Respiratory rehabilitation training data management method and system based on artificial intelligence

Abstract

The invention discloses a respiratory rehabilitation training data management method and system based on artificial intelligence, and relates to the technical field of data processing, the respiratory rehabilitation training data management method based on artificial intelligence comprises the following steps: obtaining training data of a respiratory rehabilitation person, and preprocessing the training data to obtain a multi-source data set; constructing a teacher network based on the high-performance model, training the teacher network through the multi-source data set, constructing a student network, and training the student network based on the trained teacher network; optimizing the trained student network by using a differential evolution algorithm to obtain a student network optimized model; and training a reinforcement learning agent by using data generated by the model after student network optimization, and automatically adjusting rehabilitation training parameters of the respiratory rehabilitation personnel through the reinforcement learning agent. According to the invention, computing resource and storage requirements are reduced, data-driven decision is realized, and a personalized rehabilitation training scheme is provided.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a respiratory rehabilitation training data management method and system based on artificial intelligence. Background Art

[0002] Respiratory rehabilitation training is a treatment for respiratory diseases, which aims to improve patients' respiratory function, enhance physical strength and improve their quality of life through a series of planned training and educational activities. This training is usually suitable for patients with respiratory diseases such as chronic obstructive pulmonary disease, asthma, emphysema, etc. The training content may include but is not limited to breathing technique exercises, exercise training, energy conservation techniques, nutritional guidance and psychological support.

[0003] Respiratory rehabilitation training data management refers to the process of collecting, organizing, storing and analyzing various information related to respiratory rehabilitation training. Effective data management can help doctors and patients better understand the effects of treatment and make corresponding adjustments. Applying artificial intelligence technology to respiratory rehabilitation training data management can automate and optimize the data management process.

[0004] For example, Chinese patent 202110506980.2 discloses a management method and system for respiratory rehabilitation functional training, which compares user training images with standard images, evaluates the training effect, inputs the evaluation results into the training evaluation model, and obtains the training evaluation results. However, the above method still has the following shortcomings: when training the model, this method lacks the guidance of the teacher network, the model has a high complexity, and it is easier to overfit the training data; at the same time, it is impossible to quickly and effectively adjust the parameters of the rehabilitation training program according to the rehabilitation training data.

[0005] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0006] In view of the problems in the related art, the present invention proposes a respiratory rehabilitation training data management method and system based on artificial intelligence to overcome the above-mentioned technical problems existing in the existing related technology.

[0007] To this end, the specific technical solution adopted by the present invention is as follows: According to one aspect of the present invention, a method for managing respiratory rehabilitation training data based on artificial intelligence is provided, and the method for managing respiratory rehabilitation training data based on artificial intelligence comprises the following steps: S1. Obtain training data of respiratory rehabilitation personnel, and pre-process the training data to obtain a multi-source data set.

[0008] S2. Build a teacher network based on a high-performance model, train the teacher network through multi-source data sets, build a student network, and train the student network based on the trained teacher network.

[0009] S3. Use the differential evolution algorithm to optimize the trained student network and obtain the optimized student network model.

[0010] S4. Use the data generated by the student network optimized model to train the reinforcement learning agent, and automatically adjust the rehabilitation training parameters of respiratory rehabilitation personnel through the reinforcement learning agent.

[0011] Among them, building a teacher network based on a high-performance model, training the teacher network through a multi-source data set, building a student network, and training the student network based on the trained teacher network include the following steps: S21. Select a high-performance deep learning model as the teacher network, adjust the structure of the teacher network, and set hyperparameters; S22, loading the preprocessed multi-source dataset, and using the multi-source dataset to train the teacher network, using the back propagation algorithm to update the weight of the teacher network, and evaluating the performance of the teacher network; S23. Build a student network and design the number of layers and neuron configuration of the student network; S24, using the trained teacher network to infer the training data and generate soft labels; S25. Train the learning network based on the adaptive confidence adjustment factor.

[0012] Furthermore, obtaining the training data of respiratory rehabilitation personnel and preprocessing the training data to obtain a multi-source data set includes the following steps: S11, collecting physiological data and rehabilitation training log data of respiratory rehabilitation personnel to obtain training data; S12. Determine the required label data type according to the goal of rehabilitation training, and obtain the label data; S12, removing noise and outliers from training data; S13, integrating data from different sources in the training data and standardizing the training data; S14, extracting features from training data; S15. Preprocess the collected label data and match the label data with corresponding feature data.

