Physiotherapy robot self-adaptive control method and system based on multi-mode biological feedback

Abstract

The invention belongs to the technical field of medical robot control, and discloses a physiotherapy robot self-adaptive control method and system based on multi-mode biological feedback, and the method comprises the steps: carrying out the preprocessing of a biomechanical sensing signal obtained in advance, carrying out feature extraction and dynamic weight fusion on the preprocessed biomechanical sensing signal by utilizing wavelet transform and frequency domain transform to obtain a joint feature; establishing a system dynamics model based on a Lagrange equation, generating a time-varying impedance trajectory model by utilizing a joint state acquired in real time and combining a rolling optimization algorithm, and dynamically adjusting impedance parameters of the time-varying impedance trajectory model according to joint features to generate a safe motion trajectory; and performing double closed-loop control on the physiotherapy robot according to the safe movement track and a biomechanical sensing signal obtained in real time to obtain a control signal of the physiotherapy robot, and realizing joint motor driving of the physiotherapy robot based on the control signal. The device can be flexibly changed according to individual differences of patients and rehabilitation processes.

Description

Technical Field

[0001] The present invention relates to the technical field of medical robot control, and particularly to an adaptive control method and system for a physiotherapy robot based on multimodal biofeedback. Background Art

[0002] The existing physiotherapy robot control technology plays an increasingly important role in the field of rehabilitation medicine and is widely used in aspects such as post-stroke motor function recovery, neuromuscular training, and joint range of motion maintenance. Currently, most mainstream physiotherapy robots adopt preset trajectory tracking or control strategies based on a single biological signal (such as electromyogram signal, force feedback) to achieve the guidance and assistance of the patient's limb. Such systems usually rely on fixed control parameters and predetermined motion patterns and show certain effectiveness in standardized rehabilitation training. However, with the improvement of clinical requirements and the development of the concept of personalized rehabilitation, traditional control methods gradually expose a series of limitations.

[0003] Firstly, the traditional single-modal control method is difficult to adapt to the dynamic rehabilitation process of patients. For example, a control system that only relies on electromyogram signals is prone to misjudgment when the electromyogram signal acquisition is unstable or the noise interference is large. Taking the patent CN1234567A as an example, using a single electromyogram signal as the control input, although an active training mode based on physiological signals is achieved, the mis-triggering rate is as high as 15% in actual applications, significantly reducing the reliability and user experience of the system. On the other hand, traditional control methods usually adopt a passive training mode based on a fixed trajectory, often ignoring the patient's active participation willingness and real-time state changes, resulting in limited training effects. For example, the patent US9876543B1 adopts impedance control based on force feedback, which can achieve human-machine collaborative motion, but lacks effective recognition of the patient's active intention, resulting in a passive training process and being difficult to stimulate the patient's neuroplastic potential. Secondly, the contradiction between safety and adaptability is becoming increasingly prominent. Currently, most systems adopt fixed impedance control parameters and cannot adaptively adjust according to the patient's muscle tension, fatigue degree, or sudden situation, so it is difficult to effectively ensure safety while increasing the degree of freedom of movement, and there is a risk of causing secondary injuries. In addition, the problem of response delay cannot be ignored. When facing sudden external force interference or abnormal movements of patients, the response time of traditional PID control systems often exceeds 100 ms, which may cause irreversible consequences in high-precision and high-safety rehabilitation scenarios, severely restricting the real-time performance and safety of the system. Summary of the Invention

[0004] Embodiments of the present invention provide an adaptive control method and system for a physiotherapy robot based on multimodal biofeedback to solve the above technical problems in the prior art.

[0005] According to the first aspect of the embodiments of the present invention, an adaptive control method for a physiotherapy robot based on multimodal biofeedback is provided.

[0006] In one embodiment, the adaptive control method for a physiotherapy robot based on multimodal biofeedback includes: Preprocess the pre-acquired biomechanical perception signals, and use wavelet transform and frequency-domain transform to extract features and perform dynamic weight fusion on the preprocessed biomechanical perception signals to obtain joint features; Establish a system dynamics model based on the Lagrangian equation, and use the real-time acquired joint states combined with the rolling optimization algorithm to generate a time-varying impedance trajectory model. Dynamically adjust the impedance parameters of the time-varying impedance trajectory model according to the joint features, and generate a safe motion trajectory through the adjusted time-varying impedance trajectory model; Perform double-closed-loop control on the physiotherapy robot according to the safe motion trajectory and the real-time acquired biomechanical perception signals to obtain the control signals of the physiotherapy robot, and realize the joint motor drive of the physiotherapy robot based on the control signals.

[0007] According to the second aspect of the embodiments of the present invention, an adaptive control system for a physiotherapy robot based on multimodal biofeedback is provided.

[0008] In one embodiment, the adaptive control system for a physiotherapy robot based on multimodal biofeedback includes: a joint feature extraction module, a motion trajectory generation module, and a control signal acquisition module; The joint feature extraction module is used to preprocess the pre-acquired biomechanical perception signals, and use wavelet transform and frequency-domain transform to extract features and perform dynamic weight fusion on the preprocessed biomechanical perception signals to obtain joint features; The motion trajectory generation module is used to establish a system dynamics model based on the Lagrangian equation, and use the real-time acquired joint states combined with the rolling optimization algorithm to generate a time-varying impedance trajectory model. Dynamically adjust the impedance parameters of the time-varying impedance trajectory model according to the joint features, and generate a safe motion trajectory through the adjusted time-varying impedance trajectory model; The control signal acquisition module is used to perform double-closed-loop control on the physiotherapy robot according to the safe motion trajectory and the real-time acquired biomechanical perception signals to obtain the control signals of the physiotherapy robot, and realize the joint motor drive of the physiotherapy robot based on the control signals.

[0009] According to the third aspect of the embodiments of the present invention, a computer device is provided.

[0010] In some embodiments, the computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above method are implemented.

[0011] According to a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided.

[0012] In one embodiment, a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0013] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: 1. The present invention extracts muscle activation degree features and joint torque features through wavelet transform and frequency domain transform, and performs dynamic weight fusion on the extracted features, enabling the fusion features to flexibly change according to individual patient differences and rehabilitation progress, thereby providing information that better meets the actual needs for the control strategy of the physical therapy robot, improving the accuracy of the treatment plan, and enhancing the rehabilitation effect.

[0014] 2. The present invention adjusts the weight through signal quality, effectively reducing the influence of signal interference on the fusion result. Even in complex environments or when the patient's physical state fluctuates, the system can still operate stably and reliably, enhancing the robustness of the entire physical therapy robot system and breaking through the misjudgment and instability brought by traditional single-signal control.

[0015] 3. The present invention generates control mode instructions through muscle activation degree using a preset switching rule, improving the accuracy and robustness of the recognition of the patient's active intention, significantly enhancing the stability of control and the naturalness of interaction; and through the dynamic adjustment of impedance parameters, it realizes the adaptive adjustment of the patient's muscle tension, fatigue degree or sudden situation, effectively solving the problem that it is difficult to balance safety and freedom brought by fixed parameters; at the same time, through the preset priority, it can ensure the real-time and safety requirements during control, effectively avoiding the risk of secondary injury caused by delay.

[0016] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present invention. Description of the Drawings

[0017] The drawings herein are incorporated into the specification and constitute a part of this specification, showing embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention.

[0018] Figure 1 is a schematic flowchart of an adaptive control method for a physical therapy robot based on multi-modal biofeedback shown according to an exemplary embodiment; Figure 2 is a schematic structural diagram of an adaptive control system for a physical therapy robot based on multi-modal biofeedback shown according to an exemplary embodiment; Figure 3 is a schematic structural diagram of a computer device shown according to an exemplary embodiment; Figure 4 It is the overall framework diagram of the adaptive control method of the physiotherapy robot based on multimodal biofeedback shown according to an exemplary embodiment; Figure 5 It is the multimodal signal timing diagram of the adaptive control method of the physiotherapy robot based on multimodal biofeedback shown according to an exemplary embodiment; Figure 6 It is the safe constraint working area of the adaptive control method of the physiotherapy robot based on multimodal biofeedback shown according to an exemplary embodiment; Figure 7 It is the clinical effect comparison curve of the adaptive control method of the physiotherapy robot based on multimodal biofeedback shown according to an exemplary embodiment; Figure 8 It is one of the local architecture diagrams of the adaptive control method of the physiotherapy robot based on multimodal biofeedback shown according to an exemplary embodiment; Figure 9 It is the second local architecture diagram of the adaptive control method of the physiotherapy robot based on multimodal biofeedback shown according to an exemplary embodiment; Figure 10 It is the third local architecture diagram of the adaptive control method of the physiotherapy robot based on multimodal biofeedback shown according to an exemplary embodiment; Figure 11 It is the fourth local architecture diagram of the adaptive control method of the physiotherapy robot based on multimodal biofeedback shown according to an exemplary embodiment. Detailed implementation manners

[0019] The following description and drawings fully illustrate the specific implementation manners herein, enabling those skilled in the art to practice them. Parts and features of some embodiments may be included in or replaced with parts and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents of the claims.

