A lung rehabilitation training method based on respiratory electromyogram signal feedback

Abstract

The present invention relates to a pulmonary rehabilitation training method, specifically to a pulmonary rehabilitation training method based on respiratory electromyogram signal feedback. First, the electromyogram signal acquisition system is used to capture the electrical activities of the respiratory-related muscles in real time, and signal amplification, filtering, and noise suppression technologies are adopted to optimize the signal quality. Secondly, machine learning algorithms are used to perform time-frequency analysis on the electromyogram signals, extract energy spectrum and waveform length features, and apply deep learning models including convolutional neural networks for training and classification to distinguish different types of respiratory patterns. Then, according to the classification results of the electromyogram signals, a real-time feedback system is established. This system displays the electromyogram signal intensity and respiratory quality in real time through the user interface and adjusts the difficulty and duration of the respiratory training. Finally, based on the analysis of the user's electromyogram signal characteristics and historical data, artificial intelligence algorithms are used to generate personalized pulmonary rehabilitation training programs, and the training plan is dynamically adjusted in combination with the user's physical condition and recovery progress.

Description

Technical Field

[0001] The present invention relates to a pulmonary rehabilitation training method, and more particularly to a pulmonary rehabilitation training method based on respiratory electromyogram signal feedback. Background Art

[0002] Currently, the training methods used for pulmonary rehabilitation mainly rely on traditional physical therapy techniques, including breathing exercises, muscle strength training, and aerobic exercise, etc. Although these methods are effective to a certain extent, there are various deficiencies and drawbacks, which limit their efficiency and effectiveness. First of all, traditional pulmonary rehabilitation training methods often lack sufficient personalization. These trainings are usually standardized programs formulated based on a wide range of patient groups, ignoring the differences among patients in terms of the severity of the disease, physical conditions, and recovery speed. For example, the rehabilitation needs of patients with COPD (Chronic Obstructive Pulmonary Disease) are significantly different from those of patients after lung cancer surgery, but traditional methods often adopt a "one-size-fits-all" training plan, which is not sufficient to provide the optimal training for individual specific needs. This may not only lead to poor rehabilitation effects, but also increase the risk of patient injury. Secondly, traditional methods lack a real-time feedback mechanism during implementation. When performing pulmonary rehabilitation training, coaches or therapists cannot accurately and real-time grasp the physiological responses of patients, such as the specific movement conditions of the lungs and muscles, and can only rely on observing the subjective feedback of patients to adjust the training plan. This method is not only inefficient, but also cannot ensure the safety and effectiveness of the training. For example, if a patient experiences discomfort during training, they may not be able to accurately describe the degree and nature of the discomfort, resulting in the therapist being unable to make appropriate adjustments. Finally, traditional pulmonary rehabilitation methods also have deficiencies in monitoring and evaluating the rehabilitation effects. These methods usually rely on regular medical evaluations to track the progress of patients, and these evaluations often have a long interval and cannot provide continuous progress monitoring. The lack of real-time or near-real-time progress tracking makes it difficult to adjust the rehabilitation plan, unable to timely reflect the changes and needs of patients, and may thus lead to delays in the rehabilitation process or poor effects. Summary of the Invention

[0003] The purpose of the present invention is to provide a pulmonary rehabilitation training method based on respiratory electromyogram signal feedback, so as to solve some of the drawbacks and deficiencies pointed out in the background art.

[0004] The technical solution adopted by the present invention to solve the above technical problems is as follows: A pulmonary rehabilitation training method based on respiratory electromyogram signal feedback, comprising: First, the electromyogram signal acquisition system is used to capture the electrical activities of the respiratory-related muscles in real time, and signal amplification, filtering, and noise suppression techniques are adopted to optimize the signal quality;

[0005] Secondly, use machine learning algorithms to perform time-frequency analysis on the EMG signals, extract energy spectrum and waveform length features, and apply deep learning models including convolutional neural networks for training and classification to distinguish different types of breathing patterns;

[0006] Then, based on the classification results of the EMG signals, establish a real-time feedback system that displays the EMG signal intensity and breathing quality in real time through the user interface and adjusts the difficulty and duration of the breathing training;

[0007] Finally, based on the analysis of the user's EMG signal characteristics and historical data, use artificial intelligence algorithms to generate personalized pulmonary rehabilitation training plans and dynamically adjust the training plan in combination with the user's physical condition and recovery progress.

[0008] Furthermore, the steps of the method for optimizing signal quality are as follows:

[0009] S1. First, develop an EMG sensor control module and use the formula to capture muscle activities related to breathing:

[0010]

[0011] to enhance signal selectivity, where x(t) represents the EMG signal at time t, and a, b, c are adjustment parameters;

[0012] S2. Then, apply an intelligent signal processing algorithm based on machine learning. Through the formula:

[0013]

[0014] where x(t) is the preprocessed EMG signal and k(t) is a weight function adaptively adjusted based on signal characteristics, classify the signal to distinguish different breathing patterns;

[0015] S3. Subsequently, based on the classification results of the EMG signals, construct a real-time feedback system. Through the formula:

[0016]

[0017] dynamically adjust the training difficulty, where x is the current signal feature value, and μ and λ n are training parameters to optimize the pulmonary rehabilitation training plan.

[0018] Furthermore, the methods for performing time-frequency analysis on the EMG signals include:

[0019] S1. Use multi-scale time-frequency analysis to capture the dynamic changes of the EMG signals at different time scales;

[0020] S2. Integrate the features in the time domain, frequency domain, and time-frequency domain through feature fusion technology to form a comprehensive feature vector, and apply a custom-developed deep learning model, especially a convolutional neural network optimized for the characteristics of electromyogram signals, for training and classification to distinguish different types of breathing patterns;

[0021] S3. Adjust the network architecture in the deep learning model, including the configuration of convolutional layers, pooling layers, and fully connected layers, to specifically adapt to the non-linear and non-stationary characteristics of electromyogram signals, and improve the accuracy of classifying different breathing patterns.

[0022] Furthermore, the use of the multi-scale time-frequency analysis method includes:

[0023] First, adopt a multi-scale waveform transformation with a variable window size to capture the dynamic changes of electromyogram signals at different time scales, which is suitable for analyzing the complexity of respiratory muscle activities and different breathing patterns. This transformation is achieved through the formula:

[0024] T(x,τ,s)=∫x(t)·w(s(t - τ))e -iωt dt

[0025] where x(t) represents the electromyogram signal, w is the window function, and s adjusts the window size to capture signals of different frequencies;

[0026] Then, adjust the analysis scale through an adaptive algorithm to optimize the time-frequency accuracy of signal analysis, which is controlled by the formula:

[0027]

[0028] where f(τ) is a function that changes dynamically according to the signal characteristics;

[0029] Finally, apply the technology combining empirical mode decomposition and Hilbert-Huang transform for advanced signal processing, and extract the time-frequency features of the signal through the formula:

[0030]

[0031] to provide real-time feedback and adjustment for different breathing training modes.

