Muscle injury recovery prediction method based on deep learning
Through multimodal data processing and deep learning technology, secondary accuracy calibration and dynamic weight adjustment combined with external environmental parameters are solved, and the accuracy and inefficiency of muscle injury recovery prediction in the existing technology is achieved, achieving high accuracy and high efficiency prediction effects.
Patent Information
- Application Number
- CN202510432277.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-08
- Publication Date
- 2025-07-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing deep learning-based muscle injury recovery prediction methods have shortcomings in multimodal data fusion, dynamic weight adjustment and error compensation mechanism, resulting in insufficient prediction accuracy and inefficient efficiency.
By acquiring multimodal data sets, performing time synchronization, data cleaning and standardization processing, using deep neural networks for feature extraction and fusion, combining external environment parameters for secondary accuracy calibration, and using dynamic weighting and feedback loop mechanisms to generate comprehensive evaluation indicators to optimize prediction results.
It significantly improves the accuracy and efficiency of muscle injury recovery prediction, enhances the stability and adaptability of the model, and achieves a comprehensive assessment of the muscle injury recovery process.
Smart Images
Figure CN120277329A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer technology, and specifically to a method for predicting muscle injury recovery based on deep learning. Background Art
[0002] Muscle injury is relatively common among athletes, laborers, and the elderly. Its recovery process is complex and requires scientific evaluation and prediction. Currently, the prediction of muscle injury recovery mainly relies on medical image analysis and clinical experience judgment, but there are problems such as insufficient prediction accuracy and low efficiency. With the development of deep learning technology, using deep neural networks to comprehensively analyze multi-modal data is expected to improve the accuracy and efficiency of muscle injury recovery prediction. However, existing deep learning-based methods still have deficiencies in aspects such as multi-modal data fusion, dynamic weight adjustment, and error compensation mechanisms, and further research and improvement are urgently needed. Summary of the Invention
[0003] The purpose of the present invention is to provide a method for predicting muscle injury recovery based on deep learning, which has the advantages of high prediction accuracy and high evaluation efficiency, and solves the problems of insufficient prediction accuracy and low efficiency in the prior art.
[0004] To achieve the above purpose, the present invention provides the following technical solution: A method for predicting muscle injury recovery based on deep learning, including the following steps:
[0005] S1. Data collection and preprocessing: Obtain a multi-modal data set D = {D physio , D motion , D image}
[0006] D physio is physiological signal data; D motion is motion data; D image is image data;
[0007] Process the multi-modal data set, including: time synchronization, data cleaning, and data standardization, and obtain standardized input data
[0008] S2. Preliminary measurement calibration:
[0009] Extract features from the standardized multi-modal data to obtain a fused feature vector;
[0010] Generate a preliminary calibration result C1 through linear transformation;
[0011] S3. Secondary accuracy calibration:
[0012] Obtain external environment parameters E = {E env1 , E env2 ,... Eenvm};
[0013] Extract the features of the environmental parameter E to obtain the environmental feature vector: f env = Encoder env (E), combine the preliminary calibration result C1 with the environmental feature f env , and generate the accurate calibration result C2 through the linear transformation error compensation function ε;
[0014] S4. Generation of comprehensive evaluation index: Calculate the dynamic weights ω1 and ω2;
[0015] Use the non-linear function f nonlinear Combine the dynamic weights and the calibration results to generate the comprehensive evaluation index S;
[0016] Finally, through the feedback loop mechanism, part of the information of the comprehensive evaluation index S is fed back to the preliminary measurement and calibration step to adjust the preliminary calibration result C1.
[0017] Preferably, the data acquisition and preprocessing step in S1 further includes:
[0018] Collect physiological signal data D using a heart rate monitor and an electromyogram sensor physio ;
[0019] Collect motion data D using an accelerometer and a gyroscope motion ;
[0020] Obtain image data D of the muscle injury site using an MRI or ultrasonic device image .
