A control method and system of an exoskeleton lower limb rehabilitation robot

Abstract

This application relates to the field of robot control technology, and discloses a control method and system for an exoskeleton lower limb rehabilitation robot. The method includes: generating a gait trajectory based on patient information and collecting joint interaction forces and posture angles; statistically processing the joint interaction forces to obtain fatigue indices, and quantifying the posture angles to obtain imbalance indices; using a two-layer radial basis function neural network to output fatigue compensation parameters and posture correction parameters respectively, adaptively determining the fusion weights to generate final impedance parameters; and calculating and executing the joint driving torque. This application solves the problem that existing fixed impedance control cannot simultaneously cope with both dynamic changes in patient fatigue accumulation and posture imbalance.

Description

Technical Field

[0001] This application relates to the field of robot control technology, and in particular to a control method and system for an exoskeleton lower limb rehabilitation robot. Background Technology

[0002] As an important assistive device in the field of neurorehabilitation engineering, exoskeleton lower limb rehabilitation robots assist patients with stroke, spinal cord injury, and other conditions in gait rehabilitation training through coordinated movement of the mechanical exoskeleton structure and the human lower limb. Traditional control methods for rehabilitation robots mainly include two categories: position tracking-based PID control and force feedback-based impedance control. PID control generates control commands by setting a target gait trajectory and calculating joint angle errors, while impedance control adjusts virtual stiffness and damping parameters to make the robot exhibit programmable mechanical impedance characteristics. The latter has become the mainstream control strategy for rehabilitation robots because it enables compliant human-machine interaction. Existing impedance control methods typically preset fixed stiffness and damping parameters based on the patient's muscle strength level before training begins, keeping the parameters constant throughout the training process or adjusting them only in segments according to the training phase. The joint interaction forces collected by force sensors and the posture angles collected by the inertial measurement unit are mainly used to monitor the training status and rarely directly participate in the calculation of control laws.

[0003] However, patients' muscle strength exhibits significant dynamic changes during rehabilitation training. As training time increases, muscle fatigue accumulates, leading to a gradual decline in active force exertion. Joint interaction forces show statistical characteristics of decreasing average force, increasing force variance, and rising force decay rate. Fixed stiffness parameters require patients to continuously overcome a constant elastic resistance burden, while the passive support provided by fixed damping parameters cannot adaptively increase with fatigue levels. This makes it difficult for patients to maintain effective training intensity in the later stages of fatigue, sometimes even forcing them to stop training, thus limiting training tolerance time. On the other hand, patients also face the risk of postural imbalance during training. Insufficient upper limb swing amplitude, excessive forward tilting of the trunk, and other abnormal postures may induce falls. Fixed impedance parameters lack responsiveness to postural abnormalities and cannot urgently increase damping to provide braking protection when imbalance occurs, posing a safety hazard. Summary of the Invention

[0004] This application provides a control method and system for an exoskeleton lower limb rehabilitation robot. By constructing a two-layer radial basis function neural network, the method quantifies muscle fatigue and postural imbalance and generates corresponding impedance compensation parameters. Then, based on the relative severity of the two states, the method adaptively determines the dynamic fusion weight to achieve intelligent switching of control priority. This solves the technical problem that existing fixed impedance control cannot simultaneously cope with the two dynamic changes of patient fatigue accumulation and postural imbalance, and improves the safety, comfort and training tolerance time of rehabilitation training.

[0005] In a first aspect, this application provides a control method for an exoskeleton lower limb rehabilitation robot, the control method of which includes: Step S1: Generate a personalized gait trajectory based on the patient's personal information input through the user interface interaction device, and simultaneously activate the force sensor and inertial measurement unit sensor to collect joint interaction force, upper limb swing angle and trunk tilt angle; Step S2: Statistically process the joint interaction forces within the sliding time window to obtain a quantified index of muscle fatigue, and quantify the degree of abnormality of the upper limb swing angle and trunk tilt angle to obtain a postural imbalance index. Step S3: Input the muscle fatigue quantification index into the radial basis function neural network of the fatigue response layer to output fatigue compensation stiffness parameters and fatigue compensation damping parameters. Input the posture imbalance index into the radial basis function neural network of the posture response layer to output posture correction stiffness parameters and posture correction damping parameters. Adaptively determine the dynamic fusion weight based on the relative magnitude of the muscle fatigue quantification index and the posture imbalance index. Use the dynamic fusion weight to perform weighted fusion of the fatigue compensation parameters and posture correction parameters to generate the final stiffness parameters and final damping parameters. Step S4: Calculate the driving torque of each joint based on the final stiffness parameter and final damping parameter, the target joint angle in the personalized gait trajectory and the actual joint angle, and send it to each joint drive system for execution.

[0006] Secondly, this application provides a control system for an exoskeleton lower limb rehabilitation robot, the control system of which includes: The extraction module is used to generate personalized gait trajectories based on the patient's personal information input through the user interface interaction device, and simultaneously activates force sensors and inertial measurement unit sensors to collect joint interaction forces, upper limb swing angles, and trunk tilt angles. The quantification module is used to statistically process the joint interaction forces within the sliding time window to obtain a quantified index of muscle fatigue, and to quantify the degree of abnormality of the upper limb swing angle and trunk tilt angle to obtain a postural imbalance index. The generation module is used to input the muscle fatigue quantification index into the fatigue response layer radial basis function neural network to output fatigue compensation stiffness parameters and fatigue compensation damping parameters, input the posture imbalance index into the posture response layer radial basis function neural network to output posture correction stiffness parameters and posture correction damping parameters, adaptively determine dynamic fusion weights based on the relative magnitudes of the muscle fatigue quantification index and the posture imbalance index, and perform weighted fusion of fatigue compensation parameters and posture correction parameters through the dynamic fusion weights to generate final stiffness parameters and final damping parameters. The drive module is used to calculate the drive torque of each joint based on the final stiffness parameters and final damping parameters, the target joint angle in the personalized gait trajectory and the actual joint angle, and then send it to the drive system of each joint for execution.

