A personalized dynamic rehabilitation human-computer interaction method and related device

Abstract

The application discloses a kind of personalized dynamic rehabilitation man-machine interaction method and related equipment, obtain target parameter when target patient rehabilitation training;High-density flexible electrode is attached on the corresponding muscle of target patient's limb, surface electromyogram signal is collected, surface electromyogram signal time window is obtained and motor unit discharge pulse sequence is extracted therefrom;Surface electromyogram signal time window and motor unit discharge pulse sequence are respectively input into feature extraction module, and feature fusion is carried out by deep transform neural network, and the joint movement angle and end force of the patient are predicted by regression;The average discharge frequency of motor unit and the cumulative pulse per unit time are calculated, the active participation of target patient in the training process is quantified, the motion performance of the patient is evaluated and recorded according to the average absolute percentage error between the actual motion parameters and the target motion parameters of the patient;According to the active participation and motion performance of target patient, the parameters of rehabilitation robot are dynamically adjusted, and the rehabilitation initiative of patient is improved.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation robots and human-computer interaction technology, and in particular to a personalized dynamic rehabilitation human-computer interaction method, system, terminal and computer-readable storage medium. Background Technology

[0002] With the acceleration of modernization, the incidence of stroke is rising year by year and is affecting younger people, becoming a public health problem that threatens people's health. According to relevant statistics, about 75% of stroke patients have varying degrees of functional impairment, which seriously reduces their ability to take care of themselves and makes it difficult for them to return to normal family and social life.

[0003] Traditional functional rehabilitation training requires therapists to provide manual assistance to patients according to the doctor's treatment plan. This is time-consuming and labor-intensive, and relies on the therapist's clinical experience and subjective judgment, making it difficult to accurately and objectively assess the patient's level of active participation and functional recovery.

[0004] In recent years, medical robotics technology has been widely applied in the field of motor function rehabilitation, providing patients with long-term and effective guidance and assistance for rehabilitation training. Compared with manual judgment, rehabilitation robots can more objectively, in real-time, and accurately record and evaluate patients' training status and rehabilitation effects using physiological signals such as electroencephalography (EEG) and electromyography (EMG), thereby assisting rehabilitation therapists and doctors in developing further personalized rehabilitation training plans. Therefore, rehabilitation robots can provide personalized, repeatable, and high-density rehabilitation training, with advantages such as real-time monitoring and feedback. Their development and optimization are crucial for the functional rehabilitation of patients with neuromuscular diseases.

[0005] In rehabilitation robot systems, human-computer interaction control strategies are control methods adopted to achieve specific treatment goals and adapt to patient needs, directly affecting patients' training initiative and rehabilitation outcomes. Clinical medical research shows that compared to repetitive and monotonous passive training, rehabilitation therapy with active patient participation is more effective in neural system reconstruction and motor function recovery, and can positively impact the recovery of patients' cognitive and psychological / emotional functions, thereby improving rehabilitation efficiency.

[0006] Therefore, maintaining and improving patients' active participation in the training process is a current research focus in stroke rehabilitation. Patients' motor intention, active participation, and motor performance are key factors for rehabilitation robots to achieve safe, effective, and personalized rehabilitation treatment, promoting patient participation and effort. However, existing rehabilitation robots either passively guide training according to fixed parameters set by the therapist; or they only convert motor intention into control signals while ignoring the patient's actual motor performance and active participation during training; or they only consider motor performance or active participation while ignoring the patient's motor intention, lacking comprehensive collection and feedback of information on patient active participation, which is detrimental to maintaining patients' initiative in the rehabilitation training process. Furthermore, there is a lack of accurate quantitative assessment methods for patient active participation, making it impossible to monitor changes in patient active participation and engagement in real time, and unable to dynamically adjust auxiliary control parameters according to the patient's motor ability and training progress, thus failing to meet the personalized rehabilitation needs of patients.

[0007] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0008] The main objective of this invention is to provide a personalized dynamic rehabilitation human-computer interaction method, system, terminal, and computer-readable storage medium, aiming to solve the problems of existing technologies being unable to monitor changes in the patient's active participation and engagement in real time, unable to dynamically adjust auxiliary control parameters according to the patient's motor ability and training progress, and thus unable to meet the patient's personalized rehabilitation needs.

[0009] To achieve the above objectives, the present invention provides a personalized dynamic rehabilitation human-computer interaction method, which includes the following steps:

[0010] The system acquires the target parameters for rehabilitation training of the target patient, as well as the average discharge frequency of the motor unit and the cumulative pulses per unit time corresponding to each movement, and stores them in the rehabilitation training system.

[0011] High-density flexible electrodes are attached to the corresponding muscles of the target patient's limb to collect surface electromyography (EMG) signals. The surface EMG signals are then segmented in real time using a sliding window to obtain a time window of the surface EMG signals. The motor unit discharge pulse sequence is then extracted from the time window of the surface EMG signals.

[0012] The surface electromyography signal time window and motor unit discharge pulse sequence are respectively input into the feature extraction module, and the features are fused by the deep transform neural network to regress and predict the patient's joint motion angle and end force.

[0013] The average discharge frequency and cumulative pulses per unit time of the motor unit are calculated to quantify the active participation of the target patient in the rehabilitation training process. The percentage of active participation of the patient is expressed by the ratio between the average discharge frequency and cumulative pulses per unit time during the exercise process and the target average discharge frequency and cumulative pulses. The patient's exercise performance is evaluated and recorded based on the average absolute percentage error between the patient's actual exercise parameters and the target exercise parameters.

[0014] Based on reinforcement learning algorithms, the maximum assist force and movement speed of the rehabilitation robot are dynamically adjusted according to the target patient's active participation and motor performance, so as to achieve personalized dynamic rehabilitation feedback control of the rehabilitation robot.

[0015] Optionally, the personalized dynamic rehabilitation human-computer interaction method, wherein acquiring the target parameters of the target patient during rehabilitation training, as well as the average discharge frequency of the motor unit corresponding to each movement and the cumulative pulses per unit time, and storing them in the rehabilitation training system, specifically includes:

[0016] Before the training begins, according to the rehabilitation training plan, for the first target patient with hemiplegia, the joint motion angles and end force of the healthy limbs of the first target patient when performing various daily living activities are collected and stored in the rehabilitation training system as target parameters for the first target patient's rehabilitation training.

[0017] For the second target patient with motor impairment in both limbs, the average joint motion angle and average end force of healthy people of similar age and height to the second target patient when performing various daily living activities are collected and stored in the rehabilitation training system as target parameters for the rehabilitation training of the second target patient.

[0018] The average discharge frequency of the motor unit corresponding to each movement and the cumulative pulses per unit time are stored in the rehabilitation training system to calculate the patient's active participation.

[0019] Optionally, the personalized dynamic rehabilitation human-computer interaction method, wherein extracting the motor unit discharge pulse sequence from the surface electromyography signal time window specifically includes:

[0020] Each surface electromyography signal time window is used as a sample input into a deep convolutional neural network model to extract the firing pulse sequence of the motor unit;

[0021] The deep convolutional neural network model includes four one-dimensional convolutional layers, two max-pooling layers, one fully connected layer, and a sigmoid activation function.

[0022] Optionally, the personalized dynamic rehabilitation human-computer interaction method, wherein the step of using each surface electromyography signal time window as a sample input to a deep convolutional neural network model to extract the discharge pulse sequence of the motor unit specifically includes:

[0023] Electromyographic features in the time window of surface electromyographic signals are extracted by two one-dimensional convolutional layers. Max pooling is used for feature selection and dimensionality reduction. Deep electromyographic features are then extracted by two more one-dimensional convolutional layers, and max pooling is used for final feature selection and dimensionality reduction. Finally, the output is the discharge pulse sequence of all activated motor units after passing through a fully connected layer and a sigmoid activation function. This sequence is used for motion intention recognition and active participation measurement.