[0021] Preferably, the time synchronization, data cleaning, and data standardization in the processing step of the multi-modal data set in S1 further include:
[0022] Time synchronization: Use the interpolation function Ι to align the timestamps of different modalities to obtain the synchronized data:
[0023] Data cleaning: Use the cleaning function X to remove noise and outliers and fill in the missing data to obtain the cleaned data:
[0024] Data standardization: Use the standardization function V to convert the data into a distribution with a mean of 0 and a variance of 1 to obtain the standardized data: where μ and σ are the mean and standard deviation of the data, respectively.
[0025] Preferably, the feature extraction in the preliminary measurement and calibration step in S2 includes:
[0026] Use the encoder Encoder physioExtract features from physiological signal data to obtain feature vectors:
[0027]
[0028] Use the Encoder motion Extract features from motion data to obtain feature vectors:
[0029]
[0030] Use the convolutional neural network CNN to extract features from image data to obtain feature vectors:
[0031]
[0032] Preferably, the linear transformation in the preliminary measurement and calibration step in S2 is implemented through the weight matrix W calib1 and the bias vector 1b calib1 to generate a preliminary calibration result C1, and its expression is:
[0033] C1 = W calib1 ·Concat(f physio , f motion , f image ) + b calib1 .
[0034] Preferably, the error compensation function ε in the secondary accuracy calibration step in S3 is a dynamic adjustment function based on the error between the preliminary calibration result C1 and the accurate calibration result C2 to minimize the overall calibration error, and its expression is: C2 = W calib2 ·[C1, f env + b calib2 + ε(C1, C2), where b calib2 is the bias vector 2.
[0035] Preferably, the dynamic weights ω1 and ω2 in the comprehensive evaluation index generation step in S4 are calculated by the following formula: ω1 = soft max(Wω1·C1 + bω1),
[0036] ω2 = soft max(Wω2·C2 + bω2), where Wω1, Wω2 are weight matrices, and bω1, bω2 are the bias vector 3 and the bias vector 4.
[0037] Preferably, the non - linear function f nonlinear in the comprehensive evaluation index generation step in S4 is a logarithmic function, and its expression is: S = log(ω1·C1 + ω2·C2 + ∈), where ∈ is a minimum value to prevent the input of the logarithmic function from being zero.
[0038] Preferably, the step of generating the comprehensive evaluation index further includes a final weight matrix W final and a bias vector 5b final , the final weight matrix W final and a bias vector 5b final are dynamically optimized through the training process, and their expression is: S = W final ·S + b final .
[0039] Preferably, the feedback loop mechanism in the step of generating the comprehensive evaluation index is realized by the following formula: C1′ = C1 + γ·S, where γ is a feedback coefficient used to adjust the influence degree of the comprehensive evaluation index S on the preliminary calibration result C1.
[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0041] 1. The present invention improves the input basis of the prediction model by obtaining a multi-modal data set and performing time synchronization, data cleaning, and standardization processing on it to ensure the quality and consistency of the input data.
[0042] 2. The deep neural network is used to extract features from the preprocessed multi-modal data, and the features of physiological signals, motion data, and image data are fused to improve the comprehensiveness and accuracy of feature representation.
[0043] 3. The secondary precision calibration step is introduced, and the preliminary calibration result is further optimized by combining external environmental parameters and through an error compensation function, reducing the prediction error and improving the stability and reliability of the model.
[0044] 4. The comprehensive evaluation index is generated by using the dynamic weight calculation and non-linear combination method, and the preliminary calibration result is dynamically adjusted through the feedback loop mechanism to achieve the adaptive optimization of the prediction result and significantly improve the overall prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 is the overall flowchart of the muscle injury recovery prediction method based on deep learning of the present invention;
[0046] Figure 2 is the data processing flowchart of the present invention;
[0047] Figure 3 is the feedback loop mechanism diagram of the present invention;
[0048] Figure 4 is the secondary precision calibration mechanism diagram of the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0049] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0050] Please refer to Figures 1 to 4 , the present invention provides a technical solution: a method for predicting muscle injury recovery based on deep learning. First, the present invention needs to obtain a multi-modal data set. The specific steps are as follows:
[0051] 1. Data collection and preprocessing
[0052] To achieve highly accurate prediction of muscle injury recovery, the present invention first needs to obtain a multi-modal data set. The specific steps are as follows:
[0053] 1.1 Obtain a multi-modal data set
[0054] The present invention collects a multi-modal data set including physiological signal data, motion data, and imaging data by using a variety of sensors and devices, denoted as: D = {D physio , D motion , D image}, where:
[0055] D physio is the physiological signal data collected by a heart rate monitor and an electromyogram (EMG) sensor;
[0056] D motion is the motion data collected by an accelerometer and a gyroscope;
[0057] D image is the imaging data of the muscle injury site obtained by an MRI or ultrasonic device.