[0007] The technical solution provided in this application generates personalized gait trajectories based on the patient's personal information and simultaneously activates force sensors and inertial measurement units to collect joint interaction forces, upper limb swing angles, and trunk tilt angles. This establishes a multi-source data acquisition mechanism covering both the patient's physiological characteristics and real-time motion state parameters, laying a data foundation for the subsequent accurate quantification of muscle fatigue and postural imbalance. Statistical processing of joint interaction forces within a sliding time window yields quantifiable indicators of muscle fatigue, while abnormality quantification of upper limb swing angles and trunk tilt angles yields postural imbalance indicators. This quantification transforms the original multidimensional sensor data, which was difficult to directly use for control law calculations, into scalar indicators with normalized value ranges. This eliminates dimensional differences between different physical quantities and captures the dynamic trend of fatigue accumulation through sliding window statistics, enabling the control system to simultaneously assess the patient's muscle strength and postural state based on a unified numerical framework. The core advantage of this two-layer neural network architecture lies in its ability to achieve a nonlinear mapping from quantitative indicators to impedance parameters through the Gaussian response characteristics of the radial basis function. Compared with traditional piecewise linear adjustment rules or lookup table methods, the radial basis function neural network provides a smooth and continuous parameter transition between different fatigue and imbalance levels, avoiding the sudden changes in control law and human-computer interaction that may be caused by step parameter changes. Furthermore, the setting of the neuron center point and the standard deviation of the Gaussian function reflects the targeted modeling of the progressive accumulation characteristics of fatigue and the risk characteristics of posture change.

[0008] The dynamic fusion weights are adaptively determined based on the relative magnitudes of the quantifiable indicators of muscle fatigue and postural imbalance. These weights are then used to weight and fuse fatigue compensation and posture correction parameters to generate final stiffness and damping parameters. This mechanism fundamentally solves the priority decision-making problem when fatigue adaptation and safety protection control objectives may conflict. When the patient is in a highly imbalanced state, the system automatically increases the posture response fusion weight to 0.8 to prioritize safety. When the patient is significantly fatigued but in a controllable posture state, the system automatically increases the fatigue response fusion weight to 0.7 to prioritize fatigue relief. When both states are within the normal range, the system uses a balanced weight of 0.5 to achieve comprehensive control. This adaptive weight switching strategy enables the control system to possess state perception and decision-making capabilities similar to a human therapist, allowing it to dynamically adjust the control center of gravity based on the patient's real-time state during training. Based on the final stiffness and damping parameters, the target joint angles in the personalized gait trajectory, and the actual joint angles, the driving torque of each joint is calculated and sent to the joint drive system for execution. The adaptive impedance parameters obtained from the aforementioned multi-level processing are finally transformed into torque control commands at the joint level. By compensating for position errors with the stiffness term and suppressing velocity mutations with the damping term in the impedance control formula, the robot exhibits compliant interactive characteristics that dynamically change with the patient's state on a macroscopic level. This ensures both trajectory tracking accuracy and the comfort and safety of human-machine force interaction. Compared with existing fixed impedance control methods, this significantly extends the patient's training tolerance time and reduces the risk of falls. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of an embodiment of the control method for the exoskeleton lower limb rehabilitation robot in this application. Figure 2 This is a schematic diagram illustrating the stiffness parameter adjustment characteristics of the two-layer RBF neural network in an embodiment of this application; Figure 3 This is a schematic diagram illustrating the damping parameter adjustment characteristics of a two-layer RBF neural network in an embodiment of this application. Detailed Implementation

[0011] This application provides a control method and system for an exoskeleton lower limb rehabilitation robot. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0012] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the control method for the exoskeleton lower limb rehabilitation robot in this application includes: Step S1: Generate a personalized gait trajectory based on the patient's personal information input through the user interface interaction device, and simultaneously activate the force sensor and inertial measurement unit sensor to collect joint interaction force, upper limb swing angle and trunk tilt angle; Specifically, after receiving the patient's body size parameters such as height, weight, thigh length, and calf length, as well as functional parameters such as muscle strength level and initial maximum active force benchmark, the human-computer interaction system calls the gait biomechanical model to calculate the joint angle range. The maximum hip extension angle is determined by the mapping relationship between height and standard gait database, the maximum knee flexion angle is calculated based on thigh length and femoral lever mechanics, and the ankle dorsiflexion angle is derived from calf length and foot landing cushioning requirements. The three joint angle ranges constitute an angle-time series within the gait cycle, forming a movement gait trajectory curve unique to the patient. The system synchronously sends start commands to the force sensor module and inertial measurement unit module. The strain gauge torque sensor captures the torque changes at the joints in real time through the resistance strain effect and converts them into voltage signals. The three-axis accelerometer and three-axis gyroscope built into the six-axis inertial measurement unit sensor measure linear acceleration and angular velocity, respectively. The upper limb swing angle and trunk tilt angle are calculated through attitude fusion algorithm. All sensor data are transmitted to the circular data buffer of the central computing control unit through the controller LAN bus at a period of 10 milliseconds.

[0013] Step S2: Statistically process the joint interaction forces within the sliding time window to obtain a quantitative index of muscle fatigue, and quantify the degree of abnormality of the upper limb swing angle and trunk tilt angle to obtain an index of postural imbalance. Specifically, a sliding window statistical processing method was used to analyze historical joint interaction force data to quantify the degree of muscle fatigue in patients, and anomaly detection was performed on posture angle data to assess fall risk. The sliding window was set to 30 seconds based on the physiological time constant of accumulated muscle fatigue. The window contained 3000 force sensor sampling points. The average interaction force reflected the patient's overall force exertion level during that time period, while the force variance reflected the stability of force exertion. When the patient's muscles are fatigued, the neuromuscular control ability decreases, leading to increased fluctuations in force output. The force decay rate was captured by comparing the difference in average force between the two segments before and after the window to capture the force exertion trend. The quantification of muscle fatigue was achieved by normalizing the three statistics with the initial maximum active force baseline using a weighted fusion formula. The weighting coefficients of 0.5, 0.3, and 0.2 reflected the clinical consensus that average force is a primary marker of fatigue. The posture data processing uses the Kalman filter algorithm to remove sensor noise. The upper limb swing abnormality is obtained by comparing the filtered upper limb swing angle with the lower limit of normal swing of 10 degrees. The trunk tilt abnormality is obtained by normalizing the absolute value of the filtered trunk tilt angle with the excess of the safety threshold of 5 degrees. The posture imbalance index is generated by weighted fusion of upper limb swing abnormality and trunk tilt abnormality. The weight configuration of 0.4 and 0.6 reflects the higher weight of trunk stability to fall risk.