[0024] Optionally, the personalized dynamic rehabilitation human-computer interaction method, wherein the step of inputting the surface electromyography signal time window and the motor unit discharge pulse sequence into the feature extraction module respectively, and performing feature fusion through a deep transform neural network to regress and predict the patient's joint motion angle and end force, specifically includes:

[0025] The feature extraction module, constructed using a convolutional neural network, extracts features from the time window of surface electromyography signals and the pulse sequence of motor units, respectively. The attention mechanism in the encoding module of the deep transform neural network is used to extract and filter features. Finally, the joint angle and end force are output by the fully connected layer.

[0026] The rehabilitation training system automatically selects appropriate training tasks and exercise goals based on the target patient's exercise intentions, including joint angles and end-effector forces. The system displays the target tasks and related exercise parameters on the human-computer interaction interface, guides the patient to complete the corresponding training movements through visual and voice prompts, and displays the completion status of the movements.

[0027] Optionally, the personalized dynamic rehabilitation human-computer interaction method, wherein calculating the average discharge frequency and cumulative pulses per unit time of the motor unit quantifies the active participation of the target patient during rehabilitation training, and using the ratio between the average discharge frequency and cumulative pulses per unit time during the movement process and the target average discharge frequency and cumulative pulses to represent the percentage of the patient's active participation, and assessing and recording the patient's movement performance based on the average absolute percentage error between the patient's actual movement parameters and the target movement parameters, specifically includes:

[0028] The percentage of active participation (P%) is expressed as the ratio between the patient's active participation during exercise and the target average discharge frequency and cumulative pulses.

[0029]

[0030] Where, r t r is the average discharge frequency of the moving unit during actual motion. gc is the target average discharge frequency of the motion unit. t c represents the cumulative pulses per unit time of the moving unit during actual motion. g The target cumulative pulse per unit time for the moving unit;

[0031] Joint angle θ obtained using actual motion measurements i and end force F j , and the target joint angle θ set according to the motion intention i ′ and target end force F j The mean absolute percentage error (MSE%) between the two values ​​assesses the patient's actual motor performance:

[0032]

[0033] Where i represents the i-th joint, m is the total number of joint angles measured, j represents the j-th distal end, and n is the total number of distal force measurements; the smaller the mean absolute percentage error (MSE%), the better the patient's motor performance.

[0034] Optionally, the personalized dynamic rehabilitation human-computer interaction method, wherein the step of dynamically adjusting the maximum assist force and movement speed of the rehabilitation robot based on the reinforcement learning algorithm according to the target patient's active participation and motor performance, to achieve personalized dynamic rehabilitation feedback control of the rehabilitation robot, specifically includes:

[0035] Design an evaluation function E for the movement status of the target patient during rehabilitation training, expressed as:

[0036] E=α×P%+β×(1-MSE%);

[0037] Wherein, α and β are the weighting coefficients of the percentage of active participation P% and the mean absolute percentage error MSE%, respectively, and the sum of the two is 1. Setting α = β = 0.5 indicates that the improvement of patients' active participation and motor performance is equally important.

[0038] The training parameters of the rehabilitation robot, including maximum assist force and movement speed, are adaptively adjusted based on a deep Q-network.

[0039] Furthermore, to achieve the above objectives, the present invention also provides a personalized dynamic rehabilitation human-computer interaction system, wherein the personalized dynamic rehabilitation human-computer interaction system includes:

[0040] The training task and motion target design module is used to obtain the target parameters of the target patient during rehabilitation training, as well as the average discharge frequency of the motor unit and the cumulative pulses per unit time corresponding to each movement, and store them in the rehabilitation training system.

[0041] The surface electromyography (EMG) signal acquisition and decomposition module is used to attach high-density flexible electrodes to the corresponding muscles of the target patient's limb, acquire surface EMG signals, and use a sliding window to segment the surface EMG signals in real time to obtain the surface EMG signal time window, and extract the motor unit discharge pulse sequence from the surface EMG signal time window.

[0042] The motion intention recognition module is used to input the surface electromyography signal time window and the motor unit discharge pulse sequence into the feature extraction module, and perform feature fusion through deep transform neural network to regress and predict the patient's joint motion angle and end force;

[0043] The active participation quantification and motor performance assessment module is used to calculate the average discharge frequency and cumulative pulses per unit time of motor units, quantify the active participation of the target patient in the rehabilitation training process, and use the ratio between the average discharge frequency and cumulative pulses per unit time during the exercise process and the target average discharge frequency and cumulative pulses to represent the percentage of the patient's active participation. The patient's motor performance is assessed and recorded based on the average absolute percentage error between the patient's actual motor parameters and the target motor parameters.

[0044] The personalized dynamic rehabilitation feedback control module is used to dynamically adjust the maximum assist force and movement speed of the rehabilitation robot based on the target patient's active participation and motor performance, using reinforcement learning algorithms, in order to achieve personalized dynamic rehabilitation feedback control of the rehabilitation robot.

[0045] In addition, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a personalized dynamic rehabilitation human-computer interaction program stored in the memory and executable on the processor, wherein when the personalized dynamic rehabilitation human-computer interaction program is executed by the processor, it implements the steps of the personalized dynamic rehabilitation human-computer interaction method as described above.

[0046] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a personalized dynamic rehabilitation human-computer interaction program, which, when executed by a processor, implements the steps of the personalized dynamic rehabilitation human-computer interaction method as described above.

[0047] In this invention, target parameters for rehabilitation training of the target patient, as well as the average discharge frequency and cumulative pulses per unit time of the motor unit corresponding to each movement, are acquired and stored in the rehabilitation training system. High-density flexible electrodes are attached to the corresponding muscles of the target patient's limb to collect surface electromyography (EMG) signals. The EMG signals are segmented in real time using a sliding window to obtain a time window for the EMG signals, and the discharge pulse sequence of the motor unit is extracted from the time window. The time window of the EMG signals and the discharge pulse sequence of the motor unit are respectively input into the feature extraction module, and feature fusion is performed through a deep transform neural network to regress and predict the joint motion angle and end force of the patient. The average discharge frequency and cumulative pulses per unit time of the motor unit are calculated to quantify the active participation of the target patient in the rehabilitation training process. The ratio between the average discharge frequency and cumulative pulses per unit time during the movement process and the target average discharge frequency and cumulative pulses is used to represent the percentage of active participation of the patient. The patient's movement performance is evaluated and recorded based on the average absolute percentage error between the patient's actual movement parameters and the target movement parameters. Based on a reinforcement learning algorithm, the maximum assist force and movement speed of the rehabilitation robot are dynamically adjusted according to the active participation and movement performance of the target patient to achieve personalized dynamic rehabilitation feedback control of the rehabilitation robot. This invention provides more personalized, dynamic, and precise assistance, improves the human-computer interaction experience, thereby enhancing patients' initiative in rehabilitation, promoting neurological function remodeling, and improving the effectiveness of rehabilitation treatment. Attached Figure Description

[0048] Figure 1 This is a flowchart of a preferred embodiment of the personalized dynamic rehabilitation human-computer interaction method of the present invention;

[0049] Figure 2 This is a flowchart of a personalized dynamic rehabilitation feedback control method based on motor unit discharge information, which is a preferred embodiment of the personalized dynamic rehabilitation human-computer interaction method of the present invention.

[0050] Figure 3 This is a flowchart illustrating the process of offline decomposing surface electromyography signals and calculating the average discharge frequency of the motor unit and the cumulative pulses per unit time for each movement in a preferred embodiment of the personalized dynamic rehabilitation human-computer interaction method of the present invention.