[0058] 1.2 Process the multi-modal data set
[0059] The obtained data set needs to go through a series of preprocessing steps to ensure the quality and consistency of the data. Specifically, it includes time synchronization, data cleaning, and data standardization.
[0060] Time synchronization:
[0061] Data of different modalities may have different timestamps, so these timestamps need to be aligned. Using the interpolation function Ι, align the timestamps of data of different modalities to obtain the synchronized data:
[0062] Among them, represents the data of a certain modality at the i-th time point, ti is a timestamp. The interpolation function Ι can adopt linear interpolation, polynomial interpolation or other suitable data interpolation methods to ensure the accurate alignment of different modality data in time.
[0063] Data cleaning:
[0064] The original data may contain noise, outliers or missing values and need to be cleaned. Use the cleaning function X to remove noise and outliers and fill in the missing data to obtain the cleaned data:
[0065] where, is the cleaned data. The cleaning function X specifically includes the following steps:
[0066] Noise removal: Use filtering techniques (such as low-pass filtering, median filtering, etc.) to remove high-frequency noise in physiological signals.
[0067] Outlier detection and processing: Detect and process outliers through statistical methods (such as Z-score, IQR) to ensure the stability of the data.
[0068] Missing value filling: Adopt interpolation methods or machine learning-based prediction methods to fill in the missing values to ensure the integrity of the data.
[0069] Data standardization:
[0070] In order to eliminate the dimensional difference between different data modalities and the deviation of data distribution, use the standardization function V to convert the data into a standard distribution with a mean of 0 and a variance of 1 to obtain the standardized input data: where μ and σ are the mean and standard deviation of the data respectively, and the calculation methods are as follows: where N is the total number of data points. The standardization process ensures that different modality data are on the same scale, facilitating the training and prediction of deep learning models.
[0071] Through the above preprocessing steps, the quality and consistency of multi-modal data are ensured, providing a reliable data basis for subsequent feature extraction and model training.
[0072] 2. Preliminary measurement calibration
[0073] After completing the data preprocessing, the present invention uses a deep neural network to extract and fuse features from multi-modal data to generate a preliminary calibration result C1.
[0074] 2.1 Feature extraction
[0075] In this step, different encoders are used to extract features from each modality data, specifically including:
[0076] Physiological signal feature extraction:
[0077] Use a specially designed encoder Encoder physio To extract features from the standardized physiological signal data To obtain a feature vector: Among them, Encoder physio Can adopt deep learning models such as multi-layer perceptron (MLP), convolutional neural network (CNN) or recurrent neural network (RNN), and the specific selection is based on the characteristics of the physiological signal data and task requirements.
[0078] Motion data feature extraction: Use the encoder Encoder motion To extract features from the standardized motion data To obtain a feature vector: Encoder motion Similarly, deep learning models suitable for processing time series data, such as RNN or long short-term memory network (LSTM), can be adopted to capture the temporal features in the motion data.
[0079] Image data feature extraction: Use the convolutional neural network (CNN) CNN to extract features from the image data To obtain a feature vector: CNN effectively extracts the spatial features in the image data through multi-layer convolution and pooling operations, and captures the morphological and structural changes of the muscle injury site.
[0080] 2.2 Feature fusion and preliminary calibration
[0081] Fuse the extracted multi-modal feature vectors to obtain a comprehensive feature vector: f combined = Concat(f physio , f motion , f image ), where Concat represents the concatenation operation of the feature vectors. By concatenating the feature vectors of different modalities, multi-source information is fused to enhance the comprehensiveness and accuracy of the feature representation.