[0013] Further, training the learning network based on the adaptive confidence adjustment factor includes the following steps: S251. Dynamically adjust the confidence adjustment factor parameters according to the degree of difference between categories; S252, using an adaptive confidence adjustment factor parameter to adjust the softmax layer of the student network; S253. Define the loss function of the student network, including prediction loss and distillation loss; S254, training the student network using the training set data with soft labels, and updating the weights of the student network using the back propagation algorithm and optimizer; S255, monitor the performance of the student network on the validation set; S256. Adjust the student network parameters based on the performance feedback on the validation set; S257. Evaluate the final performance of the student network on the test set.

[0014] Furthermore, dynamically adjusting the confidence adjustment factor parameters according to the degree of difference between categories includes the following steps: S2511. Use the teacher network to infer the training data, obtain the predicted probability of each category, and calculate the difference in probability distribution between different categories, use KL divergence as a metric, and output the KL divergence matrix between categories; S2512, selecting an initial confidence adjustment factor for the softmax layer; S2513, dynamically adjusting the confidence adjustment factor of each category according to the KL divergence between categories; S2514. Integrate an adaptive confidence adjustment factor into the loss function.

[0015] Furthermore, dynamically adjusting the confidence adjustment factor of each category according to the KL divergence between categories includes the following steps: S25131. For each pair of categories, if the KL divergence is greater than a preset threshold, increase the confidence adjustment factor of the pair of categories; S25132. If the KL divergence of this pair of categories is less than or equal to a preset threshold, reduce the confidence adjustment factor.

[0016] Furthermore, the loss function formula of the student network is: ; In the formula, L pred Denotes the prediction loss: L distill Represents the distillation loss: α Represents a hyperparameter.

[0017] Further, the trained student network is optimized using a differential evolution algorithm to obtain an optimized student network model, which includes the following steps: S31, define the parameters of the differential evolution algorithm and initialize the population; S32, evaluate the fitness of each individual; S33, performing mutation operation on individuals in the population to generate new candidate solutions, and performing crossover operation on the mutated individuals and the original individuals to generate new individuals; S34, select the next generation population according to the fitness value, and retain individuals with higher fitness; S35. Repeat the mutation, crossover and selection operations until the preset number of iterations is reached and the optimized student network model is obtained.

[0018] Furthermore, training a reinforcement learning agent using data generated by the optimized model of the student network, and automatically adjusting the rehabilitation training parameters of respiratory rehabilitation personnel by the reinforcement learning agent includes the following steps: S41, using the feature data generated by the model after the student network optimization as the state input of the reinforcement learning; S42, determining the action taken by the reinforcement learning agent to adjust the rehabilitation training, and designing a reward function based on the rehabilitation effect score; S43. Select a reinforcement learning algorithm and initialize the policy network and value network; S44, running the reinforcement learning agent and collecting data on state, action, reward, and next state; S45, using the data collected in S44, and updating the parameters of the policy network and the value network through the back propagation algorithm; S46. After iterative training of the reinforcement learning agent, the state of the rehabilitation personnel is monitored by the reinforcement learning agent, and an action for adjusting the rehabilitation training parameters is generated according to the current state and the trained strategy network.

[0019] Furthermore, the calculation formula of the reward function is: ; In the formula, w eff represents the weight of rehabilitation effect score, Δ R eff It indicates the change in rehabilitation effect score; w self represents the weight of the patient's self-perception score, Δ R self It indicates the amount of change in the patient's self-perception score; w ox Indicates the weight of blood oxygen saturation change, Δ R ox Indicates the change in blood oxygen saturation.

[0020] According to another aspect of the present invention, there is also provided an artificial intelligence-based respiratory rehabilitation training data management system, which includes a rehabilitation data preprocessing module, a network training module, a network optimization module and a rehabilitation training parameter adjustment module; wherein the rehabilitation data preprocessing module, the network training module, the network optimization module and the rehabilitation training parameter adjustment module are connected in sequence.

[0021] The rehabilitation data preprocessing module is used to obtain the training data of respiratory rehabilitation personnel and preprocess the training data to obtain a multi-source data set.

[0022] The network training module is used to build a teacher network based on a high-performance model, train the teacher network through a multi-source data set, build a student network, and train the student network based on the trained teacher network.

[0023] The network optimization module is used to optimize the trained student network using the differential evolution algorithm to obtain the optimized student network model.

[0024] The rehabilitation training parameter adjustment module is used to train the reinforcement learning agent using the data generated by the model after student network optimization, and automatically adjust the rehabilitation training parameters of respiratory rehabilitation personnel through the reinforcement learning agent.

