Tennis player sport injury risk early warning analysis method

Abstract

The invention provides a tennis player sport injury risk early warning analysis method, which comprises the following steps of: preprocessing acquired multi-dimensional sport data, accurately modeling time sequence dependence and dynamically focusing key features by utilizing a model architecture combining a long-short term memory network and an attention mechanism; through interpretability constraint loss function and Grad-CAM thermodynamic diagram visualization, high-risk action stages and key features are automatically extracted from a model decision path, a causal analysis interpretation report is generated, and finally real-time prediction and interpretation of damage risk levels are realized. According to the method, the accuracy of risk identification, the interpretability of model output and the adaptability of actual deployment are effectively improved, and scientific auxiliary decision making of coaches and medical teams is facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of athlete injury risk prediction, and particularly relates to a tennis athlete sports injury risk early warning analysis method. BACKGROUND

[0002] With the rapid development of intelligent sports medicine and health monitoring technology, deep learning models are increasingly widely used in tennis athlete injury risk prediction and early warning systems. Current mainstream solutions generally use multi-modal sports data collection combined with time series deep learning models (such as LSTM and convolutional neural networks) to capture key sports features of athletes during training and competition, and to automatically determine the risk level of sports injuries. These methods usually rely on data collected by large-scale wearable sensors (such as inertial measurement units, force sensors, and pressure shoe pads) and motion capture devices, and deploy sequence model-based prediction algorithms on the backend to try to improve the automation level and prediction accuracy of risk early warning. In recent years, with the advancement of artificial intelligence and edge computing technology, intelligent monitoring systems have been widely promoted in sports team training scenarios, rehabilitation medical scenarios, and real-time risk control during events, and the overall industry technology has shown a trend of transitioning from traditional experience methods to data-driven intelligent models; However, in actual application, the current injury risk prediction system generally has the problem of "black box" of the model, that is, the model training and reasoning process lacks an explainable mechanism, making it difficult for coaches, medical personnel, and other relevant personnel to effectively understand and trust the output results. Existing patents and papers focus on improving model accuracy, such as using multi-layer LSTM networks to deeply encode sports sequences, adding convolution modules to strengthen spatial feature capture, etc., but the research on the visualization and explanation mechanism of key steps and feature contributions in the model decision-making process is relatively lagging. The mainstream technology often only outputs injury risk level or probability results, and lacks clear explanations of key action stages and high-risk feature variables in the prediction process, making it difficult to provide targeted decision-making basis for training intervention, medical diagnosis, and sports rehabilitation. In addition, some traditional risk warning models have problems such as inadequate handling of different sensor data scales, action cycle recognition, label alignment, and other details, further exacerbating the limitations of model portability and universality. SUMMARY

[0003] The present application provides a tennis athlete sports injury risk early warning analysis method to solve the above technical problems.

[0004] The technical solution of the present application is as follows: a tennis athlete sports injury risk early warning analysis method, comprising: S1: Collect multi-modal motion data of tennis players during training or competition, including three-dimensional acceleration, angular velocity, ground reaction force, joint angle and motion trajectory sequence, and record corresponding motion action label and timestamp information; S2: Denoising and normalization preprocessing of the collected original motion data, using sliding window method for segmented processing of each channel signal, generating time sequence sample set, and labeling sample corresponding injury risk grade label; S3: Construct a deep neural network architecture based on long short-term memory network (LSTM) and attention mechanism fusion, wherein the attention module is used to calculate the contribution weight of each time step input feature to the final risk prediction result; S4: Input the preprocessed time sequence sample into the deep neural network for end-to-end model training, and introduce an interpretable constraint loss function in the training process, which is composed of classification loss and attention weight sparsity regularization term; S5: After the model training is completed, the gradient weighted class activation mapping (Grad-CAM) technology is used to visualize the analysis of the model decision path, and the key input features and their time distribution patterns affecting the prediction results are extracted; S6: Based on the key input features and visual analysis results, generate an explanation report in natural language form, which includes key action stages, high-risk feature variables and causal relationship analysis between them and injury risk; S7: Deploy the trained injury risk prediction model with explainability to the sports medical monitoring system, receive real-time motion data stream and output risk grade prediction results and corresponding explanation information for reference decision of coaches and medical team; S8: Continuously collect model prediction results and actual injury event feedback data during system operation, construct model performance evaluation log, and periodically update model parameters and explanation mechanism based on new data to improve long-term prediction stability and explainability consistency of the model.

[0005] The tennis player motion injury risk early warning analysis method provided by the application has the following beneficial effects: (1) The application innovatively introduces a soft attention mechanism and an explainable loss function joint optimization strategy in the LSTM-based time series deep learning framework, calculates the specific contribution of each time step feature in the multi-modal motion data to the risk prediction, and explicitly generates an attention weight distribution. This design not only enables the model to automatically focus on the key action stages and features that have physiological mechanism significance for injury risk, improving prediction accuracy, but also provides transparent and traceable explanation basis for the output of the model, significantly improving the "black box" problem of traditional deep models in motion injury risk prediction. Compared with traditional deep learning methods that simply aim to minimize loss, the output of the model is easier for coaches and medical personnel to understand and trust, greatly improving the usability of the system in practical applications; (2) The application creatively introduces a joint optimization of classification loss and attention weight sparsity regularization in the loss function, combines L1 regularization and weight clipping, and forces the model to focus on limited key features and time points, preventing the attention distribution from being too dispersed or having no discriminability. This dynamic weighting mechanism of multi-objective loss ensures that the model can maintain high risk prediction performance and ensure high explainability in the decision-making process. Under the premise of ensuring that the accuracy rate does not decrease, the model effectively reduces the risk of overfitting and redundant attention, providing a practical and applicable technical foundation for subsequent clinical intervention and sports program development; (3) The application combines the gradient weighted class activation mapping (Grad-CAM) method to visualize the entire decision-making process of the deep model, identifies and quantifies key input features and their distribution patterns in time series, and further associates sparse high-weight attention stages with physiological risk mechanisms through time series clustering and pattern extraction. Based on the model output, a structured natural language explanation report for coaches and medical personnel can be automatically generated, covering high-risk action stages, core variables and corresponding physiological causal relationships, achieving automation, standardization and professionalization of the explanation process. This mode greatly improves the business readability and medical usability of the model output, providing scientific support for improving team intervention efficiency; (4) The application designs a full-link automated process for multi-modal data acquisition, synchronization, cleaning, feature reconstruction and label alignment, ensuring high consistency and high modeling of training input data, and significantly improving the generalization ability of the model. At the same time, through efficient data format standardization and sliding window sampling, as well as automatic optimization and deployment of the model inference engine, the real-time and system stability requirements of sports medicine practical applications are fully met, realizing the online inference, explanation and feedback function closed loop of the model. BRIEF DESCRIPTION OF DRAWINGS

