority class samples without changing the original data distribution, making the data distribution more balanced, thereby improving the recognition accuracy of the model for minority class samples;

[0052] S66: Optimization and adjustment. Dynamically adjust the sampling ratio according to the class loss during training, prioritize enhancing the classes that the model has difficulty in recognizing, combine with the Tomek Links method to eliminate the noise samples in the majority classes, and improve the clarity of the classification boundary; Dynamically adjust the sampling ratio according to the class loss during training, prioritize enhancing the classes that the model has difficulty in recognizing. This dynamic adjustment method can enable the model to pay more attention to the classes that have a greater impact on the classification performance, and improve the pertinence and efficiency of training.

[0053] Preferably, when establishing the dynamic prediction model, training the dynamic prediction model with the augmented training set, and then using the dynamic prediction model to predict the injury probability in the next 7 days, it specifically includes:

[0054] S71: Data preprocessing and feature extraction. Use CNN to process the heat map data of skeletal key points, extract local spatial patterns through the convolutional layer, the input data is the exponentially weighted training load data in the past 28 days, and input the spatial feature sequence output by CNN into LSTM to capture long-term dependencies; Use CNN to process the heat map data of skeletal key points, and the convolutional layer can automatically learn and extract the local spatial patterns in the data. The heat map of skeletal key points contains rich motion posture information, and CNN can capture the key features in this information through convolutional operations, such as the position and angle of joints, etc., providing accurate input features for the subsequent injury probability prediction; The input data not only includes the heat map of skeletal key points, but also combines the exponentially weighted training load data in the past 28 days. This way of fusing multi-source data enables the model to comprehensively consider the athlete's motion posture and training load conditions, more comprehensively understand the athlete's physical state, and thus improve the accuracy of injury probability prediction;

[0055] S72: Model training and verification. Predict the injury probability and sports performance indicators simultaneously, optimize the model through the composite loss function, and reserve the data in the last 7 days as the test set to verify the generalization ability of the model in the real-time data stream; Using the composite loss function to optimize the model, the loss function can be weighted and combined according to the importance and characteristics of different tasks. This can ensure that the model pays attention to the accuracy of both injury probability prediction and sports performance indicator prediction during the optimization process, avoid the model being overly biased towards a certain task, and improve the overall performance of the model;

[0056] S73: Reinforcement learning optimization. Dynamically adjust the training parameters through multi-objective reward function design and adaptive policy iteration; Through multi-objective reward function design, multiple objectives such as the accuracy of injury probability prediction, the accuracy of sports performance indicator prediction, and model complexity can be comprehensively considered, providing a clear optimization direction for the reinforcement learning algorithm. The reinforcement learning algorithm can dynamically adjust the training parameters according to the feedback of the reward function, enabling the model to balance multiple objectives and improve the overall performance of the model.

[0057] Preferably, when dynamically adjusting training parameters through multi-objective reward function design and adaptive policy iteration, it specifically includes:

[0058] S81: State-action definition. Using real-time injury probability, sports performance indicators, training load, and fatigue index as the state space, and the action space as parameter adjustments such as increasing or decreasing training volume and inserting recovery periods. Balance risks and performance through action selection. Using real-time injury probability, sports performance indicators, training load, and fatigue index as the state space comprehensively covers the key state information of athletes during training. These information are interrelated and can comprehensively reflect the physical condition and training effect of athletes, providing rich decision-making basis for the model. For example, the injury probability can intuitively reflect the injury risk of athletes, the sports performance indicators can measure the training results of athletes, and the training load and fatigue index are helpful to understand the physical endurance and recovery of athletes;

[0059] S82: Reward function design. Adopt a dual-objective reward mechanism to reward the reduction of injury probability and the improvement of sports performance, and punish the reverse changes. At the same time, introduce an entropy regularization constraint to explore the ability and prevent premature convergence. In sports training, reducing the injury probability and improving sports performance are two important goals, but there may be certain conflicts between them. The dual-objective reward mechanism can comprehensively consider these two goals, enabling the model to find a reasonable balance point during the optimization process to meet the actual training needs;

[0060] S83: Policy iteration and deployment. Perform online policy updates based on the PPO algorithm, adjust the model parameters through real-time data feedback from wearable devices, and combine the KL early stopping mechanism to prevent policy deviation. Finally, deploy the lightweight model to edge devices. Perform online policy updates based on the PPO (Proximal Policy Optimization) algorithm, which has the advantages of fast convergence speed and good stability. Adjust the model parameters through real-time data feedback from wearable devices, enabling the model to adapt to the state changes of athletes in a timely manner and achieve dynamic optimization of training parameters.

