A lower limb rehabilitation training scheme making method and system based on electromyographic signal feedback

Abstract

The application discloses a lower limb rehabilitation training scheme making method and system based on electromyographic signal feedback and belongs to the lower limb rehabilitation guidance field, and comprises the following steps: S1, acquiring a multi-modal original data set and initializing a parameter set; S2, pre-processing the multi-modal original data set to construct a high-dimensional feature set; S3, obtaining a motion intention result and a training state parameter based on the high-dimensional feature set through an attention mechanism LSTM model; S4, dynamically outputting a target resistance torque based on the motion intention and the training state parameter; S5, designing a sliding mode controller to output a control instruction and a training action execution result based on the target resistance torque and the motion intention; and S6, outputting an optimized training scheme based on the execution result and the state parameter. The lower limb rehabilitation training scheme making method and system based on electromyographic signal feedback realize accurate perception, individualized adjustment, safe execution and full-dimensional evaluation of lower limb rehabilitation training.

Description

Technical Field

[0001] This invention relates to the field of technology, and in particular to a method and system for developing lower limb rehabilitation training programs based on electromyographic signal feedback. Background Technology

[0002] The aging of the global population, coupled with the high incidence of neurological diseases such as stroke, spinal cord injury, and cerebral palsy, has led to a year-on-year increase in the number of patients with lower limb movement disorders.

[0003] Traditional rehabilitation mainly relies on manual assistance from therapists, which can only provide basic movement guidance and lacks quantitative data support and personalized adjustments, resulting in problems such as low efficiency, resource shortage, and difficulty in quantifying effects.

[0004] Some intelligent methods typically use a single sensor (such as a simple EMG or IMU) to collect data, which cannot comprehensively capture multi-dimensional states such as muscle activation, joint movement, and body stability, resulting in large deviations in intent recognition. At the same time, the control strategy is mainly based on fixed-parameter PID, which can only achieve basic trajectory tracking and cannot cope with the dynamic needs brought about by changes in the trainee's muscle state and the switching of different training modes, resulting in large torque / angle tracking errors. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for developing lower limb rehabilitation training programs based on electromyographic signal feedback, thereby solving the aforementioned technical problems.

[0006] To achieve the above objectives, the present invention provides a method for developing a lower limb rehabilitation training program based on electromyographic signal feedback, comprising the following steps: S1. Combining trainee basic information, training mode selection, and training target parameters, simultaneously collect lower limb surface electromyography signals, motion data, and heart rate change rate data to obtain a multimodal raw dataset; and based on trainee information, training target, and training mode, obtain an initial parameter set; S2. Wavelet threshold denoising, Kalman filtering and Savitzky-Golay smoothing are performed on the multimodal raw dataset output by S1 to extract surface electromyography signal features, and motion features and pressure features are fused to construct a high-dimensional feature set. S3. Based on the high-dimensional feature set output by S2, the trainee's lower limb joint movement intention is identified through the attention mechanism LSTM model. Combined with surface electromyography signals, the movement intention results and training state parameters are obtained. S4. Based on the motion intent and training state parameters output by S3, combined with the training mode of S1, a multi-mode resistance torque basic model is constructed. The quantization factor and scaling factor of the fuzzy PID are optimized through the sparrow search algorithm, and muscle fatigue correction and safety boundary verification are incorporated to dynamically output the target resistance torque. S5. Based on the target resistance torque output by S4 and the motion intention of S3, a sliding mode controller is designed to generate initial control commands. The control commands are corrected through real-time feedback of joint angles. The control accuracy and system complexity are optimized by combining the TOS balancing mechanism. The control commands and training action execution results are then output. S6. Based on the execution results of S5 and the state parameters of S3, a quantitative evaluation index set is constructed from three dimensions: joint function, electromyographic signal consistency, and training state. The optimized training scheme is output through iterative optimization using the reinforcement learning DQN algorithm.

[0007] A system for implementing a method for developing a lower limb rehabilitation training program based on electromyographic signal feedback includes: Data acquisition module: It is used to combine trainee basic information, training mode selection and training target parameters to simultaneously collect lower limb surface electromyography signals, motion data and heart rate change rate data to obtain a multimodal raw dataset; and to obtain an initial parameter set based on trainee information, training target and training mode. The data preprocessing and fusion module is used to perform wavelet threshold denoising, Kalman filtering and Savitzky-Golay smoothing on the multimodal raw dataset, extract surface electromyography signal features, fuse motion features and pressure features, and construct a high-dimensional feature set. The intent recognition and dynamic evaluation module is used to identify the lower limb joint movement intent of trainees based on a high-dimensional feature set and through an attention mechanism LSTM model. Combined with surface electromyography signals, it obtains the movement intent results and training state parameters. The target resistance torque calculation module is used to construct a multi-mode resistance torque basic model based on the exercise intention and training state parameters, combined with the training mode. It optimizes the quantization factor and scaling factor of the fuzzy PID through the sparrow search algorithm, incorporates muscle fatigue correction and safety boundary verification, and dynamically outputs the target resistance torque. The closed-loop control module is used to design a sliding mode controller to generate initial control commands based on the target resistance torque and motion intention. It corrects the control commands through real-time feedback of joint angles, optimizes control accuracy and system complexity by combining the TOS balancing mechanism, and outputs control commands and training action execution results. The training effect evaluation and optimization module is used to construct a set of quantitative evaluation indicators based on the execution results and state parameters from three dimensions: joint function, electromyographic signal consistency, and training state. It is then iteratively optimized through the reinforcement learning DQN algorithm to output the optimized training scheme.