[0025] The beneficial effects of the present invention are: (1) The present invention provides an artificial intelligence-based respiratory rehabilitation training data management method and system, which integrates rehabilitation training data from different sources to provide more comprehensive information. Based on the knowledge distillation technology, the teacher network and the student network are trained, and the knowledge of the complex and high-performance teacher network is transferred to the smaller student network, thereby achieving model compression. Smaller models are easier to deploy in practical applications, and have lower computing resources and storage requirements.

[0026] (2) The present invention uses the differential evolution algorithm to adjust the hyperparameters of the student network, reducing the workload and time cost of manual parameter adjustment. The weight of the student network is optimized to improve the prediction accuracy and generalization ability of the student network.

[0027] (3) The present invention uses the data generated by the student network optimization model to train the reinforcement learning agent, which can achieve data-driven decision-making and provide personalized rehabilitation training programs. The reinforcement learning agent can dynamically adjust the training parameters according to the real-time status of the rehabilitation personnel to improve the rehabilitation effect. By automatically adjusting the rehabilitation training parameters through the reinforcement learning agent, manual intervention can be reduced and the efficiency and effect of rehabilitation training can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0029] Figure 1 is a flow chart of a respiratory rehabilitation training data management method based on artificial intelligence according to an embodiment of the present invention; Figure 2 The present invention is a block diagram of a respiratory rehabilitation training data management system based on artificial intelligence.

[0030] In the figure: 1. Rehabilitation data preprocessing module; 2. Network training module; 3. Network optimization module; 4. Rehabilitation training parameter adjustment module. DETAILED DESCRIPTION

[0031] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, ordinary technicians in the field should be able to understand other possible implementations and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0032] According to an embodiment of the present invention, a respiratory rehabilitation training data management method and system based on artificial intelligence are provided.

[0033] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to one embodiment of the present invention, a respiratory rehabilitation training data management method based on artificial intelligence is provided, and the respiratory rehabilitation training data management method based on artificial intelligence includes the following steps: S1. Obtain training data of respiratory rehabilitation personnel, and pre-process the training data to obtain a multi-source data set.

[0034] In a further embodiment, obtaining training data of respiratory rehabilitation personnel and preprocessing the training data to obtain a multi-source data set includes the following steps: S11. Collect the physiological data and rehabilitation training log data of respiratory rehabilitation personnel, including respiratory rate, vital capacity, heart rate, blood pressure, specific date and time of rehabilitation training, etc., to obtain training data; at the same time, collect the patient's medical records and other relevant health data.

[0035] S12. According to the goal of rehabilitation training, determine the required type of label data (such as rehabilitation effect score, doctor's diagnosis results, etc.), and obtain label data through questionnaires, doctor evaluations, rehabilitation tests, etc.

[0036] S12. Remove noise and outliers from training data.

[0037] S13. Integrate data from different sources in training data to ensure data consistency and integrity, and standardize training data to facilitate model training.

[0038] S14. Extract features from the training data, such as respiratory rate, vital capacity, etc.

[0039] S15. Preprocess the collected label data, such as standardization, encoding, etc., and match the label data with the corresponding feature data to ensure the integrity and accuracy of the multi-source data set.

[0040] S2. Build a teacher network based on a high-performance model, train the teacher network through multi-source data sets, build a student network, and train the student network based on the trained teacher network.

[0041] In a further embodiment, constructing a teacher network based on a high-performance model, training the teacher network through a multi-source data set, constructing a student network, and training the student network based on the trained teacher network includes the following steps: S21. Select a high-performance deep learning model as the teacher network, adjust the structure of the teacher network, and set hyperparameters; It should be noted that since respiratory rehabilitation training data contains complex physiological signals and multi-dimensional features, deep learning models can capture these complex patterns and relationships. For example: Convolutional Neural Networks (CNN): If the data includes visual information, such as images or videos of breathing movements, CNN can be used for feature extraction.

[0042] Recurrent Neural Networks (RNNs), especially Long Short-Term Memory Networks: For time series data, such as continuous physiological monitoring data, LSTMs are able to handle long-term dependencies in the sequence.

[0043] Consider using ensemble learning methods such as random forests or gradient boosting machines, which perform well when dealing with heterogeneous data from multiple sources.

[0044] Hybrid model: Combining CNN and LSTM, it can process spatial features and time series features at the same time and is suitable for multimodal data.

[0045] S22. Load the preprocessed multi-source dataset and use the multi-source dataset to train the teacher network. Use the back-propagation algorithm to update the weights of the teacher network, and use cross-validation and other techniques to evaluate the performance of the teacher network to ensure its accuracy and generalization ability.