[0051] Figure 4 This is a schematic diagram of a surface electromyography signal decomposition model based on a deep convolutional neural network in a preferred embodiment of the personalized dynamic rehabilitation human-computer interaction method of the present invention.

[0052] Figure 5 This is a schematic diagram of the human-computer interaction and related interfaces during the rehabilitation training process in a preferred embodiment of the personalized dynamic rehabilitation human-computer interaction method of the present invention.

[0053] Figure 6This is a schematic diagram of the structure of the convolutional neural network and the deep transform neural network in a preferred embodiment of the personalized dynamic rehabilitation human-computer interaction method of the present invention;

[0054] Figure 7 This is a flowchart illustrating the estimation of joint motion angles using inertial measurement unit data in a preferred embodiment of the personalized dynamic rehabilitation human-computer interaction method of the present invention.

[0055] Figure 8 This is a schematic diagram of the network structure of a deep Q-network in a preferred embodiment of the personalized dynamic rehabilitation human-computer interaction method of the present invention;

[0056] Figure 9 This is a schematic diagram of the training process of a deep Q-network in a preferred embodiment of the personalized dynamic rehabilitation human-computer interaction method of the present invention.

[0057] Figure 10 This is a schematic diagram of the reinforcement learning model transitioning from a simulation environment to a real environment in a preferred embodiment of the personalized dynamic rehabilitation human-computer interaction method of the present invention.

[0058] Figure 11 This is a schematic diagram illustrating the principle of a preferred embodiment of the personalized dynamic rehabilitation human-computer interaction system of the present invention;

[0059] Figure 12 This is a schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0061] Existing human-machine interaction control strategies for rehabilitation robots typically rely solely on the overall characteristics of human-machine interaction information or surface electromyography (sEMG) signals to identify movement intentions and quantify active participation. Furthermore, they adjust rehabilitation training parameters based solely on the patient's movement intentions, performance, or active participation. For example, one upper limb rehabilitation robot control method based on information fusion predicts the patient's movement intentions by considering the robot's end effector position, velocity, and the force exerted by the patient on the end effector. It also estimates environmental features based on the robot's end effector position and the force of collisions with the environment. A Kalman filter algorithm is used to fuse movement intentions and environmental features, and a Naive Bayes principle is employed to determine task completion. If the task is completed, the process terminates; otherwise, a compliance model is introduced to ensure safe task completion. Another example is an adaptive control method and system for upper limb rehabilitation robots based on game theory and sEMG. This method estimates muscle force using sEMG signals, analyzes the human-machine interaction system using game theory, derives the robot's role, and uses Nash balance to update the control rate (weighting factor) between the human and robot, enabling the robot to adaptively adjust its training mode based on the patient's movement intentions during operation. An adaptive position-constrained on-demand assisted control method and system for rehabilitation robots is based on position error transformation and human-machine interaction... Mutual torques are used to construct a motion performance function, and a continuously differentiable robot assistance level function is designed with motion performance as the input variable. This function characterizes the degree of robot assistance to the patient and serves as a weighting factor for the human-machine interaction system controller, enabling the controller to achieve on-demand assistance in different modes. For example, an intelligent active-passive hybrid training control method for rehabilitation robots obtains the difference in human-machine interaction force by calibrating the patient's joint range of motion and maximum power. This difference is used to calculate the patient's active participation during rehabilitation training and is converted into a velocity offset. The passive training joint movement speed is adaptively adjusted based on the velocity offset. Another method based on Bayesian optimization to improve on-demand assistance rehabilitation training participation uses trajectory tracking error to assess the patient's motion performance and the root mean square value of surface electromyography (EMG) signals to assess the patient's motion participation during training. An evaluation function integrating motion performance and active participation is established. The relationship between the evaluation function and the on-demand assistance strategy hyperparameters (maximum boundary assistance force) is learned through Bayesian optimization. The hyperparameters that maximize the evaluation function in the next round of on-demand assistance are found. Based on the patient's position error and the optimal hyperparameters during training, the force applied by the rehabilitation robot to the patient is adjusted in real time according to the law of the assistance force field.

[0062] Disadvantages of existing technology:

[0063] (1) Existing human-computer interaction control strategies for rehabilitation robots lack real-time or early recognition of patients’ continuous fine motor intentions, which is not conducive to timely adjustment of the control parameters of rehabilitation robots, making it difficult to quickly adapt to the patient’s constantly changing motor intentions, and failing to provide timely support and feedback, thus affecting the patient’s active participation in the rehabilitation training process.

[0064] (2) Most of the existing human-computer interaction control strategies for rehabilitation robots use the root mean square value of human-computer interaction information or surface electromyography signals to assess the patient’s active participation. This has a lag and instability (susceptible to external factors), lacks more accurate personalized active participation measurement indicators, and makes it difficult to assess and monitor the patient’s active participation in the exercise training process in real time. This results in the inability to provide timely feedback on changes in the patient’s status in rehabilitation training, making it difficult to effectively guide the patient to actively participate and affecting the rehabilitation training effect.

[0065] (3) Existing human-computer interaction control strategies for rehabilitation robots often consider the patient’s movement intention, active participation or movement performance to change training parameters. There is a lack of a method to comprehensively consider movement intention, active participation and movement performance for personalized dynamic feedback control. It is difficult to provide real-time feedback on the patient’s movement status and make targeted adjustments to training parameters.

[0066] This invention discloses a personalized dynamic rehabilitation human-computer interaction method based on motor unit discharge information. By identifying the patient's movement intention through the discharge information of the motor unit and quantifying the degree of active participation, the method automatically adjusts training parameters such as assistance force and movement speed based on real-time feedback information. It comprehensively considers movement intention, degree of active participation, and movement performance to adapt to the patient's needs and changes in movement state during rehabilitation training, providing more personalized dynamic and precise assistance support, improving the human-computer interaction experience, thereby enhancing the patient's rehabilitation initiative, promoting neural function remodeling, and improving the rehabilitation treatment effect.

[0067] The personalized dynamic rehabilitation human-computer interaction method described in the preferred embodiment of the present invention, such as... Figure 1 and Figure 2 As shown, the personalized dynamic rehabilitation human-computer interaction method includes the following steps:

[0068] Step S10: Obtain the target parameters for the target patient's rehabilitation training, as well as the average discharge frequency of the motor unit and the cumulative pulses per unit time corresponding to each movement, and store them in the rehabilitation training system.

[0069] Specifically, before the training starts, according to the rehabilitation training plan, for the first target patient with hemiplegia, the joint movement angles and end forces when the healthy side (the sound side) limb of the first target patient completes various activities of daily living are collected, stored in the rehabilitation training system, and used as target parameters for the first target patient during rehabilitation training; for the second target patient with movement disorders in both limbs, the average joint movement angles and average end forces when healthy people with similar age and height to the second target patient complete various activities of daily living are collected, stored in the rehabilitation training system, and used as target parameters for the second target patient during rehabilitation training; the average discharge frequency of motor units corresponding to each movement and cumulative pulses per unit time are stored in the rehabilitation training system, which are used to calculate the patient's active participation (the value of active participation is converted to the range of [0, 1]).

[0070] In addition, high-density flexible electrodes are attached to the surface muscles of the patient's healthy side limb to collect surface electromyography signals corresponding to each movement for more than 30s, and the surface electromyography signals are decomposed offline by a convolved blind source separation algorithm to calculate the average discharge frequency of motor units corresponding to each movement and cumulative pulses per unit time, which are stored in the rehabilitation training system for calculating the patient's active participation.

