A campus football training youth physical fitness safety monitoring method
Patent Information
- Application Number
- CN202610903622.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-23
- Publication Date
- 2026-09-15
Smart Images
Figure CN122762273A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of sports safety monitoring technology for teenagers, specifically a method for monitoring the physical fitness safety of teenagers in campus football training. Background Technology
[0002] As an important component of youth physical education, school football has been widely promoted nationwide in recent years. However, young athletes are in a critical period of rapid development of their bones, muscles, and cardiopulmonary function, and their physiological characteristics differ fundamentally from those of adult athletes, making physical fitness and safety monitoring during training particularly important.
[0003] Currently, physical fitness and safety monitoring in school football training mainly relies on three methods: First, wearable devices based on heart rate monitoring, such as smart bracelets and heart rate belts, collect exercise heart rate parameters and set fixed thresholds for early warning. These devices reflect the intensity of the body's physiological response to exercise load, rather than the actual physical intensity of the body's movement, which can easily lead to systemic biases for adolescents whose cardiopulmonary function is not yet fully developed. Second, accelerometer-based motion monitoring devices collect exercise data through accelerometer sensors, but different accelerometers generally lack sufficient measurement accuracy for low-intensity activities, making it difficult to accurately capture details such as exercise density and recovery status during training. Third, traditional physical fitness tests and coaches' manual observation, which assess athletes' condition through periodic physical fitness tests. However, the assessment cycle is long, lacks real-time accuracy, and is highly dependent on the coach's personal experience.
[0004] To address the aforementioned issues, some technical solutions have attempted to automate sports risk assessment through multi-sensor fusion and deep learning models. However, existing solutions are essentially post-event monitoring models, with the system only assessing whether the current exercise load is excessive and whether the current posture poses a risk during training. For sports injury prevention in adolescents, the most dangerous window of opportunity is not during regular training, but rather during peak height growth. During this period, bone growth far exceeds the adaptive elongation rate of muscles and tendons, leading to a temporary decrease in neuromuscular control and a significantly higher risk of sports injury. Currently, no technology can effectively predict the arrival of this rapid growth period weeks in advance; coaches can only passively adjust training schedules after significant height increase has occurred. This inherent limitation prevents existing technologies from proactively intervening in training plans before the highest-risk window, fundamentally hindering the effectiveness of sports injury prevention in adolescents.
[0005] Therefore, we propose a method for monitoring the physical fitness and safety of teenagers in campus football training. Summary of the Invention
[0006] The purpose of this invention is to provide a method for monitoring the physical fitness and safety of teenagers in campus football training, which solves the problems mentioned in the background art.
[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring the physical fitness and safety of teenagers in campus football training, comprising the following steps:
[0008] Step S1, Personalized baseline construction based on developmental grading and structured motion fingerprint: Developmental stages are divided according to the peak height growth rate of adolescents, and theoretical safety boundaries are calculated based on anthropometric parameters; Multi-channel time-series signals are collected when players complete standard test movements, and the signals are input into an encoder to extract latent vectors; The latent vectors are divided into multiple physiologically meaningful subspaces using gradient contribution value clustering to form a structured motion fingerprint; The weighted average of the safety boundaries of a predetermined number of samples with the highest similarity in the fingerprint database is used as the personalized monitoring threshold.
[0009] Step S2, Structured Distortion Analysis and Developmental Window Prediction: Calculate the distortion rate of each subspace and the directional difference between the distortion direction vector and the drift direction vector of historical injured samples in the fingerprint database; when the overall distortion rate exceeds the first threshold, or the distortion rate of any subspace exceeds the second threshold, or the directional difference is less than the third threshold, trigger a developmental warning containing distortion location and distortion direction information.
[0010] Step S3: Extract the core features that have a biomechanical causal relationship with sports injuries, and collect auxiliary physiological features;
[0011] Step S4: Risk assessment is performed using a dual-engine architecture that combines a rule engine and a time-series network engine. The rule engine encodes personalized monitoring thresholds into judgment rules, and the time-series network engine dynamically adjusts the sensitivity of the rule engine. The dual engines output a structured risk report.
[0012] Step S5: Perform quality assessment and robustness compensation on sensor data: For data with a loss duration less than or equal to the first predetermined time, kinematic constraint interpolation compensation is used; for data with a loss duration greater than the first predetermined time, the bilateral limb coupling model is dynamically updated using the distortion rate and the lost side reconstruction value is generated using the retained side data.