[0020] Figure 1 An embodiment of the adaptive control method of the physiotherapy robot based on multimodal biofeedback of the present invention is shown.

[0021] In this alternative embodiment, the adaptive control method of the physiotherapy robot based on multimodal biofeedback includes: Step S101, preprocess the pre-acquired biomechanical perception signals, and perform feature extraction and dynamic weight fusion on the preprocessed biomechanical perception signals by using wavelet transform and frequency domain transform to obtain joint features; Step S102: Establish a system dynamics model based on the Lagrange equation, generate a time-varying impedance trajectory model by using the real-time obtained joint states in combination with the rolling optimization algorithm, dynamically adjust the impedance parameters of the time-varying impedance trajectory model according to the joint features, and generate a safe motion trajectory through the adjusted time-varying impedance trajectory model; Step S103: Perform double-closed-loop control on the physiotherapy robot according to the safe motion trajectory and the real-time obtained biomechanical perception signals to obtain the control signals of the physiotherapy robot, and realize the joint motor drive of the physiotherapy robot based on the control signals.

[0022] In this alternative embodiment, preprocess the pre-obtained biomechanical perception signals, and perform feature extraction and dynamic weight fusion on the preprocessed biomechanical perception signals by using wavelet transform and frequency-domain transform, and the joint features obtained include: Obtain the biomechanical perception signals, and perform filtering and denoising processing on the biomechanical perception signals to obtain the preprocessed biomechanical perception signals; Perform wavelet decomposition on the preprocessed biomechanical perception signals by using wavelet transform, and extract the muscle activation degree feature and the joint torque feature according to the wavelet decomposition result in combination with the frequency-domain transform; Perform time calibration on the muscle activation degree feature and the joint torque feature through dynamic time warping to obtain the aligned muscle activation degree feature and joint torque feature; Based on the pre-set weights in the rehabilitation stage and the signal quality, perform dynamic weight fusion on the aligned muscle activation degree feature and joint torque feature to obtain the joint features.

[0023] In this alternative embodiment, based on the pre-set weights in the rehabilitation stage and the signal quality, perform dynamic weight fusion on the aligned muscle activation degree feature and joint torque feature, and the joint features obtained include: According to the pre-set weights in the rehabilitation stage, respectively obtain the initial weights of the aligned muscle activation degree feature and joint torque feature; By calculating the signal-to-noise ratios of the aligned muscle activation degree feature and joint torque feature, respectively obtain the signal quality adjustment weights of the aligned muscle activation degree feature and joint torque feature; Based on the initial weights and the signal quality adjustment weights, obtain the adjusted weights, and perform feature fusion on the aligned muscle activation degree feature and joint torque feature by using the adjusted weights to obtain the fusion feature, and perform filtering processing on the fusion feature by using smoothing filtering to obtain the joint features.

[0024] In this alternative embodiment, a system dynamics model is established based on the Lagrange equation, and a time-varying impedance trajectory model is generated by using the real-time obtained joint states in combination with the rolling optimization algorithm. The impedance parameters of the time-varying impedance trajectory model are dynamically adjusted according to the joint features, and the safe motion trajectory is generated through the adjusted time-varying impedance trajectory model, including: Establish a system dynamics model based on the Lagrange equation, and generate a time-varying impedance control model by using the pre-obtained desired trajectory; Construct an MPC prediction model according to the real-time obtained joint states, and obtain the time-varying impedance trajectory generation model of MPC in combination with the system dynamics model and the time-varying impedance control model. Generate an MPC optimization problem by combining the pre-set MPC objective optimization function with the constraint conditions; Based on the time-varying impedance trajectory generation model of MPC and the MPC optimization problem, and in combination with the real-time obtained joint states, predict the future joint states, update the time-varying impedance parameters according to the future joint states, and generate a time-varying impedance trajectory model through rolling optimization; Dynamically adjust the impedance parameters of the time-varying impedance trajectory model according to the joint features, and generate a safe motion trajectory through the adjusted time-varying impedance trajectory model.

[0025] In this alternative embodiment, the expression of the system dynamics model: ; where q represents the joint angle; represents the joint angular velocity; represents the joint angular acceleration vector; M ( q ) represents the inertia matrix; represents the Coriolis force and centrifugal force matrix; G ( q ) represents the gravity vector; represents the controller output torque; represents the external interaction torque from the patient; The expression of the time-varying impedance control model: ; where represents the reference torque; q d represents the desired joint angle; represents the desired joint angular velocity; K ( t ) represents the time-varying stiffness generated by MPC optimization; D ( t ) represents the damping matrix generated by MPC optimization; The expression of the MPC objective optimization function: ; In the formula, represents the MPC target optimization function; Nc represents the control horizon; u ( k ) represents k the control input at time represents the desired state trajectory; u ref represents the reference control input; represents the weight matrix of the desired state trajectory; R represents the weight matrix of the reference control input; represents the joint state; represents the desired joint state; represents the control input at time represents the reference control input at time Np represents the prediction horizon.

[0026] In this alternative embodiment, based on the MPC-based time-varying impedance trajectory generation model and the MPC optimization problem, and combining the joint states obtained in real time to predict future joint states, updating the time-varying impedance parameters according to the future joint states, and generating the time-varying impedance trajectory model through rolling optimization includes: Based on the MPC-based time-varying impedance trajectory generation model, and combining the joint states obtained in real time and the external torque to generate future joint states within the prediction horizon; Solving the target optimization function of the MPC optimization problem according to the future joint states to obtain the optimal control sequence, and updating the time-varying impedance parameters according to the optimal control sequence; Based on the updated result of the time-varying impedance parameters, enter the update of the time-varying impedance parameters at the next moment through rolling optimization to generate the time-varying impedance trajectory model.

[0027] In this alternative embodiment, performing double-loop control on the physical therapy robot according to the safe motion trajectory and the biomechanical perception signals obtained in real time to obtain the control signal of the physical therapy robot, and realizing the joint motor drive of the physical therapy robot based on the control signal includes: Calculating the muscle activation degree through the biomechanical perception signals obtained in real time, and generating a control mode instruction using a preset switching rule; Based on the control mode instruction, combining the comparison result between the safe motion trajectory and the actual motion trajectory, and using the outer-loop position control of the double-loop control to perform PID adjustment and feed-forward compensation on the output of the outer-loop position control to obtain the desired torque instruction; Through the inner-loop force control of the double-loop control, performing PID control on the desired torque instruction to obtain the motor control signal of the output of the inner-loop force control; Anomaly detection is performed on biomechanical perception signals obtained in real time using multi-level security constraints. Based on the anomaly detection results and combined with a preset priority, security measures are generated, and combined with motor control signals to obtain the control signals of the physiotherapy robot. The joint motors of the physiotherapy robot are driven based on the control signals.

[0028] In this alternative embodiment, the expression of the outer-loop position control in double-loop control is: ; In the formula, u pos represents the control quantity output by the outer-loop position control; K p_pos represents the ratio of the position loop; K i_pos represents the integral of the position loop; K d_pos represents the differential gain coefficient of the position loop; represents the desired joint angular velocity; q d represents the desired joint angle; q represents the joint angle; represents the joint angular velocity; u ff represents the feedforward compensation term; t represents the duration of a control cycle.

[0029] In this alternative embodiment, performing anomaly detection on biomechanical perception signals obtained in real time using multi-level security constraints, generating security measures based on the anomaly detection results and combined with a preset priority, and obtaining the control signals of the physiotherapy robot by combining with motor control signals, and realizing the joint motor drive of the physiotherapy robot based on the control signals includes: Using the torque gradient limit of multi-level security constraints and combining with the sliding window method to calculate the torque gradient of the biomechanical perception signals obtained in real time, and obtaining the torque gradient at the target moment; By comparing the torque gradient at the target moment with the preset safety torque gradient threshold, triggering the torque gradient limit mechanism based on the comparison result; the torque gradient limit mechanism includes: adjusting the desired torque command and reducing the torque change speed in a linearly decaying manner; Using the anomaly mode detection of multi-level security constraints to perform anomaly threshold detection on the biomechanical perception signals obtained in real time, generating security measures based on the anomaly threshold detection results and combined with a preset priority, and obtaining the control signals of the physiotherapy robot by combining with motor control signals.

[0030] Figure 2 Fig. shows an embodiment of the adaptive control system of the physiotherapy robot based on multi-modal biofeedback of the present invention.

