Rehabilitation training data processing method and device

Abstract

The invention discloses a rehabilitation training data processing method and device, and relates to the technical field of medical information data processing, and the rehabilitation training data processing method comprises the following steps: 1, data acquisition and preprocessing: collecting and preprocessing physiological, log and video data of a to-be-rehabilitated user, and integrating the physiological, log and video data into a multi-source data set; 2, high-performance model construction and knowledge distillation: selecting and training a teacher network model, constructing a student network model, and obtaining a model with relatively good performance and high calculation efficiency through knowledge distillation; step 3, differential evolution algorithm optimization: defining differential evolution algorithm parameters, and obtaining an optimized student network model through iterative optimization; and step 4, deep reinforcement learning parameter adjustment: utilizing the optimized student network model to generate feature data, and adjusting rehabilitation training parameters through reinforcement learning. The data processing efficiency and accuracy are remarkably improved, and the complexity is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of medical information data processing, and in particular to a rehabilitation training data processing method and device. Background Art

[0002] In the field of rehabilitation training data processing, collecting and analyzing multi-type data from users undergoing rehabilitation has become an important means to improve rehabilitation effects. However, existing rehabilitation training data processing methods face a series of challenges when processing these complex and diverse data.

[0003] Specifically, current data processing methods perform poorly in terms of processing efficiency and are unable to quickly respond to the processing needs of large-scale data. At the same time, the models used are often highly complex, which not only increases the technical difficulty of data processing, but also leads to excessive consumption of computing resources, affecting the overall performance and cost-effectiveness of the system. More importantly, while ensuring the accuracy of data processing, how to meet personalized needs has become an outstanding problem. Especially when faced with rehabilitation training data that contains multiple data types and has complex relationships with each other, how to effectively integrate, efficiently process and deeply analyze them in order to accurately evaluate the rehabilitation effect and optimize the rehabilitation training plan accordingly is a key issue that needs to be urgently addressed in this field. Summary of the invention

[0004] In view of this, the present invention proposes a rehabilitation training data processing method and device to solve the problems of the prior art.

[0005] The specific technical solution of the present invention is as follows: A method for processing rehabilitation training data comprises the following steps: Step 1: Data acquisition and preprocessing: collecting and preprocessing the physiological, log and video data of the users to be rehabilitated and integrating them into a multi-source data set; Step 2: High-performance model construction and knowledge distillation: Select one of the existing high-performance deep learning models as the teacher network model, train the teacher network model, design the student network model based on the rehabilitation training data, and use knowledge distillation to enable the student network model to effectively learn the knowledge of the teacher network model and adapt to the characteristics of the rehabilitation training data; Step 3: Differential evolution algorithm optimization, defining the parameters of the differential evolution algorithm, and obtaining the optimized student network model through iterative optimization; Step 4: Deep reinforcement learning parameter adjustment: use the optimized student network model to generate feature data and adjust the rehabilitation training parameters through reinforcement learning.

[0006] Specifically, in step 1, data acquisition and preprocessing, various types of data are collected from the user to be rehabilitated, including physiological data, rehabilitation training log data and rehabilitation training video data; the collected physiological data, including heart rate, blood pressure, blood oxygen saturation, respiratory rate, body temperature and vital capacity data, are integrated and calibrated; the rehabilitation training log data are logically checked and rationally analyzed, and the information on the time, intensity, training content, rest time, number of interruptions and reasons for interruptions of the rehabilitation training is recorded; image quality assessment, repair and parameter adjustment are performed on the rehabilitation training video data to improve the image quality; the physiological data, rehabilitation training log data and rehabilitation training video data are integrated into the original training data set, and noise and outlier removal, data integration, standardization processing and label data determination are performed; the preprocessed physiological data, rehabilitation training log data and rehabilitation training video data are divided into a multi-source data set, including a physiological parameter set, a training log set and a video frame image set.

[0007] Specifically, in step 2, high-performance model construction and knowledge distillation, when selecting the teacher network model, the structure and hyperparameters of the teacher network model are adjusted according to the characteristics of the rehabilitation training data and the task requirements; when training the teacher network model, the preprocessed multi-source data set is used to train the teacher network model, and the loss function is minimized through the back propagation algorithm; when constructing the student network model, its number of layers and neuron configuration are designed to adapt to the computing resource limitations in practical applications; soft labels are generated using the trained teacher network model, and are used together with the original label data to train the student network model, and through knowledge distillation, the student network model effectively learns the knowledge of the teacher network model and adapts to the characteristics of the rehabilitation training data.

