Athlete health assessment method based on big data

Abstract

The invention relates to the technical field of sports science and man-machine interaction, and discloses a big data-based athlete health assessment method, which comprises the following steps of: collecting and preprocessing multi-source heterogeneous data; generating a structured data set, and performing feature extraction and dimension reduction to obtain a time sequence feature subset and a key feature subset; constructing a multi-modal fusion health assessment model, and outputting a comprehensive health index; performing grading and dynamic early warning on the health state; carrying out self-updating and personalized calibration on the health assessment model; through multi-source data fusion, athlete physiological index data, exercise load data, environment parameter data and recovery behavior data are integrated, and the limitation of single index evaluation is broken through; meanwhile, through dynamic time sequence modeling and personalized calibration, the lagging effect of the health state is captured through LSTM, the accuracy of risk assessment is improved, and based on individual baseline data standardization features, assessment deviation caused by physiological differences of different athletes is solved.

Description

Technical Field

[0001] This invention relates to the fields of sports science and human-computer interaction technology, and more specifically to a method for athlete health assessment based on big data. Background Technology

[0002] As competitive sports develop towards higher loads and greater specialization, accurate assessment of athletes' health status has become a core issue in scientific training and sports injury prevention. Currently, athlete health assessment mainly relies on two methods: one is the single-index threshold method, which sets a fixed threshold using a single physiological indicator such as resting heart rate or an athletic performance indicator such as maximum oxygen uptake, and marks an indicator as abnormal when it exceeds the threshold; the other is the statistical model method, which establishes a linear regression or logistic regression model based on historical data and calculates a health score by weighting multiple indicators.

[0003] However, among the two methods mentioned above, the single-index threshold method does not consider the correlation between indicators and individual differences, which can easily lead to misjudgment; the statistical model method assumes that the data are linearly correlated, but it cannot capture the nonlinear relationship between exercise load and health status over time, and it does not fully explore the impact of environmental factors on health. Therefore, the existing technology for assessing athletes' health has the shortcomings of single data dimension, poor dynamic adaptability and insufficient personalization. In view of this, this invention proposes a big data-based athlete health assessment method to meet the needs of modern competitive sports for accurate assessment of athletes' health status. Summary of the Invention

[0004] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a big data-based athlete health assessment method to solve the problems existing in the background art.

[0005] This invention provides the following technical solution: a method for athlete health assessment based on big data, comprising the following steps:

[0006] Step 1: Multi-source heterogeneous data acquisition and preprocessing; Collect athlete physiological index data, exercise load data, environmental parameter data, and recovery behavior data at each time step, and perform preprocessing operations;

[0007] Step 2: Generate a structured dataset from the preprocessed data in Step 1, and perform feature extraction and dimensionality reduction to obtain a time-series feature subset and a key feature subset;

[0008] Step 3: Construct a multimodal fusion health assessment model and output a comprehensive health index; adopt a hybrid model architecture of time series modeling and feature classification, including time series feature modeling and structured feature modeling;

[0009] Step 4: Classify and dynamically warn about health status; Divide health status into five levels based on the comprehensive health index, and calculate the rate of change of the index through a sliding window to achieve dynamic early warning;

[0010] Step 5: Perform self-updating and personalized calibration of the health assessment model.

[0011] Preferably, the physiological indicators include resting heart rate, sleep heart rate, blood lactate concentration, serum creatine kinase, and sleep quality; the exercise load data includes training duration and number of repetitions; the environmental parameter data includes the athlete's ambient temperature, relative humidity, altitude, and air pressure; the recovery behavior data includes calorie intake, hydration, and massage / relaxation time; the athlete's physiological indicators are obtained through wearable devices, the exercise load data is obtained through motion sensors, the environmental parameter data is obtained through environmental monitoring instruments, and the recovery behavior data is obtained through questionnaires conducted on the athletes; the preprocessing operations include denoising, missing value imputation, and time alignment of the data.