[0061] A training quality prediction system based on training injury analysis includes:

[0062] A multi-modal data acquisition module for collecting data such as exercise volume, heart rate, and muscle load through wearable devices and multi-modal sensors, and performing real-time cumulative superposition;

[0063] A data preprocessing module for aligning data using a sliding window, removing outliers by Z-score standardization, and extracting features after structured conversion;

[0064] Causal analysis module, used to construct a causal graph based on a causal discovery algorithm, identify injury paths and verify medical rationality;

[0065] Posture risk modeling module, used to analyze time series posture data using LSTM, detect abnormal patterns and generate counterfactual intervention suggestions;

[0066] Data augmentation module, used to generate high-risk action synthetic data through a conditional diffusion model, and combine adversarial training to simulate fatigue posture offset;

[0067] Dynamic prediction module, used to predict the injury probability in the next 7 days based on the spatio-temporal features fused by CNN-LSTM, and optimize the model through multi-task output;

[0068] Reinforcement learning optimization module, which designs a multi-objective reward function and uses the PPO algorithm to dynamically adjust training parameters;

[0069] Training plan optimization module, used to adjust the load according to the ACWR threshold, generate a corrective training plan for abnormal postures, and balance risk and performance;

[0070] Adversarial training module, used to simulate posture offset under fatigue conditions and optimize the data distribution through the adversarial learning of a generator and a discriminator.

[0071] Advantages of the present invention:

[0072] 1. Cumulative superposition of multi-modal data and dynamic feature extraction: By continuously collecting the user's daily motion data and establishing a chronic load cumulative calculation model, using time series analysis methods to conduct long-term tracking and cumulative calculation of core indicators such as exercise volume, heart rate variability, and joint range of motion, a complete personal motion state database is formed; This chronic load cumulative mechanism breaks through the limitations of traditional single-training evaluation, realizes the comprehensive monitoring of the user's motion state and long-term trend analysis, can discover potential risk patterns earlier, provides a reliable data basis for personalized risk assessment and scientific training optimization, and effectively solves the problems of single data and one-sided evaluation in traditional motion monitoring systems;

[0073] 2. Causality-driven posture-load association modeling: Construct a multi-level causal graph based on a causal discovery algorithm to clarify the causal relationship between environmental factors, psychological states and motion data; Generate interpretable intervention suggestions through counterfactual reasoning, break through the limitations of traditional correlation analysis, and realize the transformation from passive monitoring to active intervention;

[0074] 3. Adversarially enhanced dynamic adaptive optimization: A generative adversarial network is designed to simulate posture deviations under fatigue conditions, combined with reinforcement learning to dynamically adjust training parameters. Through a multi-objective optimization mechanism, this significantly improves the adaptability and personalization of training plans while reducing injury risk.

[0075] 4. Diffusion Model-Based Data Augmentation and Physical Verification: A conditional diffusion model is used to generate high-risk motion data that conforms to the laws of biomechanics, addressing the problem of traditional data augmentation methods violating the laws of physics. Physical simulation is used to verify the rationality of the generated data, effectively expanding the minority class samples and improving the model's detection accuracy and generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 Shown is a flow chart of the training quality prediction method based on training damage analysis of the present invention;

[0077] Figure 2 What is shown is a schematic diagram of the structure of the training quality prediction system based on training damage analysis of the present invention. DETAILED DESCRIPTION

[0078] The present invention will be further described below with reference to the accompanying drawings and examples.