[0032] Furthermore, the construction framework of the custom-developed deep learning model includes:

[0033] S1. It is designed by integrating the features in the time domain, frequency domain, and time-frequency domain from electromyogram signals, and is achieved through the formula:

[0034]

[0035] where Respectively represent the transformation functions extracted from each domain, dealing with the diversity of breathing patterns required in pulmonary rehabilitation;

[0036] S2. Train and classify using a custom-developed convolutional neural network model optimized for the characteristics of electromyography signals. This model is executed through the formula:

[0037]

[0038] where the convolution operation captures subtle changes in the electromyography signal and distinguishes different breathing patterns including deep breathing and rapid breathing; among them, x represents the input electromyography signal; k n is the convolution kernel of the nth convolutional layer, used to extract specific features from the input signal; W nm is the weight matrix under the convolution kernel k n used to further refine the features extracted from the signal; b n is the bias term, used to adjust the bias of the convolution result; σ is the activation function, used to introduce non-linearity; the convolution operation * is used to apply the convolution kernel and weights to the signal x to extract features;

[0039] S3. Adopt a multi-modal learning strategy to enhance the model's understanding of the multi-dimensional features of electromyography signals through the formula:

[0040]

[0041] to optimize the training effect, where f i (x) represents the different modal feature extraction functions applied to the signal x; C represents the output convolutional neural network of the deep learning model; σ i is the activation function for each modality i, used to process the outputs of different modalities; λ i , μ i , are learning parameters, respectively adjusting the contribution of each modality.

[0042] Furthermore, the method for adjusting the network architecture in the deep learning model:

[0043] S1. First, it involves a convolutional layer that dynamically adjusts the filter size, using the formula:

[0044]

[0045] where θ n (s) represents the shape function of the nth convolution kernel, and σ n is the width parameter adjusted adaptively, representing the dynamic adjustment of the filter size to capture key features in the signal;

[0046] S2. Then, use depthwise separable convolution through the formula:

[0047]

[0048] Execute, where k mn is the depth convolution kernel, w n is the grouped weight matrix, b n is the bias term, which improves the feature extraction efficiency of the signal and reduces the computational complexity;

[0049] S3. Then, an adaptive pooling layer is adopted through the formula:

[0050]

[0051] to perform, where S i represents the i-th pooling region, μ i and σ i are the adaptive parameters of the center and width of the pooling window respectively, to capture and retain key features;

[0052] S4. Finally, a dynamic fully connected layer is adopted using the formula:

[0053]

[0054] where φ is the dynamic activation function, and its parameters λ j , w ji , and b j are dynamically adjusted according to the changes of the input features to optimize the response and classification accuracy of the network.

[0055] Furthermore, the steps for constructing the real-time feedback system include:

[0056] Firstly, using an advanced real-time data processing algorithm, through the formula:

[0057]

[0058] to perform, where x(t) is the electromyogram signal, and σ, ω, φ are dynamically adjusted parameters to capture the instantaneous features in the signal and generate real-time feedback;

[0059] Further, using an adaptive training algorithm, through the formula:

[0060]

[0061] to achieve, where α i , β i , γ i , λ i , δ i are adaptive parameters to dynamically adjust the difficulty and duration of the breathing training based on the real-time analysis results.

[0062] Furthermore, the generation of a personalized pulmonary rehabilitation training plan includes forming a comprehensive feature vector by combining features in the time domain, frequency domain, and time-frequency domain through a multi-dimensional feature extraction method. This process uses the formula:

[0063]

[0064] where T(x(t)) is the time-domain feature, F(x(f)) is the frequency-domain feature, H(x(t,f)) is the time-frequency domain feature, and α(t), β(f), γ(t,f) are dynamic weight functions;

[0065] Then, a personalized pulmonary rehabilitation training plan is generated using a deep reinforcement learning algorithm based on the user's electromyogram signal characteristics and historical data, adopting the formula:

[0066]

[0067] where Q(s,a) is the value function of state s and action a, r(s,a) is the reward function, and γ is the discount factor, representing the weight of future rewards;

[0068] Then, an adaptive algorithm is used to dynamically adjust the training plan in combination with the user's physical condition and recovery progress, through the formula:

[0069]

[0070] is achieved, where α i (t), β i (t), γ i (t), λ i , δ i are dynamically adjusted parameters. Through the comprehensive application of multi-dimensional feature extraction, deep reinforcement learning, and the adaptive algorithm, the analysis and personalized adjustment of the user's real-time condition and recovery progress are controlled.

[0071] Advantages of the present invention:

[0072] 1. Personalized training plan: By real-time monitoring and analyzing electromyogram signals, the present invention can generate a personalized training plan for each patient. This plan is dynamically adjusted according to the patient's current electromyogram activity and recovery progress, ensuring that the training is both appropriate and efficient.

[0073] 2. Real-time feedback and adjustment: By processing electromyogram signals through advanced algorithms, this method provides instant feedback and automatically adjusts the training intensity and duration according to the feedback results. This rapid response mechanism helps patients better control the training process and avoid overtraining and potential injuries.

[0074] 3. Enhance the scientificity and precision of training: By leveraging deep learning and reinforcement learning techniques, this method can deeply analyze the complex patterns of EMG signals, thereby more accurately identifying different breathing patterns and needs. This improves the scientificity and goal-orientation of the training program.

[0075] 4. Facilitate rapid recovery: Through the comprehensive use of time-domain, frequency-domain, and time-frequency-domain analysis, this method can comprehensively evaluate the dynamic changes of EMG signals, thereby designing a training plan that is most suitable for the patient's current state. This highly targeted training can accelerate the recovery process. Brief Description of the Drawings

[0076] Figure 1 It is a flowchart of a pulmonary rehabilitation training method based on respiratory EMG signal feedback of the present invention.

[0077] Figure 2 It is a flowchart of the method adopted for time-frequency analysis of EMG signals of the present invention. Detailed Embodiment

[0078] The following provides a detailed description of the specific embodiment of the present invention in conjunction with the drawings.

[0079] A pulmonary rehabilitation training method based on respiratory EMG signal feedback mainly involves using an EMG signal acquisition system to monitor the electrical activities of respiratory-related muscles in real time. In this process, the EMG signal acquisition system is equipped with special sensors that are attached to specific respiratory muscle areas of the user, such as the diaphragm and intercostal muscles, and can capture the electrical signals generated during respiratory movements in real time. These signals are usually affected by the electrical activities of other parts of the body, electronic device interference, or environmental noise. Therefore, signal amplification, filtering, and noise suppression techniques are required to optimize the signal quality. Signal amplification is to enhance the intensity of EMG signals so that they reach a measurable and analyzable level; filtering techniques are used to remove the interference of non-target frequencies in the signals, such as high-frequency noise or low-frequency drift; noise suppression techniques further eliminate the random noise introduced by electrical equipment or other interference sources.

