Athlete ai precise physical fitness evaluation adaptive training scheme generation method
Patent Information
- Application Number
- CN202611110718.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-24
- Publication Date
- 2026-10-23
Smart Images
Figure CN122964822A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of interdisciplinary technology of artificial intelligence and sports training, and in particular relates to a method for generating adaptive training programs for athletes based on AI-based precise physical fitness assessment. Background Technology
[0002] With the scientific and precise development of competitive sports, physical fitness assessment and personalized training have become key to improving athletes' competitive level and extending their athletic lifespan. Traditional physical fitness training models have obvious shortcomings. Their physical fitness assessments rely on coaches' subjective experience and a small number of individual indicators, resulting in a single assessment dimension and highly subjective results, which cannot comprehensively reflect the athlete's overall physical fitness status. At the same time, training programs often use general templates and rely on human experience, making it difficult to adapt to individual differences and physical shortcomings of athletes. Furthermore, the programs are statically executed and cannot be dynamically adjusted according to the real-time physical condition during training, which can easily lead to low training efficiency and even problems such as overtraining and sports injuries.
[0003] In recent years, artificial intelligence technology has been gradually applied to the field of sports training, which has to some extent made up for the shortcomings of traditional training based on experience. However, existing AI training programs still have many defects.
[0004] First, the assessment accuracy is insufficient. Most solutions only collect basic physiological data such as heart rate and steps through wearable devices and evaluate physical fitness through simple statistical analysis, which cannot accurately identify core physical fitness qualities such as strength, explosiveness, and movement patterns.
[0005] Secondly, the training programs have weak adaptability. The training plans generated by machine learning and historical data are mostly static template matching, which cannot be dynamically adjusted according to changes in the athlete's physical condition and training environment, resulting in poor scenario adaptability.
[0006] In addition, existing technologies have not yet constructed a complete closed loop of "assessment-training-feedback". Training execution data cannot be used to optimize the assessment model and scheme logic in reverse, making it difficult to achieve iterative improvement of training effects. Moreover, the model's generalization ability is insufficient, making it unable to adapt to the differentiated physical fitness requirements of different sports such as track and field, combat sports, and ball games, thus limiting its application scenarios. The insufficient depth of multi-source heterogeneous data fusion is a core technical bottleneck. Existing methods simply splice and weight multimodal data such as physiological, athletic performance, and movement posture, which cannot explore deep data correlations or dynamically adjust data weights according to the characteristics of the project, greatly limiting the improvement of physical fitness assessment accuracy.
[0007] In summary, current physical training techniques have significant shortcomings in terms of assessment accuracy, training adaptability, project generalization ability, and closed-loop optimization mechanism, which cannot meet the scientific and precise training needs of high-level competitive sports. Summary of the Invention
[0008] To address the shortcomings of existing technologies, such as single-dimensional physical fitness assessment, low multi-source data fusion leading to insufficient assessment accuracy, static training programs with poor adaptive adjustment capabilities, weak generalization of projects, and lack of closed-loop optimization mechanisms, this invention provides a method for generating adaptive training programs for athlete AI-based precise physical fitness assessment. This method utilizes multimodal deep learning, knowledge graph reasoning, and deep reinforcement learning technologies to achieve precise quantitative assessment of physical fitness and adaptive dynamic generation of training programs, thereby constructing a complete closed-loop optimization system.
[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:
[0010] A method for generating adaptive training programs for athletes based on AI-based precise physical fitness assessment includes the following steps:
[0011] S1 Multidimensional Physical Fitness Data Acquisition and Preprocessing
[0012] Athletes' comprehensive physical fitness data is collected through multiple types of sensing devices, including three main categories: physiological time-series data, structured data of athletic performance, and movement and posture data. At the same time, athletes' basic attribute data is also recorded. The collected raw data is preprocessed by cleaning, denoising, normalizing, and filling missing values to form a standardized dataset.
[0013] Physiological time-series data is collected through electromyography (EMG) sensors, heart rate monitors, blood oxygen monitoring devices, and blood lactate detection devices. This includes time-series indicators such as real-time heart rate, heart rate variability, blood oxygen saturation, surface EMG signals, and blood lactate concentration, reflecting the athlete's cardiopulmonary function and muscle condition. Structured performance data is collected through strength testing devices, speed radar, and agility testing systems. This includes quantitative indicators such as maximum strength, explosive power, reaction speed, endurance level, agility, and flexibility, reflecting the athlete's core athletic qualities. Movement and posture data is collected through 3D motion capture systems or monocular vision posture estimation algorithms. This includes data such as 3D coordinates of skeletal key points, joint range of motion, force exertion sequence, and movement trajectory, used to assess the athlete's movement pattern standardization and force exertion efficiency. Basic attribute data includes information such as the athlete's age, gender, height, weight, sport, years of training, historical injury records, and recent training load, serving as the basic constraints for assessment and program generation.