[0008] Therefore, the present invention employs the above-mentioned method and system for developing lower limb rehabilitation training programs based on electromyographic signal feedback, which has the following beneficial effects: 1. Improved perception accuracy: By integrating multimodal data such as surface electromyography (EMG) signals, IMU, and ECG signal peak time, and by improving wavelet threshold denoising and NNMF muscle coordination feature extraction, the accuracy of motion intention recognition is ≥96%, and the correlation coefficient of EMG signals is ≥0.87, providing high-quality data support for subsequent control. 2. Control Adaptive Enhancement: By adjusting the resistance torque, combined with sliding mode control and trajectory correction, the torque tracking error is ≤0.25Nm and the angle tracking error is ≤0.5°, adapting to the personalized needs of different GMFCS graded trainers in multiple training modes such as isokinetic / AAN / isochronous. 3. Comprehensive assessment guarantee: Quantitative assessment from all dimensions including joint function (ROM, tracking accuracy), electromyographic consistency (signal correlation coefficient, activation balance), and training status (fatigue index, compliance) to avoid misjudgment of effect due to a single indicator; 4. Iterative optimization of the solution: Based on reinforcement learning (DQN), a closed-loop optimization model is constructed to dynamically adjust parameters such as training duration and torque range to address training shortcomings, thereby achieving continuous improvement through "training-evaluation-optimization" and accelerating the rehabilitation process; 5. Controllable safety risks: Through electrode impedance verification, torque safety boundary verification (angle / torque change rate constraint), and fatigue perception load reduction, it ensures that there are no risks such as joint damage or excessive muscle fatigue during training. GMFCS Level 4 trainees can also use it safely. 6. Improved training compliance: Lightweight sensor deployment, real-time feedback mechanism, and personalized training intensity adaptation reduce the physical burden and psychological resistance of trainees, with an average compliance score of ≥0.7, thus improving the sustainability of rehabilitation training.

[0009] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a method for developing a lower limb rehabilitation training program based on electromyographic signal feedback, as described in this invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0012] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0013] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0014] like Figure 1 As shown, a method for developing a lower limb rehabilitation training program based on electromyographic signal feedback includes the following steps: S1. Combining trainee basic information, training mode selection, and training target parameters, simultaneously collect lower limb surface electromyography signals, motion data, and heart rate change rate data to obtain a multimodal raw dataset; and based on trainee information, training target, and training mode, obtain an initial parameter set; S2. Wavelet threshold denoising, Kalman filtering and Savitzky-Golay smoothing are performed on the multimodal raw dataset output by S1 to extract surface electromyography signal features, and motion features and pressure features are fused to construct a high-dimensional feature set. S3. Based on the high-dimensional feature set output by S2, the trainee's lower limb joint movement intention is identified through the attention mechanism LSTM model. Combined with surface electromyography signals, the movement intention results and training state parameters are obtained. S4. Based on the motion intent and training state parameters output by S3, combined with the training mode of S1, a multi-mode resistance torque basic model is constructed. Through the quantization factor and scaling factor of fuzzy PID, muscle fatigue correction and safety boundary verification are incorporated to dynamically output the target resistance torque. S5. Based on the target resistance torque output by S4 and the motion intention of S3, a sliding mode controller is designed to generate initial control commands. The control commands are corrected through real-time feedback of joint angles. The control accuracy and system complexity are optimized by combining the TOS balancing mechanism. The control commands and training action execution results are then output. S6. Based on the execution results of S5 and the state parameters of S3, a quantitative evaluation index set is constructed from three dimensions: joint function, electromyographic signal consistency, and training state. The optimized training scheme is output through iterative optimization using the reinforcement learning DQN algorithm.