[0080] After signal optimization, machine learning algorithms are then used to perform time-frequency analysis on the EMG signals, including using algorithms such as the Fast Fourier Transform (FFT) to transform time series data into the frequency domain, thereby extracting energy spectrum features, which measure the intensity of the overall energy distribution of the signal at different frequencies. In addition, waveform length features are also extracted, which are time-domain features reflecting the signal complexity and are measured by calculating the cumulative length of the signal waveform. These features are crucial for understanding the activity patterns of respiratory muscles and their changes. Next, a deep learning model, specifically a Convolutional Neural Network (CNN), is applied for training and classification. CNN is very effective in processing data with spatial and temporal hierarchical structures and can automatically detect and learn important features in the data without manual feature engineering. In this method, CNN uses the features extracted from the EMG signals to train the model to distinguish different types of respiratory patterns, such as normal breathing, deep breathing, rapid breathing, etc. In this way, the system can automatically identify and classify the user's respiratory state based on the changes in EMG signals, providing accurate data support for subsequent personalized pulmonary rehabilitation training.

[0081] Next, a real-time feedback system is established based on the classification results of the EMG signals. The intensity of the EMG signals and the user's respiratory quality are displayed in real time through the user interface. The interface provides a visual presentation of the real-time data, enabling the user and rehabilitation experts to immediately understand the current state of the respiratory activity. In addition, the system automatically adjusts the difficulty and duration of the breathing training according to the analyzed EMG signal categories (such as normal breathing, deep breathing, or rapid breathing). This adjustment is dynamic and optimized based on real-time data and preset training goals to adapt to the user's rehabilitation progress and current physical condition. For example, if the system detects that the user's breathing pattern becomes too rapid or shows difficulty, it can automatically reduce the training intensity or increase the rest time, and vice versa.

[0082] Finally, based on the analysis of the user's EMG signal characteristics and historical data, artificial intelligence algorithms are used to generate personalized pulmonary rehabilitation training programs. This method comprehensively analyzes the user's EMG signals through deep learning and pattern recognition techniques to identify their specific respiratory patterns and any abnormalities or changing trends. The algorithm takes into account the user's historical training data, evaluates the training effect, and predicts the most suitable training path. In addition, the system combines the user's current physical condition and recovery progress to dynamically adjust the training plan. This includes adjusting the difficulty, frequency, and duration of the training to ensure that the training plan can challenge the user without exceeding their physical tolerance. The system continuously monitors the user's response and makes adjustments based on real-time feedback, keeping the training program always updated to adapt to the user's rehabilitation needs.

[0083] Example 1:

[0084] In this embodiment, a certain patient experiences respiratory function decline due to chronic obstructive pulmonary disease (COPD) and requires systematic pulmonary rehabilitation training. For this purpose, the medical team plans to use electromyogram (EMG) signals to monitor and adjust the patient's training plan. This involves developing an EMG sensor control module and using a specific formula to capture respiratory-related muscle activities.

[0085] 1. Development and configuration of EMG sensors:

[0086] First, develop a high-sensitivity EMG sensor control module. These sensors are configured around the patient's diaphragm and intercostal muscles to monitor respiratory-related muscle activities in real time. The EMG signal x(t) will represent the intensity of EMG activity at time t.

[0087] 2. Signal processing formula:

[0088] The following formula is used for signal processing:

[0089]

[0090] where x(t) is the EMG signal at time t, and a, b, and c are adjustment parameters. This formula is designed to enhance the selectivity of the respiratory muscle activity signal.

[0091] 3. Parameter setting and example calculation:

[0092] Set the parameters a = 0.5, b = 0.3, c = 0.1. These parameters are adjusted to optimize the signal capture effect.

[0093] Set the measured value of the EMG signal x(t) at a certain time t to 0.05 mV, and its acceleration (second derivative) at this point to -0.01 mV / s 2 .

[0094] Substitute these values into the formula for calculation to simulate the effect of signal processing:

[0095]

[0096] This calculation will be estimated based on the real-time data of the EMG signal to determine the optimization degree of the signal and the timeliness of the feedback.

[0097] By using this signal optimization method, the quality of the EMG signal can be effectively improved, thus accurately capturing respiratory-related muscle activities. The enhanced signal selectivity helps the medical team better understand the patient's breathing pattern and muscle usage, and design a more accurate and personalized rehabilitation training plan for the patient. In addition, the real-time signal processing and feedback mechanism can dynamically adjust the difficulty and duration of the training to ensure that the training is both effective and in line with the patient's current physical condition.

[0098] The next step is to apply machine learning-based intelligent signal processing algorithms. Classification algorithms are used to distinguish different breathing patterns, such as normal breathing, deep breathing, rapid breathing, etc., so as to provide targeted training adjustments.

[0099] Use the algorithm formula:

[0100]

[0101] Classify the signal. Here, x(t) represents the electromyogram signal after preliminary filtering and noise suppression, and k(t) is the dynamic weight function, which is adaptively adjusted during the training process according to the characteristics of the signal.

[0102] Weight function setting:

[0103] Set in a specific time window T, the weight function k(t) is adaptively adjusted according to the variability of the electromyogram signal. For example, if the electromyogram signal shows high-frequency fluctuations at the beginning of training, then k(t) increases to enhance the influence of these parts, making it more sensitive to rapid changes. Specifically, the value of k(t) can be set to vary within the range of 0.1 to 0.5 to adapt to different signal intensities and variabilities.

[0104] Set during the training period, a series of electromyogram signals x(t) have been collected through electromyogram sensors. These data show periodic enhancement and weakening within the time from t = 0 to t = T, corresponding to the patterns of deep breathing and rapid breathing. Integrate this, apply dynamic weights, for example, k(t) = 0.3 during the peak of the electromyogram signal and k(t) = 0.1 during the low value period.

[0105] Calculation example: If the average value of x(t) is 0.05 millivolts and the entire T is 10 seconds, then (Schematic calculation, the actual value depends on the actual form and data of x(t) and k(t)).

[0106] Substitute this result into the classification function This value shows that according to the classification result of the electromyogram signal, the patient's current breathing pattern is of medium intensity, and the difficulty and duration of breathing training can be adjusted accordingly.

[0107] Next, use the classification result to construct a real-time feedback system that can dynamically adjust the difficulty of pulmonary rehabilitation training. This real-time feedback system performs mathematical formula operations and dynamically adjusts training parameters according to real-time signal eigenvalue to optimize the training plan and provide personalized training suggestions.

[0108] Construction of the real-time feedback system:

[0109] The system uses the formula:

[0110]

[0111] to adjust the training difficulty, where x is the current signal feature value, representing the real-time measurement value of the EMG signal obtained through the aforementioned analysis, and μ and λ n are training parameters adjusted according to historical data and training feedback.

[0112] Set μ as the preset target EMG signal intensity, and its value is set according to the patient's health condition and rehabilitation goal. For example, it is set to 0.05 mV.

[0113] λ n is the weight coefficient used to adjust the influence of terms of different orders. For example, λ 0 can be set to 0.5, λ 1 set to 0.3, and so on. The higher-order terms can gradually decrease. For example, λ 2 = 0.1, and the even higher-order terms can be set to approach zero.

[0114] Set that at a specific training moment, the real-time signal feature value x is measured as 0.06 mV.