[0008] Specifically, in step 3, the differential evolution algorithm optimization, the parameters of the differential evolution algorithm are defined, including population size, mutation factor, mutation probability and crossover probability; the population is initialized, in which each individual represents a set of possible parameter combinations of the student network model; the fitness of each individual in the population is evaluated, and the evaluation criteria are the performance indicators of the student network model on the validation set; mutation, crossover and selection operations are performed according to the evaluation results, and it is continuously iterated until the preset number of iterations is reached to obtain the optimized student network model.

[0009] Specifically, in step 4, deep reinforcement learning parameter adjustment, the optimized student network model is used to generate feature data as the state input of reinforcement learning; a reward function based on the rehabilitation effect score, the patient's self-perception score and the change in blood oxygen saturation is designed; the reinforcement learning agent is initialized, a reinforcement learning algorithm is selected, and the policy network model and the value network model are initialized based on the state input and the reward function; during the rehabilitation training process, data on the state, action, reward and next state are collected, and the parameters of the policy network model and the value network model are updated using the back propagation algorithm; the reinforcement learning agent is trained iteratively for multiple times so that it can generate actions to adjust the rehabilitation training parameters according to the current rehabilitation training state.

[0010] Specifically, assuming the original heart rate data is , calibrated heart rate data , where a and b are calibration coefficients; assuming the original blood pressure data is , systolic blood pressure is , diastolic blood pressure is , calibrated blood pressure data , where c, d, e, and f are calibration coefficients; assuming that the physiological data is , the mean is , the standard deviation is , normalized physiological data ; For physiological data ,if , then x is considered as an outlier and removed.

[0011] Specifically, in step 2, the predicted output of the teacher network model is , the true label is , loss function , where n is the number of samples; let the predicted output of the student network model be , the soft label is , the original label is , distillation loss ;,in is the weight for weighing the loss of soft labels and original labels.

[0012] Specifically, in step 3, let the number of correct predictions of the student network model on the validation set be TP, the total number of samples in the validation set be N, and the individual fitness be ; Let the individuals in the population be , the variation factor is F, and the randomly selected individuals are , , , the mutated individual ; Let the crossover probability be , after crossover , where j represents the dimension of the individual, represents the individual after mutation in the j dimension, Represents an individual in a population of j dimensions.

[0013] Specifically, in step 4, the rehabilitation effect score is set to The patient's self-perception score was , blood oxygen saturation changes to , the reward function ,in , , is the weight coefficient; let the parameters of the strategy network model be , the strategy is , where a is the action, s is the state, and the advantage function is , update formula ,in is the learning rate.

[0014] The present application also proposes a rehabilitation training data processing device, the device comprising: The data acquisition and preprocessing module is used to collect and preprocess the physiological, log and video data of the users to be rehabilitated, and integrate these data into a multi-source data set; High-performance model building and knowledge distillation module, which is used to select and train the teacher network model, build the student network model, and use knowledge distillation technology to enable the student network model to effectively learn the knowledge of the teacher network model and adapt to the characteristics of rehabilitation training data; The differential evolution algorithm optimization module is used to define the parameters of the differential evolution algorithm and obtain the optimized student network model through an iterative optimization process; Deep reinforcement learning parameter adjustment module, which is used to generate feature data using the optimized student network model and adjust rehabilitation training parameters through reinforcement learning technology The beneficial effects of the present invention are: By collecting various types of data (including physiological data, rehabilitation training log data, and rehabilitation training video data) from users to be rehabilitated, and performing data integration, calibration, preprocessing, and standardization, the accuracy and consistency of the data are ensured, providing a high-quality data foundation for subsequent analysis and model training.

[0015] Dividing the preprocessed data into multi-source data sets helps to conduct more targeted data analysis and model building for different physiological systems and rehabilitation training stages. By building high-performance teacher network models and student network models and combining knowledge distillation technology, efficient model training and knowledge inheritance are achieved, computational complexity is reduced, and the practical application ability of the model is improved.

[0016] The differential evolution algorithm is used to optimize the student network model, and the performance of the model and its ability to fit the rehabilitation training data are continuously improved through iterative mutation, crossover and selection operations. This process enables the model to more accurately reflect the actual effect of rehabilitation training and provides strong support for the formulation of personalized rehabilitation training programs.