[0013] Step S6, three-way coupled structured rejection feedback evolution: record the current fingerprint when the coach rejects the warning and the damage results within the predetermined window period after rejection; decompose the rejection cases into subspaces and update the positive and negative sample distributions of each subspace respectively; adjust the distortion warning threshold of each subspace based on the cumulative false positive rate and the cumulative false negative rate; at the same time, use the rejection cases to correct the temporal network output online and incorporate it into subsequent offline training; the above three updates are executed in parallel under the same rejection feedback signal.
[0014] Step S7: The algorithm model is deployed on an edge computing device to create a digital growth profile for each athlete, recording the structured motion fingerprint drift trajectory.
[0015] As a preferred embodiment of the present invention, the method for dividing the latent vector into multiple physiologically meaningful subspaces in step S1 is as follows: calculate the gradient contribution value of each latent dimension of the encoder to each input signal channel and form a gradient contribution matrix, and cluster the gradient contribution matrix. The number of clusters is four, corresponding to the four subspaces of coordination, explosive power, speed endurance and body control, respectively.
[0016] In a preferred embodiment of the present invention, step S1 further includes a fingerprint mapping alignment step after encoder version update: after each offline retraining of the encoder, anchor point samples that did not participate in this retraining are selected, and fingerprints are extracted using the encoder before and after the update to form anchor point pairs and calculate the mapping relationship. The historical fingerprints are then converted through the mapping relationship and used to participate in the distortion rate calculation.
[0017] In a preferred embodiment of the present invention, the overall distortion rate in step S2 is obtained by weighted summation of the distortion rates of each subspace, and the weight of each subspace is determined based on the player's historical injury record; if the player has no injury history, the average weight of players of the same age, gender, developmental stage and position in the fingerprint database is used as the initial value.
[0018] As a preferred embodiment of the present invention, the core features in step S3 include the peak value of the knee joint eversion angle, the ankle joint plantar flexion torque, the peak value of the ground reaction force, the symmetry of the force exertion of the left and right feet, the eccentric contraction ratio of the hamstring muscles and the quadriceps, and the rate of change of the trunk tilt angle. Each sensor is statically zero-point calibrated before training, and dynamic calibration is automatically triggered at certain intervals during training.
[0019] In a preferred embodiment of the present invention, the judgment logic of the rule engine in step S4 is as follows: when the real-time measurement value of any core feature exceeds the product of the personalized monitoring threshold and the warning coefficient, a warning is triggered, and the risk probability value output by the temporal network is used to dynamically adjust the warning coefficient; the temporal network is a three-layer causal dilated convolutional layer structure with a convolution kernel size of 3 and dilation rates of 1, 2 and 4 respectively.
[0020] In a preferred embodiment of the present invention, the joint output rule of step S4 and step S2 is as follows: when step S2 triggers a developmental warning but step S4 does not trigger a real-time warning, the terminal displays a developmental warning and provides suggestions for adjusting training intensity; when both warnings are triggered simultaneously and the suggestions conflict, real-time safety intervention takes priority, and the recovery plan after training is adjusted with reference to the developmental warning.
[0021] In a preferred embodiment of the present invention, the generative reconstruction in step S5 specifically involves: establishing coupling models for each subspace based on the left and right limb fingerprints collected during the player's historical healthy period; updating the parameters of the coupling models with the distortion rate of each subspace calculated in real time; generating a reconstruction value for the lost side when data on one side is lost, with the fingerprint of the side being retained as a condition; and the update rate of the coupling model is positively correlated with the distortion rate.
[0022] In a preferred embodiment of the present invention, the online correction in step S6 is to correct the current output of the temporal network using the distribution of positive and negative samples in the fingerprint database; the greater the deviation of the distortion rate of the subspace corresponding to the rejection case from the historical average distortion rate, the higher its weight in offline training; the triggering conditions for offline training are: the cumulative number of newly added rejection samples reaches a predetermined value, the number of days since the last training reaches a predetermined number, or the warning accuracy rate is lower than a predetermined value for a consecutive predetermined number of days; in the loss function of the temporal network, the weight of the rejection case is determined according to the degree of deviation between its corresponding subspace distortion rate and the historical average distortion rate.
[0023] In a preferred embodiment of the present invention, step S7 further includes tracking the early warning effect: recording the coach's decision, intervention measures and the actual injury status of the player within the predetermined tracking window after each early warning, and feeding the tracking results back to the threshold update in step S6.
[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0025] This invention, by tracking the distortion rate and direction of structured motion fingerprints in multiple physiological subspaces, can make effective predictions before the actual arrival of the rapid growth period, enabling coaches to proactively adjust training plans before the highest-risk window arrives, fundamentally solving the core defect of existing technologies that cannot predict the risk window in advance.