[0012] Preferably, the feature extraction and dimensionality reduction includes temporal feature extraction, individual baseline calibration, and feature selection; the temporal feature modeling uses a long short-term memory network to extract long-term dependencies of temporal features and outputs hidden states; the structured feature modeling uses... The model selects a subset of key features. A nonlinear transformation is performed to output a feature importance weighted score; the output hidden state is concatenated with the output feature importance weighted score, and a comprehensive health index is output through the output layer.

[0013] Preferably, the structured dataset generated in step 2 is represented as follows:

[0014] ;in, Represents a structured dataset. Indicates the first The athlete in the Multidimensional feature vectors at each time step Indicates the total time steps;

[0015] The temporal feature extraction specifically involves: segmenting the time series using a sliding window, obtaining the statistical properties and temporal correlation of features within the window, and generating a temporal feature matrix. , ;in, Represents the feature dimension; the window length of the sliding window is The sliding window step size is ;

[0016] The individual baseline calibration specifically involves: obtaining the health baseline of each athlete's historical data, and then applying it to the current characteristics. After standardization, a subset of temporal features for each athlete is obtained, represented as: ;

[0017] The feature selection specifically involves: filtering features that significantly impact health status through mutual information, and retaining a subset of key features. , represented as .

[0018] Preferably, the comprehensive health index in step 3 is expressed as:

[0019] ;in, Indicates the first The athlete in the A comprehensive health index at each time step; This represents the weight matrix of the fully connected layer; Represents the bias vector; Indicates the activation function; This represents a vector concatenation operation; express The output of the layer, express The output of the layer.

[0020] Preferably, the calculation of the exponential rate of change using a sliding window is expressed by the following formula: ;in, Indicates the rate of change of the index. Indicates the first The athlete in the A comprehensive health index at each time step. Indicates the first The athlete in the A comprehensive health index at each time step;

[0021] The health status is divided into five levels: excellent, good, moderate, poor, and critical; when the comprehensive health index meets the requirements... At that time, the health level was excellent. satisfy If the overall health index meets the requirements, an early warning will be issued; when the overall health index meets the requirements... At that time, the health level was good. satisfy If the overall health index meets the requirements, an early warning will be issued; when the overall health index meets the requirements... At that time, the health level was medium. satisfy or If the overall health index meets the requirements, an early warning will be issued; when the overall health index meets the requirements... At that time, the health level was poor, requiring intervention. satisfy or If the overall health index meets the requirements, an early warning will be issued; when the overall health index meets the requirements... If the health level is poor, training should be stopped immediately, an emergency warning should be triggered, and the team doctor should be notified to conduct an examination.

[0022] Preferably, the health assessment model mainly includes an input layer, layer, The system comprises an input layer, a fusion layer, and an output layer; the input layer is used to input a subset of temporal features and a subset of key features; the output layer is used to output a comprehensive health index.

[0023] The The layer includes an input gate, a forget gate, candidate memory units, memory unit updates, hidden state output, and an input gate;

[0024] The The layer takes a subset of key features as input, uses athlete health status labels as supervision signals, and employs mean squared error as the loss function. It iteratively trains the decision tree using a gradient boosting algorithm, initializes prediction values, calculates residuals for each iteration, and trains the 1st generation of the decision tree. The residuals of each tree are fitted, the predicted values ​​are updated, and the final output is the weighted sum of all trees;

[0025] The fusion layer is used to... Layers and The outputs of the layers are concatenated into a joint feature vector, which is then mapped to a comprehensive health index.

[0026] Preferably, the weight matrix of the forget gate is: The weight matrix of the input gate is The weight matrix of the output gate is ;

[0027] The input gate is represented as: ;in, Indicates the output of the input gate. Indicates the first The hidden state at each time step Indicates the first Input at each time step, This indicates the bias of the input gate. express Activation function;

[0028] The forget gate is represented as follows: ;in, This represents the output of the forget gate. Indicates the bias of the forget gate;

[0029] The candidate memory unit is represented as follows: ;in, Indicates the output of the candidate memory unit. Indicates the weight of the candidate memory unit. Indicates the bias of candidate memory units. Represents the hyperbolic tangent activation function;

[0030] The memory unit update is represented as follows: ;in, This indicates the updated value of the memory cell. This represents the updated value of the memory cell in the previous time step. Represents element-wise product;

[0031] The hidden state output is represented as follows: ;in, Indicates the first The hidden state at each time step Indicates the output of the output gate;

[0032] The output gate is represented as: ;in, This indicates the bias of the output gate.