[0006] Figure 1 A flowchart of a tennis player motion injury risk early warning analysis method of the application; Figure 2 This is a sub-flowchart of a sports injury risk early warning analysis method for tennis players according to the present invention; Figure 3 This is another sub-flowchart of the method for early warning analysis of sports injury risks for tennis players according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this invention provides a method for early warning and analysis of sports injury risks for tennis players, specifically including: S1: Collect multimodal motion data of tennis players during training or competition. The multimodal motion data includes three-dimensional acceleration, angular velocity, ground reaction force, joint angles and motion trajectory sequence, and record the corresponding motion action labels and timestamp information. S2: Denoise and normalize the collected raw motion data, use the sliding window method to segment the signals of each channel, generate a time series sample set, and label the damage risk level of the sample. S3: Construct a deep neural network architecture based on the fusion of Long Short-Term Memory (LSTM) network and attention mechanism, where the attention module is used to calculate the contribution weight of the input features at each time step to the final risk prediction result; S4: Input the preprocessed time series samples into the deep neural network for end-to-end model training, and introduce an interpretability constraint loss function during the training process. The loss function is composed of classification loss and attention weight sparsity regularization term. S5: After the model training is completed, the gradient weighted class activation mapping (Grad-CAM) technique is used to visualize and analyze the model decision path, and extract the key input features that affect the prediction results and their temporal distribution patterns. S6: Based on the key input features and the visual analysis results, generate an explanation report in natural language form, including key action stages, high-risk feature variables, and causal association analysis between them and injury risk; S7: Deploy the trained and explainable injury risk prediction model into the sports medical monitoring system, receive real-time sports data streams, and output risk level prediction results and corresponding explanation information for reference decision-making by coaches and medical teams; S8: Continuously collect model prediction results and actual injury event feedback data during system operation, build model performance evaluation logs, and periodically update model parameters and explanation mechanisms based on new data to improve long-term prediction stability and explainable consistency of the model.