[0015] Step S1 specifically includes the following steps: S11. Sensor Deployment and Calibration: The hybrid polymer hydrogel electrodes are attached to the target muscles of the lower limbs of the trainee, and IMU sensors are deployed on the thighs, calves and feet. Pressure sensors are also deployed on the left and right symmetrical positions of the forefoot and heels. The hybrid polymer hydrogel electrodes, IMU sensors and pressure sensors are then calibrated respectively. S12. Initialize training parameters , These represent the sensor sampling frequency, maximum safe torque, minimum safe torque, initial parameters of the training mode, data storage parameters, and peak time of the maximum ECG signal, respectively. The initial parameters of the training mode include the initial angular velocity under the isokinetic training mode (where joint movement speed is constant and resistance varies with the magnitude of force exerted; this mode is used in the muscle strengthening phase and is suitable for individuals with moderate functional impairment). The initial assistance coefficient under the AAN training mode (on-demand assistance, dynamically adjustable assistance intensity, patient-centered, suitable for all stages of rehabilitation, especially for those with moderate to severe functional impairment) Isometric training mode (fixed joint angle, trainee resists fixed resistance, used for early rehabilitation, suitable for postoperative stability training) fixed training angle Data storage parameters include data storage frequency. Real-time feedback cycle and data cache size , ; in, ; ; ; ; ; ; In the formula, This represents the basic moment coefficient, and ; Indicates the quality of trainees; This represents the GMFCS classification correction factor, and ; This represents the age modification factor; Indicates the trainee's target joint range of motion; It indicates the trainee's expected muscle activation level, which is divided into 12 levels, with the lowest being 1 and the highest being 12; S13, Multimodal data acquisition; S131. Data Acquisition Start: The trainee starts the action according to the preset training mode. The data acquisition card synchronously acquires all sensor signals and displays sensor impedance, signal amplitude, angle and heart rate change rate data in real time during the acquisition process. If the impedance exceeds the standard or the signal is saturated, the acquisition is paused and the problem is investigated. S132. Data storage: Collected data is named according to trainer ID-training mode-timestamp, written to local storage in real time, and backed up to the cloud at the same time; S133. When the training time is reached or the trainer triggers a stop command, the data collection stops and the original multimodal dataset is obtained.

[0016] The target muscles of the lower limbs of the trainee mentioned in step S11 include the rectus femoris, medial head of the gastrocnemius, peroneus longus, and tibialis posterior.

[0017] Step S2 specifically includes the following steps: S21. Targeted denoising of multimodal data: Denoising and enhancement of lower limb surface electromyography signals in multimodal raw datasets: 4th order Butterworth bandpass filter is used to remove power frequency noise and baseline drift, and then wavelet threshold denoising algorithm is used to further suppress motion artifacts; For motion data, Kalman filtering calibration is used to eliminate drift errors; For heart rate variability data, Savitzky-Golay filtering is used to smooth the variance of heart rate variability, preserving peak features while suppressing noise; S22, Multi-dimensional Feature Extraction: The time-domain, frequency-domain, time-frequency-domain, and muscle coordination features of surface electromyography (EMG) signals were extracted. The time-domain features included the root mean square (RMS), mean absolute value, and waveform length of the EMG signal. The frequency-domain features included the peak power spectral density and center frequency of the EMG signal calculated using the Welch method. The time-frequency-domain features were derived from wavelet packet decomposition of the EMG signal, extracting the energy proportions of each subband. Muscle coordination features were extracted using nonnegative matrix factorization. Extracting the peak time of the pressure signal from the pressure data and plantar pressure standard deviation ; Extracting the peak time of the electrocardiogram signal and heart rate variation variance ; S23. Feature Standardization and Temporal Sequence Construction: The multi-dimensional features extracted in step S22 are standardized using Z-score to obtain standardized features. ; S24. Weighted Fusion: An ensemble learning algorithm (random forest) is used to calculate the importance of each feature, and the weight vector is obtained after normalization. Then perform weighted fusion: ; In the formula, This indicates the fusion of high-dimensional feature sets; Indicates the first The first sample Dimensional standard features.

[0018] Step S3 specifically includes the following steps: S31. Temporal Enhancement: Extracting weighted fusion features using a sliding window. This process generates a temporal input, and then normalizes the trainee's electromyography (EMG) signals as global features, which are then concatenated to the end of each temporal window to form an enhancement sequence. ; S32, Design the modal feature attention layer; S321, Enhance the sequence They are divided into three groups: muscle synergy characteristics, electromyography characteristics, and exercise stress characteristics. S322. Map the features to the query vector using a learnable weight matrix. Key vector Sum value vector : ; ; ; In the formula, , and All represent learnable weight matrices; S323. Introduce grouping weight coefficients and calculate the query vector. With key vector similarity : ; In the formula, Indicates the transpose operation; Indicates the scaling factor; This represents the weight coefficient of the corresponding group; Indicates a block mask; S324, Weight Normalization: ; In the formula, Indicates attention weight; This represents the Softmax function; Indicates the length of the time window; The first in the attention mechanism The key vector at each time step; S325, Temporal features after attention weighting ; S33. Construct an attention-based LSTM model, adapting the input layer dimensional features of the attention-based LSTM model to the dimensionality of the high-dimensional feature set; assign weights to the muscle coordination features in the attention layer. The output layer contains the motion category and the expected angle. S34. Train an LSTM model with an attention mechanism using a high-dimensional feature set, and the loss function expression is as follows: ; in, ; ; In the formula, Indicates the total loss; Indicates the balance coefficient; Indicates classification loss; Indicates regression loss; Indicates the number of training samples; Indicates the first The true category of each sample Indicates category index, These represent the electromyographic signal output channels corresponding to lower limb joint dorsiflexion, lower limb joint flexion, lower limb joint inversion, and lower limb joint eversion, respectively. Indicates the predicted category probability; Indicates the first The prediction angle for each sample; Indicates the first The true angle of each sample; S35. Input the enhanced features to be identified into the trained attention mechanism LSTM model, and output the motion intent result. , These represent the output motion categories respectively. and expected motion angle ,and ; S36. Calculate the muscle fatigue index based on the attenuation of peak intensity of surface electromyography signals, the slowing of the slope of the flat segment, and the amplitude of heart rate rise, combined with center frequency. : ; In the formula, These represent the real-time peak electromyography intensity and the initial peak electromyography intensity, respectively. and These represent the real-time muscle level straight-segment slope and the initial muscle level straight-segment slope, respectively. and These represent real-time heart rate and initial resting heart rate, respectively. Simultaneously, motion compliance scores are calculated based on the similarity between real-time joint angles and expected angles. : ; In the formula, Indicates real-time joint angle; This indicates the expected maximum joint angle; And based on electrode-skin impedance With signal-to-noise ratio Calculate the electromyographic signal quality fraction : ; In the formula, This represents the maximum signal-to-noise ratio of the electromyographic signal; This indicates the safe threshold for electrode-skin impedance; S37, Muscle Fatigue Index Exercise compliance score and electromyographic signal quality score To obtain the training state parameters .