[0033] The technical effects and advantages of this invention are as follows:

[0034] This invention, through steps 2 and 3, facilitates the integration of athlete physiological indicators, exercise load, environmental parameters, and recovery behavior data through multi-source data fusion, overcoming the limitations of single-indicator assessment. Simultaneously, through dynamic temporal modeling and personalized calibration, it captures the lag effect of health status using LSTM, improving the accuracy of risk assessment. Furthermore, based on the standardized characteristics of individual baseline data, it addresses assessment biases caused by physiological differences among athletes. A dynamic early warning mechanism, combining index values ​​and rate of change as dual indicators, upgrades the process from status diagnosis to trend prediction, providing real-time guidance for scientific training. Attached Figure Description

[0035] Figure 1 This is a flowchart of the athlete health assessment method based on big data according to the present invention. Detailed Implementation

[0036] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The athlete health assessment method based on big data involved in the present invention is not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] like Figure 1 As shown, this invention provides a method for athlete health assessment based on big data, including the following steps:

[0038] Step 1: Multi-source heterogeneous data acquisition and preprocessing; Collect athlete physiological index data, exercise load data, environmental parameter data, and recovery behavior data at each time step, and perform preprocessing operations; the physiological index data includes, but is not limited to, resting heart rate, sleep heart rate, blood lactate concentration, serum creatine kinase, and sleep quality; the exercise load data includes, but is not limited to, training duration and number of repetitions; the environmental parameter data includes, but is not limited to, the athlete's ambient temperature, relative humidity, altitude, and air pressure; the recovery behavior data includes, but is not limited to, calorie intake, hydration, and massage / relaxation time; the athlete's physiological index data can be obtained through wearable devices, the exercise load data can be obtained through motion sensors, the environmental parameter data can be obtained through environmental monitoring instruments, and the recovery behavior data can be obtained through questionnaires; the preprocessing operations include, but are not limited to, denoising, missing value imputation, and time alignment; denoising includes wavelet transform to remove motion artifacts, missing value imputation includes moving average interpolation between adjacent time steps, and time alignment includes standardizing the sampling frequency.

[0039] Step 2: Generate a structured dataset from the preprocessed data in Step 1, and perform feature extraction and dimensionality reduction to obtain a temporal feature subset and a key feature subset; the feature extraction and dimensionality reduction includes temporal feature extraction, individual baseline calibration, and feature selection; the purpose is to capture the temporal dependence and individual differences of the data through feature extraction and dimensionality reduction.

[0040] Step 3: Construct a multimodal fusion health assessment model and output a comprehensive health index; adopt a hybrid model architecture of temporal modeling and feature classification, including temporal feature modeling and structured feature modeling; the temporal feature modeling uses a long short-term memory network to extract long-term dependencies of temporal features and outputs hidden states; the structured feature modeling uses... The model selects a subset of key features. A nonlinear transformation is performed to output a feature importance weighted score; the output hidden state is concatenated with the output feature importance weighted score, and a comprehensive health index is output through the output layer.

[0041] Step 4: Classify and dynamically warn about health status; Divide health status into five levels based on the comprehensive health index, and calculate the rate of change of the index through a sliding window to achieve dynamic early warning;

[0042] Step 5: Perform self-updating and personalized calibration of the health assessment model; regularly collect athletes' medical examination results, use them as labels, update model parameters through transfer learning, and adjust the memory gating parameters of LSTM and the learning rate of XGBoost according to individual data distribution changes to achieve personalized calibration and adaptation of the model.