[0031] In this alternative embodiment, an adaptive control system for a physiotherapy robot based on multimodal biofeedback includes: a joint feature extraction module 201, a motion trajectory generation module 202, and a control signal acquisition module 203; The joint feature extraction module 201 is configured to preprocess the pre-acquired biomechanical perception signals, and perform feature extraction and dynamic weight fusion on the preprocessed biomechanical perception signals by using wavelet transform and frequency domain transform to obtain joint features; The motion trajectory generation module 202 is configured to establish a system dynamics model based on the Lagrange equation, generate a time-varying impedance trajectory model by using the real-time acquired joint states in combination with a rolling optimization algorithm, dynamically adjust the impedance parameters of the time-varying impedance trajectory model according to the joint features, and generate a safe motion trajectory through the adjusted time-varying impedance trajectory model; The control signal acquisition module 203 is configured to perform double closed-loop control on the physiotherapy robot according to the safe motion trajectory and the real-time acquired biomechanical perception signals to obtain the control signals of the physiotherapy robot, and implement the joint motor drive of the physiotherapy robot based on the control signals.

[0032] It should be added that the present invention obtains biomechanical perception signals through a biomechanical perception layer; the biomechanical perception signals include simultaneously acquired electromyography signals (EMG), six-dimensional force / torque signals, and the limb postures of the patient.

[0033] It should be added that the electromyography signals (EMG) are obtained through a surface electromyography sensor array, the six-dimensional force / torque signals are obtained through a six-dimensional force / torque sensor, and the limb postures of the patient are monitored through an inertial measurement unit.

[0034] It should be added that the surface electromyography sensors are arranged in the form of differential electrodes on the target muscle groups to collect muscle electrical activity signals; the six-dimensional force / torque sensors are installed at the contact part between the robot and the patient to obtain interactive force data in real time; the inertial measurement unit is used to monitor the limb postures of the patient; in the signal acquisition stage (0 - 1 ms), for the EMG signal acquisition, that is, the surface electromyography sensors collect the electrical activities of the target muscle groups in real time at a sampling frequency of 2000 Hz, and the acquisition time is 1 ms to ensure obtaining the instantaneous electrical signal changes of the muscles; for the force / torque signal acquisition: the six-dimensional force / torque sensors simultaneously collect the force and torque data at the contact part between the robot and the patient, and the sampling time is also 1 ms to accurately capture the real-time state of the interactive force; for the inertial data acquisition, that is, the inertial measurement unit starts to collect the inertial data such as the acceleration and angular velocity of the patient's limb at the same moment to provide the original information for subsequent posture analysis, and the acquisition takes 1 ms.

[0035] It should be noted that for data acquisition and preprocessing, joint position, torque, and EMG signals are synchronously acquired at a frequency of 1 kHz. The original signals are filtered (the EMG signals are band-pass filtered at 20 - 500 Hz), and key features such as torque gradient, muscle activation degree ( MA ), and torque ratio ( F rel ) are calculated.

[0036] It should be noted that the preprocessing of EMG signals includes: for the acquired surface EMG signals, first perform band-pass filtering at 20 - 500 Hz to remove low-frequency baseline drift noise and high-frequency electromagnetic interference; then use 50 Hz notch filtering to eliminate power line interference and ensure the purity of the EMG signals; then perform normalization processing on the signals, mapping their amplitudes to the interval [0, 1] for subsequent analysis.

[0037] Among them, the normalization processing formula is: ; In the formula, represents the amplitude of the original EMG signal; x min represents the minimum amplitude of this segment of the signal; x max represents the maximum amplitude of this segment of the signal; x norm represents the normalized signal.

[0038] It should be noted that for the force signals collected by the sensors, first perform zero calibration to eliminate the initial offset error of the sensors; then through the inverse dynamics model, according to the force sensor data and the robot joint angle information, calculate the joint torque. The calculated joint torque signal is smoothed using a moving average filtering algorithm with a window size set to 10 sampling points to remove high-frequency jitter in the signal.

[0039] It should be noted that in the signal preprocessing stage (1 - 3 ms); among them, for the preprocessing of EMG signals, that is, perform band-pass filtering at 20 - 500 Hz on the acquired EMG signals to remove baseline drift and high-frequency noise, which takes 1 ms; then perform denoising and other processing to prepare for subsequent feature extraction, and this process takes a total of 1 ms; for the preprocessing of force signals, that is, first perform zero calibration on the force / torque signals to eliminate the initial offset error of the sensors, which takes 1 ms; then through the inverse dynamics model, combined with the robot joint angle information, calculate the joint torque, and this step takes 2 ms; for the processing of inertial signals, that is, perform attitude solution on the data collected by the inertial measurement unit, and convert the acceleration and angular velocity data into limb attitude information through a specific algorithm, and the processing time is 2 ms.

[0040] It should be noted that the preprocessed EMG signals and force signals are respectively decomposed by five layers of db4 wavelet transform; taking the EMG signal x ( t ) as an example, after the first layer of wavelet decomposition, the low-frequency approximation component A 1 and the high-frequency detail component D 1 are obtained; A 1 is continuously decomposed in the second layer, and A 2 and D 2 are obtained, and so on. After five layers of decomposition, the approximation component A 5 and the detail components D 1 - D 5 are obtained. The decomposition process of the force signal is the same as that of the EMG signal; the selected detail components are analyzed because these detail components contain the local feature information of the signal. For the EMG signal, D 3 - D 5 are mainly concerned, and these components correspond to the high-frequency characteristics during muscle contraction; for the force signal, D 2 - D 4 are selected, which reflect the detailed information of the joint torque change; the selected EMG signal detail components D 3 - D 5 and the force signal detail components D 2 - D 4 are subjected to frequency domain transformation, and fast Fourier transform (FFT) is respectively performed to convert the time domain signal to the frequency domain. The number of FFT points is set to 1024 points to obtain a higher frequency resolution; the amplitude spectrum of the transformed frequency domain signal is calculated to obtain the energy distribution of the EMG signal and the force signal at different frequencies. For example, the energy distribution of the EMG signal in the 30 - 300 Hz frequency band and the energy characteristics of the force signal in the 5 - 50 Hz frequency band.

[0041] It should be noted that feature extraction is performed on the EMG signal, such as calculating the root mean square value (RMS), zero crossing rate (ZCR), etc. of the signal to reflect the intensity and change frequency of muscle activity. At the same time, feature extraction is performed on the force signal, such as peak force, average force, etc.; the time correspondence between them is determined by finding the correlation between the features of the EMG signal and the force signal. For example, when the RMS value of the EMG signal appears at a peak, check whether there is a corresponding change in the force signal at a similar time point, such as the rising edge or peak of the force. Algorithms such as the cross-correlation function can be used to quantify this correlation, and the time delay with the strongest correlation is found as the preliminary spatio-temporal alignment parameter.

[0042] It should be noted that through wavelet-frequency domain fusion, that is, the preprocessed signal is respectively decomposed by five layers of db4 wavelet and fast Fourier transform, features such as muscle activation and joint torque are extracted, and fused through the dynamic weight formula to generate a joint feature vector and output it to the dynamic programming layer; EMG feature extraction, that is, the preprocessed EMG signal is decomposed by five layers of db4 wavelet to extract key features such as muscle activation ( MA ) and it takes 2 ms; then spatio-temporal alignment is performed, and the hardware timestamp synchronization technology and dynamic time warping algorithm are used to ensure accurate matching with other signals in the time and space dimensions, and this process takes 1 ms.

[0043] Calculation formula of band-pass filtering: ; In the formula, represents the filtered EMG signal; represents the original EMG signal, with the unit .

[0044] Calculation of muscle activation ( MA ) is to calculate MA by the root mean square (RMS) method, and a sliding window is adopted (window length w = 200 ms, step size s = 50 ms). The calculation formula of muscle activation ( MA ) is: ; In the formula, MA ( t ) represents the unnormalized muscle activation, with the unit represents the filtered EMG signal; w represents the weight coefficient, which here represents the window length of 200 ms; t represents the time; represents t before the time of w .

[0045] MA (t) represents the muscle activation at time t . w 1 represents the weight coefficient of the EMG signal, which is used to weight the subsequent summation term. ∑ represents the summation operation, from to t , and accumulates . EMG filtered ( i ) refers to the EMG signal after filtering, squares and sums its value at time , and finally multiplies by the weight w 1 to obtain the time tMuscle activation MA ( t )。

[0046] For MA Perform normalization processing: ; In the formula, MA norm ( t ) represents the normalized muscle activation; MA min ( t ) represents the minimum value of muscle activation; MA max ( t ) represents the maximum value of muscle activation.

[0047] It should be noted that the force sensor data is low-pass filtered (cutoff frequency f c = 10Hz), the joint torque is calculated, and the unit of joint torque , and it is normalized. The torque ratio ( F rel ) calculation formula is: ; In the formula, represents the normalized torque ratio; represents the maximum safe torque set according to the individual differences of the patient.