[0082] Next, generate the preliminary calibration result C1 through a linear transformation:
[0083] C1 = W calib1 · Concat(f physio , f motion , f image ) + b calib1 , where: W calib1 Is
[0084] The weight matrix for linear transformation; b calib1The pre-calibrated bias vector, numbered as bias vector 1, is to ensure its distinction in different steps.
[0085] The linear transformation maps the fused feature vector to a new space through matrix multiplication and addition of the bias vector, generating the preliminary muscle injury recovery prediction result C1. This result reflects the comprehensive influence of multi-modal features and provides a basis for subsequent secondary accuracy calibration.
[0086] Through the preliminary calibration step, the preliminary muscle injury recovery prediction result C1 is obtained, providing a basis for subsequent secondary accuracy calibration.
[0087] 3. Secondary accuracy calibration
[0088] To further improve the prediction accuracy, the present invention introduces a secondary accuracy calibration step, combines external environmental parameters, and optimizes the preliminary calibration result C1 through an error compensation function to generate the precise calibration result C2.
[0089] 3.1 Obtaining external environmental parameters
[0090] The external environmental parameters E include multiple environmental factors, such as temperature, humidity, light intensity, etc., which may affect the muscle injury recovery process. The specific acquisition method is as follows:
[0091] E = {E env1 , E env2 ,... E envm}, where E env1 , E env2 , E envm represent different environmental parameters respectively. The environmental parameters can be collected in real time through sensors or other monitoring devices to ensure the accuracy and timeliness of the data.
[0092] 3.2 Environmental feature extraction
[0093] Use the environmental parameter encoder Encoder env to extract features from the environmental parameters E, obtaining the environmental feature vector: f env = Encoder env (E), where Encoder env can adopt a deep neural network model, such as a fully connected network or a specifically designed network structure, which is specially used to extract key information from environmental parameters. The environmental feature vector f env reflects the influence of the external environment on the muscle injury recovery process and provides necessary background information for subsequent precise calibration.
[0094] 3.3 Generation of precise calibration result
[0095] Combine the preliminary calibration result C1 with the environmental feature vector fenv Combination, generating an accurate calibration result C2 through linear transformation and error compensation function ε:
[0096] C2 = W calib2 ·[C1, f env + b calib2 + ε(C1, C2), where: W calib2 is the weight matrix for linear transformation; b calib2 is the bias vector for secondary calibration, numbered as bias vector 2; ε(C1, C2) is the error compensation function, which is dynamically adjusted based on the error between the preliminary calibration result C1 and the accurate calibration result C2 to minimize the overall calibration error.
[0097] Function of the error compensation function ε
[0098] The error compensation function ε analyzes the error between the preliminary calibration result C1 and the accurate calibration result C2, and dynamically adjusts the output of the linear transformation to ensure the accuracy of the calibration result. The specific implementation method can adopt the following steps:
[0099] Error calculation:
[0100] 1. Calculate the error ΔC = C actual - C1 between the preliminary calibration result C1 and the actual recovery situation (if there is actual data as a reference);
[0101] 2. Error compensation model: Use a deep learning model (such as an error compensation network) to model the error ΔC and generate the error compensation value ε(C1, C2);
[0102] 3. Dynamic adjustment:
[0103] Add the error compensation value ε(C1, C2) to the result of the linear transformation, correct the preliminary calibration result, and generate the accurate calibration result C2.
[0104] Through the error compensation function, the secondary precision calibration step not only combines environmental parameters but also dynamically adjusts according to the actual error, further improving the accuracy of the prediction result and the stability of the model.
[0105] 4. Generation of comprehensive evaluation index
[0106] To comprehensively evaluate the muscle injury recovery situation, the present invention adopts a dynamic weight calculation and a non - linear combination method to generate a comprehensive evaluation index S, and dynamically adjusts the preliminary calibration result C1 through a feedback loop mechanism.
[0107] 4.1 Dynamic weight calculation
[0108] Calculate the dynamic weights ω1 and ω2: ω1 = soft max(Wω1·C1 + bω1),
[0109] ω2 = soft max(Wω2·C2 + bω2), where: Wω1 and Wω2 are weight matrices used to calculate dynamic weights;
[0110] bω1 and bω2 are bias vectors 3 and 4;
[0111] The soft max function is used to convert the result of the linear transformation into a probability distribution, ensuring that the sum of the weights is 1.