[0046] S23. Build a student network with a simple structure and few parameters, and design the number of layers and neuron configuration of the student network; The student network model is suitable for a lightweight network. The student network needs to be inferred quickly, so a lightweight network structure is more suitable. For example: MobileNet: A lightweight deep neural network that uses depthwise separable convolutions to reduce computation.

[0047] ShuffleNet: A lightweight network that reduces computational complexity through point group convolution and channel reorganization operations.

[0048] S24. Use the trained teacher network to infer the training data and generate soft labels. These labels contain probability distribution information. For example, if the rehabilitation effect score is a continuous scale (such as 0-10 points), the output layer of the teacher network will be a prediction of a continuous value. In this case, each prediction result can be regarded as a probability distribution, indicating the probability of the patient reaching a certain score.

[0049] S25. Training the learning network based on the adaptive confidence adjustment factor. By dynamically adjusting the confidence adjustment factor, the learning effect of the student network is optimized so that it can better imitate the teacher network.

[0050] In a further embodiment, training a learning network based on an adaptive confidence adjustment factor comprises the following steps: S251, dynamically adjusting the confidence adjustment factor parameters according to the degree of difference between categories to optimize the learning effect of the student network; It should be noted that the degree of difference between categories refers to the distribution difference of each category of samples in the training set. Some categories may be more difficult to distinguish or easier to confuse than other categories. Dynamically adjusting the confidence adjustment factor is to dynamically adjust the confidence adjustment factor according to the degree of difference between categories, so that the student network can pay more attention to the difficult-to-distinguish categories during the training process. By dynamically adjusting the confidence adjustment factor, the student network can pay more attention to the difficult-to-distinguish categories during the training process, thereby improving the overall learning effect, helping the student network to better cope with the imbalance problem in the data, and improving the robustness and generalization ability of the model.

[0051] S252. Use the adaptive confidence adjustment factor parameter to adjust the softmax layer (activation function) of the student network so that the learning network output is closer to the soft label of the teacher network. Soft labels are smoother than hard labels, which helps reduce overfitting and improve the generalization ability of the student network.

[0052] S253. Define the loss function of the student network, including prediction loss and distillation loss; Prediction loss: the loss between the student network prediction result and the true label. Distillation loss: the loss between the soft label output by the student network and the soft label output by the teacher network. By combining prediction loss and distillation loss, the student network can learn the knowledge of the true label and the teacher network at the same time, improving the overall performance.

[0053] S254, train the student network using training set data with soft labels, and update the weights of the student network using a back propagation algorithm and an optimizer; wherein, prepare the training data: load the preprocessed multi-source dataset and divide it into a training set, a validation set, and a test set.

[0054] S255. Monitor the performance of the student network on the validation set to ensure its learning effect. Evaluation can be done using indicators such as classification accuracy, precision, and recall.

[0055] S256. Adjust the student network parameters, such as learning rate, temperature parameters, optimizer, etc., based on the performance feedback on the validation set.

[0056] S257. Evaluate the final performance of the student network on the test set to ensure generalization ability.

[0057] In a further embodiment, dynamically adjusting the confidence adjustment factor parameters according to the degree of difference between categories includes the following steps: S2511. Use the teacher network to infer the training data, obtain the predicted probability of each category, and calculate the difference in probability distribution between different categories. Use KL divergence as a metric to output the KL divergence matrix between categories. KL divergence is relative entropy, which is a method used to measure the difference between two probability distributions. By calculating KL divergence, the relationship and difference between different categories can be captured, especially those that are difficult to distinguish. Guide adjustment of adjustment factors: The KL divergence matrix provides a basis for dynamically adjusting the confidence adjustment factor in subsequent steps, helping to optimize the learning effect of the student network.

[0058] S2512. Select an initial confidence adjustment factor for the softmax layer. Usually start with a smaller value to help stabilize the model training and avoid instability in model training caused by too large an initial value.

[0059] S2513. Dynamically adjust the confidence adjustment factor of each category according to the KL divergence between categories.

[0060] S2514. Integrate an adaptive confidence adjustment factor into the loss function so that it can reflect the differences between categories when calculating the loss.

[0061] In a further embodiment, dynamically adjusting the confidence adjustment factor of each category according to the inter-category KL divergence includes the following steps: S25131. For each pair of categories, if the KL divergence is greater than a preset threshold (e.g., 0.5), it means that the two categories are difficult to distinguish. Then, the confidence adjustment factor of this pair of categories is increased to make the probability distribution smoother and reduce the confidence of the student network for these difficult-to-distinguish categories. S25132. If the KL divergence of this pair of categories is less than or equal to the preset threshold, it means that the two categories are easier to distinguish. Then reduce the confidence adjustment factor to make the probability distribution sharper and increase the confidence of the student network for these easily distinguishable categories.