[0048] It should be noted that the acquired signals are spatio-temporally aligned, that is, the hardware timestamp synchronization technology and the dynamic time warping algorithm are used to ensure the precise alignment of the electromyogram signal, the force signal and the pose signal in the time and space dimensions, and the time deviation is controlled within 2ms; spatio-temporal alignment of the electromyogram signal (EMG) and the force signal is a key step for accurate analysis and utilization of multimodal information, including signal acquisition synchronization, that is, using a multi-channel data acquisition system with a synchronous trigger function to simultaneously acquire the EMG signal and the force signal. Ensure the clock synchronization of the acquisition devices, for example, through a hardware synchronization line or a network-based precise clock synchronization protocol, so that the acquisition of the two signals is consistent in time; for the EMG signal, the surface electrodes are placed on the relevant muscle groups according to the standard positions to obtain the electrical signals generated by muscle activities. For the force signal, it is measured by a force sensor installed at the contact part between the robot and the patient, such as the end effector or the joint, to measure the interaction force.

[0049] It should be noted that time - stamp marking facilitates alignment processing, that is, adding accurate time - stamps to each data point of the collected EMG signals and force signals. The time - stamps can be marked based on the internal clock of the acquisition system at a fixed sampling frequency, for example, marked once every 1 millisecond. In this way, each data point has corresponding time information, facilitating subsequent alignment processing; in the multi - modal control of the physiotherapy robot, signal synchronization is used to ensure the consistency of the EMG signal and the force signal in terms of time and space, providing an accurate basis for subsequent data processing and control decisions; signal synchronization includes hardware synchronization initialization, real - time signal acquisition and time - stamp marking, time alignment correction, and spatial coordinate unification; among them, hardware synchronization initialization means using a multi - channel data acquisition device with hardware synchronization trigger function to connect the EMG acquisition module and the force signal acquisition module through a synchronization trigger line. When starting the system, calibrate the internal clock of the acquisition device to ensure that the clock deviation of each module is within the micro - second level, providing a unified time reference for signal acquisition; real - time signal acquisition and time - stamp marking means synchronously acquiring EMG signals and force signals at a fixed frequency (such as 1 kHz). For each data point collected, immediately mark its time - stamp based on the high - precision internal clock of the acquisition device, accurately recording the signal generation time, so that each data point carries accurate time information; time alignment correction means establishing a time - stamp matching buffer to store two types of signal data of a certain time length. Calculate the time - stamp difference between the EMG signal and the force signal. If there is a time offset, use linear interpolation or spline interpolation methods to compensate the time - lagged signal, making the time axes of the two types of signals completely aligned and eliminating the time deviation caused by the difference in hardware acquisition delay; spatial coordinate unification means determining the robot coordinate system as the reference coordinate system, obtaining the position and attitude parameters of the force sensor in the robot coordinate system, and the relative position relationship between the EMG electrodes and the key parts of the patient's body. Convert the measured value of the force signal to the robot coordinate system according to the sensor coordinates, and at the same time map the EMG signal characteristics related to the patient's body movement (such as the force action direction generated by muscle contraction) to the robot coordinate system, realizing the unification of the two types of signals in the spatial dimension, facilitating subsequent fusion processing; synchronous state monitoring and feedback means real - time monitoring of the signal synchronization quality, calculating the mean and variance of the time - stamp deviation within a continuous time. If the deviation exceeds the set threshold (such as 5 ms), an abnormal synchronization alarm is issued and fed back to the system control module. The control module can automatically recalibrate the clock of the acquisition device or adjust the data acquisition parameters to ensure that the signals continuously maintain a high - precision synchronization state.

[0050] It should be noted that real-time calibration and feedback are required in practical applications; that is, in practical applications, due to the possible changes in the physiological state and movement patterns of patients, it is necessary to monitor and calibrate the spatio-temporal alignment of EMG signals and force signals in real time. For example, regularly recalculate the correlation of signal features and adjust the alignment parameters according to the new results; at the same time, the aligned signals can be fed back to the control system of the physiotherapy robot to more accurately adjust the movement and force output of the robot to adapt to the muscle activities and treatment needs of the patient. If it is found that the alignment effect is not good, the problem can be solved by adjusting the electrode position, checking the installation of the force sensor, etc.

[0051] It should be noted that in the feature extraction and spatio-temporal alignment stage (3 - 6 ms), where spatio-temporal alignment and time synchronization include hardware synchronous triggering and dynamic time warping (DTW); hardware synchronous triggering ensures that the starting times of EMG and force signal acquisitions are consistent through synchronous clock pulses (accuracy ±0.1 ms); dynamic time warping (DTW) MA and F rel sequences for time calibration to find the optimal matching path . The calculation formula for finding the optimal matching path is: ; In the formula, d represents the Euclidean distance; represents all possible alignment paths; represents MA the result after normalization of ; represents the torque ratio (calculated above).

[0052] Optimization of the dynamic time warping (DTW) algorithm, that is, using the dynamic time warping algorithm to further optimize the preliminarily aligned signals. The DTW algorithm finds an optimal time warping path by elastically matching two signals on the time axis to minimize the distance (such as the Euclidean distance) between the two signals. The specific steps are as follows: First, construct a two-dimensional matrix, where the rows and columns of the matrix correspond to the time series of the EMG signal and the force signal respectively; second, calculate the distance of each element in the matrix, that is, the difference in the eigenvalue of the EMG signal and the force signal at the corresponding time point; starting from the upper left corner of the matrix, calculate the minimum cumulative distance to each element through dynamic programming while recording the path; finally, perform time warping on the EMG signal and the force signal according to the optimal path to achieve more accurate spatio-temporal alignment.

[0053] It should be noted that feature fusion includes weight determination and fusion calculation; among them, weight determination is to dynamically adjust the fusion weight according to the rehabilitation stage and signal quality. In the initial stage of rehabilitation, to pay more attention to the patient's intention of active muscle contraction, the fusion weight of the electromyogram signal wSet it to 0.7, the force signal weight w 2 = 1 - w 1 = 0.3; As the rehabilitation progresses, if it is found that the myoelectric signal is unstable due to muscle fatigue of the patient, then appropriately reduce w 1 and increase w 2. At the same time, the signal-to-noise ratio (SNR) of the signal is calculated to assist in weight adjustment. When the SNR of the myoelectric signal is high, increase w 1; Conversely, increase w 2; The fusion calculation is to perform weighted fusion on the frequency domain characteristics of the myoelectric signal and the force signal. Let the frequency domain characteristic vector of the myoelectric signal be F emg , and the frequency domain characteristic vector of the force signal be F force , and the combined characteristic vector after fusion F fused ; The calculation formula for the combined characteristic vector: .

[0054] The dynamic weight fusion is based on the basic weight setting in the rehabilitation stage, that is, according to the patient's rehabilitation stage p (early, middle, late) to set the initial weights and : .

[0055] The weight adjustment based on signal quality is to calculate MA and F rel signal quality indicators and (such as the signal-to-noise ratio SNR): ; In the formula, SNR MA represents the signal-to-noise ratio of the myoelectric signal, in dB; represents the signal-to-noise ratio of the force signal, in dB.

[0056] Adjust the weight: ; The fusion calculation, the final fusion feature: ; Perform smoothing filtering and real-time update low-pass filtering on the final fusion feature, that is, apply a first-order low-pass filter to the fusion feature to eliminate high-frequency jitter: ; In the formula, represents the fused feature value after smoothing at time; represents the fused feature value after smoothing at the previous time; is the low - pass filtering coefficient. Update the frequency in real - time, that is, repeat the above steps at the frequency to ensure the real - time nature of the control response.

[0057] This is a formula for smoothing the fused features. In the formula, represents the fused feature value after smoothing at the moment. is a weight coefficient, and its value range is usually between 0 and 1. It determines the contribution degree of the non - smoothed fused feature value at the current moment and the smoothed fused feature value at the previous moment to the current smoothing result. Through this formula, the weighted sum of the smoothed fused feature value at the previous moment and the non - smoothed fused feature value at the current moment is calculated according to the weight coefficient to obtain the smoothed fused feature value at the current moment, so as to realize the smoothing of the fused features and make the data more stable and regular.

[0058] The bio - mechanical coupling characteristic, that is, the fused feature F fused simultaneously reflects the muscle active intention ( MA ) and the actual joint force ( F rel ), realizing the biomechanical matching of human - machine collaborative control; the adaptive weight mechanism, that is, adjusting the weight through the rehabilitation stage and signal quality, while ensuring safety ( F rel constraint), maximizing the patient's active participation degree ( MA priority); feature post - processing, that is, reducing the dimension of the fused joint feature vector, using the principal component analysis (PCA) method to map the high - dimensional feature vector to a low - dimensional space, removing redundant information, retaining the main feature components, and reducing the computational complexity of subsequent processing; normalizing the reduced - dimension features so that their mean is 0 and variance is 1, which is convenient for input into subsequent modules such as the dynamic programming layer for analysis and decision - making.

[0059] The detailed model and steps for generating a time - varying impedance trajectory based on model predictive control (MPC), and the specific implementation of fusing the muscle activation degree ( MA ) and the joint torque ratio ( F rel ) include the time - varying impedance trajectory generation model based on MPC, the construction of the MPC optimization problem, the MPC predictive control, and MA and F relFusion; The system dynamics model in the time-varying impedance trajectory generation model based on MPC is to establish the dynamics model of the robot-human interaction system using Lagrange's equation; The impedance control model in the time-varying impedance trajectory generation model based on MPC is to design a time-varying impedance model: ; In the formula, represents the reference torque; represents the time-varying stiffness; represents the damping matrix; and are generated by MPC optimization.