[0079] See also Figure 1-2 The present invention provides an embodiment: a training quality prediction method and system based on training damage analysis, comprising:

[0080] (1) Multimodal data collection and cumulative superposition

[0081] Equipment deployment: Equip athletes with wearable devices (such as smart watches and sports bracelets) and multimodal sensors (such as accelerometers, electromyography sensors, and environmental sensors) to ensure comprehensive collection of data such as exercise volume, heart rate, muscle load, training volume, training intensity, individual attributes, and environmental factors. The collected individual attributes include body mass index, body fat percentage, fat mass, visceral fat, obesity, subcutaneous fat percentage, muscle mass, skeletal muscle mass, protein, bone mass, body water content, lean body mass, basal metabolic rate, heart rate, and age.

[0082] Data collection: The equipment collects data in real time and transmits it to the data collection and processing center.

[0083] Data processing: The sliding window technique was used to align the data of exercise volume, heart rate, joint angle, gait, muscle oxygen content, and blood lactate, and Z-score standardization was used to eliminate outliers.

[0084] Feature engineering: Perform structured transformation on the data to extract key features such as dynamic stability index and energy metabolism characteristics.

[0085] Load time series analysis: Calculate chronic load using the exponentially weighted method to enhance the weight of recent data.

[0086] (2) Causal graph construction

[0087] Data preparation and variable definition: Define core variables and classify them using the 5M1E classification method.

[0088] Initialize the complete graph: Initialize all variables as a fully connected undirected graph.

[0089] Conditional independence test: Use partial correlation coefficients to judge variable independence and perform iterative edge deletion.

[0090] Directed edge rule: Identify collider structures and V-structures and label the retained causal relationship edges.

[0091] Expert verification and correction: Combine medical literature to confirm the rationality of the path and verify the causal strength through intervention experiments.

[0092] (3) Posture risk modeling

[0093] LSTM modeling: Capture long-term posture patterns through LSTM units. The input layer receives the joint angle sequence, and the hidden layer dynamically updates the memory unit state through a gating mechanism.

[0094] Abnormal pattern detection: Define the LSTM output distribution of normal postures, calculate the prediction error of the test sequence, and when the error exceeds the threshold, it is determined as an abnormal pattern.

[0095] Intervention suggestion generation: Through counterfactual reasoning, combine the user's historical data with real-time posture feedback to recommend action correction plans.

[0096] (3) Data augmentation and adversarial training

[0097] Diffusion model to generate synthetic data: Collect data on real damage scenarios and extract the spatio-temporal features of damaged actions through sensors and biomechanical models.

[0098] Conditional diffusion: Use Transformer as a denoising network to fuse conditional information during the reverse diffusion process to generate a noisy feature sequence.

[0099] Controllable generation and verification: Generate diverse high-risk action samples by adjusting conditional parameters and use biomechanical simulations to verify the physical rationality of the synthetic data.

[0100] (4) Dynamic prediction and training plan optimization

[0101] Dynamic Prediction: Establish a dynamic prediction model and train it using the augmented training set to predict the injury probability for the next 7 days.

[0102] Training Plan Optimization: Adjust the load according to the ACWR threshold, generate a corrective training plan for abnormal postures, and balance risk and performance.

[0103] (V) System Implementation Method

[0104] Multi-modal Data Preprocessing and Feature Extraction: Use CNN to process the heatmap data of skeletal key points, extract local spatial patterns through convolutional layers, and the input data is the exponentially weighted training load data for the past 28 days.

[0105] Model Training and Validation: Optimize the model through a composite loss function, reserve the last 7 days of data as the test set, and verify the generalization ability of the model in the real-time data stream.

[0106] Reinforcement Learning Optimization: Dynamically adjust the sampling ratio according to the class loss during training, prioritize enhancing the classes that are difficult for the model to recognize, and combine the Tomek Links method to eliminate noise samples in the majority class, improving the clarity of the classification boundary.

[0107] (V) System Implementation

[0108] Multi-modal Data Acquisition Module: Real-time collect data such as exercise volume, heart rate, and muscle load through wearable devices and multi-modal sensors, and perform cumulative superposition.