[0112] Function of dynamic weights
[0113] The dynamic weights ω1 and ω2 are used to measure the importance of the preliminary calibration result C1 and the precise calibration result C2 in the comprehensive evaluation. Through the soft max function, the dynamic weights can adaptively adjust the influence weights of different calibration results, and optimize the generation of the comprehensive evaluation index according to the actual data distribution and model training situation.
[0114] 4.2 Generation of comprehensive evaluation index
[0115] Use a non - linear function (logarithmic function) to combine the dynamic weights and the calibration results to generate the comprehensive evaluation index S: S = log(ω1·C1 + ω2·C2 + ∈), where ∈ is a very small value to prevent the input of the logarithmic function from being zero, usually set to 10 -8 or smaller to ensure the stability of the calculation.
[0116] Significance of the comprehensive evaluation index S
[0117] The comprehensive evaluation index S reflects the muscle injury recovery situation under the comprehensive influence of multi - modal data and environmental parameters. Through the non - linear combination of the logarithmic function, S can effectively integrate the advantages of each modal calibration result, improving the comprehensiveness and depth of the evaluation index.
[0118] 4.3 Feedback loop mechanism
[0119] Through the feedback loop mechanism, part of the information of the comprehensive evaluation index S is fed back to the preliminary measurement calibration step to adjust the preliminary calibration result C1: C1′ = C1 + γ·S, where γ is the feedback coefficient used to adjust the influence degree of the comprehensive evaluation index S on the preliminary calibration result C1, and the value of γ is optimized through the model training process to ensure that the feedback mechanism can effectively improve the accuracy of the calibration result and the adaptive ability of the model.
[0120] Advantages of the feedback loop mechanism: The feedback loop mechanism enables the model to dynamically adjust the preliminary calibration result C1 according to the comprehensive evaluation index S to achieve adaptive optimization of the prediction result. This mechanism not only improves the robustness of the model but also enhances the adaptability of the model under different environmental conditions, improving the overall prediction performance.
[0121] 4.4 Final Weight Optimization
[0122] Finally, the comprehensive evaluation index S is further optimized through the weight matrix W final and the bias vector 5b final to improve the prediction accuracy and the reliability of the evaluation index: S = W final ·S + b final , where W final is the final weight matrix for linearly transforming the comprehensive evaluation index S; 5b final is the bias vector 5, distinguished by numbers to ensure its uniqueness in different steps.
[0123] Function of Final Weight Optimization
[0124] Through the final weight optimization, the comprehensive evaluation index S can further improve the accuracy and reliability of its evaluation. The weight matrix W final and the bias vector 5b final 's training process ensures the adaptability of the comprehensive evaluation index in different data distributions and model training stages, and improves the overall prediction performance of the model.
[0125] In summary, the present invention achieves high accuracy and high efficiency in muscle injury recovery prediction through the following methods:
[0126] Multi-modal data fusion: By fusing physiological signals, motion data, and imaging data, key features in the muscle injury recovery process are comprehensively captured, enhancing the input basis of the prediction model.
[0127] Deep feature extraction: Using deep neural networks and convolutional neural networks to extract features from multi-modal data, enhancing the comprehensiveness and accuracy of feature representation.
[0128] Secondary accuracy calibration: Combining external environmental parameters and optimizing the preliminary calibration results through an error compensation function to reduce prediction errors and improve the stability and reliability of the model.
[0129] Dynamic weight and feedback mechanism: Adopting dynamic weight calculation and non-linear combination to generate a comprehensive evaluation index, and realizing the adaptive optimization of the prediction result through a feedback loop mechanism, significantly enhancing the overall prediction performance.
[0130] Error compensation mechanism: An error compensation function is introduced in the secondary accuracy calibration step, further reducing prediction errors and ensuring the stability and reliability of the model.
[0131] Comprehensive evaluation: The comprehensive evaluation index generation method realizes a comprehensive and in-depth evaluation of the muscle injury recovery situation through non-linear combination and dynamic weight calculation.