[0010] The step S1: Collecting multi-modal sports data of tennis players during training or competition, including three-dimensional acceleration, angular velocity, ground reaction force, joint angle, and motion trajectory sequence, and recording corresponding sports action labels and timestamp information. Specifically, it includes: S1.1: Based on wearable inertial measurement unit (IMU) sensors and pressure distribution measurement insoles, obtain the original sports data generated by tennis players during training or competition, including three-dimensional linear acceleration, three-dimensional angular velocity, plantar pressure distribution, and ground reaction force time series signals, to build a multi-modal input data set; Based on wearable inertial measurement unit (IMU) sensors and pressure distribution measurement insoles, collect multi-modal original sports parameters generated by tennis players during training or competition as input conditions, including three-dimensional linear acceleration, three-dimensional angular velocity, plantar pressure distribution, and ground reaction force time series signals; Using the three-axis acceleration measurement unit (range parameter: ±16g) and three-axis gyroscope (range parameter: ±2000° / s) of the IMU sensor, three-dimensional linear acceleration and three-dimensional angular velocity original signals are collected at each time step, and the analog output is converted to digital quantity through the built-in analog-to-digital conversion module, so as to be fused with other sensor data; Further, through the pressure distribution measurement insole (sensing unit array size: 8x16 nodes, sampling frequency: 500Hz), the load distribution of different areas of the foot is measured, and multi-point pressure values and their spatial distribution matrix are obtained to reflect the force state of the athlete in different postures; Further, through the cooperation of the insole and the ground force platform, the three-dimensional components of the ground reaction force are measured based on the multi-axis force sensor output, and the signal is normalized and calibrated with As a unit to provide accurate load data for subsequent biomechanical analysis; Further, the time stamp generation module is used to sample and mark the time of each channel output signal, so that the IMU data, pressure distribution and ground reaction force signal have a unified time reference, providing a basis for the time synchronization process in the next sub-step S1.2; Through the multi-sensor acquisition link, the signals of the above different physical quantities are stored as the original multi-modal input data set, realizing the construction of the basic data for subsequent feature reconstruction and risk analysis; For example, in an indoor tennis training, the IMU sensor is installed on the outside of the athlete's right ankle at a sampling rate of 200 Hz, the three-dimensional acceleration range is set to ±8g, and the three-dimensional angular velocity range is set to ±1000° / s; the pressure distribution insole has a sampling frequency of 250 Hz, and the single-point pressure measurement accuracy is The ground force plate synchronously collects three-dimensional reaction force components, with a sampling frequency of 1000 Hz and a measurement accuracy of After simultaneous collection by the above devices, 200 time steps of IMU data, 250 time steps of insole pressure matrix and 1000 time steps of reaction force data are obtained per second. Combined with the unified time stamp marking, the cross-device synchronization of the data is realized, and finally an original multi-modal input data set containing acceleration, angular velocity, pressure distribution and ground reaction force is generated, with a total data amount of about 5 GB, meeting the training needs and explainable calculation analysis of the subsequent deep learning model; S1.2: Time synchronization processing is performed on the acquired original motion data, and the data from different sensor channels are aligned based on the unified time stamp, so as to eliminate the time sequence misalignment problem caused by the difference in sampling frequency, and generate time-aligned multi-channel motion signals; S1.3: The motion capture technology based on high-speed camera system is adopted to reconstruct the joint angles and motion trajectories of the tennis player in real time, and generate the motion trajectory sequence in three-dimensional space and the angle change curve of each main joint (shoulder, elbow, wrist, hip, knee and ankle), as the key kinematic feature input; The high-speed camera system multi-base optical sensing array configuration (parameters: camera frame rate ≥200fps, resolution 1920×1080 pixels, lens viewing angle covering the complete motion range) is adopted to realize real-time capture of the three-dimensional space coordinates of the key joints of the tennis player; Further, the multi-camera calibration method (parameters: Zhang's calibration method combined with Tsai model) is used to realize the conversion of different camera coordinate systems to a unified world coordinate system, and obtain accurate camera parameter matrix and distortion coefficient, which are used for subsequent geometric correction of three-dimensional reconstruction; Further, a multi-view stereo matching algorithm (parameters: semi-global matching + cross-view matching constraint based on SURF features) is used to realize pixel-level position matching of each joint in each view image, and a three-dimensional coordinate value of the joint center in the unified world coordinate system is calculated by combining the triangulation method to form an original three-dimensional skeleton data matrix; Further, through a time sequence smoothing filter processing (algorithm: Kalman filter, state transition model based on uniform acceleration assumption), dynamic noise suppression of the joint three-dimensional coordinate sequence is realized, and stable spatial motion trajectory data is obtained, reducing the jitter and instantaneous misplacement in the high-speed motion capture process. Further, a kinematic inverse algorithm (parameters: human skeleton model, joint rotation degree of freedom constraint: shoulder 3 degrees of freedom, knee joint 1 degree of freedom) is used to calculate the angle change curve of the main joints according to the three-dimensional skeleton model, and output the shoulder, elbow, wrist, hip, knee, and ankle joint angle time sequence vector related to the hitting action. Through the above multi-modal motion capture and spatial reconstruction algorithm, the multi-channel motion signals time-synchronized in the previous step are converted into key kinematic feature data containing three-dimensional motion trajectories and joint angle time sequences, realizing spatial precision and time sequence structuring of the action features, and providing clear geometric and dynamic basis for S1.4 action label annotation. For example, in a training monitoring scene based on a high-speed camera, four high-speed cameras are configured, with a frame rate of 240 fps and a resolution of 1920x1080, arranged at the four corners of the tennis court, covering the entire serving area and the baseline area. Zhang's calibration method is used to obtain the intrinsic parameters (focal length fx=1450px, fy=1445px, optical center cx=960px, cy=540px) and extrinsic parameters (rotation matrix and translation vector) of each camera, and the multi-camera coordinates are converted to the world coordinate system. SURF feature matching is applied to match the pixel positions of the human shoulder, elbow, wrist, hip, knee, and ankle key points between multiple views, and the three-dimensional coordinates are calculated using the triangulation method, for example, the shoulder joint world coordinates during the serving stage are m. After Kalman filter smoothing, the shoulder joint trajectory jitter amplitude is reduced from 5 mm to 1 mm. Based on the human skeleton model and the rotation degree of freedom constraint, the angle change curve of the shoulder joint during the serving swing stage is calculated, the shoulder joint abduction angle is increased from degrees to degrees, and the elbow joint flexion angle is reduced from degrees to degrees. These three-dimensional trajectories and joint angle time sequences serve as high-precision kinematic features, providing reliable spatial dynamics basis for subsequent risk label annotation. S1.4: Perform action label annotation on the synchronized multi-modal motion signals and motion capture data, identify and label the types of hitting actions (such as forehand, backhand, serve, smash, etc.) performed by the athletes and their start and end time intervals based on video recording and sensor data fusion analysis, and generate a structured action label sequence; S1.5: Perform data format standardization processing on the annotated multi-modal motion data and action label sequence, store them in a unified data structure (such as HDF5 or TFRecord format) to form an original data sample set that can be used for subsequent preprocessing and model training.

[0011] The step S2: Perform denoising and normalization preprocessing on the collected original motion data, use the sliding window method to segment the signals of each channel, generate a time series sample set, and label the corresponding injury risk level label of the sample. Specifically, it includes: S2.1: Perform denoising processing based on wavelet transform on the collected original multi-modal motion data to eliminate the influence of sensor noise on joint angle and ground reaction force signals, and obtain denoised three-dimensional acceleration, angular velocity, ground reaction force, joint angle and motion trajectory sequence data; For the time-synchronized multi-modal motion data, use the wavelet transform-based denoising method (parameters: wavelet basis type, decomposition level, threshold rule) to realize multi-scale decomposition processing of three-dimensional acceleration, angular velocity, ground reaction force, joint angle and motion trajectory sequence; Further, by selecting the Daubechies-6 wavelet basis suitable for the non-stationary characteristics of the motion signal (decomposition level is 1-6 layers, preferably 4 layers), the signal is separated in the low-frequency and high-frequency subbands, and the approximate coefficients and detail coefficients of different resolution levels are obtained; Further, a soft threshold denoising strategy (parameter: threshold calculation method is Rigrsure or Minimax) is used to suppress the high-frequency noise components in the detail coefficients and retain the effective signal components containing key action features, and the denoised sensor channel data is obtained; Further, inverse wavelet reconstruction is performed on the denoised approximate coefficients and detail coefficients to restore the original signal in time domain, and the multi-channel motion signal with improved signal-to-noise ratio is generated; Further, the mean square error and signal-to-noise ratio index are used to evaluate the denoising effect, where the signal-to-noise ratio calculation formula is: where, is the signal power, is the noise power; By using wavelet transform denoising, the time-synchronized multimodal signal results from the previous step are transformed into motion data with high signal-to-noise ratio and preservation of key temporal features, thereby significantly improving the data quality in subsequent feature extraction and model training. For example, in a set of tennis training data, the sampling frequency is A Hz three-dimensional acceleration signal was used, and the Daubechies-6 wavelet basis was selected. The number of decomposition levels was set to [number missing]. Layer, threshold rule is Rigrsure, the mean square error of the reconstructed signal is reduced to that of the original. Improved signal-to-noise ratio dB. In this scenario, the noise reduction process effectively filters out high-frequency noise from sensor jitter and environmental vibration, while retaining the low-frequency smooth changes and key acceleration peaks from the start of the swing to the moment of impact. The output data exhibits higher prediction stability and generalization ability in subsequent LSTM feature encoding. S2.2: Based on the min-max normalization method, the denoised three-dimensional acceleration, angular velocity, ground reaction force, joint angle and motion trajectory sequence data are standardized to eliminate the interference of different sensor dimensions on model training and output normalized multi-channel time series data. S2.3: Based on the sliding window algorithm, the normalized multi-channel time series data is segmented and continuous motion cycle segments are extracted with a fixed time window length to generate a structured time window sample set. Each time window sample contains feature vectors of multiple time steps. S2.4: Perform label alignment operation on the structured time window sample set based on motion action labels and timestamp information to ensure that the action stage corresponding to each time window sample is accurately matched with the injury risk level label, and output a labeled time window sample set; S2.5: Based on the expert knowledge system and historical damage data, damage risk level mapping processing is performed on the labeled time window sample set to transform continuous risk indicators into discrete risk level labels, and output a structured time series sample set for model training.