[0014] Step S3: Input the muscle fatigue quantification index into the radial basis function neural network of the fatigue response layer to output the fatigue compensation stiffness parameter and fatigue compensation damping parameter. Input the posture imbalance index into the radial basis function neural network of the posture response layer to output the posture correction stiffness parameter and posture correction damping parameter. Adaptively determine the dynamic fusion weight based on the relative magnitude of the muscle fatigue quantification index and the posture imbalance index. Perform weighted fusion of the fatigue compensation parameter and the posture correction parameter through the dynamic fusion weight to generate the final stiffness parameter and the final damping parameter. Specifically, the hidden layer of the radial basis function neural network in the fatigue response layer has 5 Gaussian radial basis neurons corresponding to different fatigue levels. The uniform distribution of the neuron center points from 0.1 to 0.9 covers the entire spectrum from mild to severe fatigue. The standard deviation of the Gaussian function is 0.15 to ensure a smooth transition between adjacent fatigue levels. The activation value of each neuron is calculated using the Euclidean distance between the input muscle fatigue quantification index and the center point. The output layer maps the activation values ​​to fatigue compensation stiffness and fatigue compensation damping parameters. The stiffness coefficient decreases by up to 40% with increasing fatigue to reduce the burden on the patient to overcome elastic resistance, while the damping coefficient increases by up to 60% with increasing fatigue to enhance passive stability support. The hidden layer of the radial basis function neural network in the posture response layer has 3 neurons corresponding to mild, moderate, and severe imbalances. The output posture correction stiffness parameter is moderately increased by up to 30% during imbalance to prevent sudden joint flexion, while the posture correction damping parameter is significantly increased by up to 80% to achieve emergency braking. The dynamic fusion weights are adaptively determined in three scenarios based on the relative magnitudes of the muscle fatigue quantification index and the posture imbalance index. When the posture imbalance index is greater than 0.7, the posture response fusion weight is set to 0.8 to prioritize safety. When the muscle fatigue quantification index is greater than 0.6 and the posture imbalance index is less than 0.5, the fatigue response fusion weight is set to 0.7 to prioritize fatigue relief. When both indices are less than 0.4, the weights are each 0.5 to achieve balanced control. The final stiffness parameter and the final damping parameter are generated by linearly combining the fatigue compensation parameter and posture correction parameter output by the two layers of neural networks according to the weights using a weighted fusion formula.

[0015] Step S4: Calculate the driving torque of each joint based on the final stiffness parameters, final damping parameters, target joint angles in the personalized gait trajectory, and actual joint angles, and send it to each joint drive system for execution.

[0016] Specifically, the final impedance parameters are converted into joint drive torques and executed to achieve a compliant response of the robot to the patient's movement intentions. The joint drive torque in the impedance control formula consists of stiffness and damping terms. The stiffness term is the final stiffness parameter multiplied by the difference between the target joint angle and the actual joint angle, generating a position error compensation torque. The damping term is the final damping parameter multiplied by the difference between the target angular velocity and the actual angular velocity, generating a velocity damping torque. The sum of these two terms constitutes the joint drive torque that conforms to the patient's current muscle strength and posture. The field-oriented control algorithm converts the drive torque into current commands in the motor's dq coordinate system. The q-axis current is proportional to the torque, generating electromagnetic torque, while the d-axis current controls the excitation magnetic field strength. Three-phase current commands are generated through inverse Park transform and space vector pulse width modulation to drive the brushless motor. Controller area network communication sends control messages containing joint identifiers, drive torques, and function codes to each joint drive system. Each joint driver filters the corresponding message based on the joint identifier and executes impedance parameter adjustment or angle tracking tasks according to the function code.

[0017] In one specific embodiment, step S1 includes: The system receives the patient's height, weight, thigh length, calf length, target stride length, target gait speed, muscle strength level, and initial maximum active force baseline value through a user interface interaction device, generates a unique identification icon, and stores it. Based on height and thigh length, the maximum hip extension angle, maximum knee flexion angle, and ankle dorsiflexion angle are calculated using a gait biomechanical model. Based on the angle range of each joint, a personalized gait trajectory containing angle-time series is generated. The joint interaction force of each joint is collected by a strain gauge torque sensor with a sampling period of 10 milliseconds, and the upper limb swing angle and trunk tilt angle are collected by a six-axis inertial measurement unit sensor with a sampling period of 10 milliseconds. The joint interaction force, upper limb swing angle and trunk tilt angle are transmitted to the central computing and control unit.

[0018] Specifically, the process of generating a unique identifier icon after receiving the patient's personal information includes information verification and file creation mechanisms. The system verifies the reasonableness of input body size parameters such as height, weight, thigh length, and calf length to ensure that the values ​​are within the normal range of human physiology. The muscle strength level uses the internationally recognized 0-5 scale to calibrate the patient's muscle strength status. The initial maximum active force benchmark value is obtained through pre-training tests, which is the peak torque that the patient can generate when actively exerting force on the joints in a static state. This benchmark value serves as a normalization reference for subsequent muscle fatigue measurement. The unique identifier icon is generated by combining the first letter of the patient's name and the file creation date to ensure the uniqueness of the file. The data structure associated with the icon contains all personal information parameters and is stored in local memory. When the user clicks the icon, the system calls the corresponding parameter set to enter the training mode.

[0019] The gait biomechanical model calculates the angle range of each joint based on the patient's height and thigh length. This mechanism is based on the kinematic chain structure of the human lower limb. The maximum hip extension angle is positively correlated with height because taller individuals have a larger stride and require greater hip joint mobility. The maximum knee flexion angle is primarily determined by thigh length, as thigh length directly affects the femoral lever arm length, thus determining the torque requirement and upper limit of the safe angle during knee flexion. The ankle dorsiflexion angle is influenced by lower leg length, as lower leg length relates to the ankle joint's range of motion during the foot's landing cushioning phase. The calculated angle ranges of these three joints serve as boundary constraints. The system's built-in gait cycle template divides a complete gait cycle into a support phase and a swing phase. Within each phase, a cubic spline interpolation algorithm generates a smooth time-varying sequence of angles for each joint, forming a personalized gait trajectory curve that includes the angle-time correspondence of the hip, knee, and ankle joints within the gait cycle. The strain gauge torque sensor senses joint torque through a resistive strain gauge attached to the surface of a metallic elastomer. When the joint is subjected to force, the elastomer undergoes a slight deformation, causing a change in the resistance value of the strain gauge. The signal conditioning circuit converts the resistance change into a voltage signal and filters out high-frequency noise. A sampling period of 10 milliseconds means that 100 joint interaction force data are collected per second, meeting the dynamic response requirements in rehabilitation training. The six-axis inertial measurement unit (IMU) sensor incorporates a three-axis accelerometer to measure linear acceleration and a three-axis gyroscope to measure angular velocity. The sensor mounted on the cane calculates the upper limb swing angle based on the cane's swing trajectory, and the sensor mounted on the robot's back measures the torso's tilt angle relative to the vertical direction. All sensor data is transmitted to the data buffer of the central computing and control unit via the controller's local area network (LAN) bus.