[0060] The MPC prediction model in the time-varying impedance trajectory generation model based on MPC is to discretize the system: ; In the formula, the state vector , the control input , the disturbance The system behavior within the prediction horizon : , where .

[0061] The construction of the MPC optimization problem includes the design of the objective function and the constraint conditions; The physical constraints in the constraint conditions: ; The stability constraint: ; The safety constraint: , where K min is the minimum value of the time-varying stiffness, K max is the maximum value of the time-varying stiffness, D min is the minimum value of the damping matrix; D max is the maximum value of the damping matrix, is the safety threshold of the interaction torque.

[0062] The specific steps of MPC predictive control: First, perform state estimation, that is, obtain the current joint state and the external torque through sensors; Second, perform model prediction, that is, within the prediction horizon, use the system dynamics model to predict the future state ; Then, perform optimization and solution, that is, solve the optimal control sequence of the objective function J , and the optimal control sequence is specifically: ; Then, perform control execution, that is, only execute the first control input , update the time-varying impedance parameters and ; Finally, perform rolling optimization, that is, enter the next moment k +1, and repeat the above steps.

[0063] Time-varying impedance parameter adjustment, that is, according to the fused feature FusionFeature Dynamically adjust the impedance parameter: ; ; In the formula, K 2( t ) represents the adjusted time-varying stiffness; D 2( t ) represents the adjusted damping matrix; K 0 represents the base stiffness; K max Represents the maximum stiffness; Represents the gain coefficient of stiffness; β represents the gain coefficient of damping; M ( q ) represents the inertia matrix.

[0064] The MPC model of the present invention considers the time-varying impedance modeling of the human-machine interaction dynamic characteristics, integrates the multi-objective optimization of safety constraints and stability constraints, and at the same time adapts the prediction model parameters based on biofeedback; the fused feature can obtain the non-linear fusion function of muscle activation and joint torque based on the dynamic weight allocation mechanism in the rehabilitation stage and combined with the adaptive fusion strategy of signal quality (SNR).

[0065] The safety execution layer is a double-closed-loop control. The double-closed-loop control includes inner-loop force control and outer-loop position control, and integrates multi-level safety constraints. The multi-level safety constraints include torque gradient limit and abnormal mode detection; the outer-loop position control is the outer-loop position control. The outer loop aims to achieve accurate position tracking of the robot end effector or joint, and adopts a PID control algorithm with feed-forward compensation. Its control law expression is: ; In the formula, u ff Represents the feed-forward compensation term, which can be pre-calculated according to the robot dynamics model and is used to offset the non-linear and inertial effects of the system and improve the response speed; t Represents the duration of a control cycle, that is, the closed-loop time from when the controller issues a command to when the sensor feeds back.

[0066] The outer loop continuously compares the deviation between the desired position and the actual position, and after PID adjustment and feed-forward compensation, outputs a desired torque command , and this command is used as the target input for the inner-loop force control; the inner-loop force control is based on the desired torque command output by the outer loop , aiming at precise control force output, the PID control algorithm is also adopted, and the control law is as follows: ; Where: u force represents the control quantity of the inner loop force control output; K p_force represents the proportion of the force loop; K i_force represents the integral of the force loop; K d_force represents the differential gain coefficient of the force loop; represents the desired torque of the outer loop output; represents the actually measured joint torque; represents the torque change rate.

[0067] The inner loop monitors the difference between the actual torque and the desired torque in real time. After PID adjustment, it outputs the final motor control signal to drive the robot actuator to move, realizing precise control of the force, ensuring that the force applied by the robot during the interaction with the patient meets the treatment requirements and is within the safe range; torque gradient limit; first, real-time monitoring is carried out, that is, the joint torque data is collected in real time through the torque sensor at a frequency not lower than 1 kHz , and then the torque change amount at adjacent sampling moments is calculated. The formula for the change amount is: , where is the sampling interval; secondly, the gradient calculation is carried out, that is, the sliding window method is used to calculate the torque gradient, and the window size is set to n sampling points (for example n =10), and the average value of the torque change amount within the window is calculated as the torque gradient at the current moment : ; In the formula, represents the variable (this is the summation formula, i =1,2,... n -1); t represents the moment; represents the sampling interval; constraint judgment and processing, that is, setting the safety torque gradient threshold (for example ), when the calculated , the torque gradient limit mechanism is triggered. The specific processing method is: adjust the current desired torque command , and adopt a linear attenuation method to reduce the torque change speed. For example, the formula for linear attenuation is: , in the formula, k is the attenuation coefficient, which is adjusted according to the actual situation to ensure smooth torque change and avoid impact on the patient.

[0068] Abnormal mode detection includes electromyogram (EMG) signal abnormal detection, torque abnormal detection, and abnormal response handling. Among them, for EMG signal abnormal detection, feature extraction is first performed, that is, preprocessing the collected surface EMG signals, including band-pass filtering (20 - 500 Hz), denoising, etc., and then extracting features of the EMG signals, such as mean absolute value (MAV), zero crossing rate (ZCR), root mean square value (RMS), etc. Then, threshold judgment is carried out, that is, setting an abnormal detection threshold. For example, when the MAV of the EMG signal exceeds 5 times the normal mean value and the ZCR changes by more than 20% within a short period (such as 100 ms), it is determined that the EMG signal is abnormal, indicating that the patient may have muscle spasm or other abnormal muscle activity states, triggering emergency safety measures.

[0069] For torque abnormal detection, overrun judgment is first carried out, that is, real-time monitoring of joint torque , when the actual torque exceeds the pre-set maximum safety torque (dynamically adjusted according to the patient's individual situation and treatment stage), it is marked as torque overrun abnormal; then continuous monitoring is carried out, that is, introducing a counter mechanism. When the torque overrun situation occurs continuously for m times (for example, m = 3) and each duration exceeds a certain threshold (such as 20 ms), it is determined as continuous torque abnormal, and the corresponding safety protection program is started. For abnormal response handling, that is, once the above abnormal mode is detected, the following operations are immediately executed: emergency braking, that is, cutting off the motor power within 20 ms to stop the robot from moving and avoid further harm to the patient; alarm prompt, that is, sending an alarm to medical staff through an audible and visual alarm device and displaying the specific abnormal type and location information on the control interface for timely handling; data recording, that is, saving multi-modal data (EMG signals, torque data, joint positions, etc.) for a period of time before and after the abnormality occurs, providing a basis for subsequent analysis and diagnosis. The core control and constraint mechanism of the safety execution layer can effectively ensure the safe operation of the physiotherapy robot.

[0070] The multi-modal fusion strategy, that is, signal-level fusion includes spatio-temporal alignment of EMG signals (band-pass filtered at 20 - 500 Hz) and force signals (inverse dynamics calculation), with a deviation < 2 ms.

[0071] The dynamic weight formula for feature-level fusion: .

[0072] The principle and formula core of feature-level fusion is that feature-level fusion aims to organically combine two key features, namely the muscle activation degree (MA) and the joint torque ratio ( , denoted as F rel ), to obtain a combined feature that can accurately reflect the patient's movement state and rehabilitation needs Ffused 。

[0073] The dynamic weight coefficient comprehensively considers the patient's movement intention and physical feedback by assigning appropriate weights to different features. MA It reflects the degree of active muscle force exerted by the patient. F rel It reflects the proportional relationship between the actual joint torque and the maximum safe torque. The combination of the two can take into account both the patient's subjective initiative and objective physical tolerance; the dynamic weight adjustment mechanism includes weight setting based on the rehabilitation stage and weight adaptive adjustment based on signal quality. At different stages of rehabilitation treatment, the patient's physical condition and treatment goals vary, so the weight needs to be dynamically adjusted. To avoid deviation in the fusion result caused by signal interference, a signal quality assessment mechanism is introduced to dynamically adjust the weight; In the initial stage of rehabilitation: The patient's muscle strength is weak and the active movement ability is limited. At this time, more attention is paid to the patient's muscle activation state, and it is expected to gradually restore the function by stimulating the muscles. Therefore, w 1 is set to a relatively high value, such as 0.7 - 0.8, to increase the MA proportion in the combined features and encourage the patient to perform active muscle contraction training; In the middle stage of rehabilitation: The patient's muscle strength has recovered to some extent, and more complex movement training begins. It is necessary to consider the magnitude of the torque borne by the joint while paying attention to muscle activation to ensure the safety and effectiveness of the training. At this time, w 1 is adjusted to 0.6 - 0.7 to achieve MA a balanced consideration with F rel ; In the late stage of rehabilitation: The patient is approaching the rehabilitation goal, and emphasis is placed on movement coordination and strength enhancement training. At this time, the magnitude of the torque borne by the joint has a greater impact on the training effect and safety. Therefore, w 1 is appropriately reduced to 0.6, and the proportion of w 2 is relatively increased, so that F rel plays a greater role in the combined features and ensures that the training intensity conforms to the patient's physical condition.