[0012] Step S3: Construct a deep neural network architecture based on the fusion of Long Short-Term Memory (LSTM) network and attention mechanism, wherein the attention module is used to calculate the contribution weight of the input features at each time step to the final risk prediction result. Figure 2 As shown, it specifically includes: S3.1: A multi-layer temporal feature extraction module based on the LSTM network structure is designed to encode the preprocessed time series samples to obtain a high-level motion feature representation with temporal dependence characteristics. S3.2: Embed a soft attention mechanism module after the LSTM output layer to calculate attention weights based on the hidden state sequence of the LSTM output, so as to dynamically adjust the contribution of features at each time step to the final output; For the multi-layer LSTM hidden state sequence output by step S3.1, a weight calculation module is constructed using a soft attention mechanism (parameter: attention weight matrix). Projection vector Bias terms This enables the modeling of the correlation between the hidden state and the target prediction at each time step; Furthermore, an additive attention function (Bahdanau Attention) is applied to each time step. Perform mapping operations with the global context to calculate the unnormalized attention score. : in Rate the attention level at time step t. This is the transpose of the projection vector. The hidden state vector of the LSTM It is the hyperbolic tangent activation function; Furthermore, through exponential weighted normalization, the unnormalized score is transformed into attention weights. : in To obtain the unnormalized attention score at the k-th time step, a standardized distribution of the feature contribution at different time steps is achieved; Furthermore, based on the obtained attention weights and the hidden states at the corresponding time steps, a weighted combination is performed to form a dynamic context vector. enter: Feature aggregation used for subsequent risk prediction; By using soft attention mechanism computation and weighted combination processing, the hidden state sequence of the previous step is transformed into a context vector with explicit temporal feature contribution evaluation, thereby achieving dynamic attention to key time steps and enhancing interpretability. For example, in a set of lengths Dimensions In the LSTM hidden state sequence, set the attention weight matrix. Dimensions Projection vector Dimensions bias Dimensions For each time step , first calculate , for example, at a certain time step , and then get by Softmax normalization processing . After normalization of the entire sequence, the attention weight distribution presents sparsity, that is, only about time steps have weights higher than . After weighted combination on these time steps, the context vector is generated, which has a dimension of , and the accuracy of predicting injury risk in the subsequent classification layer is improved by %, and through the analysis of the distribution, the key action phase can be explained. This embodiment verifies the effectiveness of the soft attention mechanism in improving the prediction performance and enhancing the explainability; S3.3: Perform Softmax normalization operation on the attention weight vector to obtain the normalized weight distribution of the input features at each time step to form the basis for the explainable feature importance ranking; S3.4: Based on the normalized attention weights, perform weighted summation on the LSTM hidden states to generate a context vector as the final risk prediction input feature to improve the model's ability to identify key action phases; S3.5: Input the context vector into the fully connected classification network to generate a preliminary prediction result of the injury risk level through multiple layers of nonlinear transformation, providing a basic output for the subsequent introduction of an explainable loss function; Input the context vector generated by the weighted summation of the attention weights into the fully connected classification network as the initial input feature vector of the multi-layer nonlinear mapping, to establish a nonlinear mapping relationship between high-dimensional motion features and injury risk levels; A hierarchical fully connected structure (parameters: number of layers ≥ 2, number of nodes determined by input feature dimension and risk level category number) is adopted, and affine transformation and activation function calculation are performed in each hidden layer to gradually abstract the discriminant pattern of the context features. The activation function uses the rectified linear unit (ReLU) to ensure stable gradient propagation; , where is the output vector of the layer; is the weight matrix of the layer; is the bias vector; is the output of the previous layer; is a nonlinear activation function; Furthermore, a Softmax normalization operation is performed on the output layer to transform the final linear transformation result into a probability distribution of damage risk levels, calculated using the following formula: in, For the first The predicted probability of a risk level. The unnormalized score for the corresponding category. The unnormalized attention score at the k-th time step; Furthermore, by calculating the maximum value of each category probability and its index, preliminary risk level prediction results are obtained, and the entire prediction vector is retained to support subsequent calculation of the interpretability loss function; By using multi-layer nonlinear mapping and Softmax normalization, the context vector from the previous step is transformed into a probability distribution of damage risk level, providing a basic output for the S4 step to calculate classification loss and introduce interpretable constraints. For example, in an experiment predicting sports injury risks for professional tennis players, the context vector dimension was 128, the number of nodes in the first hidden layer was 256, and the weight matrix... Initialization uses the Xavier method, bias vector Initialize to zero; the number of nodes in the second hidden layer is 64, and the weight matrix... Initialization uses the He method, bias vector Initialize to constants The number of output layer nodes equals the number of risk level categories. The output vector is normalized using Softmax to obtain the predicted probability for each class, achieving an average accuracy of [percentage missing] on the test set. The average cross-entropy loss between the preliminary prediction results and the true labels is This provides a stable foundation for subsequent introduction of attention sparsity regularization and interpretability optimization, resulting in a stable predictive output.