[0020] In one specific embodiment, step S2, statistically processing the joint interaction forces within the sliding time window to obtain a quantitative index of muscle fatigue includes: Set the sliding time window length to 30 seconds, perform statistical calculations on the joint interaction forces within the window, and obtain the average interaction force, force variance, and force decay rate. Based on the average interaction force, force variance, force decay rate and the initial maximum active force baseline, a weighted fusion formula is used to calculate the quantification index of muscle fatigue. Kalman filtering was applied to the upper limb swing angle and the trunk tilt angle to obtain the filtered upper limb swing angle and the filtered trunk tilt angle. The upper limb swing abnormality was calculated based on the filtered upper limb swing angle, and the trunk tilt abnormality was calculated based on the filtered trunk tilt angle. The posture imbalance index was obtained by weighted fusion of the upper limb swing abnormality and the trunk tilt abnormality, with weight coefficients of 0.4 and 0.6, respectively.

[0021] Specifically, the sliding time window length is set to 30 seconds based on the physiological time characteristics of muscle fatigue accumulation in the human body. During continuous contraction of skeletal muscle, the accumulation of metabolic products such as lactic acid takes about 20 to 40 seconds to significantly affect muscle force output. The 30-second window can capture the fatigue accumulation trend while avoiding a slow response due to an excessively long window. The statistical calculation of joint interaction force within the window is obtained by summing the results from 3000 sampling points and dividing by the number of sampling points to obtain the average interaction force. The force variance is calculated by squared the difference between the force value of each sampling point and the average interaction force, summing the results, and then dividing by the number of sampling points. The force decay rate is obtained by dividing the window into two sub-intervals, the first 10 seconds and the last 10 seconds, calculating the average force of each sub-interval, and then dividing the difference by the time interval of 20 seconds to obtain the force decrease rate per unit time. In the weighted fusion formula, a weight coefficient of 0.5 is assigned to the ratio of average interaction force to the initial maximum active force benchmark, reflecting the clinical consensus that average force is the main indicator of fatigue. A weight of 0.3 is assigned to the force variance normalization term, reflecting the importance of force stability in fatigue assessment. A weight of 0.2 is assigned to the force decay rate term, capturing the dynamic trend of fatigue. The weighted sum of the three terms constitutes a quantification index of muscle fatigue with a value range of 0 to 1.

[0022] Kalman filtering processes the upper limb swing angle and trunk tilt angle by fusing the predicted and actual measured values ​​from the inertial measurement unit (IMU) sensors using a recursive algorithm. The filter maintains a state estimate and an estimation error covariance. In each sampling period, the current state is first predicted based on the system dynamics model. Then, the state estimate is updated by multiplying the residual between the measured and predicted values ​​by the Kalman gain. The process noise covariance reflects the uncertainty of the system model, and the measurement noise covariance reflects the sensor noise level. After multiple iterations, the filter converges and outputs a smoothed angle value that removes high-frequency vibration noise and zero drift. Upper limb swing anomaly is calculated by determining whether the filtered upper limb swing angle is below the normal lower limit of 10 degrees. If it is, the difference between 10 degrees and the actual angle is divided by 10 degrees as the normalized anomaly. Trunk tilt anomaly is calculated by determining whether the absolute value of the filtered trunk tilt angle exceeds the safety threshold of 5 degrees. If it exceeds, the excess is divided by 3 degrees as the normalized anomaly. When the anomaly reaches 8 degrees, the anomaly is 1, triggering an emergency stop. The postural imbalance index is generated by weighted fusion of upper limb swing abnormality and trunk tilt abnormality. The weight coefficient of 0.4 is assigned to upper limb swing abnormality to reflect the influence of upper and lower limb coordination on overall posture, and the weight coefficient of 0.6 is assigned to trunk tilt abnormality to reflect the decisive role of trunk stability as the core of human body center of gravity control in fall risk. The weighted sum of the two constitutes a postural imbalance index with a value range of 0 to 1 for subsequent neural network processing.

[0023] In one specific embodiment, step S3, inputting the muscle fatigue quantification index into the fatigue response layer radial basis function neural network to output fatigue compensation stiffness parameters and fatigue compensation damping parameters, includes: The quantification index of muscle fatigue is input into the input layer of the radial basis function neural network of the fatigue response layer. The hidden layer of the radial basis function neural network of the fatigue response layer is set with 5 Gaussian radial basis neurons, the center point vector of the neurons is 0.1, 0.3, 0.5, 0.7, 0.9, and the standard deviation parameter of the Gaussian function is 0.15. The activation values ​​of each neuron were calculated using Gaussian radial basis functions. The fatigue compensation stiffness parameter and fatigue compensation damping parameter are calculated by weighted linear combination of the output layer. The fatigue compensation stiffness parameter is equal to the reference stiffness parameter multiplied by 1 minus the weighted sum of neuron activation values ​​of 0.4 times. The fatigue compensation damping parameter is equal to the reference damping parameter multiplied by 1 plus the weighted sum of neuron activation values ​​of 0.6 times.

[0024] Specifically, the fatigue response layer radial basis function neural network adopts a three-layer structure to achieve a nonlinear mapping from muscle fatigue to impedance parameters. The input layer receives a single scalar, namely the muscle fatigue quantification index. The center point vectors of the five Gaussian radial basis neurons in the hidden layer are uniformly distributed in the interval from 0 to 1 (0.1, 0.3, 0.5, 0.7, 0.9), corresponding to five levels from mild to severe fatigue. The standard deviation parameter of Gaussian function, 0.15, determines the response width of each neuron, ensuring appropriate overlap between adjacent fatigue levels and avoiding discontinuities in the piecewise function. The activation value of each neuron is calculated using the Gaussian radial basis function, i.e., the natural exponential function. The exponential term is the square of the difference between the input muscle fatigue quantification index and the center point of the neuron, divided by twice the square of the standard deviation, and then negative. When the input value is exactly equal to a certain center point, the activation value of the neuron reaches a peak of 1. The further the input value deviates from the center point, the lower the activation value according to the Gaussian distribution. The standard deviation of 0.15 ensures that the neuron maintains a high activation level within a range of approximately 0.3 above and below the center point.