[0071] As shown in Figure 3 , Figure 3 it is a schematic flow diagram of offline decomposition of surface electromyography signals and calculation of the average discharge frequency of motor units corresponding to each movement and cumulative pulses per unit time. Firstly, channel expansion is performed on the collected surface electromyography signals, the expansion parameter R is set to 16, de-meaning is performed by subtracting the mean value of each channel from the value of the channel, then zero-phase analysis whitening transformation is performed to remove the correlation between channels to obtain the signal z; secondly, after initializing the separation matrix B, the process enters the outer loop, i represents the current number of cycles, and the number of iterations is set to 50; when the number of cycles i is less than or equal to the number of iterations, initialize the separation vector ω I (0) and ω I (-1); when the absolute value of the dot product of the transpose of the n-th separation vector ω I (n) and the (n-1)-th separation vector ω i (n-1) minus 1 is less than the tolerance Tolx, the tolerance is set to 10-8, that is, when |ω i (n) T ω i (n-1)-1|<Tolx, enter the first inner loop, execute the fixed-point algorithm, and obtain the n-th separation vector ω i (n) through orthogonalization and standardization calculation, let n = n+1, continue the first inner loop until |ω i (n) T ω i(n-1)-1|≥Tolx or n exceeds the fixed-point iteration count of 300; then, initialize the (n-1)th and nth coefficients of variation CoV. n-1 and CoV n When CoV n-1 >CoV n Then, the second inner loop is entered to estimate the i-th source, and the pulse sequence PT is estimated using peak detection and K-means clustering algorithms. n , causing CoV n-1 =CoV n Calculate the pulse sequence PT n coefficient of variation of CoV n Let the separation vectors Discharge time point t j ={t:PT n (t)=1}, J is the total number of discharge pulses; let n=n+1, continue the second inner loop, until CoV n-1 ≤CoV n Or n exceeds the fixed-point iteration count of 300; finally, when the silhouette coefficient SIL is greater than 0.9, the estimate of the i-th source is accepted, and ω is... i Add it to the separation matrix B, let i = i + 1, and continue the outer loop. When SIL ≤ 0.9, directly let i = i + 1 and continue the outer loop; when the number of loops i > the number of iterations, end the outer loop; finally, extract the discharge information of activated motor units from the surface electromyography signal through the fixed-point iterative algorithm.

[0072] This invention designs personalized training tasks and exercise goals for patients by collecting kinematic and dynamic parameters of the unaffected side, which is more in line with the patient's motor ability and rehabilitation needs. It can provide patients with reasonable rehabilitation training guidance and help restore the coordination of the patient's bilateral limbs.

[0073] Step S20: Attach high-density flexible electrodes to the corresponding muscles of the target patient's limb, collect surface electromyography (EMG) signals, and use a sliding window to segment the surface EMG signals in real time to obtain the surface EMG signal time window, and extract the motor unit discharge pulse sequence from the surface EMG signal time window.

[0074] Specifically, surface electromyography (EMG) signal acquisition and decomposition: High-density flexible electrodes are attached to the corresponding muscles of the patient's limb to acquire surface EMG signals. A sliding window is used to segment the surface EMG signals in real time, obtaining surface EMG signal time windows. In this invention, the sliding window length is set to 60 data points, and the step size is set to 20 data points. 60% of the surface EMG signal time windows and the corresponding motor unit pulse sequences are used as the training set, 20% as the validation set, and 20% as the test set. A five-fold cross-validation is used to train the online surface EMG signal decomposition model. Each surface EMG signal time window is used as a sample input to a deep convolutional neural network model to extract the discharge pulse sequence of the motor unit from the surface EMG signal time window. Wherein, as... Figure 4 As shown, Figure 4 This is a schematic diagram of the surface electromyography signal decomposition model based on a deep convolutional neural network according to the present invention. The deep convolutional neural network model includes four one-dimensional convolutional layers, two max pooling layers, one fully connected layer, and a sigmoid activation function.

[0075] The input to the deep convolutional neural network model is a time window of surface electromyography (EMG) signals. First, EMG features in the time window of surface EMG signals are extracted through two one-dimensional convolutional layers. Then, a max pooling layer is used for feature selection and dimensionality reduction. Next, two one-dimensional convolutional layers are used to extract deep EMG features, and a max pooling layer is used for final feature selection and dimensionality reduction. Finally, after passing through a fully connected layer and a sigmoid activation function, the output is the firing pulse sequence of all activated motor units, which is used for motion intention recognition and active participation measurement.

[0076] During the training of the online decomposition model, the learning rate was set to 10. -4 The weight decays to 10. -6 The model uses binary cross-entropy as the loss function and adaptive moment estimation as the optimization method. After 100 rounds of training, the optimal model is selected by the loss function value of the validation set. During rehabilitation training, the surface electromyography signals of 60 data points are input into the online decomposition model to obtain the discharge pulse sequence of all activated motor units, which is used for motor intention recognition and active participation measurement.

[0077] This invention combines surface electromyography signals and motor unit discharge information to construct a prediction model of patient joint angles and end force based on a depth transform network. According to the patient's movement intention during training, it selects appropriate target movement parameters corresponding to the healthy side, which can ensure that the patient can carry out rehabilitation training within a safe range and achieve the training goal with the assistance of a rehabilitation robot. This increases the patient's confidence in rehabilitation, avoids the tedium of fixed and repetitive training movements that lead to the rehabilitation process, and can fully mobilize and maintain the patient's initiative in training.

[0078] Step S30: Input the surface electromyography signal time window and motor unit discharge pulse sequence into the feature extraction module, and perform feature fusion through deep transform neural network to regress and predict the patient's joint motion angle and end force.

[0079] Specifically, a feature extraction module composed of a convolutional neural network extracts features from the time window of surface electromyography signals and the pulse sequence of motor units (i.e., inputs them into the movement intention recognition model), thereby predicting the patient's joint movement angles and end-effector forces. The rehabilitation training system automatically selects corresponding training tasks and movement goals based on the patient's movement intentions and rehabilitation training plan. It can select training tasks based on the correlation coefficient between the predicted joint angles and the joint angles of existing training tasks, and set corresponding movement parameters based on the patient's achievable goals. Features are extracted and filtered through the attention mechanism in the encoding module of a deep transform neural network, and finally, the joint angles and end-effector forces are output by a fully connected layer. The rehabilitation training system automatically selects appropriate training tasks and movement goals based on the target patient's movement intentions, including joint angles and end-effector forces. The target task and related movement parameters are displayed on the human-computer interaction interface. The system guides the patient to complete the corresponding training movements through visual and voice prompts and displays the completion status of the movements, such as... Figure 5 As shown, Figure 5 This is a schematic diagram of human-computer interaction and related interfaces during rehabilitation training. Visual guidance is used as an example of hand function rehabilitation training.

[0080] To monitor patients' movement intentions during rehabilitation training in real time, this invention designs a hybrid neural network structure based on convolutional neural networks and deep transform neural networks, such as... Figure 6 As shown, the model includes a feature extraction module, an encoding module, and an output module. The feature extraction module uses basic residual blocks, a soft thresholding algorithm, and an attention mechanism to extract features from surface electromyography (EMG) signals and motor unit discharge pulse sequences, respectively. These two feature sets are concatenated and fed into the input layer of the deep transform neural network for position encoding. The encoding module consists of two stacked encoder layers. Each encoder layer comprises two sub-layer connection structures: the first sub-layer connection structure includes a multi-head attention sub-layer, a normalization layer, and a residual connection; the second sub-layer connection structure includes a feedforward fully connected sub-layer, a normalization layer, and a residual connection. The concatenated features are extracted and filtered using the attention mechanism, and finally, the fully connected layer outputs the predicted joint angle and end-effector force. During the training of the motion intention recognition model, a learning rate of 10 is set. -4 The weight decays to 10. -6 Using mean absolute error as the loss function and adaptive moment estimation as the optimization method, after 100 rounds of training, the optimal model is selected based on the loss function values ​​on the validation set, ultimately yielding the motion intent recognition model; Figure 6 As shown, Figure 6This is a schematic diagram of the motion intention recognition model based on deep convolutional neural networks and deep variable speed neural networks used in this invention. The feature extraction module composed of convolutional neural networks extracts features from surface electromyography signals and motor unit pulse sequences, respectively. The attention mechanism in the deep transform neural network encoding module further extracts and filters features. Finally, the fully connected layer outputs the joint angle and end force.