[0043] In this embodiment, it should be specifically noted that step 1 sorts the various time steps and marks them as follows: ;

[0044] The structured dataset generated in step 2 is represented as follows:

[0045] ;in, Represents a structured dataset. Indicates the first The athlete in the Multidimensional feature vectors at each time step Indicates the total time steps;

[0046] The temporal feature extraction specifically involves: segmenting the time series using a sliding window, obtaining the statistical properties and temporal correlation of features within the window, and generating a temporal feature matrix. , ;in, Represents the feature dimension; the window length of the sliding window is Select The sliding window step size is 1 hour. Select Hour; the statistics of the feature can be the feature mean, variance, maximum value, and minimum value, etc.

[0047] The individual baseline calibration specifically involves: obtaining the health baseline of each athlete's historical data, and then applying it to the current characteristics. Standardization is performed to eliminate bias caused by individual physiological differences, resulting in a subset of temporal features for each athlete, represented as: The standardized formula is expressed as follows: ;in, express Standardized representation, Indicates the first The athlete in the The current features at each time step This represents the feature mean of the current feature. This represents the feature standard deviation of the current feature;

[0048] The feature selection specifically involves: filtering features that significantly impact health status through mutual information or recursive feature elimination, such as HRV sample entropy, blood lactate clearance rate, and... Peak frequency, etc., retain key feature subsets , represented as ;in, .

[0049] In this embodiment, it should be specifically noted that the comprehensive health index in step 3 is expressed as:

[0050] ;in, Indicates the first The athlete in the A comprehensive health index at each time step; This represents the weight matrix of the fully connected layer; Represents the bias vector; This represents the activation function, which maps the output to... interval, This indicates extreme unhealthiness. Indicates excellent health; This represents a vector concatenation operation; express The output of the layer, express The output of the layer.

[0051] In this embodiment, it should be specifically noted that the calculation of the exponential rate of change through the sliding window is expressed by the following formula: ;in, Indicates the rate of change of the index. Indicates the first The athlete in the A comprehensive health index at each time step. Indicates the first The athlete in the A comprehensive health index at each time step;

[0052] The health status is divided into five levels: excellent, good, moderate, poor, and critical; when the comprehensive health index meets the requirements... At that time, a health level of "excellent" indicates that all indicators are stable and there is no risk. satisfy If the overall health index declines too rapidly, an early warning will be issued; when the overall health index meets the following conditions... At that time, the health level was "good," indicating slight fluctuations in various indicators, which required monitoring. satisfy If the overall health index meets the requirements, an early warning will be issued; when the overall health index meets the requirements... At that time, the health level was medium, indicating that there were potential risks in various indicators. satisfy or If the overall health index meets the requirements, an early warning will be issued; when the overall health index meets the requirements... At that time, a poor health level indicates significant abnormalities in various indicators, requiring intervention. satisfy or If the overall health index meets the requirements, an early warning will be issued; when the overall health index meets the requirements... If the health level is poor, it indicates that there are serious abnormalities in various indicators. Training should be stopped immediately, an emergency warning should be triggered, and the team doctor should be notified to conduct an examination.

[0053] In this embodiment, it should be specifically noted that in the individual baseline calibration, the baseline time window selects data from the athlete's stable training period over the past 30 days, including characteristic phases such as the pre-competition sprint phase and the post-competition recovery phase, to ensure that the baseline can reflect the athlete's normal health status; the mean and standard deviation of each feature within the baseline window are obtained, and the feature of the current time window is converted into a standardized value relative to the baseline; for example, if the current feature is resting heart rate, the standardization formula is expressed as: ;in, express Standardized representation, Indicates the first The athlete in the The current features at each time step Indicates the first Mean resting heart rate of each athlete Indicates the first The standard deviation of resting heart rate of an athlete; through standardization, the characteristics of different dimensions and individual differences are unified into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0054] In this embodiment, it should be specifically noted that the purpose of the feature selection is to screen key health influencing factors and time-series feature matrix. It contains a large number of redundant or irrelevant features, such as ambient temperature. The hourly fluctuations in ambient temperature have little impact on health status. Therefore, feature selection is needed to reduce dimensionality while retaining the features that are most predictive of health status.