[0025] The output layer weighted linear combination maps the activation values ​​of the five neurons in the hidden layer to fatigue compensation stiffness and fatigue compensation damping parameters through preset connection weights. The connection weight vector is optimized through clinical data training to ensure that the neural network output conforms to the principles of rehabilitation treatment. In the fatigue compensation stiffness parameter calculation formula, the baseline stiffness parameter is the system preset value of 12 Nm / radian. The weighted sum of neuron activation values ​​reflects the comprehensive response intensity of neurons at each fatigue level to the current fatigue level. 1 minus 0.4 times the weighted sum of activation values ​​constitutes the stiffness adjustment coefficient. The coefficient 0.4 limits the stiffness to a maximum reduction of 60% of the baseline value. That is, when the fatigue level is 1, the weighted sum of activation values ​​approaches 1, reducing the stiffness to 7.2 Nm / radian to reduce the burden on the patient to overcome elastic resistance. In the fatigue compensation damping parameter calculation formula, the baseline damping parameter is 5 Nm / s per radian. The damping adjustment coefficient is formed by adding 0.6 times the activation value to the weighted sum of 1. The coefficient 0.6 limits the damping to increase to a maximum of 160% of the baseline value, that is, when the fatigue degree is 1, the damping increases to 8 Nm / s per radian. The increase in damping makes the joint movement speed subject to stronger viscous resistance constraints, providing more passive stability support. The coordinated adjustment of stiffness reduction and damping increase realizes the smooth transition of the robot from active drive to passive support under fatigue state.

[0026] In one specific embodiment, inputting the attitude imbalance index into the attitude response layer radial basis function neural network to output attitude correction stiffness parameters and attitude correction damping parameters includes: The posture imbalance index is input into the input layer of the posture response layer radial basis function neural network. The hidden layer of the posture response layer radial basis function neural network is set with 3 Gaussian radial basis neurons, the center point vector of the neurons is 0.3, 0.6, and 0.9, and the standard deviation parameter of the Gaussian function is 0.2. The activation values ​​of each neuron were calculated using Gaussian radial basis functions. The attitude correction stiffness parameter and attitude correction damping parameter are calculated by weighted linear combination of the output layer. The attitude correction stiffness parameter is equal to the baseline stiffness parameter multiplied by 1 plus the weighted sum of 0.3 times the neuron activation value, and the attitude correction damping parameter is equal to the baseline damping parameter multiplied by 1 plus the weighted sum of 0.8 times the neuron activation value.

[0027] Specifically, the hidden layer of the radial basis function neural network in the posture response layer has three Gaussian radial basis neurons corresponding to mild, moderate, and severe posture imbalance levels. The distribution intervals of the neuron center point vectors (0.3, 0.6, 0.9) are sparser than those in the fatigue response layer, reflecting the clinical judgment characteristics of posture imbalance grading. The standard deviation parameter of the Gaussian function (0.2) is larger than that of the fatigue response layer (0.15), resulting in a wider response range for each neuron and ensuring a smooth transition of imbalance degree at the boundaries between adjacent levels. The activation value of each neuron is also calculated using the Gaussian radial basis function. The exponent term is the square of the difference between the input posture imbalance index and the neuron's center point, divided by twice the square of the standard deviation, and then the negative value is taken. When the posture imbalance index is 0.3, the activation value of the mild imbalance neuron is 1, and the activation values ​​of other neurons are lower. When the posture imbalance index is 0.7, both moderate and severe imbalance neurons are activated, reflecting the continuous judgment of the imbalance degree. The standard deviation of 0.2 ensures that the neurons maintain a significant response within a range of approximately 0.4 above and below the center point.

[0028] The weighted linear combination of the output layer maps the activation values ​​of the three neurons to the posture correction stiffness parameter and the posture correction damping parameter. In the formula for calculating the posture correction stiffness parameter, 1 plus 0.3 times the weighted sum of the neuron activation values ​​constitutes the stiffness adjustment coefficient. The coefficient 0.3 limits the stiffness to increase to a maximum of 130% of the baseline stiffness parameter. That is, when the posture imbalance index is 1, the stiffness increases to 15.6 Nm per radian. The purpose of moderately increasing the stiffness is to improve the joint's resistance to disturbance during posture imbalance and prevent the joint from suddenly losing control and bending under external interference, which could lead to a fall. In the formula for calculating the damping parameter of posture correction, 1 plus a weighted sum of 0.8 times the neuron activation value constitutes the damping adjustment coefficient. The coefficient 0.8 is significantly greater than the stiffness coefficient 0.3, reflecting that the main strategy of posture correction is to increase damping to achieve emergency braking. When the posture imbalance index is 1, the damping is increased to 180% of the benchmark damping parameter, that is, 9 Nm / s per radian. The significantly increased damping applies strong viscous resistance to the joint angular velocity, rapidly consuming kinetic energy and decelerating the out-of-control movement to a safe state. The combination strategy of slightly increasing stiffness and significantly increasing damping ensures the stability of the joint structure and achieves rapid braking response.

[0029] Figure 2 This is a schematic diagram illustrating the stiffness parameter adjustment characteristics of a two-layer RBF neural network in an embodiment of this application. Figure 2As shown, subfigure (a) illustrates the stiffness adjustment mechanism of the fatigue response layer. When the quantification index of muscle fatigue increases from 0 to 1, the fatigue compensation stiffness parameter linearly decreases from the baseline stiffness of 12 N·m / rad to 7.2 N·m / rad, a reduction of 40%. The filled area visually represents the amount of stiffness reduction. This adjustment strategy effectively reduces the active force burden on patients under fatigue by decreasing joint elastic resistance. Subfigure (b) illustrates the stiffness adjustment mechanism of the posture response layer. When the posture imbalance index increases from 0 to 1, the posture correction stiffness parameter linearly increases from the baseline stiffness of 12 N·m / rad to 15.6 N·m / rad, an increase of 30%. The filled area represents the amount of stiffness increase. This adjustment strategy improves the resistance to disturbance and prevents sudden uncontrolled flexion of the joint by moderately increasing the stiffness of the joint structure during posture imbalance. The stiffness parameters output from the two neural network layers are dynamically fused and weighted to generate the final stiffness parameters, achieving an adaptive balance between fatigue relief and posture stability.