[0017] By designing reasonable state input, reward function and reinforcement learning agent, dynamic adjustment and optimization of rehabilitation training parameters are achieved. This innovative method not only improves the effect and safety of rehabilitation training, but also enhances the subjective experience of patients, making rehabilitation training more humane and intelligent.

[0018] The rehabilitation training data processing method and device proposed in the present invention can be extended to various rehabilitation fields, such as respiratory rehabilitation, cardiovascular rehabilitation, neurological rehabilitation, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0020] Figure 1 It is a flowchart of the rehabilitation training data processing method of the present invention; Figure 2 It is a structural schematic diagram of the rehabilitation training data processing device of the present invention. DETAILED DESCRIPTION

[0021] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0022] The present invention proposes a rehabilitation training data processing method and device. The rehabilitation training data processing device includes: a data acquisition and preprocessing module, which is used to collect and preprocess the physiological, log and video data of the user to be rehabilitated, and integrate these data into a multi-source data set; a high-performance model construction and knowledge distillation module, which is used to select and train the teacher network model, build the student network model, and obtain a model with good performance and efficient calculation through knowledge distillation technology; a differential evolution algorithm optimization module, which is used to define the parameters of the differential evolution algorithm, and obtain the optimized student network model through an iterative optimization process; a deep reinforcement learning parameter adjustment module, which is used to generate feature data using the optimized student network model, and adjust the rehabilitation training parameters through reinforcement learning technology. The rehabilitation training data processing method includes the following steps: Step 1, data acquisition and preprocessing, includes the following steps: Step 1.1, collect various types of data from the user to be rehabilitated, including: Physiological data: A variety of devices are used to collect physiological data, such as wearable health monitoring devices (smart bracelets, smart chest straps, etc.), medical-grade physiological monitoring instruments (multi-parameter monitors, etc.). The accuracy and collection frequency of these devices vary, and data integration and calibration are required to ensure the accuracy and consistency of the data. The types of physiological data collected include common physiological data that reflect body functions, such as heart rate, blood pressure, and blood oxygen saturation, as well as data such as respiratory rate, body temperature, and vital capacity. Respiratory rate is particularly important for rehabilitation personnel, as it can reflect the improvement of respiratory function; body temperature data can help determine whether there are abnormal conditions such as infection; vital capacity data directly reflects the state of lung function and is an important basis for evaluating the effectiveness of rehabilitation training.

[0023] Rehabilitation training log data: records basic information such as the time, intensity, and content of rehabilitation training. In addition, records of the training environment (such as indoor temperature and humidity) can be added. The training environment may affect the effect of rehabilitation training. For example, an environment with high humidity may affect the comfort of breathing, thereby affecting the effect of rehabilitation training. Specifically, record information such as rest time, number of interruptions, and reasons for interruptions during rehabilitation training. This information helps analyze the continuity and stability of rehabilitation training, and is of great significance for adjusting rehabilitation training plans.

[0024] Rehabilitation training video data: Multi-view video acquisition technology is used to record the rehabilitation training process from different angles. This allows for a more comprehensive observation of the movements and postures of rehabilitation personnel, avoids the loss of movement information caused by a single viewpoint, and improves the accuracy of movement and posture analysis. The collected rehabilitation training video data includes movement and posture video data during rehabilitation training, as well as panoramic videos of the training venue to obtain more environmental information. For example, the placement of rehabilitation training equipment, surrounding auxiliary facilities, etc. This information can be correlated with the rehabilitation training effect for analysis to determine the impact of environmental factors on rehabilitation training.

[0025] The above collected data are integrated into the original training data set, which serves as the basis of the entire rehabilitation training data processing flow and provides data support for subsequent operations.

[0026] Step 1.2, data preprocessing: Noise and outlier removal: For physiological data, statistical methods (such as the 3σ principle) or methods based on data distribution are used to identify and remove outliers caused by measurement errors or equipment failures, such as extremely high or low heart rate values. Filtering techniques can also be used. For example, for heart rate data, a low-pass filter can be used to remove high-frequency noise, which may be caused by electromagnetic interference from the equipment or slight human movement.

[0027] When processing abnormal values ​​in rehabilitation training log data, check for obviously illogical records (such as negative training time) and conduct rational analysis on data such as training intensity. For example, a reasonable range of training intensity is set based on the physical condition and rehabilitation stage of the rehabilitation personnel, and data outside this range is considered as abnormal values ​​and corrected or deleted.