[0062] In a further embodiment, the loss function formula of the student network is: ; In the formula, L pred Represents the prediction loss, usually using cross entropy loss: L distill Denotes the distillation loss, using KL divergence to measure the difference between the softmax output of the student network and the softmax output of the teacher network: α represents a hyperparameter used to balance the importance of prediction loss and distillation loss. α The value range is usually between 0 and 1, depending on whether you want the model to learn more from the teacher network or pay more attention to the true label.

[0063] For example, the following data was collected for 10 patients: Respiratory rate (times / minute), [18, 20, 22, 19, 21, 18, 20, 22, 21, 19].

[0064] Vital capacity (ml), [4000, 4200, 3900, 4100, 4300, 4000, 4200, 3900, 4100, 4300].

[0065] Heart rate (beats / minute), [75, 80, 78, 77, 79, 75, 80, 78, 77, 79].

[0066] Blood pressure (mmHg), [(120, 80), (125, 85), (122, 82), (118, 78), (130, 88), (120, 80), (125, 85), (122, 82), (118, 78), (130, 88)].

[0067] Specific date and time of rehabilitation training: Record the date and time of each training session separately.

[0068] Obtain rehabilitation effect scores: [7, 8, 6, 7, 9, 7, 8, 6, 7, 9].

[0069] Preprocess the above data to get: Respiratory rate, [0.1, 0.3, 0.5, 0.2, 0.4, 0.1, 0.3, 0.5, 0.4, 0.2].

[0070] Vital capacity, [0.2, 0.4, 0.1, 0.3, 0.5, 0.2, 0.4, 0.1, 0.3, 0.5].

[0071] Heart rate, [0.3, 0.6, 0.5, 0.4, 0.5, 0.3, 0.6, 0.5, 0.4, 0.5].

[0072] Blood pressure, [(0.2, 0.1), (0.4, 0.3), (0.3, 0.2), (0.1, 0.0), (0.5, 0.4), (0.2, 0.1), (0.4, 0.3), (0.3, 0.2), (0.1, 0.0), (0.5, 0.4)].

[0073] Rehabilitation effect score, [0.7, 0.8, 0.6, 0.7, 0.9, 0.7, 0.8, 0.6, 0.7, 0.9].

[0074] Use the trained teacher network for inference, generate soft labels and normalize them: Rehabilitation effect score prediction: [0.71, 0.82, 0.63, 0.75, 0.88, 0.72, 0.81, 0.64, 0.76, 0.91].

[0075] After obtaining the above data and soft labels, the back-propagation algorithm and optimizer can be used to update the student network weights to complete the training of the student network.

[0076] S3. Use the differential evolution algorithm to optimize the trained student network and obtain the optimized student network model.

[0077] In a further embodiment, optimizing the trained student network using a differential evolution algorithm to obtain an optimized student network model includes the following steps: S31. Define the parameters of the differential evolution algorithm, such as population size, mutation factor, crossover probability, etc., and initialize the population. Each individual represents a set of parameters of the student network.

[0078] S32. Perform fitness evaluation on each individual. The fitness function is the performance indicator of the student network (such as accuracy, loss, etc.), which is used to measure the pros and cons of each parameter configuration. Through fitness evaluation, we can determine which individuals are more suitable for the problem, that is, the pros and cons of the parameter configuration, and provide a basis for selection and genetic operations.

[0079] S33. Perform mutation operations on individuals in the population to generate new candidate solutions, and perform crossover operations on the mutated individuals and the original individuals to generate new individuals; mutation and crossover operations increase the diversity of the population, help the algorithm escape from the local optimum, and explore a wider solution space.

[0080] S34. Select the next generation population according to the fitness value and retain individuals with higher fitness.

[0081] S35, repeating the mutation, crossover and selection operations until the preset number of iterations is reached and the student network optimized model is obtained. After iterative optimization, a student network model with the best performance or meeting the conditions is obtained.

[0082] S4. Use the data generated by the student network optimized model to train the reinforcement learning agent, and automatically adjust the rehabilitation training parameters of respiratory rehabilitation personnel through the reinforcement learning agent to promote better rehabilitation results.