[0030] Figure 3 This is a schematic diagram illustrating the damping parameter adjustment characteristics of a two-layer RBF neural network in an embodiment of this application. Figure 3 As shown, subfigure (a) illustrates the damping adjustment mechanism of the fatigue response layer. When the quantification index of muscle fatigue increases from 0 to 1, the fatigue compensation damping parameter linearly increases from the baseline damping of 5 N·m·s / rad to 8.0 N·m·s / rad, an increase of 60%. The diamond-shaped grid-filled area represents the damping increase. This adjustment strategy provides more passive stability support and prolongs training endurance time when the patient is fatigued by increasing the viscous damping coefficient. Subfigure (b) illustrates the damping adjustment mechanism of the posture response layer. When the posture imbalance index increases from 0 to 1, the posture correction damping parameter linearly increases from the baseline damping of 5 N·m·s / rad to 9.0 N·m·s / rad, an increase of 80%. The horizontal line-filled area represents the damping increase. This adjustment strategy achieves emergency braking response in the imbalance state by significantly increasing the damping coefficient to apply strong viscous resistance to the joint angular velocity. Compared to stiffness parameters, damping parameters have a larger adjustment range, and both layers of neural networks adopt an increase strategy, reflecting the core design concept of this application to achieve compliant control and safety protection through damping adjustment. The final damping parameter generated after dynamic fusion of the two layers intelligently switches between passive support and emergency braking according to the patient's real-time status.

[0031] In one specific embodiment, a dynamic fusion weight is adaptively determined based on the relative magnitudes of the muscle fatigue quantification index and the posture imbalance index. The fatigue compensation parameters and posture correction parameters are then weighted and fused using this dynamic fusion weight to generate the final stiffness parameters and final damping parameters, including: When the attitude imbalance index is greater than 0.7, the attitude response fusion weight is set to 0.8 and the fatigue response fusion weight is set to 0.2. When the quantification index of muscle fatigue is greater than 0.6 and the postural imbalance index is less than 0.5, the postural response fusion weight is set to 0.3 and the fatigue response fusion weight is set to 0.7. When the quantification index of muscle fatigue is less than 0.4 and the postural imbalance index is less than 0.4, the postural response fusion weight is set to 0.5 and the fatigue response fusion weight is set to 0.5.

[0032] Based on the fatigue response fusion weight and the attitude response fusion weight, the final stiffness parameter and the final damping parameter are calculated using a weighted fusion formula. The final stiffness parameter is equal to the fatigue response fusion weight multiplied by the fatigue compensation stiffness parameter plus the attitude response fusion weight multiplied by the attitude correction stiffness parameter. The final damping parameter is equal to the fatigue response fusion weight multiplied by the fatigue compensation damping parameter plus the attitude response fusion weight multiplied by the attitude correction damping parameter.

[0033] Specifically, the adaptive determination mechanism of dynamic fusion weights intelligently switches control priorities by comparing the relative magnitudes of quantifiable muscle fatigue indicators and postural imbalance indicators. When the postural imbalance indicator is greater than 0.7, it indicates that the patient is in a state of severe imbalance and close to the fall threshold. At this time, the postural response fusion weight is set to 0.8 and the fatigue response fusion weight is set to 0.2, reflecting the safety priority principle. Even if the patient is in a state of high fatigue, the system still prioritizes the execution of postural correction strategies, achieving emergency braking by significantly increasing damping to avoid falls. When the quantifiable muscle fatigue indicator is greater than 0.6 and the postural imbalance indicator is less than 0.5, it indicates that the patient is in a state of significant fatigue but still has controllable posture. At this time, the fatigue response fusion weight is set to 0.7 and the postural response fusion weight is set to 0.3, reflecting the fatigue relief priority principle. The system mainly reduces the burden on the patient to overcome elastic resistance by reducing stiffness and moderately increases damping to provide passive support and extend training endurance time. When the quantification index of muscle fatigue is less than 0.4 and the postural imbalance index is less than 0.4, it indicates that the patient is in the normal or slightly abnormal range. At this time, both fusion weights are set to 0.5 to achieve a balanced control of fatigue adaptation and postural stability. The system does not favor either control target but comprehensively considers both aspects.

[0034] The weighted fusion calculation of the final stiffness parameter and the final damping parameter is generated by linearly combining the output parameters of the fatigue response layer and the attitude response layer according to the fusion weight ratio. In the calculation of the final stiffness parameter, the fatigue response fusion weight is multiplied by the fatigue compensation stiffness parameter to obtain the fatigue compensation component, and the attitude response fusion weight is multiplied by the attitude correction stiffness parameter to obtain the attitude correction component. The sum of the two components constitutes the final stiffness parameter. When the attitude imbalance index is greater than 0.7, the contribution of the attitude correction stiffness parameter reaches 80%, making the final stiffness mainly dominated by the need for attitude correction, tending to increase stiffness to prevent joint loss of control. When the muscle fatigue quantification index is greater than 0.6 and the attitude imbalance index is less than 0.5, the contribution of the fatigue compensation stiffness parameter reaches 70%, making the final stiffness mainly dominated by the need for fatigue relief, tending to reduce stiffness to reduce the burden. The final damping parameter calculation adopts the same weighted fusion logic. The fatigue compensation damping parameter and the attitude correction damping parameter are weighted according to the fusion weight and then summed. Since both fatigue compensation and attitude correction strategies require increasing damping but only by different amounts, the final damping parameter is greater than the reference damping parameter in all scenarios. The difference is that the damping increase is more significant in the attitude safety priority scenario to achieve the emergency braking effect.

[0035] In one specific embodiment, step S4 includes: Based on the final stiffness parameters and final damping parameters, the target joint angle in the personalized gait trajectory, the actual joint angle, target angular velocity, and actual angular velocity fed back by the force sensor, the driving torque of each joint is calculated through the impedance control formula. The impedance control formula is that the driving torque of each joint is equal to the final stiffness parameter multiplied by the difference between the target joint angle and the actual joint angle, plus the final damping parameter multiplied by the difference between the target angular velocity and the actual angular velocity. The drive torque of each joint is converted into a three-phase current command for the motor through a field-oriented control algorithm. The control message containing the joint identifier, the drive torque of each joint, and the function code is sent to each joint drive system for execution through the controller local area network communication.

[0036] Specifically, the impedance control formula transforms the final impedance parameter into joint driving torque through a combination of stiffness and damping terms, enabling the robot to respond compliantly to the patient's movement intentions. In the stiffness term calculation, the target joint angle is derived from the angle value corresponding to the personalized gait trajectory at the current moment. The actual joint angle is fed back in real-time by the encoder built into the force sensor module. The difference between the two, i.e., the position error, reflects the degree of deviation between the patient's actual movement and the preset trajectory. Finally, the stiffness parameter multiplied by the position error generates an elastic restoring torque that drives the joint towards the target angle. Greater stiffness results in a stronger elastic torque and a more closely the joint follows the trajectory, but also increases the resistance felt by the patient. In the damping term calculation, the target angular velocity is obtained by taking the time derivative of the angle sequence of the personalized gait trajectory. The actual angular velocity is calculated from the time difference of the encoder data. The difference between the two, i.e., the velocity error, reflects the smoothness of the movement. Finally, the damping parameter multiplied by the velocity error generates a viscous damping torque that suppresses abrupt changes in joint movement speed. Greater damping results in a stronger viscous torque and a smoother movement, but also a slower response speed. The sum of the stiffness and damping terms constitutes the driving torque of each joint. This torque takes into account the patient's current muscle fatigue and postural imbalance. The robot's assistive force is dynamically adapted in real time through dynamically adjusted stiffness and damping parameters.