[0028] For image blur and error frame processing in rehabilitation training video data, image quality assessment algorithm is used for screening, and image restoration technology is used to repair partially blurred or damaged frames. For image quality problems caused by light changes and other reasons, image contrast, brightness and other parameters are adjusted to improve image quality.

[0029] Data integration: Integrate data from different sensor devices and rehabilitation training recording systems, and associate them according to key information such as user ID or time to ensure that all types of data for each rehabilitation user can correspond. The integrated data can more comprehensively reflect the status of rehabilitation users and provide a complete data basis for subsequent processing; establish a data indexing mechanism to more quickly query and obtain various types of data for specific rehabilitation users through data indexing, thereby improving data processing efficiency. Perform a data integrity check on the integrated data to ensure that all types of data for each rehabilitation user are complete. If data is found to be missing, it can be estimated and supplemented based on other relevant data, or marked as missing data for special treatment in subsequent processing.

[0030] Standardization: For different types of data, use methods such as mean-variance standardization or minimum-maximum standardization to convert all data to the same scale. Specifically, make adaptive choices based on the characteristics and distribution of the data. For example, for physiological data, if the data distribution is close to a normal distribution, mean-variance standardization can be preferred; if the data distribution is relatively uniform, minimum-maximum standardization may be more appropriate. Standardized data can be processed more effectively during model training, avoiding model training bias caused by differences in data scales.

[0031] Label data determination: Determine label data according to the rehabilitation training goals, such as quantitative indicators for evaluating rehabilitation effects (rehabilitation degree grading), or add safety indicators for rehabilitation training as label data. For example, whether dangerous situations (such as excessive fatigue, dyspnea, etc.) have occurred during rehabilitation training, and quantify them as safety labels. Label data is used for supervised learning of subsequent models to guide the model to learn the relationship between data and rehabilitation effects, and then learn how to avoid dangerous situations and improve the safety of rehabilitation training. Different label data are determined according to different stages of rehabilitation training (such as early, middle, and late stages). For example, in the early stages of rehabilitation training, label data can focus on the basic improvement of physical function; in the middle stage, attention can be paid to the progress rate of rehabilitation training; in the late stage, the focus can be on evaluating the stability and durability of rehabilitation.

[0032] Step 1.3, data division: divide the preprocessed data into multi-source data sets: divide the data into physiological parameter sets, including various physiological index data, which can be further subdivided into cardiovascular physiological parameter subsets (including heart rate, blood pressure, etc.), respiratory physiological parameter subsets (including respiratory rate, blood oxygen saturation, etc.) and other physiological parameter subsets (such as body temperature, vital capacity, etc.). Such subdivision helps to conduct more targeted data analysis and model construction for rehabilitation training of different physiological systems; divide the data into training log sets, including record data related to the rehabilitation training process, and rehabilitation training plan data. The rehabilitation training plan data includes information such as the goal setting of rehabilitation training, the division of training stages, and the training focus of each stage; extract frame image data from the rehabilitation training video data, and classify the extracted frame images to form a video frame image set. For example, according to the type of rehabilitation training action, the frame images of the same type of action are classified into one category, which helps to analyze the characteristics of different actions more targetedly in the subsequent model processing and improve the model's recognition and analysis capabilities for rehabilitation training action postures.

[0033] In step 1, the original heart rate data is , calibrated heart rate data , where a and b are calibration coefficients; assuming the original blood pressure data is , systolic blood pressure is , diastolic blood pressure is , calibrated blood pressure data , where c, d, e, and f are calibration coefficients; assuming that the physiological data is , the mean is , the standard deviation is , normalized physiological data ; For physiological data ,if , then x is considered as an outlier and removed.

[0034] Step 2, high-performance model construction and knowledge distillation, includes the following steps: Step 2.1, teacher network model construction and training: Select the teacher network model: Select one from the existing high-performance deep learning models as the teacher network model. The specific selection depends on the characteristics of the rehabilitation training data and the task requirements. For example, if processing video frame image data, the convolutional neural network model (CNN) can be selected. CNN has a strong ability to extract image features and can automatically learn features such as postures in rehabilitation training video frame images; if processing sequence information in rehabilitation training log data, the recurrent neural network model (RNN) can be selected, especially the long short-term memory network model (LSTM) or the gated recurrent unit (GRU), which can effectively process long-term dependencies in time series data. Adjust the structure of the teacher network model according to the characteristics of the data, such as increasing or decreasing the number of network model layers, adjusting the convolution kernel size, etc. For the convolutional neural network model, if the rehabilitation training video frame image data is more complex, the number of network model layers can be appropriately increased to improve the feature extraction capability, and the convolution kernel size can be adjusted to adapt to image features of different scales. For the recurrent neural network model, adjust the number of hidden layer neurons according to the sequence length and complexity of the rehabilitation training log data. Set appropriate hyperparameters, including but not limited to learning rate and batch size. The selection of these parameters needs to be based on the characteristics of the preprocessed data. For example, if the amount of data is large, the batch size can be appropriately increased to improve training efficiency; if the data distribution is more complex, a smaller learning rate can be selected to ensure stable model convergence.