[0019] Step S4 specifically includes the following steps: S41. Construct a basic model of multi-mode resistance torque. ; Among them, the basic value of resistance torque in constant speed training mode The calculation formula is as follows: ; Baseline value of resistance torque in ANN training mode The calculation formula is as follows: ; Baseline value of resistance torque in isometric training mode The calculation formula is as follows: ; In the formula, , , These represent the gravity load coefficient, joint damping coefficient, and muscle synergy weighting coefficient, respectively. Represents gravitational acceleration; Indicates the lever arm of the lower limb joint, and ; This represents the maximum value of the muscle co-activation coefficient; Indicates real-time muscle activation level; This represents the angle adaptation function, and , Indicates the maximum functional angle of the lower limb joints; Indicates the coefficient of contraction at equal lengths; Represents the stability coefficient; S42. Real-time feedback adjustment of the basic torque based on the dynamic characteristics of different training modes; Constant velocity mode: Introducing initial angular velocity Feedback correction ensures constant angular velocity: ; AAN mode: Introduces activation rate change correction to avoid sudden torque changes. ; Equal length mode: Introduces force maintenance correction to ensure torque stability. ; In the formula, This represents the angular velocity deviation correction factor; This indicates the initial adjustment resistance torque under the modified isokinetic training mode; This represents the initial adjustment resistance torque in the modified ANN training mode; This indicates the initial adjustment resistance torque under the modified isometric training mode; Indicates the rate of change in muscle activation; Indicates the force retention coefficient; Indicates the torque of the training target of equal length; This indicates the torque feedback value at the previous moment; S44. Define the input of the fuzzy PID controller as torque error. and the rate of change of error The output is defined as the increment of the PID scaling factor. Increment of integral factor and differential factor increment ;and , , Representation pattern The torque feedback value is below; Define the following PID control law: ; In the formula, This represents the torque increment output by the PID controller; , , These represent the initial scaling factor, integral factor, and differential factor, respectively. S45, Quantization factor of fuzzy PID With scaling factor The optimized quantization factor and proportional factor are then substituted into the fuzzy PID control to obtain the secondary regulating resistance torque. ; These represent the quantization factor for torque error and the quantization factor for the rate of change of error, respectively. These represent the scaling factors for the increments of the proportional parameter, the integral parameter, and the derivative parameter, respectively. S46. Calculate the torque after correction for muscle fatigue index, exercise compliance score, and electromyographic signal quality score: ; in, ; ; In the formula, , and These represent the torque after muscle fatigue correction, the torque after motor compliance correction, and the final corrected torque, respectively. and These represent the static fatigue coefficient and the dynamic fatigue coefficient, respectively. This indicates the rate of change of the muscle fatigue index. S47, Multiple security checks; 1. Torque amplitude verification: ; 2. Joint angle correlation verification: ; 3. Torque change rate verification: ; In the formula, This represents the torque value after torque amplitude verification. This represents the torque value after joint angle correlation verification; Represents the angle adaptation function; This represents the final target resistance torque; This indicates the target torque at the previous moment; Indicates the maximum permissible rate of change of torque; S48. Output the final target resistance torque that has passed the verification. .

[0020] Step S5 specifically includes the following steps: S51. Design a sliding mode controller; Define the sliding surface , For design parameters, and ; Define the sliding mode control law: ; In the formula, This indicates the control torque output by the sliding mode controller; This represents the system's equivalent inertia; Indicates the expected joint angular acceleration; Indicates the coefficient of viscous friction; Indicates switching gain; Represents a symbolic function; S52, will Convert to motor initial control command , , Indicates the motor torque coefficient; S53. Based on feedback, the joint angle is corrected using a proportional-derivative method: ; In the formula, This indicates the revised control command; Indicates the trajectory tracking error, and ; This represents the angular velocity tracking error, and , and These represent the expected joint angular velocity and the real-time joint angular velocity, respectively. S54, The rehabilitation robot receives the corrected control commands. The lower limb joints are then driven to perform the corresponding movements, and the joint angles are collected every 10ms during the execution. and ECG signal peak time And when or At that time, an emergency deceleration command is triggered.