[0081] This invention uses motor unit discharge information to quantify the patient's active participation in the rehabilitation training process, which can more stably and accurately determine the degree of muscle activation and provide more reliable information on active participation.

[0082] Step S40: Calculate the average discharge frequency of the motor unit and the cumulative pulses per unit time to quantify the active participation of the target patient in the rehabilitation training process. The active participation of the patient is expressed by the ratio between the active participation during the exercise process and the target average discharge frequency and cumulative pulses. The patient's exercise performance is evaluated and recorded based on the average absolute percentage error between the patient's actual exercise parameters and the target exercise parameters.

[0083] Specifically, the quantification of active participation and assessment of motor performance: The motor unit discharge pulse sequence contains important information about the patient's motor behavior; this invention quantifies the patient's active participation during rehabilitation training by calculating the average discharge frequency and cumulative pulse sequence of the motor unit per unit time, and uses the ratio between the active participation during exercise and the target average discharge frequency and cumulative pulse to represent the patient's active participation percentage (P%).

[0084]

[0085] Where, r t r is the average discharge frequency of the moving unit during actual motion. g c is the target average discharge frequency of the motion unit. t c represents the cumulative pulses per unit time of the moving unit during actual motion. g The target cumulative pulse per unit time for the moving unit.

[0086] Furthermore, the joint angle θ obtained from actual motion measurements is used. i and end force F j , and the target joint angle θ set according to the motion intention i ′ and target end force F j The mean absolute percentage error (MSE%) between the two values ​​assesses the patient's actual motor performance:

[0087]

[0088] Where i represents the i-th joint, m is the total number of joint angles measured, j represents the j-th distal end, and n is the total number of distal force measurements; the smaller the mean absolute percentage error (MSE%), the better the patient's motor performance; based on this index, the patient's percentage of motor completion (1-MSE%) is displayed on the human-computer interface, such as... Figure 5 As shown.

[0089] Furthermore, to assess the patient's true motor performance, for larger joints such as the knee, the inertial measurement unit is placed tangentially or parallel to the joint's motion axis. Accelerometers measure the joint's acceleration data a(t), and gyroscopes measure the joint's angular velocity data ω(t). Figure 7 The flowchart shows the process of estimating joint motion angles using inertial measurement unit (IMU) data; the joint angles are initially estimated using numerical integration of gyroscope angular velocity data, as shown in the following formula:

[0090] θ_ω(t)=θ_ω(t-1)+ω(t)*Δt; (3)

[0091] Where is the angle estimate at time t, and is the sampling time interval. To obtain more accurate joint angle information, acceleration data is used to correct the gyroscope angle estimate. The formula for estimating the joint angle based on acceleration data is as follows:

[0092]

[0093] Where ax(t), ay(t), and az(t) are the accelerometer measurements on the X, Y, and Z axes, respectively. The numerical integral θ_ω(t) of the angular velocity data and the accelerometer correction angle θ_a(t) are fused and filtered using a Kalman filter algorithm to obtain an accurate and stable joint angle estimate θ. i For smaller joints such as finger joints, flexible electrodes are used to collect the joint angle θ. i Simultaneously, a six-axis force sensor is placed at the end of the moving limb to measure the three components of the force at the end, Fx, Fy, and Fz, and the three components of the torque, Mx, My, and Mz. The force F at the end of the limb is calculated using the sensor's measurement data. j and torque M j .

[0094] This invention evaluates the patient's motor performance by comparing the actual movement of the affected limb with the training target. It integrates active participation and motor performance, and uses a reinforcement learning algorithm to adaptively adjust the auxiliary parameters of the rehabilitation training system, thereby optimizing the patient's interactive experience and improving the rehabilitation training effect.

[0095] S50: Based on reinforcement learning algorithms, the maximum assist force and movement speed of the rehabilitation robot are dynamically adjusted according to the active participation and motor performance of the target patient, so as to realize personalized dynamic rehabilitation feedback control of the rehabilitation robot.

[0096] Specifically, the personalized dynamic rehabilitation feedback control strategy aims to improve patients' active participation and motor performance. It designs an evaluation function E for the patient's motor status during rehabilitation training, expressed as:

[0097] E=α×P%+β×(1-MSE%); (5)

[0098] Here, α and β are the weighting coefficients for the percentage of active participation (P%) and the mean absolute percentage error (MSE%), respectively, and their sum is 1. Setting α = β = 0.5 indicates that the improvement of the patient's active participation and motor performance are equally important.

[0099] The training parameters of the rehabilitation robot, including maximum assist force and movement speed, are adaptively adjusted based on a deep Q-network; for example... Figure 8 As shown, Figure 8 This is a schematic diagram of the deep Q-network structure. The input consists of state and action. After passing through hidden layers and adjusting the parameters θ, the final output is the Q-value Q(s) corresponding to each action. t ,a t A schematic diagram of the training process of a deep Q-network is shown below. Figure 9 As shown, firstly, a deep Q-network is initialized to estimate the Q-value between the environmental state and action pair, serving as the estimation network. A target value network is then created, with the same structure as the estimation network but different parameters. Next, a playback memory unit is initialized to store the previous state, action, reward, and next state, as well as to initialize the environmental model, including initializing the state and action. Finally, a training loop is performed. Based on a greedy strategy, an action is selected from the estimation network according to the current state and applied to adjust the interaction between the rehabilitation robot and the environment, observing the reward and the next state. The current state, action, reward, and next state are stored in the playback memory unit, and a batch of empirical data is randomly sampled from it. The target value network is used to calculate the target Q-value. The mean squared error loss function is used to compare the error between the estimated Q-value and the target Q-value. Gradient descent and the gradient of the loss function are used to update the parameters of the estimation network, reducing the gap between the estimated Q-value and the target Q-value. The parameters of the estimation network are periodically copied to the target value network to stabilize the calculation of the target Q-value until a stopping condition is met. The trained estimation network is used to select actions to dynamically adjust the training parameters of the rehabilitation robot, improving the patient's active participation and motor performance.

[0100] This invention dynamically adjusts the training parameters of a rehabilitation robot, including maximum assist force and movement speed, using a deep Q-network model. First, a simulation environment is created based on the patient's physiological characteristics and the nature of the rehabilitation robot and task, simulating the interaction between the patient and the robot. The patient's training is evaluated using an evaluation function E, providing feedback signals. This invention employs a deep Q-network as a reinforcement learning agent, which learns how to adjust the rehabilitation robot's training parameters based on the feedback signals from the simulation environment to maximize the evaluation function E. A reward function is designed based on the changes in the evaluation function E; when the evaluation function E increases after the training parameters are adjusted, the reward function is set to 1; otherwise, it is set to -1. To ensure the safety of the patient during rehabilitation training, a safety range is set for all training parameters; when the agent adjusts the training parameters outside the safety range, the reward function is set to -5. The reinforcement learning agent interacts with the simulation environment to train the deep Q-network, continuously adjusting the rehabilitation robot's training parameters to maximize cumulative rewards. After training, the effectiveness of the reinforcement learning model is verified in a real clinical environment under the supervision of a therapist, completing model deployment and realizing dynamic feedback control of the rehabilitation robot during rehabilitation training, improving the patient's active participation and motor performance. The training process is as follows: Figure 9 As shown. Figure 10 A schematic diagram illustrating the transition of a reinforcement learning model from a simulation environment to a real environment.