[0014] In the data preprocessing stage, wavelet transform is used to remove noise for time-series data, and linear interpolation and sliding window filling are used to handle missing values; Z-score standardization is used for structured data; and coordinate system alignment and outlier removal are performed for attitude data to ensure the consistency and usability of data from different modalities.
[0015] S2 Multimodal Fusion Precision Fitness Assessment
[0016] A multimodal feature fusion deep learning model is constructed, which extracts specific features from data of different modalities. The adaptive weighted fusion of multimodal features is achieved through a cross-modal attention mechanism. At the same time, a physical fitness knowledge graph is introduced to assist in reasoning. Finally, the model outputs physical fitness scores of each dimension, comprehensive physical fitness level, as well as identified physical fitness weaknesses and potential injury risks.
[0017] This step specifically includes:
[0018] S21 Single-Modal Feature Extraction: For physiological time-series data, a Temporal Convolutional Network (TCN) is used to extract long-term temporal dependency features and capture the dynamic changes of physiological indicators; for structured data of motor performance, a three-layer fully connected network is used to extract nonlinear mapping features and explore the correlation between various motor quality indicators; for action posture data, a Graph Convolutional Network (GCN) is used to model the skeletal topology, extract spatial posture features and temporal action features, and evaluate the rationality of action patterns and force efficiency.
[0019] S22 Cross-modal Feature Fusion: A cross-modal attention module is designed to calculate the similarity matrix between features of different modalities, generate adaptive attention weights for each modality, and perform weighted fusion of single-modal features to obtain global fused features; the attention weights can be dynamically adjusted according to the athlete's sport type, for example, increasing the weight of physiological time series data for endurance sports and increasing the weight of athletic performance and posture data for strength sports.
[0020] S23 Knowledge Graph-Assisted Reasoning: Constructs a knowledge graph in the physical fitness domain, with entities covering four major categories: physical fitness indicators, athletic qualities, injury types, and training movements. Relationships between entities are defined as positive promotion, negative impact, and risk association. The graph reasoning algorithm is used to complete the assessment dimensions, identify hidden physical fitness shortcomings that cannot be directly reflected by explicit indicators, and infer potential sports injury risks, outputting risk levels and risk locations.
[0021] S24 Multi-Task Assessment Output: The global fusion features and knowledge reasoning results are concatenated and input into the multi-task output layer. The comprehensive physical fitness level (excellent, good, qualified, weak) is output through the classification branch, and the quantitative scores of six dimensions (strength, endurance, speed, explosiveness, agility, flexibility) are output through the regression branch. At the same time, the probability of injury risk and the ranking of weaknesses are also output.
[0022] S3 Adaptive Training Scheme Generation Based on Deep Reinforcement Learning
[0023] Using physical fitness assessment results, athlete basic attributes, training cycle goals, and injury constraints as state inputs, a pre-trained deep reinforcement learning model generates personalized and executable training plans. The plans include training movement combinations, load intensity parameters, set and rest interval arrangements, key points of movement standardization, and recovery strategies.
[0024] This step uses the Proximal Policy Optimization (PPO) algorithm to build a reinforcement learning model, and the specific design is as follows:
[0025] State space: includes six dimensions of physical fitness score, comprehensive physical fitness level, athlete basic attribute vector, training goals (such as improving explosive power, improving endurance, pre-competition adjustment, etc.), and injury constraints (such as avoiding squat-type movements due to knee joint injury).
[0026] Movement space: includes continuous or discrete movements in six dimensions: training movement selection, load intensity (1RM percentage), number of repetitions per set, number of sets, rest interval, and recovery method (active recovery / passive recovery).