[0026] The distortion analysis of this invention is calculated separately according to dimensions with clear physiological significance, such as coordination, explosive power, speed endurance and body control. The warning output can clearly indicate which physiological dimension the distortion occurs in, the degree of distortion, and what type of injury the distortion direction has historically caused. Coaches can understand and trust the warning information output by the system.
[0027] This invention utilizes a subspace-level bilateral limb coupling model established based on the player's own historical healthy period, generates reconstructed values for the lost side based on the data of the retained side, and the coupling model is dynamically calibrated with the distortion rate, effectively avoiding false warnings caused by missing data.
[0028] This invention achieves deep closed-loop evolution driven by veto feedback by using a multi-path coupled veto feedback mechanism. This mechanism synchronizes the current fingerprint when the coach vetoes the warning with the fingerprint database update, subspace distortion threshold adjustment, and time-series network correction and retraining. This allows the system to continuously align with the coach's professional experience during use. Attached Figure Description
[0029] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:
[0030] Figure 1 This is a flowchart of a method for monitoring the physical fitness and safety of teenagers in campus football training according to the present invention. Detailed Implementation
[0031] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0032] like Figure 1 As shown, a method for monitoring the physical fitness and safety of teenagers in campus football training is implemented as follows:
[0033] The core of this invention lies in tracking changes in the movement patterns of adolescent athletes during their growth and development through structured motion fingerprinting, predicting the arrival of peak height growth rates in advance, and enabling personalized training safety monitoring and intervention. The invention will be described in detail below, covering aspects such as system deployment, baseline establishment, development prediction, real-time monitoring, data disaster recovery, human-machine collaboration, and record management.
[0034] I. System Deployment and Data Acquisition
[0035] In implementing this invention, data acquisition equipment is first deployed around the training field. A depth camera is installed 5 meters directly in front of the training field to capture players' movement postures and running trajectories. Each player is equipped with four nine-axis inertial measurement units, fixed to the lateral epicondyle of the femur and the lateral malleolus of the fibula on both sides, to collect joint angle change data. Each player is equipped with a soccer insole with an integrated array of piezoresistive sensors to collect plantar pressure distribution data. Three-dimensional force tables are embedded in key areas of the field (such as the penalty area and the midfield area) to collect ground reaction force data.
[0036] Before each training session, all sensors undergo static zero-point calibration: the player stands upright for several seconds, and the system records the readings of each sensor in this state as the zero-point reference. All subsequent measurements are differentially corrected based on this reference. During training, a dynamic calibration is automatically triggered at predetermined intervals: the player completes a standard squatting motion in a designated area, and the system automatically resets the zero-point reference after detecting that the motion is in place, in order to eliminate zero-point drift that may occur in the sensors during movement.
[0037] II. Establishment of Personalized Baselines
[0038] When using the system for the first time, five basic parameters are entered for each player: age, gender, height (average of measurements over three consecutive months), weight, and sitting height. The annualized height growth rate is calculated based on the height measurement data over the three consecutive months. When the annualized height growth rate is greater than 8cm / year, it is determined to be the peak period of height growth. Before that, it is the pre-peak period, and after that, it is the post-peak period.
[0039] Based on the five fundamental parameters mentioned above, a biomechanical regression model was pre-established on a sample of no fewer than 500 adolescents of the same age group to calculate the theoretical upper and lower limits of three key biomechanical indicators for the player under standard football movements: maximum knee eversion angle, maximum ankle plantar flexion torque, and eccentric contraction ratio of the hamstring and quadriceps muscles. This formed an individualized theoretical safety boundary. For players in their peak height growth period, a growth vulnerability coefficient was introduced to adjust and shrink the theoretical safety boundary. The growth vulnerability coefficient was selected based on the player's current annualized height growth rate, ranging from 1.1 to 1.3—the higher the growth rate, the larger the value, and the greater the contraction of the safety boundary.
[0040] Subsequently, each player completed three sets of standard test movements in turn: standing vertical jump, 5-meter shuttle run, and single-leg standing balance test. The system simultaneously collected kinematic data, joint angle data, and plantar pressure distribution data through a depth camera, inertial measurement unit, and pressure sensor, with a sampling frequency of 100Hz. The above multi-channel temporal signals were input into a pre-trained temporal variational autoencoder to extract 32-dimensional latent vectors. The pre-training of the temporal variational autoencoder used standard test data from no less than 1000 adolescent athletes as the training set, and unsupervised training was performed using the lower bound of evidence as the loss function until convergence.
[0041] Gradient contribution value clustering was used to divide the 32-dimensional latent vector into four physiological subspaces. The gradient contribution value of each latent dimension of the encoder to each input signal channel was calculated to form a gradient contribution matrix. This matrix was then clustered (four clusters). After clustering, each cluster corresponds to one of the four subspaces: coordination, explosive power, speed endurance, and body control. Each subspace has a dimension of 8. After clustering, a resampling method was used to verify the stability of the subspace partitioning. Latent dimensions whose frequencies fall below a predetermined value when assigned to a specific subspace were weighted less in subsequent distortion rate calculations.