[0035] Teacher network model training: Use the back propagation algorithm to input the preprocessed multi-source data set into the teacher network model for training. According to the label data in the training data, calculate the loss function (such as the cross entropy loss function for classification tasks and the mean square error loss function for regression tasks). For the evaluation of the rehabilitation training effect, if the rehabilitation degree is divided into different levels (such as mild, moderate, and severe rehabilitation), the cross entropy loss function is used; if the specific rehabilitation index value is predicted (such as the time required for the heart rate to return to the normal range), the mean square error loss function is used. The parameters of the teacher network model are continuously adjusted through back propagation to minimize the loss function. Regularly evaluate the performance of the teacher network model, and use the validation set data to calculate indicators such as accuracy and recall to determine whether there is overfitting or underfitting. When evaluating the accuracy, for classification tasks, calculate the proportion of the number of samples predicted correctly to the total number of samples; for recall, calculate the proportion of samples predicted as positive examples that are actually positive examples. Adjust the training strategy according to the evaluation results, such as adjusting the learning rate, increasing or decreasing the number of training rounds, etc. If overfitting is found, you can reduce the learning rate or increase the regularization term; if underfitting exists, you can increase the learning rate or increase the number of training rounds.

[0036] Step 2.2, student network model construction and training: Student network model construction: Design the number of layers and neuron configuration of the student network model based on factors such as the complexity of the rehabilitation training data and computing resources. If the rehabilitation training data contains multiple types and complex relationships, the number of layers of the student network model can be appropriately increased, but the limitations of computing resources must be taken into account. For example, when computing resources are limited, a shallower network model structure can be used, such as a 3-5-layer fully connected neural network model. For the neuron configuration, adjustments are made based on the dimension and number of features of the input data to ensure that the network model can fully learn the data features. The structure of the student network model may be simpler than that of the teacher network model to meet the computational efficiency requirements in practical applications. The construction of the student network model needs to take into account the actual application scenarios, inheriting the knowledge of the teacher network model while adapting to specific computing resource limitations.

[0037] Soft label generation and student network model training: The trained teacher network model is used to infer the training data and generate soft labels. The soft labels contain the probabilistic prediction results of the teacher network model on the data. For example, in the task of classifying the effect of rehabilitation training, the teacher network model may predict that the probability of a rehabilitation training sample belonging to mild rehabilitation is 0.6, the probability of moderate rehabilitation is 0.3, and the probability of severe rehabilitation is 0.1. These probability values ​​constitute the soft labels. The soft labels are used together with the original label data to train the student network model, and an adaptive confidence adjustment factor is introduced. In the early stage of training, the original label data can be given a higher weight. As the training progresses, the weight of the soft label is gradually increased according to the training situation and performance of the student network model. Through the knowledge distillation method, the student network model is trained on the basis of the teacher network model to obtain a model with good performance and efficient calculation. During the training process, the performance indicators of the student network model on the validation set, such as accuracy, mean square error, etc., are monitored, and the confidence adjustment factor is adjusted according to these indicators to ensure that the student network model can effectively learn the knowledge of the teacher network model and adapt to the characteristics of the rehabilitation training data.

[0038] In step 2, let the predicted output of the teacher network model be , the true label is , loss function , where n is the number of samples; let the predicted output of the student network model be , the soft label is , the original label is , the value of the original label is the same as the value of the true label, and the distillation loss ,in is the weight for weighing the loss of soft labels and original labels.

[0039] Step 3, differential evolution algorithm optimization, includes the following steps: Step 3.1, algorithm parameter definition and population initialization: First, define the parameters of the differential evolution algorithm. Parameter types include: population size, which determines the number of individuals participating in the evolution. This number is determined based on the complexity of the student network model and the optimization goal; mutation factor, which is used to control the intensity of the mutation operation. The choice of its value directly affects the exploration ability of the algorithm; mutation probability, which determines the possibility of individuals performing mutation operations. By adjusting this probability, the local search and global search capabilities of the algorithm can be balanced; crossover probability, which determines the possibility of individuals performing crossover operations. The setting of this probability affects the diversity of individuals.