[0109] Data Preprocessing Module: Align data using a sliding window, eliminate outliers through Z-score standardization, extract features after structured transformation, input to CNN to process the heatmap data of skeletal key points, and output the spatial feature sequence to LSTM to capture long-term posture patterns.

[0110] Causal Analysis Module: Construct a causal graph based on the causal discovery algorithm, identify injury paths and verify medical rationality, and verify the causal strength through intervention experiments.

[0111] Posture Risk Modeling Module: Use LSTM to analyze time series posture data, detect abnormal patterns, and generate intervention suggestions through counterfactual reasoning.

[0112] Data Augmentation Module: Generate synthetic data of high-risk actions through a conditional diffusion model, and combine adversarial training to simulate fatigue posture deviation.

[0113] Dynamic Prediction Module: Based on the CNN-LSTM fusion of spatio-temporal features, predict the injury probability for the next 7 days, and optimize the model through multi-task output.

[0114] Reinforcement Learning Optimization Module: By designing a multi-objective reward function and using the PPO algorithm to dynamically adjust training parameters, a lightweight model is deployed to intelligent devices to achieve low-latency real-time prediction and feedback.

[0115] System Architecture

[0116] Multi-modal Data Acquisition Module: Data such as exercise volume, heart rate, and muscle load are collected through wearable devices and multi-modal sensors and are cumulatively superimposed in real time.

[0117] Data Preprocessing Module: The data is aligned using a sliding window, outliers are removed by Z-score normalization, and features are extracted after structured transformation.

[0118] Causal Analysis Module: A causal graph is constructed based on a causal discovery algorithm to identify damage paths and verify medical rationality.

[0119] Posture Risk Modeling Module: LSTM is used to analyze time series posture data, detect abnormal patterns, and generate intervention suggestions through counterfactual reasoning.

[0120] Data Augmentation Module: A diffusion model is used to generate synthetic data of high-risk actions to expand the training set, and adversarial training is used to simulate posture offsets under fatigue conditions.

[0121] Dynamic Prediction Module: Based on the CNN-LSTM fusion of spatio-temporal features, the injury probability for the next 7 days is predicted, and the model is optimized through multi-task output.

[0122] Reinforcement Learning Optimization Module: By designing a multi-objective reward function and using the PPO algorithm to dynamically adjust training parameters.

[0123] Training Plan Optimization Module: The load is adjusted according to the ACWR threshold, and a corrective training plan is generated for abnormal postures to balance risk and performance.

[0124] Example 1

[0125] 1. Application Scenario

[0126] For the daily training of professional marathon runners, through multi-modal data monitoring and dynamic prediction, the training plan is optimized and the risk of knee injuries is reduced.

[0127] 2. Implementation Steps and Data Flow

[0128] (1) Data Acquisition and Preprocessing

[0129] Device Configuration:

[0130] Wearable devices: Garmin HRM-Pro heart rate belt (collects heart rate and cadence), Shimmer3 IMU sensor (joint angles, gait); Environmental sensors: Portable weather station (temperature, humidity), treadmill force plate (ground reaction force).

[0131] Data processing:

[0132] Sliding window alignment: Align heart rate (200 Hz) and gait data (100 Hz) with a 10-second window, and use Z-score to remove outliers (e.g., heart rate > 220 - age).

[0133] Feature extraction: Calculate the dynamic stability index (coefficient of variation of knee varus angle) and energy metabolism characteristics (blood lactate - heart rate ratio).

[0134] (2) Causal graph construction

[0135] Variable definition:

[0136] Cause layer: Ambient temperature (25°C), sleep quality (PSQI score = 5).

[0137] Intermediate layer: Cadence (180 steps / minute), peak knee joint pressure (calculated by OpenSim simulation).

[0138] Result layer: Patellofemoral pain syndrome (injury endpoint).

[0139] Causal discovery:

[0140] Use the PC algorithm to find the path: High temperature → Decrease in cadence → Increase in knee joint pressure → Injury, verified by an intervention experiment (cooling vest restores cadence and reduces knee joint pressure by 12%).