[0013] Step S4: The preprocessed time series samples are input into the deep neural network for end-to-end model training. An interpretability constraint loss function is introduced during training, consisting of a classification loss and an attention weight sparsity regularization term. Figure 3 As shown, it specifically includes: S4.1: Input the preprocessed time series samples into the constructed deep neural network architecture that integrates the long short-term memory network and the attention mechanism to extract the temporal dependencies of the input features and generate the initial prediction output; S4.2: Based on the difference between the predicted results of the model output and the true injury risk level label, calculate the classification loss term, and use the cross-entropy loss function to quantitatively evaluate the prediction error of the model to guide the gradient update direction of the model parameters; S4.3: Perform sparsity regularization processing on the attention weight distribution output by the attention module, and use the L1 regularization strategy to constrain the attention weight vector to promote the model to focus on the input features that play a key role in injury prediction, forming an interpretable decision basis; Perform sparsity measurement calculation on the attention weight vector output by the attention mechanism module under the current batch of samples, and use the L1 norm to evaluate the sparsity degree of the weight distribution; Based on the above sparsity measurement results, use the L1 regularization strategy (parameters: regularization coefficient ) to construct a sparsity constraint term, and add to the loss function to control the sparsity degree of the attention weight vector and guide the attention mechanism to focus on a small number of high-contribution time step features during parameter update; Further, through the weight clipping method (parameters: threshold ), set the weight coefficients below to zero to achieve explicit sparsification and reduce the interference of non-contributing features on the model; Further, based on the physiological correlation filtering rule, map the non-zero weights after sparsification to the kinematic feature variables, and exclude features unrelated to injury mechanisms to improve the interpretive effectiveness of the sparsity regularization term; By combining gradient backpropagation, adjust the gradient contribution proportion of the sparsity constraint term in the total loss, so that the model maintains prediction accuracy while forming a weight distribution with clear feature focus; Through the synergistic effect of L1 regularization and physiological correlation filtering, the original attention weight distribution in the previous step is converted into a sparse and highly correlated feature importance matrix, achieving the technical effect of balancing prediction performance and interpretability; For example, for an input feature sequence of length time steps, obtain the attention weight vector , where the maximum weight is , the minimum weight is , the regularization coefficient is set to = , the threshold is set to = , and the L1 norm is calculated as . Add regularization penalty term and set the weights of 14 time steps to zero during the weight pruning process. After the physiological correlation mapping, 8 time step features related to shoulder impact and knee load were retained, and the prediction accuracy on the validation set remained at , proving that the sparsity regularization strategy improves the model's focus on effective features while maintaining the prediction performance; S4.4: Linearly combine the classification loss term and the attention weight sparsity regularization term to construct an explainability constrained loss function for joint optimization, to simultaneously optimize the prediction performance and explainability indicators during model training; S4.5: Based on the constructed explainability constrained loss function, iteratively optimize the model parameters in the deep neural network through the backpropagation algorithm to obtain a tennis player injury risk warning model with high prediction accuracy and explainability.

[0014] The step S5: After the model training is completed, the gradient weighted class activation mapping (Grad-CAM) technology is used to visualize the model decision path and extract the key input features and their time distribution patterns that affect the prediction results. Specifically, it includes: S5.1: Perform forward propagation calculation on the trained deep neural network model based on LSTM and attention mechanism fusion to obtain the output prediction results and corresponding feature map data of the model on the validation set, where the feature map data includes the hidden state output and attention weight distribution of each time step, providing input basis for subsequent Grad-CAM analysis; S5.2: Based on the feature map data, the gradient weighted class activation mapping (Grad-CAM) technology is used to perform backpropagation operation on the last layer of convolutional features of the model, calculate the gradient information of each time step feature map to the output prediction result, to obtain the sensitivity weight coefficient of each time step feature map to the final classification result, and form a time sensitivity heat map as the visualization basis for the model decision path; S5.3: Normalize the time sensitivity heat map and combine the multi-modal motion data in the original input time series to perform feature mapping alignment to generate an explainability heat map matrix consistent with the original input dimension, where the heat value of each time step represents the influence intensity of the input feature at that time on the final prediction result of the model, serving as the basis for key input feature identification; S5.4: Based on the explainability heat map matrix, use the threshold segmentation algorithm to extract time steps with heat values higher than the set threshold to identify input feature sequences that play a key role in model prediction, forming a key feature time window set for subsequent generation of natural language explanation reports and positioning analysis of high-risk action stages; ​S5.5: performing a time distribution pattern extraction operation on the input features in the set of key feature time windows, using a time series clustering algorithm to classify the key time windows into patterns, identifying high-risk action periods with similar feature distributions, and outputting the feature sequences corresponding to each cluster center and their time distribution patterns to provide interpretable support at the dynamic behavior level; For the input features in the set of key feature time windows, a time distribution pattern extraction method (parameters: input feature sequence set, time step length, feature dimension) is used to encode and analyze the key feature sequences in different time windows in terms of time attributes; Further, a dynamic time warping (DTW) algorithm (parameters: distance metric method is Euclidean distance, normalized time scale) is used to align the feature time sequences between different time windows, eliminate the influence of action execution speed differences on pattern recognition, and obtain standardized feature time curves; Further, a K-means-based time series clustering algorithm (parameters: the number of clusters K is determined according to the elbow rule and Silhouette coefficient) is used to classify the key feature time curves after time alignment into patterns, and the center feature sequence of each cluster is generated as a representative pattern; Further, the similarity between each time window and the cluster center is calculated by a pattern consistency measurement function, as follows: wherein, is the pattern consistency score of the ith time window, is the time window length, is the distance measurement value between time step t and the cluster center; Further, time windows with a pattern consistency score higher than a set threshold are marked as high-risk action periods, and the feature sequences corresponding to each cluster center and their time distribution patterns are outputted; Through the time series clustering algorithm and time alignment processing method, the key feature time window sequence in the previous step is converted into a set of high-risk action periods with interpretable pattern labels, achieving interpretable support at the dynamic behavior level; For example, in a set of real professional tennis player sports data application scenarios, the input condition is a set of collected key feature time windows, containing 120 time windows, each with a length of 50 time steps and a feature dimension of 20. In the time pattern extraction stage, the DTW algorithm is used to time-align all time windows, with a normalized scale of 50 steps and a Euclidean distance metric. After alignment, the K-means clustering algorithm is used for pattern classification, with K=5 determined according to the elbow rule and Silhouette coefficient, and finally 5 cluster center patterns are obtained. In the pattern consistency calculation, =50, the score threshold is set to 0.8, and the windows with similarity higher than the threshold are marked as high-risk action periods, such as emergency stop turning and landing buffer stage. The final output result includes five cluster mode labels, the corresponding feature time distribution of each mode, and the marked list of high-risk action periods. The accuracy of identification in the verification set reaches 92%, which significantly improves the clarity and practicality of the behavior mode output of the model explanation.