[0021] In step S53, a TOS balancing mechanism is introduced to perform a quantitative evaluation of the accuracy-complexity of the controller corresponding to the output control command: ; In the formula, Indicates a balanced score; Indicates the current tracking accuracy; Indicates the reference accuracy rate; This represents the slope of the sigmoid curve; Indicates the percentage change in parameters; Indicates the boundary of the TOS; like Then reduce the proportional-derivative correction parameter until... Maintain the adjusted proportional-derivative correction parameter correction control command.

[0022] Step S6 specifically includes the following steps: S61. Constructing a set of quantitative evaluation indicators , These represent joint function indicators, electromyographic signal consistency indicators, and training status indicators, respectively. Among them, joint function indicators Including joint range of motion Tracking accuracy Torque tracking error : ; ; ; In the formula, Indicates training duration; Indicates the number of torque sampling points; Indicates the first The actual output torque feedback value of the next sample; Indicates the first The target resistance torque for the next sample; Electromyographic signal consistency index Including the correlation coefficient between surface electromyography (EMG) signals and motor signals and muscle activation balance : ; ; In the formula, and This represents the cross-correlation coefficient between electromyographic signal B and electrode signal A; Representing the muscle synergy matrix The One collaborative pattern vector; Training status indicators Including average fatigue index and average compliance score : ; ; In the formula, Indicates the sampling time interval; Indicates the first Muscle fatigue index at each sampling time; Indicates the first Motion compliance score at each sampling time; S62, Based on a set of quantitative evaluation indicators Determine the level of training effectiveness: excellent: , , , , , and ; good: , , , , , ,and ; qualified: , , , , , ,and ; Otherwise, it will be deemed unqualified; S63. Construct a reinforcement learning model; Its state space Action space The reward function expression is as follows: ; In the formula, This represents the reward value for reinforcement learning; Indicates the range of motion of the target joint; S64. The DQN algorithm is used to optimize the action space, and the iterative formula is as follows: ; In the formula, Representing state Next action Action value function; Indicates the immediate reward of the current state-action pair; Indicates the discount factor; Indicates the next state All actions The maximum action value function; Iterate until a set number of iterations are reached, then select the action value function. The largest movement is the optimal movement. ; S65. Output optimization results: Output the optimal action. This translates into suggestions for adjusting training duration, resistance torque range, and training mode.

[0023] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for developing a lower limb rehabilitation training program based on electromyographic signal feedback, characterized in that: Includes the following steps: S1. Combining the trainee's basic information, training mode selection, and training target parameters, simultaneously collect lower limb surface electromyography signals, motion data, and heart rate change rate data to obtain a multimodal raw dataset; Based on the trainer information, training objectives, and training modes, an initial parameter set is obtained; S2. Wavelet threshold denoising, Kalman filtering and Savitzky-Golay smoothing are performed on the multimodal raw dataset output by S1 to extract surface electromyography signal features, and motion features and pressure features are fused to construct a high-dimensional feature set. S3. Based on the high-dimensional feature set output by S2, the trainee's lower limb joint movement intention is identified through the attention mechanism LSTM model. Combined with surface electromyography signals, the movement intention results and training state parameters are obtained. S4. Based on the motion intent and training state parameters output by S3, combined with the training mode of S1, a multi-mode resistance torque basic model is constructed. Through the quantization factor and scaling factor of fuzzy PID, muscle fatigue correction and safety boundary verification are incorporated to dynamically output the target resistance torque. S5. Based on the target resistance torque output by S4 and the motion intention of S3, a sliding mode controller is designed to generate initial control commands. The control commands are corrected through real-time feedback of joint angles. The control accuracy and system complexity are optimized by combining the TOS balancing mechanism. The control commands and training action execution results are then output. S6. Based on the execution results of S5 and the state parameters of S3, a quantitative evaluation index set is constructed from three dimensions: joint function, electromyographic signal consistency, and training state. The optimized training scheme is output through iterative optimization using the reinforcement learning DQN algorithm.