[0115] Substitute these values into the formula for calculation:

[0116]

[0117] The result of this calculation is:

[0118] R(x) = 0.5·1 + 0.3·0.01 + 0.1·0.00005 = 0.5 + 0.003 + 0.0000005 ≈ 0.503

[0119] This result provides an adjustment coefficient to guide the adjustment of the training difficulty, such as by adjusting the duration or intensity of the breathing exercise. Through this real-time feedback system, the training difficulty can be dynamically adjusted according to the actual EMG signal intensity and the patient's performance, ensuring that the training plan always matches the patient's current ability and rehabilitation needs.

[0120] Example 2:

[0121] In this example, multi-scale time-frequency analysis is used to capture the dynamic changes of the EMG signal at different time scales. This method is particularly suitable for analyzing the complexity of respiratory muscle activity and identifying different respiratory patterns.

[0122] Implementation of multi-scale time-frequency analysis:

[0123] Multiscale time-frequency analysis is achieved by using multiscale waveform transforms with variable window sizes. This transform utilizes an adjusted window function w that changes with the scale parameter s to adapt to the characteristics of the electromyogram (EMG) signal x(t) at different time points τ and frequencies ω. By varying the window size s, the analysis can more finely capture signal components that change rapidly or slowly, suitable for different breathing patterns.

[0124] This method is implemented through the following formula:

[0125]

[0126] where x(t) represents the EMG signal, w is the window function that varies according to the scale s, τ is the time shift, and ω is the frequency. This formula allows for local time-frequency analysis of the signal, capable of revealing the detailed characteristics of the signal at specific times and frequencies.

[0127] Suppose a series of EMG signal data is being analyzed, which represents the muscle activity of a COPD patient during pulmonary rehabilitation training. Assume the window function w is selected as a Gaussian window, and the range of s can vary from 0.1 second to 1 second to adapt to muscle contractions and relaxations at different speeds.

[0128] Select a specific moment τ = 2 s and a medium scale s = 0.5 s for local analysis.

[0129] Calculate ω as the main frequency of the muscle activity, assumed to be 5 Hz.

[0130] Set the value of x(t) at t = 2 s to be 0.03 mV.

[0131] Substitute into the formula for calculation:

[0132] This calculation will give the time-frequency analysis result at τ = 2 s and s = 0.5 s, revealing the characteristics of the EMG signal at this specific moment and scale. Through this analysis, training experts can gain a detailed understanding of the patient's muscle electrical activity during different breathing training phases, thereby adjusting the training plan to better suit the patient's actual situation, such as adjusting the frequency, intensity, or duration of the breathing training.

[0133] Next, adjust the analysis scale through an adaptive algorithm to optimize the time-frequency accuracy of the signal analysis. The system dynamically optimizes the analysis parameters according to the real-time characteristics of the EMG signal, thereby ensuring the maximum accuracy and relevance of the signal analysis.

[0134] The adaptive algorithm uses the formula:

[0135]

[0136] Among them, T(x,τ,f(τ)) is the result of multi-scale waveform transformation at time offset τ and dynamically adjusted analysis frequency f(τ). This formula calculates the ratio of signal energy obtained by analyzing through a specific adaptive frequency function f(τ), relative to the analysis on the standard time scale τ. By maximizing this ratio, the algorithm can find the optimal analysis scale and frequency to capture key signal features.

[0137] After the electromyogram signal x(t) undergoes multi-scale waveform transformation processing, the measured data at a specific time offset τ shows that the signal exhibits different intensities at different frequency scales. It is set that the dynamic adjustment range of f(τ) is from 0.1 Hz to 10 Hz, which reflects the change from slow to fast breathing patterns.

[0138] Select a specific moment τ = 2 s and calculate |T(x, 2 s, 5 Hz)| 2 value and |T(x, 2 s, 2 s)| 2 value at the standard scale. Set the former to 0.04 unit energy and the latter to 0.02 unit energy.

[0139] Calculate according to the formula:

[0140]

[0141] This result indicates that analyzing the signal at an adjusted frequency of 5 Hz has more energy than analyzing it on the standard time scale, indicating that this is a more effective analysis frequency.

[0142] By implementing this adaptive algorithm, the system can automatically adjust the analysis time scale and frequency according to the real-time performance and changes of the electromyogram signal.

[0143] Finally, the combination technology of empirical mode decomposition (EMD) and Hilbert-Huang transform (HHT) is adopted for advanced signal processing. This technology combination can effectively extract the time-frequency characteristics of the electromyogram signal and provide real-time feedback and adjust the training mode according to these characteristics.

[0144] Empirical Mode Decomposition and Hilbert-Huang Transform:

[0145] Empirical Mode Decomposition (EMD) is an analysis method applicable to non-linear and non-stationary time series data. It first decomposes the data into multiple Intrinsic Mode Functions (IMFs), and each IMF should have a clear physical meaning, such as a specific frequency range. Subsequently, the Hilbert transform is applied to each IMF to extract its instantaneous frequency and amplitude, thereby constituting the time-frequency characteristics. Specifically, the formula:

[0146]

[0147] Describes how to reconstruct the time - frequency representation of a signal by combining the amplitudes and phases of each IMF, where the IMF n (t) is the nth intrinsic mode function and θ n (t) is the corresponding instantaneous phase.

[0148] Suppose that after performing EMD on the electromyogram of a certain patient, three significant IMF components are obtained, corresponding to different breathing frequencies respectively. Each IMF represents a specific breathing pattern (such as normal breathing, deep breathing, rapid breathing).

[0149] Suppose the first IMF represents the normal breathing pattern, specifically expressed as: IMF 1 (t)=0.05cos(2πt);

[0150] The second IMF represents deep breathing, IMF 2 (t)=0.08cos(0.5πt);

[0151] The third IMF represents rapid breathing, IMF 3 (t)=0.03cos(4πt).

[0152] Substituting these IMFs into the Hilbert - Huang transform formula, the signal reconstruction value H(x) at each moment can be calculated to evaluate the current breathing pattern:

[0153] H(x)=0.05cos(2πt)+0.08cos(0.5πt)+0.03cos(4πt)

[0154] This formula can be used to evaluate the patient's breathing status in real - time and dynamically adjust the training intensity according to signal changes.

[0155] By applying empirical mode decomposition and Hilbert - Huang transform, the time - frequency characteristics of electromyogram signals can be accurately extracted, and based on this, targeted breathing training feedback and adjustment can be provided for patients.

[0156] Example 3:

[0157] In this example, feature fusion technology is used to integrate features in the time domain, frequency domain, and time - frequency domain, and a deep - learning model developed for the characteristics of electromyogram signals is used for training and classification to identify different types of breathing patterns. It can accurately analyze and reflect the subtle changes in the patient's muscle activities, thereby providing a personalized rehabilitation training plan.

[0158] Application of feature fusion technology:

[0159] Feature fusion technology is used to integrate multi - dimensional data from electromyogram signals. By combining the time domain The intensity and frequency domain of electromyographic activity The signal frequency distribution and time-frequency domain The signal time-frequency analysis is performed to form a comprehensive feature vector. These feature vectors are calculated through the formula:

[0160]

[0161] where α(x), β(x), and γ(x) are adjustment parameters used to balance the contributions of each domain, and δ controls the scale and smoothness of the entire expression, ensuring the effective fusion of information extracted from multiple data sources.