[0027] Reward function: A multi-objective weighted reward function is adopted, the formula is as follows: ,in Rewards for improved training effectiveness are calculated based on the performance improvement rate under similar training conditions. To reward injury risk control, positive rewards are given for load and movement within safe thresholds, and negative penalties are given for exceeding the thresholds; The load fit reward is calculated based on the athlete's real-time physiological feedback and the degree of fit with the load; weighting coefficient. , , Adjust dynamically according to the type of sport; for example, increase the intensity of strength-based sports. Weighting: During rehabilitation training, the β weight is increased. After the model is pre-trained with a large amount of historical training data, it can output the optimal training plan for the input state, ensuring that the plan improves physical fitness while controlling the risk of injury and adapting to the individual condition of the athlete.
[0028] S4 closed-loop feedback and model iterative optimization construct a complete closed loop of "evaluation-generation-execution-feedback-optimization", which is divided into two levels: online solution adjustment and offline model optimization.
[0029] Online adjustment: During training, wearable devices collect physiological data such as heart rate and electromyography, as well as movement posture data of athletes in real time, and calculate the deviation between the actual load and the preset load. When the deviation exceeds the set threshold, the intensity, number of sets or rest time of subsequent training are automatically adjusted to achieve dynamic adaptive adjustment during the training process. At the same time, injury risk indicators are monitored in real time, and the load is automatically reduced or low-risk movements are switched when the risk value increases.
[0030] Offline optimization: At fixed intervals (e.g., weekly), historical training and evaluation data within the period are summarized, and the multimodal fusion evaluation model and reinforcement learning scheme generation model are fine-tuned and the model parameters are updated; at the same time, the entities and relationships of the physical fitness knowledge graph are expanded to continuously improve the evaluation accuracy and scheme generation quality of the model, forming a positive iterative optimization mechanism.
[0031] Furthermore, this method supports multi-sports adaptation. By embedding sports feature vectors, the physical requirements, evaluation standards, and training logic of different sports are encoded into feature vectors. After inputting these vectors into the model, the sports adaptation mode can be quickly switched, supporting applications in various sports such as track and field, ball games, combat sports, and swimming.
[0032] The beneficial effects of the present invention after adopting the above structure are as follows:
[0033] (1) By using multimodal data fusion and cross-modal attention mechanism, we can deeply explore the correlation information of three types of data: physiological, sports performance and posture. Combined with the implicit reasoning ability of knowledge graph, we can achieve accurate quantitative assessment of physical fitness in six dimensions. Compared with single data assessment methods, the accuracy rate is improved, and we can effectively identify hidden physical fitness shortcomings and potential injury risks.
[0034] (2) Deep reinforcement learning is used to generate training programs, which can generate highly personalized programs based on the individual athlete’s condition, training goals and injury constraints; at the same time, real-time training feedback is combined to realize online dynamic adjustment, which solves the problem of poor adaptability of traditional static programs.
[0035] (3) Through the project feature embedding mechanism, it can quickly adapt to the physical fitness requirements and training logic of different sports projects, without having to retrain the model for a single project, which greatly expands the application scenarios.
[0036] (4) A complete closed loop from data collection, evaluation, scheme generation to feedback optimization has been constructed. The adaptability of a single training session is ensured by online adjustment, and the model capability is continuously improved by offline iteration, forming a positive cycle of scientific training, effectively improving training efficiency and reducing the risk of sports injury. Attached Figure Description
[0037] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.
[0038] Figure 1 This is a wireframe diagram providing an overview of the overall technical solution of the present invention;
[0039] Figure 2 This is a detailed wireframe diagram of the multimodal fusion body performance evaluation of the present invention;
[0040] Figure 3 Generate a wireframe diagram for the PPO reinforcement learning scheme of this invention;
[0041] Figure 4 This is a wireframe diagram of the closed-loop feedback and iterative optimization of the present invention. Detailed Implementation
[0042] Step 1: Multi-dimensional physical fitness data collection and preprocessing
[0043] The data acquisition terminals used in this embodiment mainly include a smart wearable bracelet, a 4-channel surface electromyography sensor, an 8-camera 3D motion capture system, a strength testing platform, and a chest heart rate monitor. In addition, the athlete's basic attribute data is entered through the supporting athlete information management system.
[0044] 1.1 Data Collection Content
[0045] Physiological time-series data: The heart rate and blood oxygen saturation were sampled at a frequency of 1Hz, and the surface electromyography signal was sampled at a frequency of 1000Hz. The sampling sites covered four core muscle groups: quadriceps, hamstrings, biceps, and gastrocnemius. Blood lactate concentration data before and after training were recorded simultaneously. The total duration of a single test was 30 minutes, fully covering the four stages of rest, warm-up, specific testing, and recovery.