[0083] In a further embodiment, training a reinforcement learning agent using data generated by the student network optimized model, and automatically adjusting the rehabilitation training parameters of respiratory rehabilitation personnel by the reinforcement learning agent includes the following steps: S41. The feature data generated by the student network optimized model is used as the state input of reinforcement learning. The input and output data together constitute the data generated by the student network optimized model, such as respiratory rate, vital capacity, rehabilitation effect score, patient self-perception score, change in blood oxygen saturation, etc. The generated data is used as the state input of the reinforcement learning agent to reflect the current health status and rehabilitation progress of the rehabilitation personnel. The data generated by the student network can more accurately describe the current state of the rehabilitation personnel and help the reinforcement learning agent make more reasonable decisions. The feature data provides rich information, which helps the reinforcement learning agent to understand and analyze the rehabilitation process more comprehensively.

[0084] S42. Determine the actions taken by the reinforcement learning agent to adjust the rehabilitation training, such as training intensity (high, medium, low), training frequency (daily, every other day, etc.), training duration (minutes), etc., and design a reward function based on the rehabilitation effect score to encourage the agent to take actions that are conducive to the rehabilitation effect.

[0085] S43. Select a reinforcement learning algorithm, such as a deep Q network, proximal policy optimization, etc., and initialize the policy network and value network to estimate the value of each state-action pair and formulate a strategy.

[0086] S44. Run the reinforcement learning agent and collect data on the state, action, reward, and next state at each time step.

[0087] S45. Use the data collected in S44 and update the parameters of the policy network and the value network through the back-propagation algorithm. The specific update method depends on the selected reinforcement learning algorithm. For example, the deep Q network uses the Q-learning update method, and the proximal policy optimization uses the policy gradient update method.

[0088] S46. After iterative training of the reinforcement learning agent, the state of the rehabilitation personnel is monitored by the reinforcement learning agent, and an action for adjusting the rehabilitation training parameters is generated according to the current state and the trained strategy network.

[0089] Through reinforcement learning agents, rehabilitation training parameters can be monitored and adjusted in real time to ensure the effectiveness and personalization of the training process. Using the trained policy network, the agent can make optimal training adjustments based on the real-time status of the rehabilitation personnel to improve the rehabilitation effect.

[0090] In a further embodiment, the reward function is calculated as: ; In the formula, R represents the total reward function, which is used to evaluate the effectiveness of the actions taken by the reinforcement learning agent; w eff represents the weight of rehabilitation effect score, Δ R eff It indicates the change in rehabilitation effect score; w self represents the weight of the patient's self-perception score, Δ R self It indicates the amount of change in the patient's self-perception score; w ox Indicates the weight of blood oxygen saturation change, Δ R ox Indicates the change in blood oxygen saturation.

[0091] Suppose, we obtain the feature data generated by the student network after optimization: Respiratory rate: 14. Vital capacity: 3.5. Rehabilitation effect score: 7.1. Patient self-perception score: 6.8. Blood oxygen saturation: 95%. Determine the action: training intensity (high, medium, low), training frequency (daily, every other day, etc.), training duration (30 minutes, 40 minutes, etc.). The weight of the rehabilitation effect score is 0.5, the weight of the patient self-perception score is 0.3, and the weight of the change in blood oxygen saturation is 0.2.

[0092] Time step 1, initial state: Respiratory rate, 14.

[0093] Vital capacity, 3.5.

[0094] Rehabilitation effect score, 7.1.

[0095] Patient self-perception score, 6.8.

[0096] Blood oxygen saturation, 95%.

[0097] action: Training intensity: medium.

[0098] Training frequency, daily.

[0099] Training duration: 30 minutes.

[0100] Next state: Respiratory rate, 15.

[0101] Vital capacity, 3.6.

[0102] Rehabilitation effect score, 7.3.

[0103] Patient self-perception score, 7.0.

[0104] Blood oxygen saturation, 95.5%.

[0105] Reward calculation, R=0.26.

[0106] Using the data collected at time step 1, the parameters of the policy network and value network are updated through the back propagation algorithm. The reinforcement learning agent continues to run and collects more time step data to adjust the reinforcement learning agent in real time. The reinforcement learning agent inputs the policy network according to the current state. The policy network outputs the Q value of each possible action and selects the action with the highest Q value as the current adjustment suggestion. The Q value represents the long-term expected return of taking a specific action in a specific state, which is estimated by the iterative reinforcement learning algorithm.

[0107] like Figure 2As shown, according to another embodiment of the present invention, a respiratory rehabilitation training data management system based on artificial intelligence is also provided, and the respiratory rehabilitation training data management system based on artificial intelligence includes a rehabilitation data preprocessing module 1, a network training module 2, a network optimization module 3 and a rehabilitation training parameter adjustment module 4; wherein, the rehabilitation data preprocessing module 1, the network training module 2, the network optimization module 3 and the rehabilitation training parameter adjustment module 4 are connected in sequence.