[0040] Next, the population is initialized according to the defined parameters. Each individual in the population represents a set of possible parameter combinations of the student network model. The initialization of the population serves as the starting point for the optimization of the differential evolution algorithm and provides an initial parameter space for subsequent operations.

[0041] Step 3.2, individual assessment and action: After the population is initialized, individual evaluation is performed. Specifically, the fitness of each individual in the population is evaluated, and the evaluation criteria are the performance indicators of the student network model on the validation set, such as accuracy, mean square error, etc. These performance indicators reflect the ability of the parameter combination represented by each individual to fit the rehabilitation training data.

[0042] After the fitness evaluation is completed, the subsequent mutation, crossover and selection operations are performed according to the evaluation results. The mutation operation randomly changes the parameters of individuals in the population to generate new candidate solutions. This operation can increase the diversity of the population and improve the exploration ability of the algorithm. The crossover operation exchanges some parameters of the mutated individuals with other individuals to generate new individuals. The crossover operation helps to combine the advantages of different individuals and form a better parameter combination. The selection operation retains individuals with higher fitness according to the fitness evaluation results. This operation ensures that the population continues to develop in a better direction during the evolution process.

[0043] Repeat the above mutation, crossover and selection operations, and select according to the fitness evaluation results after each operation. This process is iterated continuously until the preset number of iterations is reached.

[0044] Step 3.3, obtain the optimized student network model: After multiple iterations, the optimized student network model was finally obtained. This model has better performance and the ability to fit the rehabilitation training data. Through the optimization of the differential evolution algorithm, the parameters of the student network model were effectively adjusted, making it more suitable for processing rehabilitation training data.

[0045] In step 3, let the number of correct predictions of the student network model on the validation set be TP, the total number of samples in the validation set be N, and the individual fitness be ; Let the individuals in the population be , the variation factor is F, and the randomly selected individuals are , , , the mutated individual ; Let the crossover probability be , after crossover , where j represents the dimension of the individual, represents the individual after mutation in the j dimension, Represents an individual in a population of j dimensions.

[0046] Step 4, deep reinforcement learning parameter adjustment, includes the following steps: Step 4.1, state input determination: The student network model optimized by the differential evolution algorithm is used to generate feature data. These feature data integrate various information in the rehabilitation training data, including but not limited to physiological parameter features (such as heart rate, blood pressure, etc.), training log features (such as training time, number of training sessions, etc.) and video image features (such as patient's movement posture, expression changes, etc.).

[0047] The feature data generated above is used as the state input of reinforcement learning. Since the student network model has been optimized by the differential evolution algorithm and has a better fitting ability for rehabilitation training data, the feature data generated by it can more comprehensively and accurately reflect the current state of rehabilitation training and provide more reliable state input information for reinforcement learning.

[0048] Step 4.2, reward function design: Design a reward function based on the rehabilitation effect score. This reward function includes the change in the rehabilitation effect score, which is used to measure the improvement in the degree of rehabilitation.

[0049] The change in the patient's self-perception score is added to the reward function to reflect the improvement in the patient's subjective feeling.

[0050] Consider the weight of changes in blood oxygen saturation, because blood oxygen saturation is an important physiological indicator reflecting the state of rehabilitation. By setting reasonable weights, the reward function is closely related to the goals and key indicators of rehabilitation training. For example, if blood oxygen saturation is significantly improved during rehabilitation training, and the improvement reaches a certain threshold, then a larger reward weight is given in the reward function, because this may mean that the patient's physical function has been positively improved during rehabilitation training.

[0051] Through the above design, it is ensured that the reinforcement learning agent can obtain corresponding reward feedback according to the actual effect of rehabilitation training, thereby guiding it to learn in a direction that is conducive to improving the rehabilitation effect.

[0052] Step 4.3, reinforcement learning agent initialization and training: Select a suitable reinforcement learning algorithm (such as the deep Q network model DQN, the proximal policy optimization algorithm PPO, etc.), and initialize the policy network model and the value network model based on the previously determined state input and the designed reward function. The policy network model is used to generate actions, that is, to adjust the decision of rehabilitation training parameters; the value network model is used to evaluate the state value.