[0046] Athletic performance data: A total of 6 core performance indicators, namely 1RM squat maximum strength, 30-meter sprint speed, standing vertical jump height, YOYO interval endurance running performance, T-run agility performance, and sit-and-reach flexibility performance.
[0047] Movement posture data: Collected by a 3D motion capture system. When athletes complete standard test movements such as squats, vertical jumps, and sprints, the system simultaneously captures skeletal data at a sampling frequency of 200Hz, outputting the 3D coordinates of 25 core skeletal key points, and then calculating posture parameters such as knee joint angle, hip joint angle, trunk tilt angle, and ankle dorsiflexion angle based on this.
[0048] Basic attribute data: The information entered includes the athlete's age of 24, male, event is men's sprint, training years of 6 years, no history of major injuries, recent weekly training volume of 12 hours, and the current training cycle goal is to improve aerobic endurance and maintain explosive power level.
[0049] 1.2 Data Preprocessing
[0050] The processing of physiological time series data is divided into three steps: First, 5-level decomposition and reconstruction is performed using db4 wavelet to remove baseline drift and 50Hz power frequency noise; then, missing values are processed using the sliding window method, with the window size set to 5 sampling points; finally, min-max normalization is performed on each index to uniformly map the data to the [0,1] interval.
[0051] Structured data such as athletic performance are standardized using Z-score, which is calculated by referencing norm data of athletes of the same age and sport. The formula is z=(x−μ) / σ, where μ is the mean of the group and σ is the standard deviation of the group, thereby eliminating the dimensional differences between different indicators.
[0052] The processing flow of motion posture data is as follows: First, align the world coordinate system with the human body local coordinate system, and establish the human body coordinate system with the center point of the pelvis as the origin; then, use the isolated forest algorithm to remove abnormal posture frames and eliminate noise in the motion capture process; finally, calculate the joint angle sequence and force timing features of each motion cycle, and extract parameters such as peak angle, force duration, and force transmission efficiency of the motion cycle.
[0053] Step 2: Precise physical fitness assessment using multimodal fusion
[0054] This embodiment establishes a corresponding multimodal fusion evaluation model, and the specific implementation method is as follows:
[0055] 2.1 Single-modal feature extraction network
[0056] Physiological temporal features were extracted using a Temporal Convolutional Network (TCN). The network consists of three dilated causal convolutional layers with dilation coefficients of 1, 2, and 4, and a kernel size of 3. Each layer is followed by a ReLU activation function and a layer normalization module. The input dimension is [batch size, 1800, 8], corresponding to eight physiological indicators sampled at 30 minutes and 1Hz. The final output is a 128-dimensional temporal feature vector, which is used to capture the dynamic changes and fatigue characteristics of physiological indicators.
[0057] The athletic performance features are processed through a 3-layer fully connected network with hidden layer dimensions of 64 and 32 respectively. The activation function is ReLU, and the final output is a 32-dimensional structured feature vector to explore the nonlinear correlation between various athletic qualities.
[0058] Action posture features are extracted using a 2-layer graph convolutional network (GCN). An adjacency matrix is constructed using 25 skeletal keypoints as nodes and natural skeletal connections as edges to extract spatial topological features. At the same time, a temporal convolutional layer is superimposed to capture the temporal variation features of the action. Finally, a 64-dimensional posture feature vector is output, which reflects the standardization of the action pattern and the efficiency of force exertion.
[0059] 2.2 Cross-modal attention fusion module
[0060] First, the feature vectors of the three unimodal features are mapped to a unified 64-dimensional feature space through a linear layer, and then fed into the cross-modal attention module. The module first calculates the cosine similarity matrix between the three modal features, generates the attention weight matrix through the softmax function, and then performs a weighted summation on the features of each modality to obtain a 128-dimensional global fusion feature.
[0061] The athletes in this embodiment are sprinters, which is an explosive power-dominated sport. The system will automatically adjust the attention weight baseline, set the weight of athletic performance characteristics to 0.4, the weight of posture characteristics to 0.35, and the weight of physiological characteristics to 0.25, highlighting the evaluation proportion of explosive power quality and movement patterns, and adapting to the characteristics of the sport.