[0015] The step S6: based on the key input features and the visualization analysis result, an explanation report in natural language form is generated, which includes key action stages, high-risk feature variables and causal correlation analysis between them and injury risk. Specifically, it includes: S6.1: The key input features and their time distribution patterns extracted from the model decision path are structured and coded to form a feature-time-impact intensity triple data structure that can be parsed; S6.2: Based on the feature variables and time distribution information in the triple data structure, combined with the motion action label and the physiological knowledge base, the key action stages corresponding to the high-risk feature variables are identified, such as the swing stage, the landing buffer stage or the emergency stop turning stage; S6.3: According to the activation intensity and distribution pattern of the feature variables in the Grad-CAM visualization analysis result, the contribution of each high-risk feature variable to the injury risk prediction is calculated, and it is mapped to the causal correlation intensity level in the natural language description; S6.4: Based on the contribution level and action stage information, a structured natural language description is generated by combining template driving and rule reasoning, covering key action stages, high-risk feature variables and causal correlation analysis between them and injury risk; In terms of input conditions, the contribution level of high-risk feature variables and the corresponding key action stage information have been calculated by the previous sub-step, including feature name, action stage identification and causal correlation intensity level, etc. Structured data; The template-driven natural language generation method (parameters: pre-defined explanation template set) is used to realize the formatted mapping of key action stages and high-risk feature variables, and the feature-action-contribution triple is used as the input data set for template filling; Further, through the rule reasoning algorithm (parameters: physiological mechanism knowledge rule base), the logical association between different data dimensions is realized, and the rule matching is used to establish the causal relationship between high-risk feature variables and the potential physiological risk mechanisms triggered by them, and generate an intermediate reasoning result set; Further, a weight priority-based sorting algorithm (parameters: contribution level, time window duration) is adopted to sort and optimize the intermediate inference result set, ensuring that the action stage features with high contribution level and long duration are presented first in the natural language description. Further, a syntax structure generation algorithm (parameters: part-of-speech tagging model, dependency syntax tree construction rules) is adopted to assemble the template filling results and inference result set into a set of natural language sentences with professional terms and logical coherence, combining feature variables, action stages, and causal association information. Through language optimization processing (parameters: term consistency dictionary, redundancy elimination rules), the sentence set from the previous step is converted into a structured, unambiguous, and compliant natural language description for the written expression specifications in the field of sports medicine, achieving professional adoptability of the explanation report in human-machine collaborative decision-making scenarios. For example, in a tennis player injury risk analysis task, the input features are shoulder joint angular velocity (unit: rad / s), knee joint flexion and extension displacement (unit: mm), and ground reaction force peak value (unit: N), with their contribution levels being high, medium, and high, respectively, and the corresponding action stages being the swing stage, the sudden stop and turn stage, and the landing and buffering stage. The template-driven method calls the predefined template "During the action stage [action stage], the change of [feature variable] is significant, with a contribution level of [contribution level] to the prediction of injury risk," and fills in the three data in turn. The rule-based reasoning combines the rules in the knowledge base that high angular velocity of the shoulder joint is associated with risk of rotator cuff injury, sudden change in knee joint flexion and extension displacement is associated with risk of ligament stretch injury, and abnormal peak value of ground reaction force is associated with risk of foot impact injury, to generate the corresponding causal association description. The sorting algorithm presents the shoulder joint angular velocity, ground reaction force peak value, and knee joint flexion and extension displacement in order according to the level order of high, medium, and high. The syntax structure generation algorithm generates "During the swing stage, the significant increase in shoulder joint angular velocity contributes a high level to the prediction of injury risk, which may cause rotator cuff injury risk; during the landing and buffering stage, the abnormal increase in ground reaction force peak value contributes a high level to the prediction of injury risk, which may cause foot impact injury risk; during the sudden stop and turn stage, the sudden change in knee joint flexion and extension displacement contributes a medium level to the prediction of injury risk, which may cause ligament stretch injury risk." The language optimization processing eliminates redundant phrases and unifies terms, standardizing terms such as "rotator cuff injury risk" and "foot impact injury risk" into medical standard expressions, and finally outputs a structured natural language description that can be directly incorporated into the explanation report; S6.5: Integrate the generated natural language description into a complete explanation report document, and attach the risk level and confidence information predicted by the model for reference by the coach and medical team in decision-making.