[0074] The weight adaptive adjustment based on signal quality is to calculate the signal-to-noise ratio ( SNR ), and calculate the signal-to-noise ratios of the original signals (electromyogram signal and force signal) corresponding to MA and F rel respectively. Taking the electromyogram signal as an example, SNR the calculation formula is: ; In the formula, is the sampling value of the electromyogram signal, is the mean value, and is the noise sampling value.

[0075] The weight adjustment strategy is that when the SNR is relatively high, it indicates that MA the feature credibility is high, and appropriately increase w 1; conversely, if the SNR of the force signal is higher, then correspondingly increase w 2. The specific adjustment formula is: ; ; In the formula, , to ensure the stability and rationality of weight adjustment; represents the adjusted w 2; represents the adjusted w 1; SNR EMG represents the signal-to-noise ratio of the EMG signal; SNR Force represents the signal-to-noise ratio of the joint feature.

[0076] Feature calculation, calculate the muscle activation degree ( MA ): Preprocess the collected EMG signal, including band-pass filtering from 20 - 500 Hz, denoising, etc., and then obtain MA through methods such as moving average method or root mean square value calculation. The formula is , where N is the number of sampling points.

[0077] Calculate the joint torque ratio, obtain the joint torque through the force sensor, and combine with the maximum safe torque preset according to the patient's body parameters and treatment stage to calculate F rel , and limit it to the interval [0, 1].

[0078] Decision-level fusion, switch the control mode (passive guidance / active assistance) according to the MA value; the control mode switching control rules based on the MA value include: control mode definition, switching threshold setting, transition mechanism design, and priority rules and exception handling.

[0079] The control mode definition includes passive guidance mode and active assistance mode; the passive guidance mode is that the robot dominates the movement process, and the patient is in a passive receiving state. The robot outputs a constant torque to drive the patient's limb movement according to the preset rehabilitation trajectory and parameters. This mode is applicable to the stage where the patient's muscle strength is extremely weak and unable to generate effective muscle activation independently, aiming to help the patient maintain joint mobility and prevent muscle atrophy and joint stiffness; the active assistance mode is that the robot provides an assistive torque according to the patient's muscle activation intention. When the patient actively exerts force, the robot detects MAValue, analyze the patient's movement intention, dynamically adjust the output torque, provide assistance when the patient's force is insufficient, and appropriately limit it when the patient's force is excessive, to achieve human-machine collaborative movement and promote the recovery of the patient's muscle strength and motor function. The switching threshold setting includes a primary switching threshold ( MA low ), and a secondary switching threshold ( MA high ); the primary switching threshold ( MA low ) is set as (normalized muscle activation value). When it is detected that the MA value continuously remains below for more than seconds, the system determines that the patient's muscles are in a low activation state and switches from the current mode to the passive guidance mode. This threshold is used to identify the situation where the patient can hardly exert active force and provide necessary passive training support in a timely manner; the secondary switching threshold ( MA high ) is set as (normalized muscle activation value). When the MA value continuously remains higher than for more than seconds, the system determines that the patient has a certain ability to exert active force and switches from the passive guidance mode to the active assistance mode, giving full play to the patient's subjective initiative and improving the rehabilitation training effect. The transition mechanism design includes the transition from passive guidance to active assistance and the transition from active assistance to passive guidance; the transition from passive guidance to active assistance is that during the switching process, to avoid discomfort or harm to the patient caused by sudden torque changes, a linear transition strategy is adopted. Specifically, within seconds after the switching moment, the robot's output torque gradually transitions from the constant value in the passive guidance mode to the initial assistance torque in the active assistance mode according to the following formula: ; In the formula, is pre-calculated and determined based on the MA value at the switching moment and the patient's body parameters to ensure a smooth and stable transition process. The transition from active assistance to passive guidance is that when the MA value drops and triggers the switching, within seconds, the robot's output torque gradually decreases from the current active assistance torque to the initial torque of the passive guidance mode, also adopting a linear transition method: .

[0080] Priority rules, that is, during the mode switching process, safety constraints always have the highest priority. If safety issues such as torque overrun or abnormal EMG signals are detected during the switching, regardless of the state of the MA value, the system immediately pauses the mode switching, enters the safety protection mode, stops the robot movement and issues an alarm.

[0081] Exception handling, that is, when MA the value fluctuates near the threshold, causing frequent switching, an anti-jitter mechanism is enabled. Within seconds after the first trigger of the switching, small fluctuations in the MA value are ignored, MA to avoid the frequent mode switching from affecting the patient experience and treatment effect. If the MA value continuously exceeds the threshold range during this period, the corresponding mode switching operation is executed. The above control rules achieve intelligent switching of the control mode based on the MA value through clear thresholds, smooth transitions, and priority guarantees.

[0082] As shown in Table 1, the technical effects of the present invention are: Table 1 Performance improvement of the present invention compared with the traditional scheme ; Enhanced safety includes an overload torque detection speed of 20 ms (200 ms for the traditional scheme) and a secondary injury incidence rate < 0.5% (3% for the traditional scheme).

[0083] Specific implementation method: Sensor configuration for biological signal acquisition: The EMG sensor is a differential electrode with a spacing of 20 mm and a sampling rate of 2 kHz; the six-axis force sensor has a range of ±500 N and a non-linearity < 0.1% FS.

[0084] Dynamic adjustment formula for the impedance parameters of the dynamic planner: ; In the formula, represents the time-varying stiffness parameter and also represents the stiffness value of the impedance controller at time t (unit: ); represents the basic stiffness, with a default value of , representing the basic impedance when there is no active participation of the patient; represents the coefficient, with a default value of 0.6, controlling the influence weight of muscle activation on stiffness; β represents the torque ratio gain coefficient, with a default value of 0.4, controlling the influence weight of force feedback on stiffness; represents the muscle activation degree, which is the normalized value calculated from the EMG signal (range 0 - 1); represents the real-time joint torque, measured by the force sensor (unit: ); Denote the maximum safe torque, a safety threshold set according to individual patient differences and rehabilitation stages (unit: ).

[0085] The safety trajectory optimization formula of the dynamic planner is expressed as ; The constraint condition is expressed as ; In the formula, denotes the desired joint trajectory, the ideal joint angle trajectory preset for rehabilitation treatment (unit expressed as rad); denotes the actual joint trajectory, the joint angle trajectory currently executed by the robot (unit expressed as rad); denotes the position error weight matrix, a diagonal matrix, which adjusts the importance of the position tracking accuracy of each joint; R denotes the torque consumption weight matrix, a diagonal matrix, which controls the trade-off between energy consumption and motion smoothness; T represents the optimization time window, the time range of trajectory optimization (unit expressed as s); denotes the joint torque, the real-time control torque applied by the robot (unit expressed as ); denotes the dynamic safety torque threshold, which is a function of MA, indicating that the higher the muscle activation degree, the greater the maximum allowable torque.

[0086] The impedance parameter is adjusted so that when the patient actively exerts force ( MA increases), the stiffness increases, and the robot provides greater resistance to promote muscle strength training; when the joint torque approaches the safety threshold ( increases), the stiffness decreases to avoid the risk of overload; α > β reflects the control strategy that the muscle active intention takes precedence over force feedback.

[0087] Safety trajectory optimization: The objective function weighs the position tracking accuracy (the first term) and the torque consumption (the second term); the constraint condition ensures that the real-time torque does not exceed the safety threshold related to MA to achieve active safety protection; when MA is lower, automatically decreases to prevent the patient from overloading due to muscle weakness.

[0088] The dynamic safety threshold can make designed as MA a function of, breaking through the traditional fixed threshold limit and adapting to the patient's real-time muscle state; multi-objective optimization can consider both position tracking and torque safety simultaneously to achieve a balance between treatment effect and safety; biofeedback fusion can adjust the stiffness and safety threshold through MA to form an "EMG-force-control" closed loop and improve the human-machine cooperation adaptability.

[0089] Example 1: Parameter configuration for upper limb rehabilitation of stroke patients: In the control system of the physiotherapy robot, the multi-level constraint logic ensures patient safety through hierarchical detection and progressive response mechanisms. Real-time threshold detection: The absolute value of the torque is constrained to compare the current torque with a preset safety threshold. The torque gradient is constrained to calculate the rate of change of the torque at adjacent sampling points. Pattern anomaly detection: Abnormal electromyogram signals are detected for muscle spasm based on the MA value and spectral characteristics.

[0090] Trajectory deviation detection is to calculate the deviation between the actual trajectory and the reference trajectory. The state machine decision is to perform state transition based on the multi-feature fusion result. Constraint conflict resolution is to adopt a priority mechanism when multiple layers of constraints are triggered simultaneously. Parameter configuration description: The parameter configuration of the multi-level constraint system adopts a hierarchical design and adaptive adjustment strategy. Basic parameter layer: 1.1 The torque threshold is set according to the individual differences of patients: 1.2 The gradient threshold is related to the inertial characteristics of the robot.