[0042] The structured fingerprint extracted in the above process is expressed as follows: ,in , , These are subspace vectors representing coordination, explosive power, speed endurance, and body control, each with a dimension of 8. This structured fingerprint construction allows subsequent distortion analysis to be performed on dimensions with clear physiological significance, providing coaches with understandable early warning information.
[0043] The system maintains a fingerprint database, storing each player's historical fingerprints and corresponding security boundaries. The similarity between the current fingerprint and the j-th historical fingerprint in the database is calculated using cosine similarity:
[0044]
[0045] Where z is the current fingerprint vector. Let be the j-th historical fingerprint vector in the fingerprint database. The cosine similarity value ranges from -1 to 1, with values closer to 1 indicating higher similarity. The K most similar samples (K ranging from 10 to 30) are selected, and a weighted average of the similarity values is calculated based on the safety boundaries of each sample to obtain the personalized monitoring thresholds for each of the player's core features.
[0046]
[0047] in For the calibrated personalized monitoring threshold, For the first Safety boundary values for each historical sample.
[0048] When the fingerprint database update triggers the offline retraining of the temporal variational autoencoder, after the retraining is completed, anchor samples in the fingerprint database that did not participate in this retraining are selected. Fingerprints are extracted using the encoders before and after the update to form anchor pairs and the mapping relationship is calculated. The historical fingerprints are then transformed into the updated encoder space through this mapping relationship before participating in the subsequent distortion rate calculation, ensuring the comparability of fingerprints between different encoder versions.
[0049] III. Developmental Window Prediction
[0050] The system stores the structured fingerprint of each player's first test as a baseline. Thereafter, the player is retested at regular intervals (usually weekly), and the current structured fingerprint is extracted.
[0051] Calculate the distortion rate of each of the four subspaces between the current time step and four weeks ago. The distortion rate of the m-th subspace is... Calculate as follows:
[0052]
[0053] in This is the fingerprint of the m-th subspace at the current time. This is the fingerprint of the m-th subspace four weeks ago. The first sample of the same age, sex, and developmental stage in the fingerprint database The standard deviation of the fingerprint in each subspace. The symbol ||·|| represents the Euclidean distance between vectors. This calculation uses a four-week time window to transform the original variation of the fingerprint into a standardized distortion rate based on the population standard deviation, making the distortion rates comparable between different players and different subspaces.
[0054] The weights w_m of each subspace are determined based on the gradient contribution of the injury types that have occurred in the player's historical training records. Specifically, for each soft tissue injury that occurred in the player's historical training records, the structured fingerprint of the most recent test before the injury was extracted. The gradient contribution of the distortion rate of each subspace relative to the injury type was calculated, and the standardized gradient contribution of each injury type was used as the weight of each subspace. The value of each subspace weight w_m ranges from 0.15 to 0.40, and the sum of the weights of the four subspaces is 1. If the player has no injury history, the average subspace weights of players of the same age, gender, developmental stage, and playing position in the fingerprint database are used as the initial value.
[0055] Weighted composite distortion rate It is obtained by weighted summation of the distortion rates of each subspace:
[0056]
[0057] Simultaneously calculate the distortion direction vector :
[0058]
[0059] in For full-space fingerprint recognition at the current moment, This is a full-space fingerprint from four weeks ago. Calculate the angle between this distortion direction vector and the drift direction vectors of all historical injury samples in the fingerprint database. :
[0060]
[0061] in For the fingerprint database The drift direction vectors of the historical injury samples are used, with the symbol "·" representing the dot product operation. The minimum angle θ between the current distortion direction vector and the current historical injury sample is taken. min =min j (θ j .
[0062] Three-dimensional joint determination mechanism: when δ weighted Exceeding the first threshold λ1, or any subspace δ sub If m exceeds the corresponding second threshold λ2,m, or θ min When the value is less than the third threshold λ3, a developmental warning is triggered. λ1 is taken as the 85th percentile of the overall distortion rate of samples of the same age, sex, and developmental stage in the fingerprint database. λ2,m is taken as the 90th percentile of the distortion rate of each subspace. The value of λ3 ranges from 15° to 30°, and the specific value is determined by the 5th percentile of the angle distribution of the direction vectors of all historical injury samples in the fingerprint database.