[0141] (3) Posture risk modeling

[0142] LSTM anomaly detection:

[0143] Input: Hip-knee-ankle three-dimensional angle time series (sampling rate 100 Hz).

[0144] Threshold setting: Root mean square error (RMSE) of the prediction error for normal running posture < 0.15, and it is detected that the knee varus angle at landing exceeds the threshold (RMSE = 0.22).

[0145] Counterfactual advice: Generate an intervention plan of "reduce stride length by 5 cm", and the simulation shows that the peak knee joint pressure is reduced by 18%.

[0146] (4) Data augmentation

[0147] Diffusion model generates high-risk actions:

[0148] Condition input: Spatiotemporal characteristics of excessive striding (stride length / height > 1.5).

[0149] Generated samples: Synthesize 10 groups of landing actions with genu varum angle > 10°, and verify the rationality of joint forces through OpenSim.

[0150] (5) Dynamic prediction

[0151] CNN-LSTM model:

[0152] Input: Chronic load (exponentially weighted) and knee joint thermogram in the past 28 days.

[0153] Output: Injury probability in the next 7 days (0 - 1), and the predicted probability for the athlete on the 3rd day reaches 0.68 (actual pain occurs on the 4th day).

[0154] (6) Training optimization

[0155] Reinforcement learning adjustment:

[0156] Actions: Reduce the next-day training volume from 15 km to 10 km and insert water recovery training.

[0157] Reward function: Decrease in injury probability (+0.3 reward), and at the same time keep VO2max unchanged (+0.1 reward).

[0158] 3. System deployment and effects

[0159] Edge device: Deploy a lightweight model to the athlete's Garmin watch for real-time warning of high-risk actions (delay < 50 ms).

[0160] Effect verification: During the 6-month period, the injury rate in the experimental group decreased by 42%, and the training performance (5 km pacing) increased by 3.2%.

[0161] 4. Key innovation points

[0162] Causality-driven intervention: Clearly define the "stride frequency - temperature - injury" chain through counterfactual reasoning, which is superior to traditional correlation analysis.

[0163] Generative data augmentation: Biomechanically reasonable data synthesized by the diffusion model increases the number of minority-class injury samples by 5 times.

[0164] Dynamic adaptability: The PPO algorithm automatically adjusts the ACWR threshold every 7 days (from 0.8 → 1.2 to adapt to the season stage).

[0165] This embodiment demonstrates the closed-loop application of the technical solution in a real sports scenario, from data collection to full-process implementation of dynamic optimization.

[0166] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the above embodiments, and various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art.

Claims

1. A training quality prediction method based on training injury analysis, characterized in that: It includes the following steps: S11: Multimodal data collection and cumulative superposition. Wearable devices and multimodal sensors are used to collect the user's motion data, and the collected data is cumulatively superimposed. Among them, the motion data includes exercise volume, heart rate, muscle load, training volume, training intensity, individual attributes, and environmental factors; S12: Causal graph construction. The daily motion data, mental state, and environmental factors are used to construct a causal graph through a causal discovery algorithm to identify potential injury paths; S13: Pose risk modeling. LSTM is used to analyze time-series pose data, detect abnormal patterns, and generate intervention suggestions through counterfactual reasoning; S14: Data augmentation. A diffusion model is used to generate synthetic data of high-risk actions to expand the training set, and adversarial training is used to simulate pose offsets in a fatigued state; S15: Dynamic prediction. A dynamic prediction model is established, and the expanded training set is used to train the dynamic prediction model. Then, the dynamic prediction model is used to predict the injury probability in the next 7 days; S16: Training plan optimization. The load is adjusted according to the ACWR threshold, and corrective training is generated for abnormal poses.

2. The training quality prediction method based on training injury analysis according to claim 1, wherein: When cumulatively superimposing the collected data, it specifically includes: S21: Data processing. First, data alignment is performed on the exercise volume, heart rate, joint angle, gait, muscle oxygen content, and blood lactic acid content through a sliding window technique, and then outliers are removed using Z-score standardization; S22: Feature engineering. The data is structurally transformed, and the original data is organized into a table form. Then, feature extraction is performed on the processed data. Among them, the extracted features include a dynamic stability index and an energy metabolism feature; S23: Load time-series analysis. The exponential weighting method is used to calculate the chronic load and enhance the weight of recent data.