[0101] This invention comprehensively considers three key factors in the rehabilitation training process: the patient's intention to exercise, active participation, and motor performance. It can provide more personalized robot-assisted rehabilitation training based on the patient's active wishes, thereby improving the patient's initiative in participating in exercise training and increasing the efficiency of rehabilitation treatment.

[0102] This invention is based on the principle of mirror rehabilitation. It designs training tasks and goals according to the kinematic and dynamic parameters of the patient's healthy limb, and evaluates the patient's motor performance by comparing the actual movement of the patient's affected limb with the training goals. This provides patients with more personalized rehabilitation training guidance and is conducive to the rehabilitation of the motor function of the patient's affected limb to the point of coordination with the healthy side.

[0103] This invention sets target exercise parameters corresponding to the healthy side according to the patient's exercise intention, enabling the patient to achieve more reasonable training goals, improving the patient's initiative in participating in rehabilitation training, facilitating the realization of active rehabilitation, promoting the remodeling of the patient's neurological function, and accelerating the rehabilitation progress.

[0104] This invention recognizes motor intention by fusing raw surface electromyography signals and motor unit discharge features. Compared with motor intention recognition methods that only use time-frequency domain features, motor unit discharge features provide neural drive-related information for motor intention recognition, which helps to perform more stable and precise continuous motor intention recognition.

[0105] This invention quantifies the patient's active participation in rehabilitation training by using the average discharge frequency of the motor unit and the cumulative pulses per unit time. In order to eliminate the influence of individual differences on the quantitative indicators, the active participation is normalized by using the average discharge frequency of the patient's healthy side and the cumulative pulses per unit time, and expressed as a percentage, which can provide more intuitive and accurate feedback to the patient and the rehabilitation training system.

[0106] This invention comprehensively considers the patient's motor intention, active participation, and actual motor performance to achieve dynamic rehabilitation feedback control of the rehabilitation training system, providing patients with training guidance that better matches their motor abilities and rehabilitation needs. Furthermore, it uses reinforcement learning algorithms to dynamically adjust the auxiliary parameters of the rehabilitation training system, thereby improving and maintaining the patient's initiative in rehabilitation training and promoting the effective remodeling of neurological function.

[0107] This invention aims to provide patients with personalized and intelligent robot-assisted rehabilitation training by comprehensively considering motor intention, active participation, and motor performance, thereby enhancing patients' initiative in participating in motor rehabilitation and promoting effective remodeling of their neurological function. This invention collects the joint angles and distal forces of the patient's unaffected side and designs rehabilitation training tasks and motor goals according to the rehabilitation training plan, recording the active participation of the unaffected side when performing different movements. Based on the principle of mirror rehabilitation, it ensures the recovery of bilateral limb coordination. First, surface electromyography (EMG) signals are decomposed into motor unit discharge pulse sequences using a deep convolutional neural network. Second, a deep transform neural network is used to fuse the motor unit discharge pulse sequences and surface EMG signal features to predict the patient's joint movement angles and distal forces, enabling real-time monitoring of the patient's continuous motor intention and adaptively selecting training tasks and goals for the patient. Simultaneously, the average discharge frequency and cumulative pulse sequence of the motor units are used to quantify the patient's active participation during rehabilitation training, and the active participation of the healthy side is used for standardization to achieve stable and personalized assessment of active participation. Finally, the patient's motor performance is evaluated and recorded based on the average absolute error percentage between the patient's actual motor parameters and the target motor parameters. Finally, based on reinforcement learning algorithms, the training parameters of the rehabilitation robot, such as the maximum assist force and movement speed, are dynamically adjusted according to the patient's active participation and motor performance. This enables personalized dynamic rehabilitation feedback control of the rehabilitation robot, improving rehabilitation efficiency and accelerating rehabilitation progress.

[0108] Furthermore, such as Figure 11 As shown, based on the above-described personalized dynamic rehabilitation human-computer interaction method, the present invention also provides a personalized dynamic rehabilitation human-computer interaction system, wherein the personalized dynamic rehabilitation human-computer interaction system includes:

[0109] The training task and motion target design module 51 is used to obtain the target parameters of the target patient during rehabilitation training, as well as the average discharge frequency of the motor unit and the cumulative pulse per unit time corresponding to each movement, and store them in the rehabilitation training system.

[0110] The surface electromyography (EMG) signal acquisition and decomposition module 52 is used to attach high-density flexible electrodes to the corresponding muscles of the target patient's limb, acquire surface EMG signals, and use a sliding window to segment the surface EMG signals in real time to obtain the surface EMG signal time window, and extract the motor unit discharge pulse sequence from the surface EMG signal time window.

[0111] The motion intention recognition module 53 is used to input the surface electromyography signal time window and the motor unit discharge pulse sequence into the feature extraction module respectively, and perform feature fusion through deep transform neural network to regress and predict the patient's joint motion angle and end force.

[0112] The active participation quantification and motor performance assessment module 54 is used to calculate the average discharge frequency and cumulative pulses per unit time of the motor unit, quantify the active participation of the target patient in the rehabilitation training process, and use the ratio between the average discharge frequency and cumulative pulses per unit time during the exercise process and the target average discharge frequency and cumulative pulses to represent the percentage of the patient's active participation. The patient's motor performance is assessed and recorded based on the average absolute percentage error between the patient's actual motor parameters and the target motor parameters.

[0113] The personalized dynamic rehabilitation feedback control module 55 is used to dynamically adjust the maximum assist force and movement speed of the rehabilitation robot based on the active participation and motor performance of the target patient, using a reinforcement learning algorithm, so as to achieve personalized dynamic rehabilitation feedback control of the rehabilitation robot.

[0114] Furthermore, such as Figure 12 As shown, based on the above-mentioned personalized dynamic rehabilitation human-computer interaction method and system, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 12 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0115] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a personalized dynamic rehabilitation human-computer interaction program 40, which can be executed by the processor 10 to implement the personalized dynamic rehabilitation human-computer interaction method of this application.

[0116] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the personalized dynamic rehabilitation human-computer interaction method.

[0117] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0118] In one embodiment, when the processor 10 executes the personalized dynamic rehabilitation human-computer interaction program 40 in the memory 20, the following steps are performed:

[0119] The system acquires the target parameters for rehabilitation training of the target patient, as well as the average discharge frequency of the motor unit and the cumulative pulses per unit time corresponding to each movement, and stores them in the rehabilitation training system.

[0120] High-density flexible electrodes are attached to the corresponding muscles of the target patient's limb to collect surface electromyography (EMG) signals. The surface EMG signals are then segmented in real time using a sliding window to obtain a time window of the surface EMG signals. The motor unit discharge pulse sequence is then extracted from the time window of the surface EMG signals.

[0121] The surface electromyography signal time window and motor unit discharge pulse sequence are respectively input into the feature extraction module, and the features are fused by the deep transform neural network to regress and predict the patient's joint motion angle and end force.

[0122] The average discharge frequency and cumulative pulses per unit time of the motor unit are calculated to quantify the active participation of the target patient in the rehabilitation training process. The percentage of active participation of the patient is expressed by the ratio between the average discharge frequency and cumulative pulses per unit time during the exercise process and the target average discharge frequency and cumulative pulses. The patient's exercise performance is evaluated and recorded based on the average absolute percentage error between the patient's actual exercise parameters and the target exercise parameters.