[0162] Custom-developed deep learning model construction framework:

[0163] A customized convolutional neural network is adopted to adapt to the characteristics of electromyographic signals. The architecture of the model takes into account the non-linear and non-stationary characteristics of electromyographic signals. The network structure includes multiple convolutional layers, pooling layers, and fully connected layers, and each layer is optimized for specific characteristics of electromyographic signals to improve the accuracy and efficiency of classification.

[0164] It is set that in the processing of specific patient data, α(x) = 0.3, β(x) = 0.5, γ(x) = 0.2, and δ = 2. Such settings are to emphasize the frequency domain features because in electromyographic signal analysis, frequency features are often crucial for distinguishing different types of breathing patterns. Through this set of parameters, the influence of the time and frequency features of the signal can be balanced.

[0165] It is set that after the electromyographic signal data collected within a period of time is processed, According to the above formula and parameters, the calculation result of integrating the feature vectors will be:

[0166]

[0167] This calculation result will be used to train the deep learning model and further used for classifying different breathing patterns.

[0168] The training and classification of signals are carried out using a custom-developed convolutional neural network (CNN) optimized for the characteristics of electromyographic signals. This deep learning model combines advanced mathematics and machine learning techniques and is particularly suitable for processing and analyzing complex bio-signals such as electromyographic signals.

[0169] Model architecture:

[0170] This CNN model processes the input electromyographic signal x through stacked convolutional layers, and each layer is designed to capture specific features in the signal. The model uses the formula:

[0171]

[0172] Among them, k n is the convolution kernel of the nth convolutional layer, which is used to extract specific features from the input signal; W nm is the weight matrix under the convolution kernel k n and is used to further refine the features extracted from the signal; b n is the bias term, which is used to adjust the bias of the convolution result; σ is the activation function, which is used to introduce non-linearity; the convolution operation * is used to apply the convolution kernel and weights to the signal x to extract features.

[0173] It is set that the network design includes 3 convolutional layers, and each layer contains a different number of convolution kernels:

[0174] The first layer: 5 convolution kernels, each with a size of 3×3, which are used to capture the primary features of the signal.

[0175] The second layer: 10 convolution kernels, each with a size of 3×3, which are used to capture more complex features.

[0176] The third layer: 15 convolution kernels, each with a size of 3×3, which are used for deep feature extraction.

[0177] It is set that the weight W nm of each convolution kernel ranges within [-0.1, 0.1], and the bias b n ranges within [-0.05, 0.05].

[0178] It is set that in the real-time analysis of the electromyogram signal x(t), the input x is one-dimensional time series data. Through the convolution operation of each layer, the signal is converted into a series of feature maps, and these feature maps are passed to the next layer after passing through the activation function and further pooling operations. Finally, the output layer will classify based on the features extracted from the last layer to determine the breathing pattern (such as deep breathing or rapid breathing).

[0179] Through this method, the complex patterns of electromyogram signals can be effectively decoded and classified, so as to provide a customized training plan for patients according to their specific breathing conditions.

[0180] Furthermore, a multi-modal learning strategy is adopted to optimize the training effect, aiming to improve the understanding of the multi-dimensional features of electromyogram signals by the deep learning model and optimize the training effect through this understanding. The specific implementation plan includes using formulas to integrate the feature extraction and analysis results under different modalities, so as to more precisely adjust the training program.

[0181] The multi-modal learning strategy is implemented through the following formula:

[0182]

[0183] Among them, f i(x) represents different modal feature extraction functions applied to signal x, which involves analysis in the time domain, frequency domain, and time-frequency domain; C represents a deep learning model, namely a convolutional neural network, used to further process the output of these feature extractions; σ i is a specific activation function for each modality i, used to process the outputs of different modalities; λ i , μ i , are learning parameters used to adjust the importance, mean, and variance of the contributions of each modality.

[0184] Three modalities are set corresponding to three different types of respiratory pattern analysis:

[0185] λ 1 = 0.5, λ 2 = 0.3, λ 3 = 0.2 to adjust the contributions of each modality;

[0186] μ 1 = 0.05, μ 2 = 0.03, μ 3 = 0.02 represent the target outputs expected for each modality;

[0187] represents the variance of the outputs of each modality;

[0188] The features extracted from the real-time electromyogram signal x by the three modalities are fed into a customized CNN model and output respectively:

[0189] C(f 1 (x)) = 0.048

[0190] C(f 2 (x)) = 0.032

[0191] C(f 3 (x)) = 0.018

[0192] Substitute these values into the formula of the multi-modal learning strategy for calculation:

[0193]

[0194] This calculation helps to optimize the training scheme by integrating the analysis results of different modalities, ensuring that the training program can adapt to the specific needs and real-time performance of the patient.

[0195] Example 4:

[0196] In this embodiment, the network architecture is adjusted in the deep learning model to specifically adapt to the non - linear and non - stationary characteristics of the electromyogram (EMG) signals and improve the classification accuracy of different breathing patterns. By using a convolutional layer with dynamically adjusted filter sizes, the complex changes in the EMG signals can be effectively addressed.

[0197] Strategy for network architecture adjustment:

[0198] This adjustment of the deep learning model is mainly implemented through a convolutional layer with dynamically adjusted filter sizes, enabling the model to more flexibly adapt to the local features of the signals. This adjustment utilizes the formula:

[0199]

[0200] where θ n (s) represents the shape function of the n - th convolutional kernel, and σ n is the width parameter for adaptive adjustment, representing the dynamic adjustment of the filter size. This formula can adjust the response of the filter according to the local features in the EMG signals, thereby more precisely capturing the key signal features.

[0201] It is set that there are 10 different convolutional layers in the network, and the filter size σ n of each layer can be dynamically adjusted according to the frequency content of the signal. The adjustment range of the filter size is set from 0.1 s to 1 s, which allows the network to capture both very rapid breathing changes and relatively slow changes.

[0202] It is set that in a certain training instance, the EMG signal x(t) exhibits rapid peak changes within 2 s. σ n is adjusted to 0.2 s to facilitate capturing such rapid changes. For a specific θ n (s), if it is designed as a Gaussian function, it can be set as:

[0203]

[0204] Then, an integration operation is performed using the above parameters and the signal:

[0205]

[0206] This operation will be optimized for the rapid - changing peaks, enabling the network to more effectively identify and classify rapid breathing patterns. Through this method, the deep learning model can not only adapt to the complexity of the EMG signals but also improve the accuracy of breathing pattern classification by precisely capturing time - space features.

[0207] To further optimize the efficiency of signal processing and improve the accuracy of feature extraction, depthwise separable convolution technology is adopted. The network can reduce the computational complexity while maintaining efficient feature extraction ability when processing complex electromyogram signals.