3. The training quality prediction method based on training injury analysis according to claim 2, characterized in that: When constructing a causal graph through a causal discovery algorithm with daily motion data, mental state, and environmental factors to identify potential injury paths, it specifically includes: S31: Data preparation and variable definition. The defined core variables include motion data, mental state, environmental factors, and injury endpoints, and they are classified using the 5M1E classification method; S32: Initialize the complete graph. All variables are initialized as a fully connected undirected graph; S33: Conditional independence test. The partial correlation coefficient is used to judge variable independence, and iterative edge deletion is performed. The rule for iterative edge deletion is: if variable X and Y are independent under the given condition set Z, then the edge X - Y is deleted; S34: Directed edge rules, including collider structure identification and V-structure verification. The principle of collider structure identification is: if there is a structure X → Z ← Y and X and Y are independent, then it is directed as X → Z ← Y; S35: Output the undirected graph skeleton and mark the retained causal relationship edges; S36: Causal graph construction and path identification. Hierarchical partitioning is performed and the causal graph is drawn using the Mermaid syntax, and then critical path analysis is performed; S37: Expert verification and correction. The rationality of the path is confirmed by combining medical literature, and the causal strength is verified through intervention experiments.

4. A training quality prediction method based on training injury analysis according to claim 3, characterized in that: When performing causal graph construction and path identification, the causal graph is divided into the following levels: A11: Cause layer, including environmental factors and psychological states. Among them, environmental factors include slope, temperature, and road surface type, and psychological states include stress index and sleep quality score; A12: Intermediate layer, including motion data and physiological indicators. Motion data includes gait asymmetry and step frequency, and physiological indicators include knee joint pressure and creatine concentration; A13: Result layer, which is the injury endpoint.

5. The training quality prediction method based on training injury analysis according to claim 4, characterized in that: When using LSTM to analyze time series pose data, detect abnormal patterns, and generate intervention suggestions through counterfactual reasoning, it specifically includes: S41: LSTM modeling. Capture long-term pose patterns through LSTM units. The input layer receives the joint angle sequence. The hidden layer dynamically updates the memory unit state through the gating mechanism, and defines the LSTM output distribution of normal poses, calculates the prediction error of the test sequence, and determines it as an abnormal pattern when the error exceeds the threshold; S42: Calculation of key risk indicators. Calculate the key indicators for training injuries based on the identified ones, including combining kinematic equations and ground reaction force data to calculate the peak knee joint pressure; S43: Application of causal graph. Screen causal paths through conditional independence tests and retain significant paths; S44: Counterfactual intervention simulation. Set counterfactual conditions and simulate the pose data distribution after intervention through a generative model; S45: Dynamic recommendation generation. Combine the user's historical data and real-time pose feedback to recommend action correction plans.

6. The training quality prediction method based on training injury analysis according to claim 5, wherein: When using a diffusion model to generate synthetic data of high-risk actions to expand the training set, it specifically includes: S51: Data preparation. Collect real injury scenario data and corresponding conditional labels, and extract the spatio-temporal features of injury actions through sensors and biomechanical models as conditional inputs; S52: Conditional diffusion. Use Transformer as the denoising network, fuse conditional information during the reverse diffusion process, gradually add Gaussian noise to the original action data to generate a noiseified feature sequence, and then generate synthetic actions through conditional guidance; S53: Controllable generation and verification. Generate diverse high-risk action samples by adjusting conditional parameters, and use biomechanical simulation to verify the physical rationality of the synthetic data to ensure that the joint forces meet medical standards.