[0123] Based on reinforcement learning algorithms, the maximum assist force and movement speed of the rehabilitation robot are dynamically adjusted according to the target patient's active participation and motor performance, so as to achieve personalized dynamic rehabilitation feedback control of the rehabilitation robot.

[0124] The acquisition of target parameters during rehabilitation training for the target patient, as well as the average discharge frequency of the motor unit and the cumulative pulses per unit time corresponding to each movement, and storage of these parameters in the rehabilitation training system, specifically includes:

[0125] Before the training begins, according to the rehabilitation training plan, for the first target patient with hemiplegia, the joint motion angles and end force of the healthy limbs of the first target patient when performing various daily living activities are collected and stored in the rehabilitation training system as target parameters for the first target patient's rehabilitation training.

[0126] For the second target patient with motor impairment in both limbs, the average joint motion angle and average end force of healthy people of similar age and height to the second target patient when performing various daily living activities are collected and stored in the rehabilitation training system as target parameters for the rehabilitation training of the second target patient.

[0127] The average discharge frequency of the motor unit corresponding to each movement and the cumulative pulses per unit time are stored in the rehabilitation training system to calculate the patient's active participation.

[0128] Specifically, the extraction of the motor unit discharge pulse sequence from the surface electromyography signal time window includes:

[0129] Each surface electromyography signal time window is used as a sample input into a deep convolutional neural network model to extract the firing pulse sequence of the motor unit;

[0130] The deep convolutional neural network model includes four one-dimensional convolutional layers, two max-pooling layers, one fully connected layer, and a sigmoid activation function.

[0131] Specifically, the step of using each surface electromyography signal time window as a sample input to a deep convolutional neural network model to extract the discharge pulse sequence of the motor unit includes:

[0132] Electromyographic features in the time window of surface electromyographic signals are extracted by two one-dimensional convolutional layers. Max pooling is used for feature selection and dimensionality reduction. Deep electromyographic features are then extracted by two more one-dimensional convolutional layers, and max pooling is used for final feature selection and dimensionality reduction. Finally, the output is the discharge pulse sequence of all activated motor units after passing through a fully connected layer and a sigmoid activation function. This sequence is used for motion intention recognition and active participation measurement.

[0133] Specifically, the step of inputting the surface electromyography signal time window and motor unit discharge pulse sequence into the feature extraction module, and performing feature fusion through a deep transform neural network to regress and predict the patient's joint motion angle and end force, includes:

[0134] The feature extraction module, constructed using a convolutional neural network, extracts features from the time window of surface electromyography signals and the pulse sequence of motor units, respectively. The attention mechanism in the encoding module of the deep transform neural network is used to extract and filter features. Finally, the joint angle and end force are output by the fully connected layer.

[0135] The rehabilitation training system automatically selects appropriate training tasks and exercise goals based on the target patient's exercise intentions, including joint angles and end-effector forces. The system displays the target tasks and related exercise parameters on the human-computer interaction interface, guides the patient to complete the corresponding training movements through visual and voice prompts, and displays the completion status of the movements.

[0136] The calculation of the average discharge frequency and cumulative pulses per unit time of the motor unit quantifies the active participation of the target patient during rehabilitation training. The ratio of the average discharge frequency and cumulative pulses per unit time during exercise to the target average discharge frequency and cumulative pulses represents the percentage of the patient's active participation. The patient's motor performance is assessed and recorded based on the average absolute percentage error between the patient's actual motor parameters and the target motor parameters. Specifically, this includes:

[0137] The percentage of active participation (P%) is expressed as the ratio between the patient's active participation during exercise and the target average discharge frequency and cumulative pulses.

[0138]

[0139] Where, r t r is the average discharge frequency of the moving unit during actual motion. g c is the target average discharge frequency of the motion unit. t c represents the cumulative pulses per unit time of the moving unit during actual motion. gThe target cumulative pulse per unit time for the moving unit;

[0140] Joint angle θ obtained using actual motion measurements i and end force F j , and the target joint angle θ set according to the motion intention i ′ and target end force F j The mean absolute percentage error (MSE%) between the two values ​​assesses the patient's actual motor performance:

[0141]

[0142] Where i represents the i-th joint, m is the total number of joint angles measured, j represents the j-th distal end, and n is the total number of distal force measurements; the smaller the mean absolute percentage error (MSE%), the better the patient's motor performance.

[0143] Specifically, the method of dynamically adjusting the maximum assist force and movement speed of the rehabilitation robot based on the reinforcement learning algorithm and the target patient's active participation and motor performance to achieve personalized dynamic rehabilitation feedback control of the rehabilitation robot includes:

[0144] Design an evaluation function E for the movement status of the target patient during rehabilitation training, expressed as:

[0145] E=α×P%+β×(1-MSE%);

[0146] Wherein, α and β are the weighting coefficients of the percentage of active participation P% and the mean absolute percentage error MSE%, respectively, and the sum of the two is 1. Setting α = β = 0.5 indicates that the improvement of patients' active participation and motor performance is equally important.

[0147] The training parameters of the rehabilitation robot, including maximum assist force and movement speed, are adaptively adjusted based on a deep Q-network.

[0148] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a personalized dynamic rehabilitation human-computer interaction program, and the personalized dynamic rehabilitation human-computer interaction program, when executed by a processor, implements the steps of the personalized dynamic rehabilitation human-computer interaction method as described above.

[0149] In summary, this invention provides a personalized dynamic rehabilitation human-computer interaction method and related equipment. The method includes: acquiring target parameters for rehabilitation training of the target patient, as well as the average discharge frequency of the motor unit and the cumulative pulses per unit time corresponding to each movement, and storing them in the rehabilitation training system; attaching high-density flexible electrodes to the corresponding muscles of the target patient's limb, collecting surface electromyography (EMG) signals, and using a sliding window to segment the surface EMG signals in real time to obtain a surface EMG signal time window, and extracting the motor unit discharge pulse sequence from the surface EMG signal time window; inputting the surface EMG signal time window and the motor unit discharge pulse sequence into a feature extraction module, and performing feature fusion through a deep transform neural network. This invention combines regression analysis to predict the joint motion angles and end-effector forces of patients; calculates the average discharge frequency and cumulative pulses per unit time of the motor units to quantify the active participation of the target patient during rehabilitation training; and uses the ratio between the average discharge frequency and cumulative pulses per unit time during movement and the target average discharge frequency and cumulative pulses to represent the percentage of active participation. The patient's motor performance is assessed and recorded based on the mean absolute percentage error between the patient's actual motor parameters and the target motor parameters. Based on a reinforcement learning algorithm, the maximum assist force and movement speed of the rehabilitation robot are dynamically adjusted according to the target patient's active participation and motor performance to achieve personalized dynamic rehabilitation feedback control of the rehabilitation robot. This invention provides more personalized, dynamic, and precise auxiliary support, improves the human-computer interaction experience, thereby increasing the patient's rehabilitation initiative, promoting neural function remodeling, and improving the effectiveness of rehabilitation treatment.