[0055] The feature selection, taking the screening of features that significantly affect health status through mutual information as an example, specifically involves:

[0056] Mutual information is used to measure the nonlinear correlation between features and a target variable, where the target variable is a health status label, such as healthy, sub-healthy, or injured; the formula is expressed as:

[0057] ;in, Representation of features With target variable mutual information value, Representation of features Information entropy Represent the target variable Information entropy Represents joint entropy; the larger the mutual information value, the stronger the feature. For target variable The stronger the predictive ability;

[0058] Obtain the mutual information values ​​of all time-series features and health status labels, and retain features whose mutual information values ​​meet a threshold as key features, forming a subset of key features; the threshold can be set to... In this embodiment, we select Those skilled in the art can also modify the threshold value according to the actual situation. If the mutual information value is greater than If it is true, then keep it; otherwise, don't keep it.

[0059] In this embodiment, it should be specifically noted that the health assessment model mainly includes an input layer, layer, The system comprises an input layer, a fusion layer, and an output layer; the input layer is used to input a subset of temporal features and a subset of key features; the output layer is used to output a comprehensive health index.

[0060] The The layer will have a length of The temporal feature subset as The input sequence, set The hidden layer has 64 neurons. The biases of the forget gate, input gate, and output gate are initialized to 0. The weight matrix of the forget gate is... The weight matrix of the input gate is The weight matrix of the output gate is Randomly initialize the weight matrix;

[0061] The The layer includes an input gate, a forget gate, candidate memory units, memory unit updates, hidden state output, and an input gate;

[0062] The input gate is represented as: ;in, Indicates the output of the input gate. Indicates the first The hidden state at each time step Indicates the first Input at each time step, This indicates the bias of the input gate. express Activation function; the input gate is used to control the input ratio of new information;

[0063] The forget gate is represented as follows: ;in, This represents the output of the forget gate. The bias of the forget gate is indicated; the forget gate is used to control the proportion of historical information retained.

[0064] The candidate memory unit is represented as follows: ;in, Indicates the output of the candidate memory unit. Indicates the weight of the candidate memory unit. Indicates the bias of candidate memory units. This represents the hyperbolic tangent activation function; the candidate memory units are used to generate new memories;

[0065] The memory unit update is represented as follows: ;in, This indicates the updated value of the memory cell. This represents the updated value of the memory cell in the previous time step. This represents element-wise multiplication; the memory unit update is used to merge historical memory with new memory.

[0066] The hidden state output is represented as follows: ;in, Indicates the first The hidden state at each time step The output of the output gate is indicated; the hidden state output is used to characterize the temporal features of the current time.

[0067] The output gate is represented as: ;in, This indicates the bias of the output gate; the output gate is used to control the output ratio of the memory unit.

[0068] The final output of the layer step Hidden state This is used to characterize long-term dependencies of temporal features.

[0069] In this embodiment, it should be specifically noted that the... The layer takes a subset of key features as input, uses athlete health status labels as supervision signals, and employs mean squared error as the loss function. It iteratively trains multiple decision trees using a gradient boosting algorithm, initializes prediction values, calculates residuals for each iteration, and trains the th layer. The residuals of each tree are fitted, the predicted values ​​are updated, and the final output is the weighted sum of all trees;

[0070] The fusion layer is used to... Layers and The outputs of the layers are concatenated into a joint feature vector, which is then mapped to a comprehensive health index.

[0071] pass Layers capture the long-term dynamic dependence of athletes' health status, through By mining the nonlinear relationships of key indicators in the layer and integrating them through the fusion layer, a comprehensive health index is output. This achieves efficient fusion of multimodal data, solves the limitation of a single model that can only capture the correlation of a single type of feature, and significantly improves the comprehensiveness and accuracy of health assessment.