2. The method for developing a lower limb rehabilitation training program based on electromyographic signal feedback according to claim 1, characterized in that: Step S1 specifically includes the following steps: S11. Sensor Deployment and Calibration: Hybrid polymer hydrogel electrodes are attached to the target muscles of the lower limbs of the trainee, and IMU sensors are deployed on the thighs, calves and feet. Pressure sensors are also deployed on the left and right symmetrical positions of the forefoot and heels. The signals collected by the hybrid polymer hydrogel electrodes, IMU sensors and pressure sensors are then calibrated. S12. Initialize training parameters , , These represent the sensor sampling frequency, maximum safe torque, minimum safe torque, initial parameters of the training mode, data storage parameters, and the peak time of the maximum ECG signal, respectively. The initial parameters of the training mode include the initial angular velocity in the isokinetic training mode. Initial auxiliary coefficients in AAN training mode Fixed training angle in isometric training mode Data storage parameters include data storage frequency. Real-time feedback cycle and data cache size , ; in, ； ； ； ,and ; ； ； In the formula, This represents the basic moment coefficient, and ; Indicates the trainee's weight / mass. This represents the GMFCS classification correction factor, and ; This represents the age modification factor; Indicates the trainee's target joint range of motion; Indicates the trainee's expected muscle activation level; S13, Multimodal data acquisition; S131. Data Acquisition Start: The trainee starts the action according to the preset training mode. The data acquisition card synchronously acquires all sensor signals and displays sensor impedance, signal amplitude, angle and heart rate change rate data in real time during the acquisition process. If the impedance exceeds the standard or the signal is saturated, the acquisition is paused and the problem is investigated. S132. Data storage: Collected data is named according to trainer ID-training mode-timestamp, written to local storage in real time, and backed up to the cloud at the same time; S133. When the training time is reached or the trainer triggers a stop command, the data collection stops and the original multimodal dataset is obtained.

3. The method for developing a lower limb rehabilitation training program based on electromyographic signal feedback according to claim 2, characterized in that: The target muscles of the lower limbs of the trainee mentioned in step S11 include the rectus femoris, medial head of the gastrocnemius, peroneus longus, and tibialis posterior.

4. The method for developing a lower limb rehabilitation training program based on electromyographic signal feedback according to claim 2, characterized in that: Step S2 specifically includes the following steps: S21. Targeted denoising of multimodal data: Denoising and enhancement of lower limb surface electromyography signals in multimodal raw datasets: 4th order Butterworth bandpass filter is used to remove power frequency noise and baseline drift, and then wavelet threshold denoising algorithm is used to further suppress motion artifacts; For motion data, Kalman filtering calibration is used to eliminate drift errors; For heart rate variability data, Savitzky-Golay filtering is used to smooth the variance of heart rate variability, preserving peak features while suppressing noise; S22, Multi-dimensional Feature Extraction: The time-domain, frequency-domain, time-frequency-domain, and muscle coordination features of surface electromyography (EMG) signals were extracted. The time-domain features included the root mean square (RMS), mean absolute value, and waveform length of the EMG signal. The frequency-domain features included the peak power spectral density and center frequency of the EMG signal calculated using the Welch method. The time-frequency-domain features were derived from wavelet packet decomposition of the EMG signal, extracting the energy proportions of each subband. Muscle coordination features were extracted using nonnegative matrix factorization. Extracting joint angular velocities from motion data ; Extracting the peak time of the pressure signal from the pressure data and plantar pressure standard deviation ; Extracting the peak time of the electrocardiogram signal and heart rate variation variance ; S23. Feature Standardization and Temporal Sequence Construction: The multi-dimensional features extracted in step S22 are standardized using Z-score to obtain standardized features. ; S24. Weighted Fusion: An ensemble learning algorithm is used to calculate the importance of each feature, and a weight vector is obtained after normalization. Then perform weighted fusion: ； In the formula, This indicates the fusion of high-dimensional feature sets; Indicates the first The first sample Dimensional standard features.

5. The method for developing a lower limb rehabilitation training program based on electromyographic signal feedback according to claim 4, characterized in that: Step S3 specifically includes the following steps: S31. Temporal Enhancement: Extracting weighted fusion features using a sliding window. This process generates a temporal input, and then normalizes the trainee's electromyography (EMG) signals as global features, which are then concatenated to the end of each temporal window to form an enhancement sequence. ; S32, Design the modal feature attention layer; S321, Enhance the sequence It is divided into three groups: muscle synergy characteristics, electromyography characteristics, and exercise stress characteristics. S322. Map the features to the query vector using a learnable weight matrix. Key vector Sum value vector : ； ； ； In the formula, , and All represent learnable weight matrices; S323. Introduce grouping weight coefficients and calculate the query vector. With key vector similarity : ； In the formula, Indicates the transpose operation; Indicates the scaling factor; This represents the weight coefficient of the corresponding group; Indicates a block mask; S324, Weight Normalization: ； In the formula, Indicates attention weight; Represents the Softmax function; Indicates the length of the time window; The first in the attention mechanism The key vector at each time step; S325, Temporal features after attention weighting ; S33. Construct an attention-based LSTM model, adapting the input layer dimensional features of the attention-based LSTM model to the dimensionality of the high-dimensional feature set; assign weights to the muscle coordination features in the attention layer. The output layer contains the motion category and the expected angle. S34. Train an LSTM model with an attention mechanism using a high-dimensional feature set, and the loss function expression is as follows: ； in, ； ； In the formula, Indicates the total loss; Indicates the balance coefficient; Indicates classification loss; Indicates regression loss; Indicates the number of training samples; Indicates the first The true category of each sample Indicates category index, These represent the electromyographic signal output channels corresponding to lower limb joint dorsiflexion, lower limb joint flexion, lower limb joint inversion, and lower limb joint eversion, respectively. Indicates the predicted category probability; Indicates the first The prediction angle for each sample; Indicates the first The true angle of each sample; S35. Input the enhanced features to be identified into the trained attention mechanism LSTM model, and output the motion intent result. , These represent the output motion categories respectively. and expected motion angle ,and ; S36. Calculate the muscle fatigue index based on the attenuation of peak intensity of surface electromyography signals, the slowing of the slope of the flat segment, and the amplitude of heart rate rise, combined with center frequency. : ； In the formula, and These represent the real-time peak electromyography (EMG) intensity and the initial peak EMG intensity, respectively. and These represent the real-time muscle level straight-segment slope and the initial muscle level straight-segment slope, respectively. and These represent real-time heart rate and initial resting heart rate, respectively. Simultaneously, motion compliance scores are calculated based on the similarity between real-time joint angles and expected angles. : ； In the formula, Indicates real-time joint angle; This indicates the expected maximum joint angle; And based on electrode-skin impedance With signal-to-noise ratio Calculate the quality fraction of electromyographic signals : ； In the formula, This represents the maximum signal-to-noise ratio of the electromyographic signal; This indicates the safe threshold for electrode-skin impedance; S37, Muscle Fatigue Index Exercise compliance score and electromyographic signal quality score To obtain the training state parameters .