[0208] Depthwise separable convolution is performed in two steps: first, each input channel is processed by depthwise convolution, and then the outputs of these channels are combined through pointwise convolution (1x1 convolution). This step-by-step processing greatly reduces the number of parameters and the computational burden. The specific formula is:

[0209]

[0210] where k mn is the depthwise convolution kernel for processing the depth features of the signal; w n is the weight matrix corresponding to the nth channel for combining these features; b n is the bias term.

[0211] For example, with 3 input channels and 2 depthwise convolution kernels per channel (M = 2, N = 3), the size of each convolution kernel is set to 3×3:

[0212] Set the weights of each depthwise convolution kernel k mn to be randomly initialized in the range [-0.1, 0.1].

[0213] The grouped weight matrix w n for the weights of each channel is also randomly initialized in the range [-0.1, 0.1].

[0214] The bias term b n is initially set to a small positive number such as 0.01 to avoid the nonlinear effect of the activation function near 0.

[0215] Set the input signal x to be a real-time data stream obtained from an electromyogram sensor, and its shape conforms to the network input requirements. After being processed by depthwise separable convolution, the signals of each channel are independently analyzed and recombined to form a new feature map. For the example calculation:

[0216]

[0217] This calculation will generate different characteristic responses for each breathing pattern. For example, the performances of deep breathing and rapid breathing patterns on the feature map will be significantly different, thus helping the model to classify more accurately and guide rehabilitation training. Through the application of this depthwise separable convolution technology, the lung rehabilitation training method can process electromyogram signals more efficiently, while reducing the consumption of hardware resources and improving the processing speed, making the rehabilitation training more precise and responsive.

[0218] Furthermore, an adaptive pooling layer is adopted to optimize feature extraction. By dynamically adjusting the size and position of the pooling window, key features in the signal can be captured and retained more precisely, which is particularly important when dealing with non-stationary and highly variable EMG signals.

[0219] The design of the adaptive pooling layer allows the size and position of the pooling window to be dynamically adjusted according to the characteristics of the signal, optimizing the local feature extraction of the signal. This is achieved through the following formula:

[0220]

[0221] where S i represents the i-th pooling region, and μ i and σ i are the adaptive parameters of the center and width of this region respectively. This method can adjust the pooling strategy according to the dynamic changes of the signal, so as to better retain useful information and remove irrelevant noise.

[0222] It is set that there are 5 pooling regions in the network design, i.e., I = 5, and the parameters of each region are dynamically adjusted according to the signal content. For example, the value range of μ i can be set between 0% and 100% of the signal length, and the value range of σ i is between 0.1 s and 1 s to adapt to EMG activities at different speeds.

[0223] Suppose at a certain moment, the EMG signal x(t) shows different activity intensities in five different regions, and the pooling window needs to be dynamically adjusted according to the center position and change speed of these activities. Suppose the μ i of each region are 1 s, 2 s, 3 s, 4 s, 5 s respectively, and σ i are all set to 0.5 s. According to these parameters, the pooling results are calculated for each region using the formula:

[0224]

[0225] This calculation method can weight the peaks of the signal in each region, highlight important features, and reduce the influence of background noise. Through the application of this adaptive pooling strategy, the deep learning model can process complex EMG signals more effectively, especially when the signal exhibits significant temporal variations or non-linear characteristics.

[0226] Finally, a dynamic fully connected layer is adopted, and a formula is used to adjust the parameters of the network to cope with the complexity of EMG signals. By dynamically adjusting the parameters and weights of the activation function, the response of the model to the changes of EMG signals can be optimized, thereby improving the accuracy and efficiency of classification.

[0227] The design of the dynamic fully connected layer is to enable the network to adjust its parameters according to the characteristics of the input signal, so as to better identify different breathing patterns. This is achieved through the following formula:

[0228]

[0229] where φ is the dynamic activation function, which can be ReLU, sigmoid or other advanced functions, and the specific choice depends on the characteristics of the signal; λ j , w ji , and b j are dynamically adjusted parameters, including the weights and bias terms of each neuron.

[0230] Design a network that contains, for example, 100 neurons (J = 100), and each neuron has 10 connection weights for the input features (I = 10). For the specific values of the parameters:

[0231] λ j varies in the range of [0.1, 2.0] and is adjusted according to specific training tasks and the desired output sensitivity.

[0232] w ji can be set to range from [-0.5, 0.5] to cover a wide range of input effects.

[0233] b j is usually in the range of [-0.1, 0.1] and is used to fine-tune the output threshold.

[0234] Set the input feature vector x = 0.05, -0.07, 0.15,..., 0.01 (a total of 10 input features) at a certain moment. For the first neuron, its weight and bias are initialized as:

[0235] w 1i = [0.1, -0.1, 0.05, …, -0.05]

[0236] b 1 = 0.05

[0237] λ 1 = 1.5

[0238] Using ReLU as the activation function, the output of this neuron is calculated as:

[0239]

[0240] If φ is ReLU, that is, φ(z) = max(0, z), the output is 0.25.

[0241] In this way, the dynamic fully connected layer not only improves the model's sensitivity to signal changes, but also can classify and respond to different breathing patterns more precisely.

[0242] Example 5:

[0243] In this example, a real-time feedback system is constructed, aiming to instantaneously adjust the training plan to adapt to the patient's current state. The system uses a real-time data processing algorithm, which uses mathematical formulas to analyze the electromyogram (EMG) signals and generate real-time feedback, which helps to optimize the training effect and improve the patient's rehabilitation efficiency.

[0244] The following formula of the real-time feedback system is used for real-time data processing:

[0245]

[0246] where x(t) is the EMG signal measured in real time, and σ, ω, and φ are dynamically adjusted parameters, representing the width of the filtering window, the angular frequency of signal processing, and the phase respectively. These parameters can be adjusted according to the characteristics of the EMG signal and the required feedback sensitivity.

[0247] Set σ to 0.1 s, so that signal changes within a shorter time scale can be captured, which is suitable for fast-response EMG activities.

[0248] Set ω to 2π×10, that is, the signal processing frequency is 10 Hz, which is a typical frequency setting and can capture EMG signals related to common breathing rates.

[0249] φ can be set to π / 4 to provide an initial phase, which helps to adjust the starting point of signal analysis.

[0250] Set at a certain moment t = 2 s, the value of the EMG signal x(t) is 0.05 mV. According to the formula, the following calculations can be performed:

[0251]

[0252] This integration will be carried out within the local area of the signal, highlighting the signal characteristics at the moment t = 2 s. Numerical integration methods need to be used for actual calculations, which can be carried out using calculation software or special algorithms. Through this method, the real-time feedback system can provide precise training feedback according to the instantaneous changes of the EMG signal, thus realizing highly personalized rehabilitation training.

[0253] An adaptive training algorithm is adopted, which can dynamically adjust the difficulty and duration of training according to the real-time analysis results of EMG signals, thus realizing highly personalized and precise training.

[0254] Design principle of the adaptive training algorithm:

[0255] This algorithm utilizes advanced mathematical models to analyze EMG signals and adjusts training parameters according to signal characteristics. The formula is as follows:

[0256]

[0257] Where x i (τ) is the real-time EMG signal, and α i , β i , γ i , λ i , and δ i are adaptive parameters. These parameters are adjusted according to the real-time analysis results of EMG signals to optimize the training effect.