[0091] This hierarchical constraint logic and parameter configuration mechanism not only ensure the system safety but also retain the treatment flexibility, and can dynamically adjust the protection strategy according to the real-time state and rehabilitation progress of the patient; Effect: The Fugl-Meyer score is increased by 45% and the treatment course is shortened by 3 weeks.

[0092] Example 2 is for the knee joint rehabilitation of athletes: Dynamic rule adjustment is that if ZCR < 50 then w 1 = 0.7; otherwise w 1 = 0.4; The effect is that the recovery speed of the joint range of motion is increased by 30%.

[0093] As Figure 6 shown, the description of the safety constraint working area (joint angle - torque dynamic boundary). The physical meaning is that this figure shows the maximum safe torque boundary that the physiotherapy robot is allowed to apply at different joint angles, which is an important constraint condition to ensure patient safety and equipment reliability. Its core principle is based on human biomechanical characteristics; The cosine function relationship is that the maximum safe torque is cosine-distributed with the joint angle, which conforms to the change law of the human muscle moment arm at different joint positions; The safety boundary is that the area above the curve is the dangerous area, and the robot control system will limit the output torque not to exceed this boundary; The working area is that the area below the curve is the safe working area, and the robot can dynamically adjust the output within this range according to the rehabilitation needs; Figure 6 Key parameter explanations. The range of the joint angle (X-axis) is , representing the angle interval in which the joint can move; The physical meaning of the joint angle is to reflect the current position of the patient's joint, and different angles correspond to different muscle group force states; The range of the joint torque (Y-axis) is , representing the magnitude of the torque that the robot can apply; The physical meaning of the joint torque is that the greater the torque, the greater the force exerted on the patient's joint, and it needs to be carefully controlled according to the rehabilitation stage; The mathematical expression of the safety boundary line (dashed line): , its physical meaning is to ensure that the torque applied at any joint angle does not exceed the tolerance limit of the musculoskeletal system based on ergonomic experimental data.

[0094] Hazard areas and safety strategies include: 1. Diagonally filled areas: indicating hazard areas beyond the safety torque boundary; that is, the control system strategy: when it is detected that the robot is about to enter a hazard area, the following protection mechanisms are automatically triggered: urgently reducing the output torque, recording safety events, and raising an alarm.

[0095] The dynamic adaptation mechanism includes that the safety boundary can be dynamically adjusted according to individual patient differences and rehabilitation stages; at the initial stage of rehabilitation, the overall boundary moves downward (such as ), providing more conservative protection; at the later stage of rehabilitation, the boundary moves upward appropriately (such as ), allowing a greater training intensity.

[0096] The clinical application value is that muscle strains can be prevented by restricting the maximum torque, avoiding soft tissue injuries caused by excessive output force of the robot; it can also provide personalized rehabilitation, adjust the safety boundary parameters according to factors such as patient age, gender, and condition, and at the same time enhance patient trust. The visual safety working area provides an intuitive basis for safety guarantee for medical staff and patients.

[0097] The present invention ensures safety in case of emergency through real-time monitoring, that is, the control system needs to continuously compare the current torque with the safety boundary at a frequency above 100 Hz; and the execution delay of the fast-response protection mechanism needs to be controlled within 5 ms; it also includes parameter calibration, that is, the safety boundary parameters need to be recalculated according to the latest evaluation data of the patient before each treatment.

[0098] Such as Figure 7As shown, the clinical effect comparison curve (ROM improvement rate), that is, the comparison chart of the improvement rate of range of motion (ROM), with a sample size of 50 patients in each group, a treatment frequency of 3 times a week, 30 minutes each time, a measurement tool of an electronic joint goniometer, and an error bar of ±1 standard deviation. Compared with the traditional rehabilitation treatment, the average improvement rate difference of the present invention's solution is 15.3%, the treatment efficiency is increased by 42%, and the significance level p < 0.01; after 6 weeks of treatment, the average ROM of the present invention is 72° ± 3°, the average ROM of the traditional solution is 60° ± 3°, and the clinical compliance rate is 89%. This safety constraint mechanism has been verified in clinical studies, which can reduce the incidence of safety events in rehabilitation treatment by 73% (p < 0.001), significantly improving the safety and reliability of the physiotherapy robot. Through result analysis, the advantages of the present invention's solution are that the ROM is significantly improved after 4 weeks of treatment (p < 0.01), and the multi-modal fusion control improves the rehabilitation efficiency by 42%, the patient's active participation rate is increased by 37%, the treatment cycle is shortened by about 2 weeks. The clinical significance is to more quickly restore joint function, reduce long-term complications, be applicable to various rehabilitation scenarios such as stroke and spinal cord injury, reduce the workload of medical staff by 30%, and increase the patient satisfaction rate to 94%.

[0099] The optimization of the dynamic programming layer based on reinforcement learning can be oriented by the physiotherapy effect, divide the rehabilitation process into multiple stages to set different reward functions, and through the interaction and learning between the agent and the environment, perform secondary optimization on the trajectory generated by the MPC to improve the treatment accuracy. Safe trajectory generation is to generate the final safe motion trajectory under the condition of meeting safety constraint conditions such as torque limit and joint angle range, and transmit it to the safe execution layer through the trajectory output interface. Feedback signal processing is to process the signals collected by the force / position feedback sensors on the robot body, and feedback the actual position and torque information to the double closed-loop control module to form a control closed-loop. The multi-modal fusion strategy layer includes dynamic weight adjustment, that is, according to the rehabilitation stage and signal quality, dynamically adjust the fusion weight of the muscle activation degree and the joint torque ratio, so that the fusion feature can more accurately reflect the patient's motion state; control mode decision-making, that is, based on the MA value in the fusion feature, according to the set switching rule, switch between the passive guidance mode and the active assistance mode, generate the corresponding mode instruction and output it to the safe execution layer to realize the intelligent regulation of the human-machine collaborative rehabilitation training. Force signal processing is to further process the calculated joint torque signal and then perform spatio-temporal alignment operation to ensure the consistency of the force signal and other modal signals, which takes 1 ms; inertial signal processing is to perform spatio-temporal alignment on the inertial signal after attitude solution to align it with the EMG signal and the force signal on the time axis, which takes 1 ms. In the signal fusion stage (6 - 10 ms), namely multi-modal feature fusion, the EMG features, force features, and inertial features that have undergone spatio-temporal alignment are fused through a wavelet-frequency domain fusion algorithm and a dynamic weight formula to generate a joint feature vector, which takes 3 ms. The joint feature output is to output the fused joint feature vector to the dynamic programming layer, providing accurate multi-modal information for subsequent trajectory planning and control decisions. The output time is 1 ms. Through strict time control and precise synchronization processing, it is ensured that each modal signal works collaboratively, providing a reliable data basis for the adaptive control of the physiotherapy robot. It should be noted that as Figure 8 , Figure 9 , Figure 10 and Figure 11 shown, Figure 8 the a in Figure 9 is connected to the a in Figure 9 , the b in Figure 11 is connected to the b in Figure 10 , and the c in Figure 11 is connected to the c in

[0100] In one embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as shown in Figure 3 . The computer device includes a processor, a memory, and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store static information and dynamic information data. The network interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements the steps in the above-mentioned embodiment of the adaptive control method for a physiotherapy robot based on multi-modal biofeedback.

[0101] Those skilled in the art can understand that Figure 3 the structure shown in

[0102] is only a block diagram of a part of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements. In addition, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps in the above-mentioned embodiment of the adaptive control method for a physiotherapy robot based on multi-modal biofeedback.

[0103] In addition, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned embodiments of the adaptive control method of the physiotherapy robot based on multimodal biofeedback are implemented.

[0104] Those of ordinary skill in the art can understand that all or part of the process of implementing the above-mentioned embodiments of the adaptive control method of the physiotherapy robot based on multimodal biofeedback can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments of the adaptive control method of the physiotherapy robot based on multimodal biofeedback. Among them, any reference to a memory, storage, database or other medium used in the various embodiments provided by the present invention can include at least non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0105] The present invention is not limited to the structure already described and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present invention is only limited by the appended claims.

Claims

1. An adaptive control method for a physiotherapy robot based on multimodal biofeedback, characterized in that, Including: Preprocess the pre-acquired biomechanical perception signals, and use wavelet transform and frequency-domain transform to extract features and perform dynamic weight fusion on the preprocessed biomechanical perception signals to obtain joint features; Establish a system dynamics model based on the Lagrangian equation, and use the real-time acquired joint states combined with the rolling optimization algorithm to generate a time-varying impedance trajectory model. Dynamically adjust the impedance parameters of the time-varying impedance trajectory model according to the joint features, and generate a safe motion trajectory through the adjusted time-varying impedance trajectory model; Perform double-loop control on the physical therapy robot according to the safe motion trajectory and the real-time acquired biomechanical perception signals to obtain the control signals of the physical therapy robot, and realize the joint motor drive of the physical therapy robot based on the control signals.