[0072] In this embodiment, it should be specifically noted that the... The layers are specifically:

[0073] The mean of the health status labels corresponding to all data is used as the initial predicted value to reflect the overall level and provide a reasonable starting point for subsequent iterations, avoiding excessive initial bias in the model. Simultaneously, model parameters are initialized, including tree depth, learning rate, and regularization parameters. The tree depth is a limit on the maximum depth of a single decision tree to prevent overfitting. The learning rate controls the contribution weight of each tree to the final prediction result, improving the model's generalization ability by reducing the update magnitude at each step. The regularization parameters include... Parameters and Parameters, the The parameters are used to control the sparsity of the leaf node weights. The parameter is used to control the smoothness of the leaf node weights;

[0074] Decision trees are generated through multiple iterations. ; Total number of decision trees; based on initial predictions. Calculate the residuals of the samples, where the samples are the subset of key features; the residuals of the samples are expressed by the formula: ;in, Indicates the first The first round residual of the sample, i.e. the first... The residuals of the key feature subsets corresponding to each athlete in the first iteration. Indicates the first The true health status labels of each athlete; using a subset of key features as input, and residuals... Construct a decision tree for the target value. The tree construction process uses a greedy algorithm to select the optimal splitting node. For each feature in the key feature subset, it iterates through all possible splitting points, calculates the gain after splitting, selects the feature with the largest gain and the splitting point, and splits the current node into left and right child nodes. This process is repeated until the maximum depth is reached or further splitting is not possible. The predicted value of the first tree is then used to construct the tree. The weighted summation is applied to the initial predicted value to obtain the predicted value after the first iteration: ;in, This represents the predicted value after the first iteration. Indicates the learning rate. This is used to limit the impact of a single tree; Indicates the first Key feature subset data corresponding to each athlete; in the first... During round iteration, based on the previous Predicted value of the wheel Calculate the new residual The formula is expressed as:

[0075] Using a subset of key features as input and residuals as input... Construct a decision tree for the target value. Repeat the above process of splitting nodes and building the tree to generate the first... Trees, update predictions: The iteration stops when any set condition is met; the set conditions include: reaching the maximum predicted number of trees. The loss function no longer decreases significantly and the iteration time exceeds the preset upper limit; after After round of iterations, The layer outputs the final predicted value. That is, feature importance weighted score ; ; A larger value indicates a stronger positive contribution of the feature to health status, while a smaller value indicates a stronger negative contribution of the feature to health status.

[0076] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0077] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A big data based method for assessing the health of an athlete, the method comprising: The method comprises the following steps: ​ Step 1, multi-source heterogeneous data collection and preprocessing; collecting and preprocessing physiological index data, exercise load data, environmental parameter data and recovery behavior data of athletes at each time step; Step 2, generating a structured data set from the preprocessed data in step 1, and performing feature extraction and dimension reduction to obtain a time series feature subset and a key feature subset; Step 3, constructing a multi-modal fusion health assessment model to output a comprehensive health index; a hybrid model architecture of time series modeling and feature classification is adopted, including time series feature modeling and structured feature modeling; Step 4, grading and dynamic early warning of health status; According to the comprehensive health index, the health status is divided into five levels, and the index change rate is calculated through a sliding window to realize dynamic early warning; Step 5, self-updating and personalized calibration of the health assessment model.

2. The method of claim 1, wherein: The physiological index data includes resting heart rate, sleep heart rate, blood lactic acid concentration, serum creatine kinase and sleep quality; the exercise load data includes training duration and action repetition times; the environmental parameter data includes the temperature, relative humidity, altitude and air pressure of the environment where the athlete is located; the recovery behavior data includes nutrient intake heat, water intake and massage relaxation time; the physiological index data of the athlete is obtained through the wearable device of the athlete, the exercise load data is obtained through the motion sensor, the environmental parameter data is obtained through the environmental monitor, and the recovery behavior data is obtained through the questionnaire survey of the athlete; the preprocessing operation includes denoising, missing value filling and time alignment of the data.

3. The method of claim 2, wherein: The feature extraction and dimension reduction include time series feature extraction, individual baseline calibration and feature selection; the long-term dependence relationship of the time series feature is extracted using a long short-term memory network for time series feature modeling, and a hidden state is output; The structured features modeling uses The model on the screened key feature subset Performing a nonlinear transformation, outputting a feature importance weighting score; The output hidden state and the output feature importance weighted score are spliced, and the comprehensive health index is output through the output layer.