[0132] Traceability and parameter sources: The parameters in all formulas have clear sources and definitions, ensuring that the acquisition of these parameters during the calculation process is reasonable and traceable, thereby enhancing the transparency and interpretability of the model.
[0133] With these technical features and advantages, the present invention has significant advantages in improving the accuracy and efficiency of predicting muscle injury recovery, and can effectively solve the problems of insufficient prediction accuracy and low efficiency existing in the prior art.
[0134] Example 1: Prediction of Athletes' Muscle Injury Recovery
[0135] After high-intensity training, athletes often suffer from muscle injuries of varying degrees. Accurately predicting the recovery time and effect is crucial for formulating reasonable training and rehabilitation plans. Existing methods mainly rely on doctors' experience and simple medical image analysis, lacking systematicness and accuracy. This example aims to achieve precise prediction of athletes' muscle injury recovery through the method of the present invention, utilizing multi-modal data and deep learning techniques.
[0136] Implementation process:
[0137] 1. Data collection and preprocessing: Use a heart rate monitor and an electromyogram (EMG) sensor to collect the physiological signal data D of the athlete during training physio ; Use an accelerometer and a gyroscope to collect motion data D motion ; Obtain the image data D of the athlete's muscle injury site through an MRI device image ;
[0138] Perform time synchronization, data cleaning, and standardization processing on the collected multi-modal data to obtain a standardized data set
[0139] 2. Preliminary measurement calibration:
[0140] Use an Encoder physio to extract the physiological signal feature f physio ;
[0141] Use an Encoder motion to extract the motion data feature f motion ;
[0142] Use a CNN to extract the image data feature f image ;
[0143] Concatenate the extracted feature vectors to obtain a fused feature vector f combined ;
[0144] Generate a preliminary calibration result C1 through linear transformation.
[0145] 3. Secondary Precision Calibration:
[0146] Obtain the environmental parameter E of the training site;
[0147] Use Encoder env Extract the environmental feature f env ;
[0148] Combine the preliminary calibration result C1 and the environmental feature f env , and generate the precise calibration result C2 through linear transformation and error compensation function.
[0149] 4. Generation of Comprehensive Evaluation Index:
[0150] Calculate the dynamic weights ω1 and ω2;
[0151] Use the logarithmic function to generate the comprehensive evaluation index S;
[0152] Adjust the preliminary calibration result C1 through the feedback loop mechanism to generate the final evaluation index S.
[0153] To verify the advantages of the present invention over the prior art, a number of comparative experiments were conducted. The following table shows the comparison results of the present invention and the prior art in multiple performance indicators:
[0154]
[0155]
[0156] Experiment Description
[0157] Experiment Objects: 50 athletes and 50 elderly people were selected as experiment objects, and the muscle injury recovery prediction was applied using the method of the present invention and the prior art method respectively.
[0158] Data Collection: The method of the present invention collected multi-modal data (physiological signals, motion data, image data) and environmental parameters, while the prior art only collected single-modal data (such as image data).
[0159] Evaluation Index: By comparing the prediction results with the actual recovery situation, calculate the prediction accuracy and evaluation efficiency. At the same time, evaluate the adaptability and stability of the evaluation model.
[0160] Analysis of Experiment Results
[0161] Through the comparative experiment, the results show that the method of the present invention is superior to the prior art method in multiple aspects such as prediction accuracy, evaluation efficiency, data processing ability, dynamic weight adjustment, error compensation mechanism, model adaptability, generation of comprehensive evaluation index and adaptive optimization mechanism, demonstrating significant technical advantages.