[0037] The field-oriented control algorithm converts the driving torque of each joint into three-phase current commands for the motor, achieving a precise mapping from torque to current. The algorithm first divides the driving torque by the motor torque constant to obtain the target value of the q-axis current in the dq coordinate system. The q-axis current, proportional to the electromagnetic torque, is the main component generating the driving torque. The d-axis current control excitation magnetic field strength is typically set to zero to maximize torque output efficiency. The current command in the dq coordinate system is converted into a current command in the three-phase stationary coordinate system using inverse Park transform. Then, space vector pulse width modulation (SPWM) technology is used to generate the switching signals for the three-phase inverter, driving the stator windings of the brushless motor to generate a rotating magnetic field. The controller area network (LAN) communication uses a message mechanism to transmit control commands. The message data segment includes a joint identifier field to specify the hip, knee, or ankle joint; a driving torque field for each joint carrying the specific torque value; and a function code field identifying the task type, such as impedance parameter adjustment or angle tracking execution. After receiving the message, each joint drive system filters out the corresponding joint control command through the joint identifier, parses the task type according to the function code, and drives the motor to perform the corresponding action, realizing distributed coordinated control of the multi-joint drive system by the central computing control unit.

[0038] The control method of the exoskeleton lower limb rehabilitation robot in the embodiments of this application has been described above. The control system of the exoskeleton lower limb rehabilitation robot in the embodiments of this application is described below. One embodiment of the control system of the exoskeleton lower limb rehabilitation robot in the embodiments of this application includes: The extraction module is used to generate personalized gait trajectories based on the patient's personal information input through the user interface interaction device, and simultaneously activates force sensors and inertial measurement unit sensors to collect joint interaction forces, upper limb swing angles, and trunk tilt angles. The quantification module is used to statistically process the joint interaction forces within the sliding time window to obtain a quantified index of muscle fatigue, and to quantify the degree of abnormality of the upper limb swing angle and trunk tilt angle to obtain a postural imbalance index. The generation module is used to input the muscle fatigue quantification index into the fatigue response layer radial basis function neural network to output fatigue compensation stiffness parameters and fatigue compensation damping parameters, input the posture imbalance index into the posture response layer radial basis function neural network to output posture correction stiffness parameters and posture correction damping parameters, adaptively determine dynamic fusion weights based on the relative magnitudes of the muscle fatigue quantification index and the posture imbalance index, and perform weighted fusion of fatigue compensation parameters and posture correction parameters through the dynamic fusion weights to generate final stiffness parameters and final damping parameters. The drive module is used to calculate the drive torque of each joint based on the final stiffness parameters and final damping parameters, the target joint angle in the personalized gait trajectory and the actual joint angle, and then send it to the drive system of each joint for execution.

[0039] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A control method for an exoskeleton lower limb rehabilitation robot, characterized in that, The method includes: Step S1: Generate a personalized gait trajectory based on the patient's personal information input through the user interface interaction device, and simultaneously activate the force sensor and inertial measurement unit sensor to collect joint interaction force, upper limb swing angle and trunk tilt angle; Step S2: Statistically process the joint interaction forces within the sliding time window to obtain a quantified index of muscle fatigue, and quantify the degree of abnormality of the upper limb swing angle and trunk tilt angle to obtain a postural imbalance index. Step S3: Input the muscle fatigue quantification index into the radial basis function neural network of the fatigue response layer to output fatigue compensation stiffness parameters and fatigue compensation damping parameters. Input the posture imbalance index into the radial basis function neural network of the posture response layer to output posture correction stiffness parameters and posture correction damping parameters. Adaptively determine the dynamic fusion weight based on the relative magnitude of the muscle fatigue quantification index and the posture imbalance index. Use the dynamic fusion weight to perform weighted fusion of the fatigue compensation parameters and posture correction parameters to generate the final stiffness parameters and final damping parameters. Step S4: Calculate the driving torque of each joint based on the final stiffness parameter and final damping parameter, the target joint angle in the personalized gait trajectory and the actual joint angle, and send it to each joint drive system for execution.

2. The control method for the exoskeleton lower limb rehabilitation robot according to claim 1, characterized in that, Step S1 includes: The system receives the patient's height, weight, thigh length, calf length, target stride length, target gait speed, muscle strength level, and initial maximum active force baseline value through a user interface interaction device, generates a unique identification icon, and stores it. Based on the height and thigh length, the maximum hip extension angle, maximum knee flexion angle, and ankle dorsiflexion angle are calculated using a gait biomechanical model. A personalized gait trajectory containing angle-time series is generated based on the angle range of each joint. The joint interaction force of each joint is collected by a strain gauge torque sensor with a sampling period of 10 milliseconds, and the upper limb swing angle and trunk tilt angle are collected by a six-axis inertial measurement unit sensor with a sampling period of 10 milliseconds. The joint interaction force, upper limb swing angle and trunk tilt angle are transmitted to the central computing and control unit.

3. The control method for the exoskeleton lower limb rehabilitation robot according to claim 2, characterized in that, In step S2, the statistical processing of the joint interaction forces within the sliding time window to obtain a quantitative index of muscle fatigue includes: Set the sliding time window length to 30 seconds, and perform statistical calculations on the joint interaction forces within the window to obtain the average interaction force, force variance, and force decay rate. Based on the average interaction force, force variance, force decay rate and the initial maximum active force benchmark value, the muscle fatigue quantification index is calculated by a weighted fusion formula. The upper limb swing angle and trunk tilt angle are processed by Kalman filtering to obtain the filtered upper limb swing angle and the filtered trunk tilt angle. The upper limb swing abnormality is calculated based on the filtered upper limb swing angle, and the trunk tilt abnormality is calculated based on the filtered trunk tilt angle. The posture imbalance index is obtained by weighted fusion of the upper limb swing abnormality and the trunk tilt abnormality, where the weight coefficients in the weighted fusion are 0.4 and 0.6, respectively.