[0063] The developmental early warning output is a structured, anatomical report, including the trigger condition type (overall distortion exceeding the standard, subspace distortion exceeding the standard, or direction matching historical injury samples), the specific distorted subspace name and distortion degree, the historical injury type matched by the distortion direction, and suggested intervention measures. For example, when the coordination subspace distortion rate exceeds the standard and the distortion direction matches the trajectory of a previous hamstring strain sample, the system will indicate "The coordination subspace distortion rate reaches 1.82 standard deviations, and the direction matches the trajectory of a previous hamstring strain sample. It is predicted that the patient will enter a period of rapid height growth in the next 4 to 5 weeks, and the risk of hamstring strain is increased," and provide specific training adjustment suggestions.
[0064] IV. Real-time Feature Extraction and Dual-Engine Risk Assessment
[0065] During training, the system continuously collects six core features at a sampling frequency of 100Hz: peak knee valgus angle (calculated in real time by the Euler angle difference of the inertial measurement unit at the lateral epicondyle of the femur and the lateral malleolus of the fibula), ankle plantar flexion torque, peak ground reaction force (collected by a three-dimensional force measuring platform embedded in the field), symmetry of force exertion between the left and right feet (calculated by comparing data from the left and right feet using insole pressure sensors), eccentric contraction ratio of the hamstring muscles and quadriceps, and rate of change of trunk tilt angle.
[0066] Simultaneously collect three auxiliary physiological characteristics: heart rate variability, running distance, and speed distribution. These auxiliary physiological characteristics do not participate in the primary decision-making regarding safety risks and are only used for macro-management of training load.
[0067] The dual-engine risk assessment architecture operates in parallel:
[0068] The rule engine encodes personalized monitoring thresholds into a set of interpretable decision rules. The decision logic is as follows: when the real-time measurement value Mi(t) of any core feature exceeds a certain threshold... Exceeding the corresponding personalized monitoring threshold T i With the early warning coefficient α i The warning is triggered when the product of α and β is reached. The warning coefficient α is... i The value ranges from 1.1 to 1.5. Different warning coefficients can be set for the core features corresponding to different types of injuries. For features related to anterior cruciate ligament injuries (such as knee valgus angle), a lower value is used to improve sensitivity, while a higher value is used for other types of injuries. Each trigger rule is accompanied by a preset biomechanical explanation text, enabling coaches to understand the physiological basis of the warning.
[0069] The temporal network engine employs a three-layer causal dilated convolutional layer structure (kernel size 3, dilation rates 1, 2, and 4 respectively). The input is the core feature temporal data from the past 30 seconds, and the output is the risk probability value P. AI (t The time-series network does not directly output early warning decisions; instead, its output risk probability values are used to dynamically adjust the early warning coefficients of the rule engine.
[0070] α i (t)=α i0 ×(1-β×P AI (t))
[0071] Where α i0 Let P be the baseline warning coefficient corresponding to the i-th core feature, β be the sensitivity adjustment coefficient (ranging from 0 to 0.3), and P be the baseline warning coefficient. AI (t This represents the risk probability value output by the time-series network at the current moment. A higher risk probability value results in a lower warning coefficient and greater sensitivity of the system to player fatigue. The training data for the time-series network is independently assessed by at least two sports medicine physicians with intermediate or higher professional titles; in case of discrepancies, a third attending physician will make the final decision.
[0072] The combined output of the rules engine and the time-series network engine is a structured risk report, which includes the risk type, risk level, biomechanical evidence description, and recommended intervention measures.
[0073] V. Combined Output of Developmental Early Warning and Real-time Early Warning
[0074] When the developmental window prediction triggers a developmental warning but the dual-engine risk assessment does not trigger a real-time warning, the coach terminal displays the developmental warning content along with quantitative adjustment suggestions for the current training intensity.
[0075] When developmental window prediction and dual-engine risk assessment trigger warnings simultaneously and intervention recommendations conflict, the system prioritizes real-time safety interventions. After training, adjustments are made to the recovery plan based on the developmental warnings.
[0076] VI. Data Quality Assessment and Robustness Compensation
[0077] Calculate the Signal Quality Index (SQI) for each sensor data packet:
[0078] SQI=λ signal ×S signal +λ noise ×(1-S noise )+λ artifact ×(1-S artifact )
[0079] Where S signal S represents the normalized signal strength value (between 0 and 1, with larger values indicating stronger signals). noise S represents the normalized noise level value (between 0 and 1, with higher values indicating higher noise). artifact λ is the normalized motion artifact severity value (between 0 and 1, with larger values indicating more severe artifacts). signal , λ noise , λ artifact These are the corresponding weighting coefficients, and the sum of the three is 1. λ signal , λ noise , λ artifact The specific value is determined by normalizing the signal-to-noise ratio calibrated at the factory for each sensor. When the signal quality index is lower than the preset threshold, the corresponding data frame is marked as unusable.