[0016] The step S7: deploying the injury risk prediction model trained and with explainability to the sports medical monitoring system, receiving real-time sports data stream and outputting risk level prediction results and corresponding explanation information for reference decision of coaches and medical team. Specifically, it includes: S7.1: encapsulating the injury risk prediction model trained and verified with high prediction accuracy and explainability into a deployable model file format, the model file including network structure definition, parameter weight and inference configuration information, to adapt to the computing resource environment of the sports medical monitoring system; S7.2: based on the hardware resources and real-time requirements of the deployment environment, performing inference engine optimization on the model file, using TensorRT or OpenVINO tool chain to fuse, quantize and schedule optimize the model execution graph, to obtain low delay and high throughput inference ability; S7.3: constructing a real-time data access pipeline, preprocessing and feature serialization of multi-modal sports data stream from wearable sensors or video analysis module, generating real-time feature vector sequence conforming to model input format as input condition of model inference; S7.4: based on the optimized inference model, performing forward propagation calculation on the real-time feature vector sequence, outputting injury risk level prediction results of the current sports action segment, and synchronously calculating attention weight distribution of each time step to support generation of subsequent explanation information; S7.5: based on the mapping relationship between the attention weight distribution and the input feature sequence, using gradient weighted class activation mapping (Grad-CAM) technology to perform real-time visual analysis on the model decision path, extracting key input features and their time distribution patterns affecting the current prediction results, and generating structured explanation feature vector; S7.6: based on the structured explanation feature vector and the preset sports medical knowledge base, performing natural language generation algorithm to generate risk explanation report for coaches and medical team, the explanation report including key action stage, high-risk feature variable and causal relationship analysis between them and injury risk, to improve the explainability and decision adoption rate of model output; The input condition is the structured explanation feature vector generated by the previous sub-step S7.5, which contains key input features, time distribution patterns and their influence intensity parameters, and has completed alignment processing with multi-modal sports data sequence, while the preset sports medical knowledge base stores the physiological causal relationship mapping information between action stage, feature variable and injury type; using feature semantic matching algorithm (parameters: word embedding vector dimension , matching threshold ), to realize the semantic alignment function of structured explanation feature vector and knowledge base entries, to calculate the matching score by cosine similarity and to filter out the physiological mechanism entries meeting the conditions; Further, through the causal correlation strength calculation method (parameters: significance level , the fitting model type is logistic regression), the contribution of high-risk feature variables to damage risk is quantified, and the causal correlation coefficient matrix is obtained, and the calculation formula is Among them is the logistic regression coefficient vector; Further, through the template-driven natural language generation algorithm (parameters: template number , rule matching priority descending), the key action stage, feature variable name, causal coefficient level are mapped to Chinese sentences, and three-paragraph description text containing action background, feature explanation and risk prompt is generated; Through the explanation information aggregation strategy, the Chinese description generated by the previous action is merged in chronological order, and the risk level and confidence information predicted by the model are appended at the end to obtain the complete risk explanation report data structure, including report metadata, text paragraphs and additional indicators; For example, in a tennis training scene, the structured explanation feature vector contains three key input features: knee flexion angular velocity peak ° / s, foot lateral pressure peak N, hip abduction angle change rate ° / s, and the corresponding time window is concentrated in the take-off and landing stage. The knowledge base entry shows that there is a strong correlation between knee flexion angular velocity and meniscus injury risk, pressure peak and ankle sprain risk, and hip abduction change rate and waist muscle strain. The semantic matching calculates the cosine similarity of the three as , , , all of which exceed the threshold. The logistic regression fitting obtains the causal coefficient vector , and after normalization, the correlation strength level is high, medium, medium. Template filling generates the description: "During take-off and landing, the peak value of knee flexion angular velocity is high, which causes high risk load to the meniscus; the increase of foot lateral pressure peak indicates that the ankle is in a medium risk state; the increase of hip abduction angle change rate increases the risk stress of the waist muscle group". The additional output model predicts the risk level to be high, and the confidence %. The report provides clear action stage and physiological risk correspondence to the coach and medical team, facilitating the development of targeted intervention strategies; S7.7: The risk level prediction result and the corresponding natural language explanation report are pushed to the coach and medical team through a visual interface or mobile terminal, realizing real-time monitoring and auxiliary decision support of sports injury risk, so as to improve the application value of the model in sports medicine scenarios.

[0017] The step S8: continuously collecting model prediction results and actual injury event feedback data during system operation, constructing model performance evaluation log, and periodically updating model parameters and explanation mechanism based on new data, so as to improve the long-term prediction stability and explainability consistency of the model. Specifically, it includes: S8.1: Collecting the prediction results output by the model during operation and the corresponding injury event feedback data to construct a structured model performance evaluation log; S8.2: Based on the collected prediction results and true labels, calculate the performance indicators of the model in different time windows, including accuracy, recall rate, F1 score and explainability consistency score, to evaluate the real-time prediction effect and explanation stability of the model; S8.3: Perform time series analysis on the model performance evaluation log using sliding window statistics method to identify model performance decline trend and explainability drift interval to determine model update trigger condition; S8.4: Based on the newly collected sports data and labeled samples, fine-tune the deep neural network model parameters using incremental learning strategy to improve the generalization ability and prediction robustness of the model under new data distribution; S8.5: After updating the model parameters, recalibrate the model explanation mechanism based on the attention weight distribution and Grad-CAM visualization results to ensure that the explanation information output by the model is consistent with the actual physiological mechanism, and improve the credibility and adoptability of the model in long-term operation.

[0018] So far, the technical solutions of the present application have been described in conjunction with the preferred embodiments shown in the drawings, but those skilled in the art can easily understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to related technical features without departing from the principles of the present application, and the technical solutions after these changes or replacements will fall within the protection scope of the present application.

[0019] The above description is only the preferred embodiments of the present application and is not intended to limit the present application; for those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and rules of the present application shall be included in the protection scope of the present application.

Claims

1. A method for early warning analysis of tennis player sports injury risk, characterized in that, The method comprises the following steps: S1: Collecting multi-modal motion data of a tennis player during training or competition, and recording corresponding motion action labels and timestamp information; S2: Preprocessing the collected raw motion data, segmenting each channel signal using a sliding window method, generating a time series sample set, and labeling the sample's corresponding injury risk level label; S3: Constructing a deep neural network architecture based on the fusion of long short-term memory networks and attention mechanisms, wherein the attention module is used to calculate the contribution weight of each time step input feature to the final risk prediction result; S4: Inputting the time series sample set into the deep neural network for end-to-end model training, and introducing an explainability constraint loss function during training, which is composed of a classification loss and an attention weight sparsity regularization term; S5: After the model training is completed, the gradient weighted class activation mapping technique is used to visualize and analyze the model decision path, and extract the key input features and their time distribution patterns that affect the prediction result; S6: Based on the key input features and the visualization analysis results, a natural language form of explanation report is generated, which includes key action stages, high-risk feature variables, and causal relationship analysis between them and injury risk; S7: The trained injury risk prediction model with explainability is deployed in the sports medical monitoring system, receives real-time motion data stream and outputs risk level prediction results and corresponding explanation information.