[0150] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0151] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0152] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A personalized dynamic rehabilitation human-computer interaction method, characterized in that, The personalized dynamic rehabilitation human-computer interaction method includes: The system acquires the target parameters for rehabilitation training of the target patient, as well as the average discharge frequency of the motor unit and the cumulative pulses per unit time corresponding to each movement, and stores them in the rehabilitation training system. High-density flexible electrodes are attached to the corresponding muscles of the target patient's limb to collect surface electromyography (EMG) signals. The surface EMG signals are then segmented in real time using a sliding window to obtain a time window of the surface EMG signals. The motor unit discharge pulse sequence is then extracted from the time window of the surface EMG signals. The extraction of the motor unit discharge pulse sequence from the surface electromyography signal time window specifically includes: Each surface electromyography signal time window is used as a sample input into a deep convolutional neural network model to extract the firing pulse sequence of the motor unit; The deep convolutional neural network model includes four one-dimensional convolutional layers, two max pooling layers, one fully connected layer, and a sigmoid activation function. The method of using each surface electromyography (EMG) signal time window as a sample input to a deep convolutional neural network model to extract the discharge pulse sequence of motor units specifically includes: extracting EMG features from the surface EMG signal time window through two one-dimensional convolutional layers, using a max pooling layer for feature selection and dimensionality reduction, then extracting deep EMG features through two more one-dimensional convolutional layers, and using a max pooling layer for final feature selection and dimensionality reduction, and finally outputting the discharge pulse sequence of all activated motor units through a fully connected layer and a sigmoid activation function for motion intention recognition and active participation measurement; The surface electromyography signal time window and motor unit discharge pulse sequence are respectively input into the feature extraction module, and feature fusion is performed through deep transform neural network to regress and predict the patient's joint motion angle and end force. The average discharge frequency and cumulative pulses per unit time of the motor unit are calculated to quantify the active participation of the target patient in the rehabilitation training process. The percentage of active participation of the patient is expressed by the ratio between the average discharge frequency and cumulative pulses per unit time during the exercise process and the target average discharge frequency and cumulative pulses. The patient's exercise performance is evaluated and recorded based on the average absolute percentage error between the patient's actual exercise parameters and the target exercise parameters. The calculation of the average discharge frequency and cumulative pulses per unit time of the motor unit quantifies the active participation of the target patient during rehabilitation training. The percentage of active participation is expressed as the ratio between the average discharge frequency and cumulative pulses per unit time during exercise and the target average discharge frequency and cumulative pulses. The patient's motor performance is assessed and recorded based on the mean absolute percentage error between the patient's actual motor parameters and the target motor parameters. Specifically, this includes: The percentage of active participation in a patient is expressed as the ratio between the active participation during exercise and the target average discharge frequency and cumulative pulses. : ； in, This represents the average discharge frequency of the moving unit during actual motion. The target average discharge frequency of the moving unit. This refers to the cumulative pulses per unit time of the moving unit during actual motion. The target cumulative pulse per unit time for the moving unit; Joint angles obtained using actual motion measurements and end force , and the target joint angle set according to the motion intention and target end force Mean absolute percentage error between Assess the patient's actual motor performance: ； Where i represents the i-th joint, m is the total number of measured joint angles, j represents the j-th distal end, and n is the total number of measured distal end forces; mean absolute percentage error The smaller the size, the better the patient's motor performance; Based on reinforcement learning algorithms, the maximum assist force and movement speed of the rehabilitation robot are dynamically adjusted according to the target patient's active participation and motor performance, so as to achieve personalized dynamic rehabilitation feedback control of the rehabilitation robot.

2. The personalized dynamic rehabilitation human-computer interaction method according to claim 1, characterized in that, The acquisition of target parameters during rehabilitation training for the target patient, as well as the average discharge frequency of the motor unit and the cumulative pulses per unit time corresponding to each movement, and storage of these parameters in the rehabilitation training system, specifically includes: Before the training begins, according to the rehabilitation training plan, for the first target patient with hemiplegia, the joint motion angles and end force of the healthy limbs of the first target patient when performing various daily living activities are collected and stored in the rehabilitation training system as target parameters for the first target patient's rehabilitation training. For the second target patient with motor impairment in both limbs, the average joint motion angle and average end force of healthy people of similar age and height to the second target patient when performing various daily living activities are collected and stored in the rehabilitation training system as target parameters for the rehabilitation training of the second target patient. The average discharge frequency of the motor unit corresponding to each movement and the cumulative pulses per unit time are stored in the rehabilitation training system to calculate the patient's active participation.

3. The personalized dynamic rehabilitation human-computer interaction method according to claim 1, characterized in that, The step involves inputting the surface electromyography signal time window and the motor unit discharge pulse sequence into the feature extraction module, and then fusing the features through a deep transform neural network to regress and predict the patient's joint motion angle and end force. Specifically, this includes: The feature extraction module, constructed using a convolutional neural network, extracts features from the time window of surface electromyography signals and the pulse sequence of motor units, respectively. The attention mechanism in the encoding module of the deep transform neural network is used to extract and filter features. Finally, the joint angle and end force are output by the fully connected layer. The rehabilitation training system automatically selects appropriate training tasks and exercise goals based on the target patient's exercise intentions, including joint angles and end-effector forces. The system displays the target tasks and related exercise parameters on the human-computer interaction interface, guides the patient to complete the corresponding training movements through visual and voice prompts, and displays the completion status of the movements.

4. The personalized dynamic rehabilitation human-machine interaction method of claim 1, wherein, The method, based on reinforcement learning algorithms, dynamically adjusts the maximum assist force and movement speed of the rehabilitation robot according to the target patient's active participation and motor performance, to achieve personalized dynamic rehabilitation feedback control of the rehabilitation robot. Specifically, this includes: Design a function to evaluate the movement status of target patients during rehabilitation training. , represented as: ； in, and Percentage of active participation and mean absolute percentage error The weighting coefficients, the sum of which is 1, are set. This indicates that the improvement in the patient's active participation and motor performance is equally important; The training parameters of the rehabilitation robot, including maximum assist force and movement speed, are adaptively adjusted based on a deep Q-network.

5. A personalized dynamic rehabilitation human-machine interaction system, characterized in that, The personalized dynamic rehabilitation human-computer interaction system is used to implement the personalized dynamic rehabilitation human-computer interaction method according to any one of claims 1-4, and the personalized dynamic rehabilitation human-computer interaction system includes: The training task and motion target design module is used to obtain the target parameters of the target patient during rehabilitation training, as well as the average discharge frequency of the motor unit and the cumulative pulses per unit time corresponding to each movement, and store them in the rehabilitation training system. The surface electromyography (EMG) signal acquisition and decomposition module is used to attach high-density flexible electrodes to the corresponding muscles of the target patient's limb, acquire surface EMG signals, and use a sliding window to segment the surface EMG signals in real time to obtain the surface EMG signal time window, and extract the motor unit discharge pulse sequence from the surface EMG signal time window. The motion intention recognition module is used to input the surface electromyography signal time window and the motor unit discharge pulse sequence into the feature extraction module, and perform feature fusion through deep transform neural network to regress and predict the patient's joint motion angle and end force; The active participation quantification and motor performance assessment module is used to calculate the average discharge frequency and cumulative pulses per unit time of motor units, quantify the active participation of the target patient in the rehabilitation training process, and use the ratio between the average discharge frequency and cumulative pulses per unit time during the exercise process and the target average discharge frequency and cumulative pulses to represent the percentage of the patient's active participation. The patient's motor performance is assessed and recorded based on the average absolute percentage error between the patient's actual motor parameters and the target motor parameters. The personalized dynamic rehabilitation feedback control module is used to dynamically adjust the maximum assist force and movement speed of the rehabilitation robot based on the target patient's active participation and motor performance, using reinforcement learning algorithms, in order to achieve personalized dynamic rehabilitation feedback control of the rehabilitation robot.

6. A terminal, characterized by comprising: The terminal includes: a memory, a processor, and a personalized dynamic rehabilitation human-computer interaction program stored in the memory and executable on the processor. When the personalized dynamic rehabilitation human-computer interaction program is executed by the processor, it implements the steps of the personalized dynamic rehabilitation human-computer interaction method as described in any one of claims 1-4.

7. A computer readable storage medium characterized in that, The computer-readable storage medium stores a personalized dynamic rehabilitation human-computer interaction program, which, when executed by a processor, implements the steps of the personalized dynamic rehabilitation human-computer interaction method as described in any one of claims 1-4.

Images