[0062] 2.3 Construction and Reasoning of Physical Fitness Knowledge Graph
[0063] The physical fitness knowledge graph constructed in this embodiment contains 126 entities and 312 relationships. The entities are divided into 4 main categories:
[0064] Physical fitness indicators include 38 items such as resting heart rate, blood lactate threshold, and vertical jump height.
[0065] Athletic qualities: covering 6 major categories including maximum strength, aerobic endurance, reaction speed, and relative explosive power, with a total of 24 specific qualities;
[0066] Injury type category: Includes 32 common sports injuries such as muscle strain, patellar tendinitis, and fatigue periostitis;
[0067] Training exercises include 26 basic exercises such as back squats, Romanian deadlifts, interval running, and box jumps.
[0068] There are four main types of relationships: positive influence, negative influence, risk association, and movement adaptation. For example, quadriceps strength positively influences squat performance, foot dorsiflexion limitation negatively influences squat movement pattern, weak hamstring strength is associated with the risk of hamstring strain, and deadlift movement adaptation improves posterior chain strength.
[0069] The reasoning process uses the TransE algorithm for knowledge embedding reasoning, inputting the explicit indicator vectors obtained from the assessment and outputting the athlete's implicit weaknesses and injury risk probabilities.
[0070] In this embodiment, it was deduced that the athlete's hamstring concentric strength is relatively weak, accompanied by a mild risk of strain. This conclusion will be added to the evaluation results as supplementary content.
[0071] 2.4 Multi-task output layer
[0072] After concatenating the global fusion features with the 16-dimensional knowledge reasoning features, the result is input into a multi-task output layer, which contains three parallel branches:
[0073] Level classification branch: using softmax activation, outputting the probability of four levels: excellent, good, qualified, and weak. In this embodiment, the athlete's overall physical fitness level is judged as "good".
[0074] Dimensional scoring branches: Linear activation is used to output six dimensions of scores from 0 to 100. The results in this example are: strength 82 points, speed 88 points, endurance 65 points, explosive power 90 points, agility 76 points, and flexibility 68 points.
[0075] Risk output branch: Using sigmoid activation, the probability of injury risk for each part is output. In this embodiment, the probability of hamstring strain is 0.32, which is a low risk level.
[0076] The system ultimately outputs a complete physical fitness assessment report, which clearly identifies the core weaknesses as aerobic endurance and lower limb flexibility, and the potential risk as hamstring strain.
[0077] Step 3: Generation of Adaptive Training Scheme Based on Deep Reinforcement Learning
[0078] This embodiment uses the PPO algorithm to build a reinforcement learning scheme generation model, and the specific implementation is as follows:
[0079] 3.1 State Space Design
[0080] The state vector has a total of 48 dimensions, including: 6-dimensional rating (6 dimensions), comprehensive level encoding (4-dimensional one-hot encoding), basic attribute encoding (12 dimensions), training objective encoding (8 dimensions), and injury constraint encoding (18 dimensions). In this embodiment, the training objective encoding corresponds to the training objective of "improving endurance level, taking into account the maintenance of explosive power, and controlling the risk of hamstring injury".
[0081] 3.2 Motion Space Design
[0082] The action space is a mixture of discrete and continuous spaces:
[0083] Discrete actions: Select 6-8 training actions from a library of 26 basic actions, and choose active / passive recovery methods;
[0084] Continuous exercise: including load intensity (0.4-0.951RM), number of repetitions per set (1-20), number of sets (2-8), and rest interval (30-300 seconds).
[0085] 3.3 Reward Function Design
[0086] The reward function weights are set in accordance with the characteristics of the sprint event, with corresponding parameters of α=0.5, β=0.3, and γ=0.2.
[0087] To reward performance, the physical fitness improvement is calculated based on historical data of similar training programs. The greater the expected improvement in endurance, the higher the reward, while a decrease in explosive power will result in a corresponding penalty.
[0088] For safety incentives, positive rewards are given when the load on hamstring-related movements is within the safe threshold, and deductions are made proportionally when the load exceeds the threshold; additional rewards are given for avoiding high-risk movements.
[0089] To ensure appropriate rewards, the system calculates the matching degree between physiological indicators such as post-training blood lactate and heart rate and the training load. Positive rewards are given when the indicators are within a reasonable range, and the corresponding rewards are deducted when the indicators deviate too much.