[0258] The system settings are set with three groups of adaptive parameters, and each group processes different signal characteristics respectively:

[0259] α i is adjusted within the range of [0.1, 1.0], affecting the sensitivity of the signal derivative term.

[0260] β i The range, [0.01, 0.1], determines the smoothness of the signal and the memory decay rate.

[0261] γ i is within the range of [0.05, 0.5], adjusting the contribution of the periodic component.

[0262] λ i is set in the range of [1, 10], affecting the frequency of the periodic function.

[0263] δ i is adjusted within the range of [0, π], used to adjust the phase difference.

[0264] Assume that at a certain moment t, the EMG signal x i (τ) shows a specific activity pattern. For simplicity of calculation, consider the signal pattern x 1 (τ) = 0.1sin(2πτ):

[0265]

[0266] This formula dynamically adjusts the training parameters by separately calculating and integrating the effects of signal attenuation and periodic changes.

[0267] Example 6:

[0268] In this example, a personalized training plan is generated. Through comprehensive multi-dimensional feature extraction in the time domain, frequency domain, and time-frequency domain, in-depth analysis of the patient's EMG signals is achieved. By applying mathematical formulas, this method can form a comprehensive feature vector and provide a customized training plan for each patient.

[0269] Multi-dimensional Feature Extraction Method:

[0270] This method uses the following formula for multi-dimensional feature extraction:

[0271]

[0272] Where T(x(t)) represents the time-domain feature, F(x(f)) represents the frequency-domain feature, and H(x(t,f)) represents the time-frequency domain feature. These features are extracted from the electromyogram signal and can reflect the basic and complex properties of the signal. The dynamic weight functions α(t), β(f), and γ(t,f) are used to adjust the influence of different features to optimize the weight distribution in the feature extraction process.

[0273] Set the value of σ to 0.5. This parameter controls the width of the Gaussian window and affects the local sensitivity of feature extraction.

[0274] τ can be regarded as the central moment of analysis. For real-time analysis, this can be the current time or an important time point.

[0275] The dynamic weights α(t), β(f), γ(t,f) vary within the range of [0.1, 1.0] according to the specific application scenario and signal characteristics.

[0276] Suppose the electromyogram signal x(t) at a certain moment t shows a spike in the time domain, the main frequency in the frequency domain is 10 Hz, and the time-frequency domain analysis reveals the change of the signal energy over time:

[0277] T(x(t)) = sin(2π·10·t);

[0278] F(x(f)) = δ(f - 10) (δ is the Dirac function, indicating a peak at a frequency of 10 Hz);

[0279] H(x(t,f)) is expressed as sin(2π·10·t) × δ(f - 10);

[0280] Apply the formula to calculate the feature vector:

[0281]

[0282] This integral will focus on the signal characteristics at t = τ, providing detailed information about the current breathing pattern.

[0283] Adopt a deep reinforcement learning algorithm to generate a personalized training plan. By evaluating the electromyogram signal characteristics and historical data of the user, the training program is dynamically adjusted to maximize the rehabilitation effect. Deep reinforcement learning mainly optimizes the long-term reward by learning the optimal action strategy.

[0284] Implementation Strategies of Deep Reinforcement Learning:

[0285] The deep reinforcement learning model calculates the value of an action through the following formula:

[0286]

[0287] where Q(s,a) is the value function for the given state s and action a; r(s,a) is the immediate reward obtained by executing action a in state s; γ is the discount factor used to adjust the impact of future rewards; s ′ and a ′ represent the subsequent state and action respectively.

[0288] γ is usually set between 0.9 and 0.99, which can ensure that the model does not ignore the long-term benefits while considering the immediate reward.

[0289] The reward function r(s,a) is designed based on the improvement degree of the electromyogram signal and the rehabilitation goal. For example, if a certain training action can effectively relieve symptoms or enhance muscle function, a positive reward should be given; conversely, if the action causes symptom deterioration or is ineffective, a negative reward should be given.

[0290] Suppose within a certain training period, the user performs a specific breathing training action, and the monitored electromyogram signal shows enhanced muscle activity, indicating good training effect:

[0291] Suppose the state s represents the electromyogram state before training, and the action a is the performed breathing training action.

[0292] The immediate reward r(s,a) = 5, indicating that this action has a significant positive impact on muscle recovery.

[0293] Suppose the discount factor γ = 0.95.

[0294] Suppose through model prediction, the predicted value Q(s ′ of the best action a ′ in the subsequent state s ′ ,a ′ ) = 50.

[0295] Substitute these values into the reinforcement learning formula to calculate the value of the current action:

[0296] Q(s,a) = 5 + 0.95×50 = 52.5

[0297] This indicates that executing this breathing training action has a high overall value in the current state.

[0298] The training plan is dynamically adjusted using an adaptive algorithm in combination with the user's physical condition and recovery progress, which involves real-time analysis of the user's electromyogram (EMG) signals and adjustment of the training intensity and mode according to the analysis results, so as to ensure that the training plan not only adapts to the user's current health status but also meets the requirements of their recovery progress.

[0299] Implementation strategy of the adaptive algorithm:

[0300] The adaptive training plan is achieved through the following formula:

[0301]

[0302] where x i (τ) represents the i-th dimensional feature extracted from the user's EMG signal, and α i (t), β i (t), γ i (t), λ i , and δ i are dynamically adjusted parameters, which are adjusted according to the user's real-time physical condition and recovery progress.

[0303] It is set that the system monitors and adjusts five different training parameters, N = 5.

[0304] Parameter adjustment range: α i (t) is between [0.1, 1.0], controlling the sensitivity of the derivative term; β i (t) is between [0.01, 0.1], adjusting the attenuation rate; γ i (t) is between [0.05, 0.5], adjusting the influence of the periodic term; λ i is between [1, 10], adjusting the frequency of the periodic function; δ i is between [0, π], controlling the phase.

[0305] A specific training period is set, in which the user's EMG signal x i (τ) exhibits obvious periodic or attenuation characteristics for each dimension i. Taking one dimension as an example, α 1 (t) = 0.5, β 1 (t) = 0.05, γ 1 (t) = 0.3, λ 1 = 5, δ 1 = π / 4.

[0306] Calculations are performed using the above parameters:

[0307]

[0308] This formula will adjust the training intensity and duration based on real-time data to ensure that the training meets the specific recovery needs of the user.