[0080] When the duration of continuous unavailable data is less than or equal to 3 seconds, kinematic constraint interpolation is used for compensation: joint angle changes are maintained by cubic spline interpolation to preserve second-order continuity, and gait phase is maintained by Fourier series fitting to preserve periodicity.
[0081] When the duration of continuous unavailable data exceeds 3 seconds, generative reconstruction driven by the dynamic coupling model is initiated. Coupled models are established for each subspace of the left and right limb fingerprints collected during the player's historical healthy period. The coupled model for the m-th subspace is expressed as a joint Gaussian distribution p(z left ,m,z right ,m)=N(μ m ,Σ m The distortion rate δ of each subspace, calculated in real time during the developmental window prediction, will be used. sub m is used as prior information to dynamically update the covariance matrix of the coupled model:
[0082] Σ m(t)=(1-γ(t))×Σ m (t-1)+γ(t)×(z left ,m-μ left ,m)(z right ,m-μ right ,m)ᵀ
[0083] γ(t) = γ0 × (1 + δ sub ,m)
[0084] Where γ0 is the baseline learning rate (ranging from 0.05 to 0.15), δ sub ,m is the distortion rate of the m-th subspace at the current time, z left m and z right ,m represents the m-th subspace fingerprint of the left and right limbs at the current moment, respectively, and μ left ,m and μ right , m, represent the mean values of the m-th subspace fingerprints of the left and right limbs during the historical healthy period, respectively. A higher distortion rate results in a higher learning rate and a faster convergence of the coupled model to the current developmental state. When sensor data for one limb is lost, the fingerprint z of each subspace on the remaining limb is preserved. left m is a condition, and the posterior distribution p(z) of the fingerprints of each subspace on the lost side is generated by a conditional variational autoencoder. right ,m|z left ,m The expected value of the posterior distribution is taken as the reconstructed value.
[0085] If data from both sensors is lost simultaneously, the system switches to rule-only engine mode, marks the confidence level as low, and sends a message to the coach terminal saying "Insufficient data, manual observation recommended".
[0086] VII. Three-way Coupled Negation Feedback Evolution
[0087] When a coach vetoes a system warning, the system records the current structured fingerprint z at the time of the veto. reject and the time for veto t reject And track the player’s actual injury outcome during a 60-day tracking window.
[0088] First-line update – fingerprint database update: Decompose the rejected case into four subsamples z according to four subspaces. sub ,m reject The label of each subsample m The label is determined based on whether the corresponding type of injury occurs during the follow-up period: if the corresponding physiological dimension of the subspace experiences the corresponding type of injury within a 60-day window after rejection, then the label is... m =1 (positive sample), if not occurred, then label m=0 (negative sample). Each subsample is added to the corresponding partition of each subspace in the fingerprint database.
[0089] The second update—subspace threshold update—adjusts the distortion warning sub-threshold λ2,m for each subspace based on the cumulative distribution function of positive and negative samples in each subspace. Let FP... m Let be the cumulative false alarm rate (the proportion of negative samples that trigger an alert) in the m-th subspace. When FP m When it exceeds 50%, increase λ2,m by 10%; let FN m Let FN be the cumulative false negative rate (the proportion of positive samples that do not trigger an alert). m When the threshold exceeds 30%, λ2,m is reduced by 10%. The adjustments to the thresholds of each subspace are independent of each other.
[0090] Third-way update – online correction and offline weight configuration: Utilizing the distribution of positive and negative samples in each subspace of the fingerprint database, adjust the risk probability value P of the current output of the temporal network. AI (t Perform Bayesian posterior correction:
[0091] P corrected =P AI (t)×P(label m |z sub ,m reject ) / P(label m )
[0092] Where P(label) m |z sub ,m reject The posterior probability P(label) is estimated from the positive and negative sample distributions of this subspace in the fingerprint database. m This represents the prior probability (the overall incidence of this type of damage in the database). This modification allows the system to benefit instantly from rejection feedback between two offline training sessions.
[0093] Simultaneously, rejected cases are incorporated into the differentiated weight configuration for subsequent offline training of the temporal network. The weight η of this case in the loss function is determined as follows:
[0094] η = η0 × (1 + ρ × |δ sub ,m-δ sub ,m avg | / δ sub ,m avg )
[0095] Where η0 is the baseline weight (with a value of 1.0), ρ is the correction coefficient (with a value of 0.5 to 1.5), and δ sub,m is the distortion rate of the subspace corresponding to the veto case, δ sub ,m avg This represents the historical average distortion rate of this player's subspace. The greater the deviation of the distortion rate of the subspace corresponding to a rejection case from the player's historical average distortion rate, the higher the weight of that case in the offline training loss function. The offline training trigger condition for the temporal network is one of the following three conditions: the cumulative number of newly added rejection feedback samples in the fingerprint database reaches 30, 7 days have passed since the last offline training, or the system's warning accuracy rate is below 85% for 3 consecutive days.