6. The method for developing a lower limb rehabilitation training program based on electromyographic signal feedback according to claim 5, characterized in that: Step S4 Specifically, the following steps are included: S41. Construct a basic model of multi-mode resistance torque. ; Among them, the basic value of resistance torque in constant speed training mode The calculation formula is as follows: ； Baseline value of resistance torque in ANN training mode The calculation formula is as follows: ； Baseline value of resistance torque in isometric training mode The calculation formula is as follows: ； In the formula, , , These represent the gravity load coefficient, joint damping coefficient, and muscle synergy weighting coefficient, respectively. Represents gravitational acceleration; Indicates the lever arm of the lower limb joint, and ; This represents the maximum value of the muscle co-activation coefficient; Indicates real-time muscle activation level; This represents the angle adaptation function, and , Indicates the maximum functional angle of the lower limb joints; Indicates the coefficient of contraction at equal lengths; Represents the stability coefficient; S42. Real-time feedback adjustment of the basic torque based on the dynamic characteristics of different training modes; Constant velocity mode: Introducing initial angular velocity Feedback correction ensures constant angular velocity: ； AAN mode: Introduces activation rate change correction to avoid sudden torque changes. ； Equal length mode: Introduces force maintenance correction to ensure torque stability. ； In the formula, This represents the angular velocity deviation correction factor; This indicates the initial adjustment resistance torque under the modified isokinetic training mode; This represents the initial adjustment resistance torque in the modified ANN training mode; This indicates the initial adjustment resistance torque under the modified isometric training mode; Indicates the rate of change in muscle activation; Indicates the force retention coefficient; Indicates the torque of the training target of equal length; This indicates the torque feedback value at the previous moment; S44. Define the input of the fuzzy PID controller as torque error. and the rate of change of error The output is defined as the increment of the PID scaling factor. Increment of integral factor and differential factor increment ;and , , Representation pattern The torque feedback value is below; Define the following PID control law: ； In the formula, This represents the torque increment output by the PID controller. , , These represent the initial scaling factor, integral factor, and differential factor, respectively. S45, Quantization factor of fuzzy PID With scaling factor The optimized quantization factor and proportional factor are then substituted into the fuzzy PID control to obtain the secondary regulating resistance torque. ; These represent the quantization factor for torque error and the quantization factor for the rate of change of error, respectively. These represent the scaling factors for the increments of the proportional parameter, the integral parameter, and the derivative parameter, respectively. S46. Calculate the torque after correction for muscle fatigue index, exercise compliance score, and electromyographic signal quality score: ； in, ； ； In the formula, , and These represent the torque after muscle fatigue correction, the torque after motor compliance correction, and the final corrected torque, respectively. and These represent the static fatigue coefficient and the dynamic fatigue coefficient, respectively. This indicates the rate of change of the muscle fatigue index. S47, Multiple security checks; Torque amplitude verification: ; Joint angle correlation verification: ; Torque change rate verification: ； In the formula, This represents the torque value after torque amplitude verification. This represents the torque value after joint angle correlation verification; Represents the angle adaptation function; This represents the final target resistance torque; This indicates the target torque at the previous moment; Indicates the maximum permissible rate of change of torque; S48. Output the final target resistance torque that has passed the verification. .