[0108] The rehabilitation data preprocessing module 1 is used to obtain the training data of respiratory rehabilitation personnel and preprocess the training data to obtain a multi-source data set.

[0109] The network training module 2 is used to build a teacher network based on a high-performance model, train the teacher network through a multi-source data set, build a student network, and train the student network based on the trained teacher network.

[0110] The network optimization module 3 is used to optimize the trained student network using a differential evolution algorithm to obtain an optimized student network model.

[0111] The rehabilitation training parameter adjustment module 4 is used to train the reinforcement learning agent using the data generated by the model after the student network optimization, and automatically adjust the rehabilitation training parameters of the respiratory rehabilitation personnel through the reinforcement learning agent.

[0112] In summary, the present invention provides a respiratory rehabilitation training data management method and system based on artificial intelligence, which provides more comprehensive information by integrating rehabilitation training data from different sources. Based on the knowledge distillation technology, the teacher network and the student network are trained, and the knowledge of the complex and high-performance teacher network is transferred to the smaller student network, thereby realizing model compression. Smaller models are easier to deploy in practical applications, and the computing resources and storage requirements are also lower. The present invention uses a differential evolution algorithm to adjust the hyperparameters of the student network to reduce the workload and time cost of manual parameter adjustment. The weight of the student network is optimized to improve the prediction accuracy and generalization ability of the student network. The present invention uses the data generated by the optimized model of the student network to train a reinforcement learning agent, which can realize data-driven decision-making and provide personalized rehabilitation training programs. The reinforcement learning agent can dynamically adjust the training parameters according to the real-time status of the rehabilitation personnel to improve the rehabilitation effect. The automatic adjustment of rehabilitation training parameters by the reinforcement learning agent can reduce manual intervention and improve the efficiency and effect of rehabilitation training.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A respiratory rehabilitation training data management method based on artificial intelligence, characterized in that: The respiratory rehabilitation training data management method based on artificial intelligence includes the following steps: S1. Obtain training data of respiratory rehabilitation personnel, and pre-process the training data to obtain a multi-source data set; S2. Build a teacher network based on a high-performance model, train the teacher network through multi-source data sets, build a student network, and train the student network based on the trained teacher network; S3, using the differential evolution algorithm to optimize the trained student network to obtain the optimized student network model; S4, using the data generated by the student network optimized model to train a reinforcement learning agent, and automatically adjusting the rehabilitation training parameters of respiratory rehabilitation personnel through the reinforcement learning agent; The steps of constructing a teacher network based on a high-performance model, training the teacher network through a multi-source data set, constructing a student network, and training the student network based on the trained teacher network include the following steps: S21. Select a high-performance deep learning model as the teacher network, adjust the structure of the teacher network, and set hyperparameters; S22, loading the preprocessed multi-source dataset, and using the multi-source dataset to train the teacher network, using the back propagation algorithm to update the weight of the teacher network, and evaluating the performance of the teacher network; S23. Build a student network and design the number of layers and neuron configuration of the student network; S24, using the trained teacher network to infer the training data and generate soft labels; S25. Train the learning network based on the adaptive confidence adjustment factor.

2. A respiratory rehabilitation training data management method based on artificial intelligence according to claim 1, characterized in that: The method of obtaining the training data of respiratory rehabilitation personnel and preprocessing the training data to obtain a multi-source data set includes the following steps: S11, collecting physiological data and rehabilitation training log data of respiratory rehabilitation personnel to obtain training data; S12. Determine the required label data type according to the goal of rehabilitation training, and obtain the label data; S12, removing noise and outliers from training data; S13, integrating data from different sources in the training data and standardizing the training data; S14, extracting features from training data; S15. Preprocess the collected label data and match the label data with corresponding feature data.

3. A respiratory rehabilitation training data management method based on artificial intelligence according to claim 2, characterized in that: The training of the learning network based on the adaptive confidence adjustment factor comprises the following steps: S251. Dynamically adjust the confidence adjustment factor parameters according to the degree of difference between categories; S252, using an adaptive confidence adjustment factor parameter to adjust the softmax layer of the student network; S253. Define the loss function of the student network, including prediction loss and distillation loss; S254, training the student network using the training set data with soft labels, and updating the weights of the student network using the back propagation algorithm and optimizer; S255, monitor the performance of the student network on the validation set; S256. Adjust the student network parameters based on the performance feedback on the validation set; S257. Evaluate the final performance of the student network on the test set.