[0053] Run the reinforcement learning agent to collect data on state, action, reward, and next state during rehabilitation training. Use the back propagation algorithm to update the parameters of the policy network model and the value network model based on the collected data, so that the reinforcement learning agent can learn the best action strategy.

[0054] The reinforcement learning agent is trained multiple times. As the training progresses, the reinforcement learning agent can generate more accurate actions to adjust the rehabilitation training parameters according to the current rehabilitation training status. For example, for patients with faster rehabilitation progress, the reinforcement learning agent may generate actions such as increasing the training intensity or changing the training method to achieve a personalized rehabilitation training plan.

[0055] In step 4, the rehabilitation effect score is set to The patient's self-perception score was , blood oxygen saturation changes to , the reward function ,in , , is the weight coefficient; let the parameters of the strategy network model be , the strategy is , where a is the action, s is the state, and the advantage function is , update formula ,in is the learning rate.

[0056] The beneficial effects of the present invention are: By comprehensively collecting information on rehabilitation users from multiple dimensions (physiological data, rehabilitation training log data, rehabilitation training video data) and performing sophisticated preprocessing (including noise and outlier removal, data integration, standardization, and label data determination), the accuracy and consistency of the data are ensured, providing a solid foundation for subsequent analysis and model training.

[0057] A high-performance deep learning model (teacher network model) is used for preliminary training, and a student network model with higher computational efficiency is constructed through knowledge distillation technology, which not only retains the powerful learning ability of the teacher network model, but also adapts to the computing resource limitations in practical applications. At the same time, for different types of rehabilitation training data (such as video frame images, time series data, etc.), the most appropriate network model structure and hyperparameters are selected to improve the model's pertinence and training efficiency.

[0058] The differential evolution algorithm is used to optimize the parameters of the student network model and automatically adjust the model's parameter combination to better fit the rehabilitation training data, thereby improving the model's performance and generalization ability. This process does not require manual intervention, thus achieving intelligent and automated model optimization.

[0059] Combined with deep reinforcement learning technology, according to the current state of rehabilitation training (including physiological parameters, training logs and video image features, etc.) and reward function (based on rehabilitation effect score, patient self-perception score and blood oxygen saturation change, etc.), the rehabilitation training parameters are intelligently adjusted to provide patients with personalized rehabilitation training programs. This not only improves the pertinence and effectiveness of rehabilitation training, but also enhances the patient's participation and comfort.

[0060] By real-time monitoring and analyzing data changes during rehabilitation training, problems in rehabilitation training, such as insufficient or excessive training intensity, can be discovered and corrected in a timely manner. At the same time, using label data and model prediction results to quantitatively evaluate the effects of rehabilitation training provides rehabilitation personnel with an objective and accurate basis for evaluation, which helps to timely adjust and optimize rehabilitation training plans.

[0061] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for processing rehabilitation training data, characterized in that: The following steps are involved: Step 1, data acquisition and preprocessing, collect and preprocess the physiological, log and video data of the users to be rehabilitated and integrate them into a multi-source data set; Step 2: High-performance model construction and knowledge distillation: Select one of the existing high-performance deep learning models as the teacher network model, train the teacher network model, design the student network model based on the rehabilitation training data, and use knowledge distillation to enable the student network model to effectively learn the knowledge of the teacher network model and adapt to the characteristics of the rehabilitation training data; Step 3: Differential evolution algorithm optimization, defining the parameters of the differential evolution algorithm, and obtaining the optimized student network model through iterative optimization; Step 4: Deep reinforcement learning parameter adjustment: use the optimized student network model to generate feature data and adjust the rehabilitation training parameters through reinforcement learning.

2. The rehabilitation training data processing method according to claim 1, characterized in that: In the step 1 of data acquisition and preprocessing, various types of data are collected from the user to be rehabilitated, including physiological data, rehabilitation training log data, and rehabilitation training video data; Integrate and calibrate the collected physiological data, including heart rate, blood pressure, blood oxygen saturation, respiratory rate, body temperature and vital capacity; Conduct logic checks and rationality analysis on rehabilitation training log data, and record information on rehabilitation training time, intensity, training content, rest time, number of interruptions, and reasons for interruptions; Perform image quality assessment, restoration and parameter adjustment on rehabilitation training video data to improve image quality; The physiological data, rehabilitation training log data and rehabilitation training video data are integrated into an original training data set, and noise and outlier removal, data integration, standardization processing and label data determination are performed; the preprocessed physiological data, rehabilitation training log data and rehabilitation training video data are divided into a multi-source data set, including a physiological parameter set, a training log set and a video frame image set.