[0090] 3.4 Model Training and Solution Output
[0091] The model is pre-trained using 100,000 training data points from professional athletes. The policy network is iteratively optimized using the PPO algorithm, with a pruning coefficient of 0.2 and a learning rate of 0.0003. After inputting the state vector of this embodiment, the training scheme output by the policy network is as follows:
[0092] Training Theme: Improve aerobic endurance + maintain explosive power + hamstring risk control. Combination of exercises: Aerobic interval running, box jumps, Romanian deadlifts, dynamic hip stretches, core planks, and eccentric hamstring training. Load parameters: Interval running intensity: 85% of maximum heart rate, 8 sets, 3 minutes per set, 2-minute rest between sets; Box jumps: 60% of maximum vertical jump height, 4 sets x 8 reps, 90-second rest; Romanian deadlifts: 50% of 1RM, 4 sets x 10 reps, 120-second rest; Eccentric hamstring training: 3 sets x 12 reps, 60-second rest. Recovery strategy: 15 minutes of active relaxation + static stretching after training, focusing on stretching the hamstrings and hip flexors, combined with foam rolling relaxation.
[0093] The plan also includes key points for proper form, clarifying the knee angle and trunk tilt requirements for the Romanian deadlift to reduce the risk of hamstring strain.
[0094] Step 4: Closed-loop feedback and model iterative optimization
[0095] 4.1 Online dynamic adjustment
[0096] During training, the system collects athletes' heart rate and electromyography (EMG) data in real time. When the system detects that the athlete's heart rate is consistently above 90% of the maximum heart rate and the median frequency of EMG decreases by more than 15%, it determines that neuromuscular fatigue has accumulated. At this time, the system will automatically extend the interval time of the next two sets of interval running by 30 seconds and reduce the running speed of one set to ensure that the load is within a controllable range.
[0097] Throughout the training, the electromyographic signals and movement posture of the hamstrings will be monitored in real time. When there is asymmetry in force exertion or deformation of movement, an early warning will be issued in time, prompting athletes to adjust their movement posture. If necessary, the training movement will be replaced with a lower-risk movement to control the risk of injury.
[0098] 4.2 Offline Model Iteration
[0099] Every weekend, we summarize the training and retest data of all athletes for the week. We then use transfer learning to fine-tune the evaluation model and reinforcement learning model, setting the base learning rate to 0.0001 and the iteration rounds to 10. We also employ an early stopping mechanism to prevent model overfitting and add newly discovered index correlations and action adaptation relationships to the knowledge graph to continuously optimize reasoning capabilities.
[0100] Through such continuous iteration, the model's evaluation accuracy can be improved by 1-2% per month, and the load adaptability of the program will also gradually improve, forming a positive optimization cycle. In this embodiment, after 3 months of systematic training, the athlete's aerobic endurance level improved by 12%, and the explosive power remained at more than 95% of the original level. No hamstring injury occurred during the training period, which verified the actual effectiveness of this method.
[0101] Those skilled in the art will understand that this invention is not limited to sprint events. By adjusting the feature embedding vectors of the events, it can be quickly adapted to various sports such as basketball, weightlifting, and swimming, and has a wide range of applications.
[0102] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents. In conclusion, if those skilled in the art, inspired by this description, design similar structural methods and embodiments without departing from the inventive spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A method for generating adaptive training programs for athletes based on AI-based precise physical fitness assessment, characterized in that, Includes the following steps: S1. Multi-dimensional physical fitness data collection and preprocessing: Collect three types of physical fitness-related data from athletes: physiological time-series data, structured data of athletic performance, and movement posture data. At the same time, input athletes' basic attribute data. Perform preprocessing on the raw data of different modalities to form a standardized dataset. S2. Precise physical fitness assessment through multimodal fusion: Single-modal features are extracted from the three types of data to obtain corresponding single-modal features. Adaptive attention weights for each modality are calculated through a cross-modal attention mechanism. The single-modal features are weighted and fused to obtain global fusion features. A knowledge graph in the physical fitness domain is introduced to assist in reasoning and obtain knowledge reasoning results. The global fusion features and knowledge reasoning results are combined to perform multi-task assessment output and obtain the physical fitness assessment results. S3. Adaptive training program generation based on deep reinforcement learning: Using physical fitness assessment results, athlete basic attributes, training cycle goals and injury constraints as state inputs, a personalized training program containing training movement combinations, load intensity parameters, set interval arrangements and recovery strategies is generated through a pre-trained deep reinforcement learning model. S4. Closed-loop feedback and model iterative optimization: During training, athletes' physiological data and movement posture data are collected in real time. The training plan is dynamically adjusted online according to changes in load deviation and injury risk. Historical training and assessment data are periodically summarized, and the assessment model, plan generation model and physical fitness knowledge graph are iteratively optimized offline.