4. A respiratory rehabilitation training data management method based on artificial intelligence according to claim 3, characterized in that: The step of dynamically adjusting the confidence adjustment factor parameters according to the degree of difference between categories comprises the following steps: S2511. Use the teacher network to infer the training data, obtain the predicted probability of each category, and calculate the difference in probability distribution between different categories, use KL divergence as a metric, and output the KL divergence matrix between categories; S2512, selecting an initial confidence adjustment factor for the softmax layer; S2513, dynamically adjusting the confidence adjustment factor of each category according to the KL divergence between categories; S2514. Integrate an adaptive confidence adjustment factor into the loss function.

5. A respiratory rehabilitation training data management method based on artificial intelligence according to claim 4, characterized in that: The step of dynamically adjusting the confidence adjustment factor of each category according to the KL divergence between categories includes the following steps: S25131. For each pair of categories, if the KL divergence is greater than a preset threshold, increase the confidence adjustment factor of the pair of categories; S25132. If the KL divergence of this pair of categories is less than or equal to a preset threshold, reduce the confidence adjustment factor.

6. A respiratory rehabilitation training data management method based on artificial intelligence according to claim 5, characterized in that: The loss function formula of the student network is: ； In the formula, L pred Denotes the prediction loss: L distill Represents the distillation loss: α Represents a hyperparameter.

7. The method for managing respiratory rehabilitation training data based on artificial intelligence according to claim 5, characterized in that: The method of optimizing the trained student network by using the differential evolution algorithm to obtain the optimized student network model comprises the following steps: S31, define the parameters of the differential evolution algorithm and initialize the population; S32, evaluate the fitness of each individual; S33, performing mutation operation on individuals in the population to generate new candidate solutions, and performing crossover operation on the mutated individuals and the original individuals to generate new individuals; S34, select the next generation population according to the fitness value, and retain individuals with higher fitness; S35. Repeat the mutation, crossover and selection operations until the preset number of iterations is reached and the optimized student network model is obtained.

8. A respiratory rehabilitation training data management method based on artificial intelligence according to claim 7, characterized in that: The method of training a reinforcement learning agent using data generated by the model after the student network optimization, and automatically adjusting the rehabilitation training parameters of respiratory rehabilitation personnel by the reinforcement learning agent includes the following steps: S41, using the feature data generated by the model after the student network optimization as the state input of the reinforcement learning; S42, determining the action taken by the reinforcement learning agent to adjust the rehabilitation training, and designing a reward function based on the rehabilitation effect score; S43. Select a reinforcement learning algorithm and initialize the policy network and value network; S44, running the reinforcement learning agent and collecting data on state, action, reward, and next state; S45, using the data collected in S44, and updating the parameters of the policy network and the value network through the back propagation algorithm; S46. After iterative training of the reinforcement learning agent, the state of the rehabilitation personnel is monitored by the reinforcement learning agent, and an action for adjusting the rehabilitation training parameters is generated according to the current state and the trained strategy network.

9. The method for managing respiratory rehabilitation training data based on artificial intelligence according to claim 8, characterized in that: The calculation formula of the reward function is: ； In the formula, w eff represents the weight of rehabilitation effect score, Δ R eff It indicates the change in rehabilitation effect score; w self represents the weight of the patient's self-perception score, Δ R self It indicates the amount of change in the patient's self-perception score; w ox Indicates the weight of blood oxygen saturation change, Δ R ox Indicates the change in blood oxygen saturation.

10. A respiratory rehabilitation training data management system based on artificial intelligence, used to implement the respiratory rehabilitation training data management method based on artificial intelligence according to any one of claims 1 to 9, characterized in that: The respiratory rehabilitation training data management system based on artificial intelligence includes a rehabilitation data preprocessing module, a network training module, a network optimization module and a rehabilitation training parameter adjustment module; Wherein, the rehabilitation data preprocessing module, the network training module, the network optimization module and the rehabilitation training parameter adjustment module are connected in sequence; The rehabilitation data preprocessing module is used to obtain the training data of respiratory rehabilitation personnel and preprocess the training data to obtain a multi-source data set; The network training module is used to build a teacher network based on a high-performance model, train the teacher network through a multi-source data set, build a student network, and train the student network based on the trained teacher network; The network optimization module is used to optimize the trained student network using a differential evolution algorithm to obtain an optimized student network model; The rehabilitation training parameter adjustment module is used to train the reinforcement learning agent using the data generated by the model after the student network optimization, and automatically adjust the rehabilitation training parameters of the respiratory rehabilitation personnel through the reinforcement learning agent.

Images