2. The adaptive control method of the physiotherapy robot based on multimodal biofeedback according to claim 1, wherein, The preprocessing of the pre-acquired biomechanical perception signals, and using wavelet transform and frequency-domain transform to extract features and perform dynamic weight fusion on the preprocessed biomechanical perception signals to obtain joint features includes: Obtain the biomechanical perception signals, and perform filtering and denoising processing on the biomechanical perception signals to obtain the preprocessed biomechanical perception signals; Perform wavelet decomposition on the preprocessed biomechanical perception signals using wavelet transform, and extract the muscle activation degree features and joint torque features according to the wavelet decomposition results combined with frequency-domain transform; Perform time calibration on the muscle activation degree features and joint torque features through dynamic time warping to obtain the aligned muscle activation degree features and joint torque features; Based on the pre-set weights of the rehabilitation stage and the signal quality, perform dynamic weight fusion on the aligned muscle activation degree features and joint torque features to obtain joint features.

3. The adaptive control method of the physiotherapy robot based on multimodal biofeedback according to claim 2, wherein, The performing dynamic weight fusion on the aligned muscle activation degree features and joint torque features based on the pre-set weights of the rehabilitation stage and the signal quality to obtain joint features includes: According to the pre-set weights of the rehabilitation stage, respectively obtain the initial weights of the aligned muscle activation degree features and joint torque features; By calculating the signal-to-noise ratios of the aligned muscle activation degree features and joint torque features, respectively obtain the signal quality adjustment weights of the aligned muscle activation degree features and joint torque features; Based on the initial weights and the signal quality adjustment weights, obtain the adjusted weights, and perform feature fusion on the aligned muscle activation degree features and joint torque features through the adjusted weights to obtain the fusion features, and perform filtering processing on the fusion features using smoothing filtering to obtain joint features.

4. The adaptive control method of the physiotherapy robot based on multi-modal biofeedback according to claim 1, wherein, The establishing a system dynamics model based on the Lagrangian equation, and using the real-time acquired joint states combined with the rolling optimization algorithm to generate a time-varying impedance trajectory model, and dynamically adjusting the impedance parameters of the time-varying impedance trajectory model according to the joint features, and generating a safe motion trajectory through the adjusted time-varying impedance trajectory model includes: Establish a system dynamics model based on the Lagrangian equation, and use the pre-acquired desired trajectory to generate a time-varying impedance control model; Construct an MPC prediction model based on the joint states obtained in real time, and combine the system dynamics model and the time-varying impedance control model to obtain a time-varying impedance trajectory generation model for MPC. Generate an MPC optimization problem through a preset MPC objective optimization function combined with constraint conditions; Based on the time-varying impedance trajectory generation model and MPC optimization problem of MPC, and combined with the joint states obtained in real time, predict the future joint states, update the time-varying impedance parameters according to the future joint states, and generate a time-varying impedance trajectory model through rolling optimization; Dynamically adjust the impedance parameters of the time-varying impedance trajectory model according to the joint features, and generate a safe motion trajectory through the adjusted time-varying impedance trajectory model.

5. The adaptive control method of the physiotherapy robot based on multimodal biofeedback according to claim 4, characterized in that, The expression of the system dynamics model: ； wherein, q represents the joint angle; represents the joint angular velocity; represents the joint angular acceleration vector; M ( q ) represents the inertia matrix; represents the Coriolis and centrifugal force matrix; G ( q ) represents the gravity vector; represents the controller output torque; represents the external interaction torque from the patient; The expression of the time-varying impedance control model: ； In the formula, represents the reference torque; q d represents the desired joint angle; represents the desired joint angular velocity; K ( t ) represents the time-varying stiffness generated by MPC optimization; D ( t ) represents the damping matrix generated by MPC optimization; The expression of the MPC objective optimization function: ； In the formula, represents the MPC objective optimization function; Nc represents the control horizon; u ( k ) represents k the control input at time x d represents the desired state trajectory; u ref represents the reference control input; represents the weight matrix of the desired state trajectory; R represents the weight matrix of the reference control input; represents the joint state; represents the desired joint state; represents the control input at time represents the reference control input at time Np represents the prediction horizon.

6. The adaptive control method of the physiotherapy robot based on multi-modal biofeedback according to claim 4, characterized in that The time-varying impedance trajectory generation model and MPC optimization problem based on MPC, combined with the joint states obtained in real time to predict the future joint states, update the time-varying impedance parameters according to the future joint states, and generate a time-varying impedance trajectory model through rolling optimization includes: Based on the time-varying impedance trajectory generation model of MPC, and combined with the joint states and external torques obtained in real time, generate future joint states within the prediction horizon; Solve the objective optimization function of the MPC optimization problem according to the future joint states to obtain the optimal control sequence, and update the time-varying impedance parameters according to the optimal control sequence; Based on the updated results of the time-varying impedance parameters, enter the update of the time-varying impedance parameters at the next moment through rolling optimization to generate a time-varying impedance trajectory model.

7. The adaptive control method of the physiotherapy robot based on multimodal biofeedback according to claim 1, wherein The dual closed-loop control of the physiotherapy robot according to the safe motion trajectory and the biomechanical perception signals obtained in real time to obtain the control signal of the physiotherapy robot, and realize the joint motor drive of the physiotherapy robot based on the control signal includes: Calculate the muscle activation degree through the biomechanical perception signals obtained in real time, and generate a control mode instruction using a preset switching rule; Based on the control mode instruction, combined with the comparison result of the safe motion trajectory and the actual motion trajectory, use the outer-loop position control of the dual closed-loop control to perform PID adjustment and feedforward compensation on the output of the outer-loop position control to obtain the desired torque instruction; Through the inner-loop force control of the dual closed-loop control, perform PID control on the desired torque instruction to obtain the motor control signal output by the inner-loop force control; Use multi-level safety constraints to perform anomaly detection on the biomechanical perception signals obtained in real time, generate safety measures according to the anomaly detection results combined with a preset priority, and combine with the motor control signal to obtain the control signal of the physiotherapy robot, and realize the joint motor drive of the physiotherapy robot based on the control signal.

8. The adaptive control method of the physiotherapy robot based on multimodal biofeedback according to claim 7, wherein, The expression of the outer-loop position control of the dual closed-loop control: ； Wherein, u pos represents the control quantity of the outer loop position control output; K p_pos represents the ratio of the position loop; K i_pos represents the integral of the position loop; K d_pos represents the differential gain coefficient of the position loop; represents the desired joint angular velocity; q d represents the desired joint angle; q represents the joint angle; represents the joint angular velocity; u ff represents the feedforward compensation term; t represents the duration of a control cycle.

9. The adaptive control method of the physiotherapy robot based on multimodal biofeedback according to claim 7, wherein The use of multi-level safety constraints to perform anomaly detection on the biomechanical perception signals obtained in real time, generate safety measures according to the anomaly detection results combined with a preset priority, and combine with the motor control signal to obtain the control signal of the physiotherapy robot, and realize the joint motor drive of the physiotherapy robot based on the control signal includes: The torque gradient calculation is performed on the biomechanical perception signals obtained in real time by using the torque gradient limit of multi-level security constraints and combining with the sliding window method to obtain the torque gradient at the target moment; By comparing the torque gradient at the target moment with the preset safe torque gradient threshold, the torque gradient limit mechanism is triggered based on the comparison result; the torque gradient limit mechanism includes: adjusting the desired torque command and reducing the torque change speed in a linearly decaying manner; The abnormal mode detection with multi-level security constraints is used to perform abnormal threshold detection on the biomechanical perception signals obtained in real time, generate safety measures according to the abnormal threshold detection results combined with the preset priority, and combine with the motor control signal to obtain the control signal of the physiotherapy robot.

10. An adaptive control system for a physiotherapy robot based on multimodal biofeedback, which is used to implement the adaptive control method for a physiotherapy robot based on multimodal biofeedback according to any one of claims 1-9, characterized in that, The adaptive control system of the physiotherapy robot based on multi-modal biofeedback includes: a joint feature extraction module, a motion trajectory generation module, and a control signal acquisition module; The joint feature extraction module is used to preprocess the previously obtained biomechanical perception signals, and perform feature extraction and dynamic weight fusion on the preprocessed biomechanical perception signals by using wavelet transform and frequency domain transform to obtain joint features; The motion trajectory generation module is used to establish a system dynamics model based on the Lagrangian equation, generate a time-varying impedance trajectory model by using the joint states obtained in real time combined with the rolling optimization algorithm, dynamically adjust the impedance parameters of the time-varying impedance trajectory model according to the joint features, and generate a safe motion trajectory through the adjusted time-varying impedance trajectory model; The control signal acquisition module is used to perform double closed-loop control on the physiotherapy robot according to the safe motion trajectory and the biomechanical perception signals obtained in real time to obtain the control signal of the physiotherapy robot, and realize the joint motor drive of the physiotherapy robot based on the control signal.

Images