4. The method of claim 3, wherein: The structured data set generated in step 2 is represented as: ; wherein, denotes a structured dataset, denotes the th player's multi-dimensional feature vector at the th time step, denotes the total number of time steps; The time sequence feature extraction specifically comprises: dividing a time sequence by using a sliding window, obtaining a statistic quantity and a time sequence correlation of features in the window, and generating a time sequence feature matrix , ; wherein, denotes a feature dimension; a window length of the sliding window is , and a sliding window step is ; The individual baseline calibration specifically comprises: obtaining a health baseline of historical data of each athlete, and performing normalization on current features to obtain a time sequence feature subset of each athlete, denoted as: ​ The feature selection specifically includes: screening features that have significant influence on the health state through mutual information, and retaining a key feature subset , is expressed as .

5. The method of claim 4, wherein: The comprehensive health index in step 3 is represented as: ; wherein, represents the overall health index of the th athlete at the th time step; represents the fully connected layer weight matrix; represents the bias vector; represents the activation function; represents the vector concatenation operation; represents the output of the th layer, the output of the th layer.

6. The method of claim 5, wherein: The exponential change rate is calculated by a sliding window is expressed by the formula: ; wherein, represents the exponential change rate, represents the comprehensive health index of the i-th athlete at the j-th time step, represents the comprehensive health index of the i-th athlete at the j-th time step, represents the comprehensive health index of the i-th athlete at the j-th time step, represents the comprehensive health index of the i-th athlete at the j-th time step, represents the comprehensive health index of the i-th athlete at the j-th time step, represents the comprehensive health index of the i-th athlete at the j-th time step. The health status is divided into five levels: excellent, good, moderate, poor, and critical; when the comprehensive health index meets the requirements... At that time, the health level was excellent. satisfy If the overall health index meets the requirements, an early warning will be issued; when the overall health index meets the requirements... At that time, the health level was good. satisfy If the overall health index meets the requirements, an early warning will be issued; when the overall health index meets the requirements... At that time, the health level was medium. satisfy or If the overall health index meets the requirements, an early warning will be issued; when the overall health index meets the requirements... At that time, the health level was poor, requiring intervention. satisfy or If the overall health index meets the requirements, an early warning will be issued; when the overall health index meets the requirements... If the health level is poor, training should be stopped immediately, an emergency warning should be triggered, and the team doctor should be notified to conduct an examination.

7. The method of claim 6, wherein: The health assessment model mainly comprises an input layer, a hidden layer, a fusion layer and an output layer; the input layer is used for inputting a time sequence feature subset and a key feature subset; and the output layer is used for outputting a comprehensive health index. The The layers include an input gate, a forget gate, a candidate memory cell, a memory cell update, a hidden state output, and an input gate. The The layer takes a subset of key features as input, takes the athlete health status label as a supervisory signal, uses mean squared error as a loss function, iteratively trains decision trees through gradient boosting algorithm, initializes a prediction value, calculates a residual for each iteration, trains a tree to fit the residual, updates the prediction value, and finally outputs the weighted sum of all trees; The fusion layer is used to splice the output results of the layers into a joint feature vector, and map the joint feature vector to a comprehensive health index. the output results of the layers into a joint feature vector, and map the joint feature vector to a comprehensive health index.

8. The method of claim 7, wherein: The weight matrix of the forget gate is The weight matrix of the input gate is The weight matrix of the output gate is ; The input gate is represented as: ; where, represents the output of the input gate, represents the hidden state at the th time step, represents the input at the th time step, represents the bias of the input gate, represents an activation function; The forget gate is represented as: ; wherein, represents an output of the forget gate, represents a bias of the forget gate; The candidate memory unit is represented as: ; wherein, represents an output of the candidate memory unit, represents a weight of the candidate memory unit, represents a bias of the candidate memory unit, represents a hyperbolic tangent activation function; The memory cell update is represented as: ; wherein, represents the memory cell update value, represents the memory cell update value of the previous time step, represents the element-wise product; The hidden state output is represented as: ; wherein, represents the hidden state of the th time step, represents the output of the output gate; The output gate is represented as: ; wherein, represents the bias of the output gate.

Images