[0309] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and what is described in the above embodiments and the specification is only to illustrate the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will also have various changes and improvements, and these changes and improvements fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A pulmonary rehabilitation training method based on respiratory electromyographic signal feedback, characterized in that The following steps are involved: First, the electrical activity of breathing-related muscles is captured in real time through the electromyographic signal acquisition system, and signal amplification, filtering and noise suppression techniques are used to optimize signal quality. Secondly, we use machine learning algorithms to perform time-frequency analysis on the electromyographic signals, extract the energy spectrum and waveform length features, and apply deep learning models including convolutional neural networks for training and classification to distinguish different types of breathing patterns. Then, based on the classification results of the electromyographic signals, a real-time feedback system is established. The system displays the electromyographic signal strength and breathing quality in real time through the user interface, and adjusts the difficulty and duration of breathing training; Finally, based on the user's electromyographic signal characteristics and historical data analysis, an artificial intelligence algorithm is used to generate a personalized pulmonary rehabilitation training program, and the training plan is dynamically adjusted based on the user's physical condition and recovery progress; The real-time feedback system construction steps include: First, we use advanced real-time data processing algorithms, through the formula: , where x(t) is the electromyographic signal, σ, ω, φ are dynamically adjusted parameters to capture the instantaneous features in the signal and generate real-time feedback; Further, using the adaptive training algorithm, through the formula: Implementation, where α i , β i , γ i , i , δ i For adaptive parameters, the difficulty and duration of breathing training are dynamically adjusted based on real-time analysis results; The generating of the personalized pulmonary rehabilitation training program includes forming a comprehensive feature vector by combining the features of the time domain, the frequency domain and the time-frequency domain through a multi-dimensional feature extraction method. The process uses the formula: Where T(x(t)) is the time domain feature, F(x(f)) is the frequency domain feature, H(x(t,f)) is the time-frequency domain feature, α(t), β(f), γ(t,f) are dynamic weight functions; Then, a deep reinforcement learning algorithm is used to generate a personalized pulmonary rehabilitation training program based on the user's electromyographic signal characteristics and historical data, using the formula: Where Q(s,a) is the value function of state s and action a, r(s,a) is the reward function, and γ is the discount factor, which represents the weight of future rewards; Then, the training plan is dynamically adjusted using an adaptive algorithm based on the user's physical condition and recovery progress, using the formula: Implementation, where α i (t), β i (t), γ i (t), λ i , δ i It is a dynamically adjusted parameter that controls the analysis and personalized adjustment of the user's real-time status and recovery progress through the comprehensive application of multi-dimensional feature extraction, deep reinforcement learning and adaptive algorithms.

2. A pulmonary rehabilitation training method based on respiratory myoelectric signal feedback according to claim 1, characterized in that The steps of the method for optimizing signal quality are as follows: S1. First, develop the myoelectric sensor control module and use the formula to capture the muscle activity related to breathing: To enhance signal selectivity, where x(t) represents the electromyographic signal at time t, and a, b, and c are adjustment parameters; S2, then apply the intelligent signal processing algorithm based on machine learning, through the formula: Where x(t) is the preprocessed electromyographic signal, k(t) is the weight function adaptively adjusted based on the signal characteristics to classify the signal and distinguish different breathing patterns; S3. Then, according to the classification results of the electromyographic signals, a real-time feedback system is constructed, using the formula: Dynamically adjust the training difficulty, where x is the current signal feature value, μ and λ n To optimize the pulmonary rehabilitation training program for training parameters.

3. A pulmonary rehabilitation training method based on respiratory myoelectric signal feedback according to claim 1, characterized in that The method for performing time-frequency analysis on the electromyographic signal comprises: S1. Use multi-scale time-frequency analysis to capture the dynamic changes of EMG signals at different time scales. S2. The features of time domain, frequency domain and time-frequency domain are integrated to form a comprehensive feature vector through feature fusion technology, and a custom-developed deep learning model, especially a convolutional neural network optimized for electromyographic signal characteristics, is used for training and classification to distinguish different types of breathing patterns; S3. By adjusting the network architecture in the deep learning model, including the configuration of convolutional layers, pooling layers, and fully connected layers, it is particularly adapted to the nonlinear and non-stationary characteristics of electromyographic signals and improves the accuracy of classifying different breathing patterns.

4. A pulmonary rehabilitation training method based on respiratory myoelectric signal feedback according to claim 3, characterized in that The multi-scale time-frequency analysis method includes: First, a multi-scale waveform transform with a variable window size is used to capture the dynamic changes of the electromyographic signal at different time scales, which is suitable for analyzing the complexity of respiratory muscle activity and different breathing patterns. The transform is expressed by the formula: T(x,τ,s)=∫x(t)·w(s(t-τ))e -iωt dt Implementation, where x(t) represents the electromyographic signal, w is the window function, and s adjusts the window size to capture signals of different frequencies; Then the analysis scale is adjusted through an adaptive algorithm to optimize the time-frequency accuracy of signal analysis, according to the formula: control, where f(τ) is a function that changes dynamically according to the signal characteristics; Finally, the empirical mode decomposition and Hilbert-Huang transform technology are combined for advanced signal processing, through the formula: Extract the time-frequency features of the signal to provide real-time feedback and adjustments for different breathing training modes.

5. A pulmonary rehabilitation training method based on respiratory myoelectric signal feedback according to claim 3, characterized in that The custom developed deep learning model building framework includes: S1, designed by integrating the time domain, frequency domain and time-frequency domain features from the electromyographic signal, through the formula: to achieve, where They represent the transformation functions extracted from each domain, respectively, to handle the diversity of breathing patterns required in pulmonary rehabilitation; S2. Training and classification are performed using a custom-developed convolutional neural network model optimized for EMG signal characteristics, which is implemented using the formula: The convolution operation captures subtle changes in the EMG signal and distinguishes different breathing patterns including deep breathing and rapid breathing; where x represents the input EMG signal; k n is the convolution kernel of the nth convolutional layer, which is used to extract specific features from the input signal; W nm is the convolution kernel k n The weight matrix under is used to further refine the features extracted from the signal; b n is a bias term used to adjust the bias of the convolution result; σ is an activation function used to introduce nonlinearity; the convolution operation * is used to apply the convolution kernel and weights to the signal x to extract features; S3. Use a multimodal learning strategy to improve the model's understanding of the multidimensional features of electromyographic signals, using the formula: To optimize the training effect, where f i (x) represents the different modal feature extraction functions applied to the signal x; C represents the convolutional neural network output of the deep learning model; σ i is the activation function for each modality i, used to process the output of different modalities; i ,μ i , are learning parameters that tune each modal contribution separately.

6. A pulmonary rehabilitation training method based on respiratory myoelectric signal feedback according to claim 3, characterized in that Method for adjusting the network architecture in the deep learning model: S1. First, the convolution layer involves dynamically adjusting the filter size, using the formula: where θ n (s) represents the shape function of the nth convolution kernel, σ n The adaptive width parameter represents the dynamic adjustment of the filter size to capture the key features in the signal; S2, then use the depth-separable convolution formula: Execute, where k mn is the depth convolution kernel, w n is the group weight matrix, b n It is a bias term that improves the efficiency of signal feature extraction and reduces computational complexity; S3, then use the adaptive pooling layer through the formula: Carry out, where S i represents the i-th pooling area, μ i and σ i are adaptive parameters of the center and width of the pooling window to capture and retain key features; S4. Finally, the dynamic fully connected layer is used using the formula: where φ is a dynamic activation function with parameter λ j ,w ji , and b j Dynamically adjust according to changes in input features to optimize the network's response and classification accuracy.

Images