[0162] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for predicting muscle injury recovery based on deep learning, characterized in that, It includes the following steps: S1. Data collection and preprocessing: Obtain a multimodal dataset D = {D physio , D motion , D image} where D physio is physiological signal data; D motion is motion data; D image is image data. Process the multi-modal data set, including: time synchronization, data cleaning, and data standardization, and obtain the standardized input data S2. Preliminary measurement calibration: Extract features from the standardized multi-modal data to obtain a fused feature vector; Generate a preliminary calibration result C1 through linear transformation; S3. Secondary accuracy calibration: Obtain external environment parameters E = {E env1 , E env2 ,... E envm}; Extract the features of environmental parameter E to obtain the environmental feature vector: f env = Encoder env (E), combine the preliminary calibration result C1 with the environmental feature f env , and generate the accurate calibration result C2 through the linear transformation error compensation function ε; S4. Generation of comprehensive evaluation indicators: Calculate dynamic weights ω1 and ω2; Use the non-linear function f nonlinear Combine the dynamic weights and the calibration results to generate a comprehensive evaluation index S; Finally, in the feedback loop mechanism, part of the information of the comprehensive evaluation indicator S is fed back to the preliminary measurement calibration step to adjust the preliminary calibration result C1.
2. The method for predicting muscle injury recovery based on deep learning according to claim 1, wherein: The data acquisition and preprocessing step in S1 further includes: Collect physiological signal data D using a heart rate monitor and an electromyogram sensor physio ; Collect motion data D using an accelerometer and a gyroscope motion ; Obtain image data D of the muscle injury site using an MRI or ultrasound device image .
3. The muscle injury recovery prediction method based on deep learning according to claim 1, wherein: The time synchronization, data cleaning, and data standardization in the processing step of the multi-modal data set in S1 further include: Time synchronization: Use the interpolation function Ι to align the timestamps of data in different modalities and obtain the synchronized data: Data cleaning: Use cleaning function X to remove noise and outliers, fill in missing data, and obtain the cleaned data: Data standardization: The data is converted into a distribution with a mean of 0 and a variance of 1 using the standardization function V to obtain standardized data: where μ and σ are the mean and standard deviation of the data, respectively.
4. The method for predicting muscle injury recovery based on deep learning according to claim 1, characterized in that: The feature extraction in the preliminary measurement calibration step in S2 includes: Use an Encoder physio Extract features from the physiological signal data to obtain feature vectors: Use an Encoder motion Extract features from the motion data to obtain feature vectors: Use a convolutional neural network CNN to extract features from the image data to obtain a feature vector:
5. The method for predicting muscle injury recovery based on deep learning according to claim 1, wherein: The linear transformation in the preliminary measurement and calibration step in S2 is implemented through the weight matrix W calib1 and the bias vector 1b calib1 to generate a preliminary calibration result C1, and its expression is: C1 = W calib1 ·Concat(f physio , f motion , f image ) + b calib1 。 6. The method for predicting muscle injury recovery based on deep learning according to claim 1, wherein: The error compensation function ε in the secondary precision calibration step in S3 is a dynamic adjustment function based on the error between the preliminary calibration result C1 and the precise calibration result C2 to minimize the overall calibration error, and its expression is: C2 = W calib2 ·[C1, f env + b calib2 + ε(C1, C2), where b calib2 is the bias vector 2.
7. The method for predicting muscle injury recovery based on deep learning according to claim 1, wherein: The dynamic weights ω1 and ω2 in the comprehensive evaluation indicator generation step in S4 are calculated by the following formula: ω1 = soft max(Wω1·C1 + bω1), ω2 = soft max(Wω2·C2 + bω2), where Wω1, Wω2 are weight matrices, and bω1, bω2 are bias vectors 3 and bias vectors 4.
8. The method for predicting muscle injury recovery based on deep learning according to claim 1, wherein: The non-linear function f in the comprehensive evaluation index generation step in S4 nonlinear is a logarithmic function, and its expression is: S = log(ω1·C1 + ω2·C2 + ∈), where ∈ is a minimum value to prevent the input of the logarithmic function from being zero.
9. The muscle injury recovery prediction method based on deep learning according to claim 1, characterized in that: The step of generating the comprehensive evaluation index further includes a final weight matrix W final and a bias vector 5b final , the final weight matrix W final and the bias vector 5b final are dynamically optimized through the training process, and its expression is: S = W final ·S + b final .
10. The method for predicting muscle injury recovery based on deep learning according to claim 1, characterized in that: The feedback loop mechanism in the comprehensive evaluation indicator generation step is implemented by the following formula: C1′ = C1 + γ·S, where γ is a feedback coefficient used to adjust the influence degree of the comprehensive evaluation indicator S on the preliminary calibration result C1.