7. The method for developing a lower limb rehabilitation training program based on electromyographic signal feedback according to claim 6, characterized in that: Step S5 specifically includes the following steps: S51. Design a sliding mode controller; Define the sliding surface , For design parameters, and ; Define the sliding mode control law: ； In the formula, This indicates the control torque output by the sliding mode controller; This represents the system's equivalent inertia; Indicates the expected joint angular acceleration; Indicates the coefficient of viscous friction; Indicates switching gain; Represents a symbolic function; S52, will Convert to motor initial control command , , Indicates the motor torque coefficient; S53. Based on feedback, the joint angle is corrected using a proportional-derivative method: ； In the formula, This indicates the revised control command; Indicates the trajectory tracking error, and ; This represents the angular velocity tracking error, and , and These represent the expected joint angular velocity and the real-time joint angular velocity, respectively. S54, The rehabilitation robot receives the corrected control commands. The lower limb joints are then driven to perform the corresponding movements, and the joint angles are collected every 10ms during the execution. and ECG signal peak time And when or At that time, an emergency deceleration command is triggered.

8. The method for developing a lower limb rehabilitation training program based on electromyographic signal feedback according to claim 7, characterized in that: In step S53, a TOS balancing mechanism is introduced to perform a quantitative evaluation of the accuracy-complexity of the controller corresponding to the output control command: ； In the formula, Indicates a balanced score; Indicates the current tracking accuracy; Indicates the reference accuracy rate; This represents the slope of the sigmoid curve; Indicates the percentage change in parameters; Indicates the boundary of the TOS; like Then reduce the proportional-derivative correction parameter until... Maintain the adjusted proportional-derivative correction parameter correction control command.

9. The method for developing a lower limb rehabilitation training program based on electromyographic signal feedback according to claim 8, characterized in that: Step S6 specifically includes the following steps: S61. Constructing a set of quantitative evaluation indicators , These represent joint function indicators, electromyographic signal consistency indicators, and training status indicators, respectively. Among them, joint function indicators Including joint range of motion Tracking accuracy Torque tracking error : ； ； ； In the formula, Indicates training duration; Indicates the number of torque sampling points; Indicates the first The actual output torque feedback value of the next sample; Indicates the first The target resistance torque for the next sample; Electromyographic signal consistency index Including the correlation coefficient between surface electromyography (EMG) signals and motor signals and muscle activation balance : ； ； In the formula, and This represents the cross-correlation coefficient between electromyographic signal B and electrode signal A; Representing the muscle synergy matrix The One collaborative pattern vector; Training status indicators Including average fatigue index and average compliance score : ； ； In the formula, Indicates the sampling time interval; Indicates the first Muscle fatigue index at each sampling time; Indicates the first Motion compliance score at each sampling time; S62, Based on a set of quantitative evaluation indicators Determine the level of training effectiveness: excellent: , , , , , and ; good: , , , , , ,and ; qualified: , , , , , ,and ; Otherwise, it is deemed unqualified; S63. Construct a reinforcement learning model; Its state space Action space The reward function expression is as follows: ； In the formula, This represents the reward value for reinforcement learning; Indicates the range of motion of the target joint; S64. The DQN algorithm is used to optimize the action space, and the iterative formula is as follows: ； In the formula, Representing state Next action Action value function; Indicates the immediate reward of the current state-action pair; Indicates the discount factor; Indicates the next state All actions The maximum action value function; Iterate until a set number of iterations are reached, then select the action value function. The largest movement is the optimal movement. ; S65. Output optimization results: Output the optimal action. This translates into suggestions for adjusting training duration, resistance torque range, and training mode.

10. A system for performing the method for developing a lower limb rehabilitation training program based on electromyographic signal feedback as described in any one of claims 1-9, characterized in that: include: Data acquisition module: used to combine trainee basic information, training mode selection and training target parameters, and simultaneously collect lower limb surface electromyography signals, motion data and heart rate change rate data to obtain multimodal raw dataset; Based on the trainer information, training objectives, and training modes, an initial parameter set is obtained; The data preprocessing and fusion module is used to perform wavelet threshold denoising, Kalman filtering and Savitzky-Golay smoothing on the multimodal raw dataset, extract surface electromyography signal features, fuse motion features and pressure features, and construct a high-dimensional feature set. The intent recognition and dynamic evaluation module is used to identify the lower limb joint movement intent of trainees based on a high-dimensional feature set and through an attention mechanism LSTM model. Combined with surface electromyography signals, it obtains the movement intent results and training state parameters. The target resistance torque calculation module is used to construct a multi-mode resistance torque basic model based on exercise intention and training state parameters, combined with training mode. It incorporates muscle fatigue correction and safety boundary verification through the quantization factor and scaling factor of fuzzy PID, and dynamically outputs the target resistance torque. The closed-loop control module is used to design a sliding mode controller to generate initial control commands based on the target resistance torque and motion intention. It corrects the control commands through real-time feedback of joint angles, optimizes control accuracy and system complexity by combining the TOS balancing mechanism, and outputs control commands and training action execution results. The training effect evaluation and optimization module is used to construct a set of quantitative evaluation indicators based on the execution results and state parameters from three dimensions: joint function, electromyographic signal consistency, and training state. It iteratively optimizes parameters such as training duration and torque range through the reinforcement learning DQN algorithm and outputs the optimized training plan.

Images