7. A training quality prediction method based on training injury analysis according to claim 6, characterized in that: When simulating pose offsets in a fatigued state through adversarial training, it specifically includes: S61: Fatigue data collection and preprocessing. Collect motion data in a real fatigued state and label the fatigue index; S62: Model architecture design, including a generator and a discriminator. Among them, the input of the generator is random noise and a fatigue label, and the output is simulated fatigued pose offset data; the input of the discriminator is real and generated pose data, and the output is the Wasserstein distance score, and a gradient penalty term is used to constrain the Lipschitz continuity of the discriminator; S63: Adversarial training. Use the RMSProp optimizer to alternately optimize the generator and the discriminator; S64: Class division. Statistically analyze the frequency of injury types and identify the minority classes that need to be enhanced; S65: Oversampling process. For each minority class sample, calculate its k-nearest neighbors, select neighboring samples based on the Euclidean distance, and randomly interpolate between the minority class sample and its neighbors to generate new samples; S66: Optimization and adjustment. Dynamically adjust the sampling ratio according to the class loss during training, prioritize enhancing the classes that the model has difficulty recognizing, and combine the Tomek Links method to eliminate noise samples in the majority class to improve the clarity of the classification boundary.

8. The training quality prediction method based on training injury analysis according to claim 7, wherein: When establishing a dynamic prediction model, training the dynamic prediction model using the augmented training set, and then using the dynamic prediction model to predict the injury probability for the next 7 days, it specifically includes: S71: Data preprocessing and feature extraction. Use CNN to process the skeletal key point heat map data, extract local spatial patterns through convolutional layers, the input data is the exponentially weighted training load data for the past 28 days, and input the spatial feature sequence output by CNN into LSTM to capture long-term dependencies; S72: Model training and validation. Predict both the injury probability and the sports performance indicators simultaneously, optimize the model through a composite loss function, and retain the data for the last 7 days as the test set to verify the generalization ability of the model in the real-time data stream; S73: Reinforcement learning optimization. Dynamically adjust the training parameters through multi-objective reward function design and adaptive policy iteration.

9. The training quality prediction method based on training injury analysis according to claim 8, wherein: When dynamically adjusting the training parameters through multi-objective reward function design and adaptive policy iteration, it specifically includes: S81: State-action definition. Use the real-time injury probability, sports performance indicators, training load, and fatigue index as the state space, and the action space is parameter adjustments such as increasing or decreasing the training volume and inserting recovery periods. Balance risk and performance through action selection; S82: Reward function design. Adopt a dual-objective reward mechanism to reward the reduction of injury probability and the improvement of sports performance, and punish the reverse changes. At the same time, introduce entropy regularization to constrain the policy exploration ability to prevent premature convergence; S83: Policy iteration and deployment. Perform online policy updates based on the PPO algorithm, adjust the model parameters through real-time data feedback from wearable devices, and combine the KL early stopping mechanism to prevent policy deviation. Finally, deploy the lightweight model to edge devices.

10. A training quality prediction system based on training injury analysis according to claim 9, characterized in that: It includes: A multi-modal data acquisition module, which is used to collect data such as exercise volume, heart rate, and muscle load through wearable devices and multi-modal sensors, and perform real-time cumulative superposition; A data preprocessing module, which is used to align data using a sliding window, eliminate outliers by Z-score standardization, and extract features after structured transformation; A causal analysis module, which is used to construct a causal graph based on a causal discovery algorithm, identify injury paths and verify medical rationality; A posture risk modeling module, which is used to analyze time series posture data using LSTM, detect abnormal patterns, and generate counterfactual intervention suggestions; A data augmentation module, which is used to generate high-risk action synthetic data through a conditional diffusion model, and combine adversarial training to simulate fatigue posture deviation; A dynamic prediction module, which is used to predict the injury probability for the next 7 days based on the CNN-LSTM fused spatio-temporal features, and optimize the model through multi-task output; A reinforcement learning optimization module, which dynamically adjusts the training parameters by designing a multi-objective reward function and using the PPO algorithm; A training plan optimization module, which is used to adjust the load according to the ACWR threshold, generate a corrective training plan for abnormal postures, and balance risk and performance; An adversarial training module, which is used to simulate posture offsets in a fatigued state and optimize the data distribution through adversarial learning between a generator and a discriminator.

Images