[0096] The three updates mentioned above are executed in parallel under the same veto feedback signal, forming a deep closed-loop evolution mechanism of "one veto, three updates".
[0097] VIII. Edge Deployment and Record Management
[0098] All algorithm models are deployed on edge computing devices (tablets or embedded computing devices) near the training field, with the total number of parameters kept below 500,000. Data processing is completed locally, without involving the network transmission of athletes' physiological data, effectively protecting personal privacy.
[0099] A digital growth profile is created for each athlete, recording the drift trajectory of each subspace of their structured athletic fingerprint since joining the team, the results of each risk assessment, early warning records, and coach intervention records. Coaches can access the growth profile of any player at any time to understand their development process and risk trends.
[0100] The system records the coach's decision (adoption or rejection) after each warning is issued, the intervention measures taken after adoption, and the actual injury occurrence of the player within the 30-day follow-up window after intervention. The follow-up verification results are fed back to the threshold update mechanism. If a certain type of warning is proven false in multiple follow-up verifications, the system automatically lowers its priority; if a certain type of warning is proven effective in multiple follow-up verifications, the system continues to maintain or increase its sensitivity, achieving closed-loop verification and continuous optimization.
[0101] IX. Differentiated Implementation at Different Developmental Stages
[0102] For players in their pre-peak period (annualized height growth rate less than 8cm / year), the system mainly performs baseline establishment, feature extraction, and real-time risk assessment. The developmental window prediction is monitored at a low frequency (distortion rate is calculated every two weeks) to accumulate fingerprint data from the early developmental stage as a benchmark for subsequent comparisons.
[0103] For players in their peak growth phase (annualized height growth rate greater than 8cm / year), the system automatically increases the monitoring frequency of the developmental window prediction to once a week and incorporates the growth vulnerability coefficient into the calculation of the theoretical safety boundary, thus shrinking the overall monitoring threshold for players in this stage. Simultaneously, the system proactively alerts the coach on the coach's terminal that the player has entered their peak height growth period and suggests increasing the proportion of coordination training and controlling the total amount of high-intensity sprint training.
[0104] For players in their post-peak period (where their annual height growth rate drops to below 8cm / year), the system gradually restores the normal monitoring frequency and incorporates the fingerprint drift trajectory accumulated throughout the peak period into the player's digital growth profile as reference data for subsequent talent selection and long-term tracking.
[0105] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or basic characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.
[0106] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A method for monitoring the physical fitness and safety of teenagers in campus football training, characterized in that: The methods and steps include the following: Step S1, Personalized baseline construction based on developmental grading and structured motion fingerprint: Developmental stages are divided according to the peak growth rate of adolescent height, and theoretical safety boundaries are calculated based on anthropometric parameters; Multi-channel temporal signals are collected when players complete standard test movements, and the signals are input into an encoder to extract latent vectors; The latent vectors are divided into multiple physiologically meaningful subspaces using gradient contribution value clustering to form a structured motion fingerprint. The weighted average of the security boundaries of a predetermined number of samples with the highest similarity in the fingerprint database is used as the personalized monitoring threshold; Step S2, Structured Distortion Analysis and Developmental Window Prediction: Calculate the distortion rate of each subspace and the directional difference between the distortion direction vector and the drift direction vector of historical injured samples in the fingerprint database; when the overall distortion rate exceeds the first threshold, or the distortion rate of any subspace exceeds the second threshold, or the directional difference is less than the third threshold, trigger a developmental warning containing distortion location and distortion direction information. Step S3: Extract the core features that have a biomechanical causal relationship with sports injuries, and collect auxiliary physiological features; Step S4: Risk assessment is performed using a dual-engine architecture that combines a rule engine and a time-series network engine. The rule engine encodes personalized monitoring thresholds into judgment rules, and the time-series network engine dynamically adjusts the sensitivity of the rule engine. The dual engines output a structured risk report. Step S5: Perform quality assessment and robustness compensation on sensor data: For data with a loss duration less than or equal to the first predetermined time, kinematic constraint interpolation compensation is used; for data with a loss duration greater than the first predetermined time, the bilateral limb coupling model is dynamically updated using the distortion rate and the lost side reconstruction value is generated using the retained side data. Step S6, three-way coupled structured rejection feedback evolution: record the current fingerprint when the coach rejects the warning and the damage results within the predetermined window period after rejection; decompose the rejection cases into subspaces and update the positive and negative sample distributions of each subspace respectively; adjust the distortion warning threshold of each subspace based on the cumulative false positive rate and the cumulative false negative rate; at the same time, use the rejection cases to correct the temporal network output online and incorporate it into subsequent offline training; the above three updates are executed in parallel under the same rejection feedback signal. Step S7: The algorithm model is deployed on an edge computing device to create a digital growth profile for each athlete, recording the structured motion fingerprint drift trajectory.