2. The method according to claim 1, wherein, The step S8 further comprises: S8: Continuously collecting model prediction results and actual injury event feedback data during system operation, constructing model performance evaluation logs, and periodically updating model parameters and explanation mechanisms based on new data.

3. The method of claim 1, wherein the method comprises: The step S1 specifically comprises: Based on wearable inertial measurement unit sensors and pressure distribution measurement insoles, raw motion data generated by tennis players during training or competition is obtained, and a multi-modal input data set is constructed; Time synchronization processing is performed on the raw motion data, and data from different sensor channels is aligned based on a unified timestamp to generate time-aligned multi-channel motion signals; An action capture technology based on a high-speed camera system is used to reconstruct the joint angles and motion trajectories of the tennis player in real time, generating a motion trajectory sequence in three-dimensional space and angle change curves of each main joint; After synchronization, the multi-modal motion signals and action capture data are labeled, and based on video recording and sensor data fusion analysis, the type of shot executed by the player and its start and end time interval are identified and labeled, generating a structured action label sequence; The labeled multi-modal motion data and the action label sequence are standardized in data format, stored in a unified data structure, and formed into a raw data sample set.

4. The method according to claim 1, wherein, The multi-modal motion data includes three-dimensional acceleration, angular velocity, ground reaction force, joint angle, and motion trajectory sequence.

5. The method of claim 1, wherein the method is characterized by: The step S2 specifically comprises: Performing wavelet transform-based denoising processing on the collected raw multi-modal motion data to obtain denoised multi-modal motion data; The denoised multi-modal motion data is normalized based on a min-max normalization method, and normalized multi-channel time series data is output; The normalized multi-channel time series data is segmented based on a sliding window algorithm, and a structured time window sample set is generated by extracting continuous motion cycle segments with a fixed time window length, each time window sample containing a feature vector of multiple time steps; The structured time window sample set is labeled based on motion action labels and timestamp information, and a labeled time window sample set is output; The labeled time window sample set is mapped to a damage risk level based on expert knowledge systems and historical damage data, and a structured time sequence sample set is output by converting continuous risk indicators into discrete risk level labels.

6. The method according to claim 5, wherein, In step S2, Daubechies-6 wavelet basis is used for denoising of each channel signal, and the number of decomposition layers is 1-6, and the denoising threshold rule is selected as Rigrsure or Minimax.

7. The method of claim 1, wherein the method is characterized by: The step S3 specifically includes: A multi-layer time sequence feature extraction module is designed based on the LSTM network structure, and the preprocessed time sequence sample is encoded to obtain a high-level motion feature representation; A soft attention mechanism module is embedded after the LSTM output layer, and the attention weight is calculated based on the hidden state sequence output by the LSTM, and the contribution of each time step feature to the final output is dynamically adjusted; The attention weight vector is subjected to Softmax normalization operation to obtain the normalized weight distribution of each time step input feature, and an interpretable feature importance ranking basis is formed; The LSTM hidden state is weighted and summed based on the normalized attention weight to generate a context vector as the final risk prediction input feature; The context vector is input into the fully connected classification network to generate a preliminary prediction result of the damage risk level through multi-layer nonlinear transformation.

8. The method of claim 1, wherein the method is characterized by: The step S4 specifically includes: The preprocessed time sequence sample is input into the constructed deep neural network architecture integrating LSTM and attention mechanism to extract the time sequence dependency of the input feature and generate an initial prediction output; Based on the difference between the initial prediction output and the real damage risk level label, a classification loss term is calculated, and a cross-entropy loss function is used to quantitatively evaluate the model prediction error; The attention weight distribution output by the attention module is subjected to sparsity regularization processing, and the L1 regularization strategy is used to constrain the attention weight vector; The classification loss term and the attention weight sparsity regularization term are linearly combined to construct an interpretable constraint loss function for joint optimization; Based on the interpretable constraint loss function, the model parameters in the deep neural network are iteratively optimized through the back propagation algorithm to obtain a tennis player damage risk warning model. 9.The tennis player movement injury risk early warning analysis method according to claim 1, characterized in that, The step S5 specifically includes: Forward propagation calculation is performed on the trained deep neural network model to obtain the output prediction result of the model on the validation set and the corresponding feature spectrum data; Based on the feature map data, a gradient weighted class activation mapping technique is used to perform a back propagation operation on the last layer of convolution feature maps of the model, to calculate gradient information of each time step feature map for the output prediction result, to obtain sensitivity weight coefficients of each time step feature map for the final classification result, and to form a time sensitivity heat map; The time sensitivity heat map is normalized, combined with multi-modal motion data in the original input time sequence, and a feature mapping alignment operation is performed to generate an explainability heat map matrix; Based on the explainability heat map matrix, a threshold segmentation algorithm is used to extract time steps with heat values higher than a set threshold, to identify input feature sequences that play a key role in model prediction, and to form a key feature time window set; The input features in the key feature time window set are subjected to a time distribution pattern extraction operation, a time series clustering algorithm is used to classify the key time windows, to identify high-risk action periods with similar feature distributions, and to output the feature sequences and their time distribution patterns corresponding to each cluster center.

10. The method of claim 1, wherein the method is characterized by: The step S6 further includes using a template-driven natural language generation algorithm to fill in the feature contribution triplets, action phase information and physiological knowledge base rule mapping logic to generate an intermediate reasoning result set, and using a weight priority-based sorting algorithm to sort and output the intermediate reasoning result set according to the contribution degree and duration. The step S6 further includes using a template-driven natural language generation algorithm to fill in the feature contribution triplets, action phase information and physiological knowledge base rule mapping logic to generate an intermediate reasoning result set, and using a weight priority-based sorting algorithm to sort and output the intermediate reasoning result set according to the contribution degree and duration.

Images