3. The rehabilitation training data processing method according to claim 1, characterized in that: In the step 2 of high-performance model construction and knowledge distillation, when selecting the teacher network model, the structure and hyperparameters of the teacher network model are adjusted according to the characteristics of the rehabilitation training data and the task requirements; when training the teacher network model, the preprocessed multi-source data set is used to train the teacher network model, and the loss function is minimized by the back propagation algorithm; when constructing the student network model, its number of layers and neuron configuration are designed to adapt to the computing resource limitations in practical applications; soft labels are generated using the trained teacher network model, and are used together with the original label data to train the student network model, and through knowledge distillation, the student network model effectively learns the knowledge of the teacher network model and adapts to the characteristics of the rehabilitation training data.

4. The rehabilitation training data processing method according to claim 1, characterized in that: In the step 3 of differential evolution algorithm optimization, the parameters of the differential evolution algorithm are defined, including population size, mutation factor, mutation probability and crossover probability; the population is initialized, in which each individual represents a set of possible parameter combinations of the student network model; the fitness of each individual in the population is evaluated, and the evaluation criteria are the performance indicators of the student network model on the validation set; mutation, crossover and selection operations are performed according to the evaluation results, and it is continuously iterated until a preset number of iterations is reached to obtain the optimized student network model.

5. The rehabilitation training data processing method according to claim 1, characterized in that: In the step 4 of deep reinforcement learning parameter adjustment, the optimized student network model is used to generate feature data as the state input of reinforcement learning; a reward function based on the rehabilitation effect score, the patient's self-perception score and the change of blood oxygen saturation is designed; the reinforcement learning agent is initialized, a reinforcement learning algorithm is selected, and the policy network model and the value network model are initialized based on the state input and the reward function; during the rehabilitation training process, data on the state, action, reward and next state are collected, and the parameters of the policy network model and the value network model are updated using the back propagation algorithm; The reinforcement learning agent is trained for multiple iterations so that it can generate actions to adjust rehabilitation training parameters based on the current rehabilitation training state.

6. The rehabilitation training data processing method according to claim 2, characterized in that: In step 1, the original heart rate data is assumed to be , calibrated heart rate data , where a and b are calibration coefficients; assuming the original blood pressure data is , systolic blood pressure is , diastolic pressure is , calibrated blood pressure data , where c, d, e, and f are calibration coefficients; assuming that the physiological data is , the mean is , the standard deviation is , normalized physiological data ; For physiological data ,if , then x is considered as an outlier and removed.

7. The rehabilitation training data processing method according to claim 3, characterized in that: In step 2, the predicted output of the teacher network model is assumed to be , the true label is , loss function , where n is the number of samples; let the predicted output of the student network model be , the soft label is , the original label is , distillation loss ;in is the weight for weighing the loss of soft labels and original labels.

8. The rehabilitation training data processing method according to claim 4, characterized in that: In step 3, the number of correct predictions of the student network model on the validation set is TP, the total number of samples in the validation set is N, and the individual fitness is ; Let the individuals in the population be , the variation factor is F, and the randomly selected individuals are , , , the mutated individual ; Let the crossover probability be , after crossover , where j represents the dimension of the individual, represents the individual after mutation in the j dimension, Represents an individual in a population of j dimensions.

9. The rehabilitation training data processing method according to claim 5, characterized in that: In step 4, the rehabilitation effect score is set to The patient's self-perception score was , blood oxygen saturation changes to , the reward function ,in , , is the weight coefficient; let the parameters of the strategy network model be , the strategy is , where a is the action, s is the state, and the advantage function is , update formula ,in is the learning rate.

10. A rehabilitation training data processing device, using the rehabilitation training data processing method according to claim 1, characterized in that: The device comprises: The data acquisition and preprocessing module is used to collect and preprocess the physiological, log and video data of the users to be rehabilitated, and integrate these data into a multi-source data set; High-performance model building and knowledge distillation module, which is used to select and train the teacher network model, build the student network model, and use knowledge distillation technology to enable the student network model to effectively learn the knowledge of the teacher network model and adapt to the characteristics of rehabilitation training data; The differential evolution algorithm optimization module is used to define the parameters of the differential evolution algorithm and obtain the optimized student network model through an iterative optimization process; The deep reinforcement learning parameter adjustment module is used to generate feature data using the optimized student network model and adjust the rehabilitation training parameters through reinforcement learning technology.

Images