2. The method for generating an adaptive training program for precise AI-based physical fitness assessment of athletes according to claim 1, characterized in that, In step S1: The physiological time-series data includes at least one of real-time heart rate, heart rate variability, blood oxygen saturation, surface electromyography signal, and blood lactate concentration; the structured data of athletic performance includes at least one of maximum strength, explosive power, reaction speed, endurance level, agility, and flexibility; the movement posture data includes at least one of three-dimensional coordinates of skeletal key points, joint range of motion, force exertion sequence, and movement trajectory; and the basic attribute data includes at least one of age, gender, height and weight, sport, years of training, historical injury records, and recent training load. Preprocessing methods include: denoising and missing value imputation for time-series data, standardization for structured data, and coordinate system alignment and outlier removal for attitude data.
3. The method for generating an adaptive training program for precise AI-based physical fitness assessment of athletes according to claim 1, characterized in that, In step S2, the specific method for single-modal feature extraction is as follows: for physiological time-series data, a temporal convolutional network is used to extract long temporal dependency features; for structured motion performance data, a fully connected network is used to extract nonlinear mapping features; and for action posture data, a graph convolutional network is used to model the skeletal topology and extract spatial posture features and temporal action features.
4. The method for generating an adaptive training program for precise AI-based physical fitness assessment of athletes according to claim 1, characterized in that, In step S2, the adaptive attention weights for each modality are dynamically adjusted according to the type of sport the athlete belongs to: for endurance sports, the attention weights corresponding to physiological time series data are increased, and for strength sports, the attention weights corresponding to athletic performance data and movement posture data are increased.
5. The method for generating an adaptive training program for precise AI-based physical fitness assessment of athletes according to claim 1, characterized in that, In step S2: The entities in the physical fitness domain knowledge graph cover four categories: physical fitness indicators, athletic qualities, injury types, and training movements. Positive promotion, negative impact, and risk association relationships are defined between entities. Hidden physical fitness deficiencies and potential sports injury risks are identified through graph reasoning algorithms. The multi-task assessment output includes three parallel branches: the classification branch outputs the overall physical fitness level, the regression branch outputs a multi-dimensional physical fitness quantitative score, and the risk output branch outputs the probability of injury and the ranking of physical fitness weaknesses.
6. The method for generating an adaptive training program for precise AI-based physical fitness assessment of athletes according to claim 1, characterized in that, In step S3, the deep reinforcement learning model is constructed using a proximal policy optimization algorithm. Its state space includes multi-dimensional physical fitness scores, comprehensive physical fitness levels, athlete basic attribute vectors, training objectives, and injury constraints. Its action space includes six dimensions: training action selection, load intensity, number of repetitions per set, number of sets, rest interval, and recovery method.
7. The method for generating an adaptive training program for precise AI-based physical fitness assessment of athletes according to claim 6, characterized in that, In step S3, the deep reinforcement learning model adopts a multi-objective weighted reward function, which includes a training effect improvement reward, an injury risk control reward, and a load fit reward. The weight coefficient of each reward is dynamically adjusted according to the type of sport.
8. The method for generating an adaptive training program for precise AI-based physical fitness assessment of athletes according to claim 1, characterized in that, In step S4, the rules for online dynamic adjustment are as follows: when the deviation between the actual load and the preset load exceeds the set threshold, the intensity, number of sets or interval time of subsequent training are adjusted; when the injury risk index increases, the training load is automatically reduced or switched to low-risk training actions.
9. The method for generating an adaptive training program for precise AI-based physical fitness assessment of athletes according to claim 1, characterized in that, In step S4, the offline iterative optimization specifically involves: using historical training data and evaluation data within the period to fine-tune the parameters of the multimodal fusion evaluation model and the reinforcement learning scheme generation model, while expanding the entities and relationships of the physical fitness knowledge graph.
10. The method for generating an adaptive training program for precise AI-based physical fitness assessment of athletes according to claim 1, characterized in that, The method encodes the physical requirements, evaluation standards, and training logic of different sports into feature vectors by embedding project features into vectors. After inputting the vectors into the model, the corresponding project adaptation mode is switched, supporting applications of multiple sports.