2. The method for monitoring the physical fitness and safety of teenagers in campus football training according to claim 1, characterized in that: The method for dividing the latent vector into multiple physiologically meaningful subspaces in step S1 is as follows: calculate the gradient contribution value of each latent dimension of the encoder to each input signal channel and form a gradient contribution matrix. Then, cluster the gradient contribution matrix into four subspaces, corresponding to coordination, explosive power, speed endurance, and body control.
3. The method for monitoring the physical fitness and safety of teenagers in campus football training according to claim 2, characterized in that: Step S1 also includes a fingerprint mapping alignment step after encoder version update: after each offline retraining of the encoder, anchor point samples that did not participate in this retraining are selected, and fingerprints are extracted using the encoder before and after the update to form anchor point pairs and calculate the mapping relationship. The historical fingerprints are then transformed through the mapping relationship before participating in the distortion rate calculation.
4. The method for monitoring the physical fitness and safety of teenagers in campus football training according to claim 1, characterized in that: In step S2, the overall distortion rate is obtained by weighted summation of the distortion rates of each subspace, and the weight of each subspace is determined based on the player's historical injury record; if the player has no injury history, the average weight of players of the same age, gender, developmental stage and position in the fingerprint database is used as the initial value.
5. The method for monitoring the physical fitness and safety of teenagers in campus football training according to claim 1, characterized in that: The core features in step S3 include the peak knee eversion angle, ankle plantar flexion torque, peak ground reaction force, symmetry of force exertion on the left and right feet, eccentric contraction ratio of the hamstring muscles and quadriceps, and rate of change of trunk tilt angle. Each sensor is statically zero-point calibrated before training, and dynamic calibration is automatically triggered at certain intervals during training.
6. The method for monitoring the physical fitness and safety of teenagers in campus football training according to claim 1, characterized in that: The judgment logic of the rule engine in step S4 is as follows: when the real-time measurement value of any core feature exceeds the product of the personalized monitoring threshold and the warning coefficient, a warning is triggered. The risk probability value output by the temporal network is used to dynamically adjust the warning coefficient. The temporal network is a three-layer causal dilated convolutional layer structure with a kernel size of 3 and dilation rates of 1, 2, and 4.
7. The method for monitoring the physical fitness and safety of teenagers in campus football training according to claim 1, characterized in that: The joint output rule for step S4 and step S2 is as follows: when step S2 triggers a developmental warning but step S4 does not trigger a real-time warning, the terminal displays a developmental warning and provides suggestions for adjusting training intensity; when both warnings are triggered simultaneously and the suggestions conflict, real-time safety intervention takes priority, and the recovery plan after training is adjusted with reference to the developmental warning.
8. The method for monitoring the physical fitness and safety of teenagers in campus football training according to claim 1, characterized in that: The generative reconstruction in step S5 specifically involves: establishing coupling models for each subspace based on the left and right limb fingerprints collected during the player's historical healthy period; updating the parameters of the coupling models with the distortion rate of each subspace calculated in real time; generating a reconstruction value for the lost side when data on one side is lost, based on the condition of retaining the fingerprint on the side; and the update rate of the coupling model is positively correlated with the distortion rate.
9. A method for monitoring the physical fitness and safety of teenagers in campus football training according to claim 1, characterized in that: In step S6, online correction involves using the distribution of positive and negative samples in the fingerprint database to correct the current output of the temporal network. The greater the deviation of the distortion rate of the subspace corresponding to the rejection case from the historical average distortion rate, the higher its weight in offline training. The triggering conditions for offline training are: the cumulative number of newly added rejection samples reaches a predetermined value, the number of days since the last training reaches a predetermined number, or the warning accuracy rate is lower than a predetermined value for a consecutive predetermined number of days. In the loss function of the temporal network, the weight of the rejection case is determined according to the degree of deviation between its corresponding subspace distortion rate and the historical average distortion rate.
10. A method for monitoring the physical fitness and safety of teenagers in campus football training according to claim 1, characterized in that: Step S7 also includes tracking the early warning effect: recording the coach's decisions, intervention measures, and the actual injury status of the player within the predetermined tracking window after each early warning, and feeding the tracking results back to the threshold update in step S6.