4. The control method for the exoskeleton lower limb rehabilitation robot according to claim 3, characterized in that, In step S3, inputting the muscle fatigue quantification index into the fatigue response layer radial basis function neural network to output fatigue compensation stiffness parameters and fatigue compensation damping parameters includes: The muscle fatigue quantification index is input into the input layer of the fatigue response layer radial basis function neural network. The hidden layer of the fatigue response layer radial basis function neural network is set with 5 Gaussian radial basis neurons, the neuron center point vectors are 0.1, 0.3, 0.5, 0.7, and 0.9, and the Gaussian function standard deviation parameter is 0.

15. The activation values ​​of each neuron were calculated using Gaussian radial basis functions. The fatigue compensation stiffness parameter and fatigue compensation damping parameter are calculated by weighted linear combination of the output layer. The fatigue compensation stiffness parameter is equal to the reference stiffness parameter multiplied by 1 minus the weighted sum of neuron activation values ​​by 0.4 times. The fatigue compensation damping parameter is equal to the reference damping parameter multiplied by 1 plus the weighted sum of neuron activation values ​​by 0.6 times.

5. The control method for the exoskeleton lower limb rehabilitation robot according to claim 4, characterized in that, In step S3, inputting the attitude imbalance index into the radial basis function neural network of the attitude response layer to output the attitude correction stiffness parameter and the attitude correction damping parameter includes: The posture imbalance index is input into the input layer of the posture response layer radial basis function neural network. The hidden layer of the posture response layer radial basis function neural network is set with 3 Gaussian radial basis neurons, the center point vectors of the neurons are 0.3, 0.6, and 0.9, and the standard deviation parameter of the Gaussian function is 0.

2. The activation values ​​of each neuron were calculated using Gaussian radial basis functions. The attitude correction stiffness parameter and attitude correction damping parameter are calculated by weighted linear combination of the output layer. The attitude correction stiffness parameter is equal to the reference stiffness parameter multiplied by 1 plus a weighted sum of 0.3 times the neuron activation value. The attitude correction damping parameter is equal to the reference damping parameter multiplied by 1 plus a weighted sum of 0.8 times the neuron activation value.

6. The control method for the exoskeleton lower limb rehabilitation robot according to claim 5, characterized in that, In step S3, the step of adaptively determining the dynamic fusion weight based on the relative magnitude of the muscle fatigue quantification index and the posture imbalance index, and then weighting and fusing the fatigue compensation parameters and posture correction parameters using the dynamic fusion weight to generate the final stiffness parameters and final damping parameters includes: When the attitude imbalance index is greater than 0.7, the attitude response fusion weight is set to 0.8 and the fatigue response fusion weight is set to 0.

2. When the muscle fatigue quantification index is greater than 0.6 and the posture imbalance index is less than 0.5, the posture response fusion weight is set to 0.3 and the fatigue response fusion weight is set to 0.

7. When the muscle fatigue quantification index is less than 0.4 and the posture imbalance index is less than 0.4, the posture response fusion weight is set to 0.5 and the fatigue response fusion weight is set to 0.

5. Based on the fatigue response fusion weight and the attitude response fusion weight, the final stiffness parameter and the final damping parameter are calculated using a weighted fusion formula. The final stiffness parameter is equal to the fatigue response fusion weight multiplied by the fatigue compensation stiffness parameter plus the attitude response fusion weight multiplied by the attitude correction stiffness parameter. The final damping parameter is equal to the fatigue response fusion weight multiplied by the fatigue compensation damping parameter plus the attitude response fusion weight multiplied by the attitude correction damping parameter.

7. The control method for the exoskeleton lower limb rehabilitation robot according to claim 1, characterized in that, Step S4 includes: Based on the final stiffness parameter and final damping parameter, the target joint angle in the personalized gait trajectory, the actual joint angle, target angular velocity, and actual angular velocity fed back by the force sensor, the driving torque of each joint is calculated using an impedance control formula. The impedance control formula is that the driving torque of each joint is equal to the final stiffness parameter multiplied by the difference between the target joint angle and the actual joint angle, plus the final damping parameter multiplied by the difference between the target angular velocity and the actual angular velocity. The driving torque of each joint is converted into a three-phase current command for the motor through a field-oriented control algorithm. The control message containing the joint identifier, the driving torque of each joint, and the function code is sent to the driving system of each joint for execution through the controller local area network communication.

8. A control system for an exoskeleton lower limb rehabilitation robot, characterized in that, A control method for implementing the exoskeleton lower limb rehabilitation robot as described in any one of claims 1-7, wherein the control system of the exoskeleton lower limb rehabilitation robot comprises: The extraction module is used to generate personalized gait trajectories based on the patient's personal information input through the user interface interaction device, and simultaneously activates force sensors and inertial measurement unit sensors to collect joint interaction forces, upper limb swing angles, and trunk tilt angles. The quantification module is used to statistically process the joint interaction forces within the sliding time window to obtain a quantified index of muscle fatigue, and to quantify the degree of abnormality of the upper limb swing angle and trunk tilt angle to obtain a postural imbalance index. The generation module is used to input the muscle fatigue quantification index into the fatigue response layer radial basis function neural network to output fatigue compensation stiffness parameters and fatigue compensation damping parameters, input the posture imbalance index into the posture response layer radial basis function neural network to output posture correction stiffness parameters and posture correction damping parameters, adaptively determine dynamic fusion weights based on the relative magnitudes of the muscle fatigue quantification index and the posture imbalance index, and perform weighted fusion of fatigue compensation parameters and posture correction parameters through the dynamic fusion weights to generate final stiffness parameters and final damping parameters. The drive module is used to calculate the drive torque of each joint based on the final stiffness parameters and final damping parameters, the target joint angle in the personalized gait trajectory and the actual joint angle, and then send it to the drive system of each joint for execution.

9. The system according to claim 8, characterized in that, Personalized gait trajectories are generated based on the patient's personal information input through the user interface device. Simultaneously, force sensors and inertial measurement units collect data on joint interaction forces, upper limb swing angles, and trunk tilt angles, including: The system receives the patient's height, weight, thigh length, calf length, target stride length, target gait speed, muscle strength level, and initial maximum active force baseline value through a user interface interaction device, generates a unique identification icon, and stores it. Based on the height and thigh length, the maximum hip extension angle, maximum knee flexion angle, and ankle dorsiflexion angle are calculated using a gait biomechanical model. A personalized gait trajectory containing angle-time series is generated based on the angle range of each joint. The joint interaction force of each joint is collected by a strain gauge torque sensor with a sampling period of 10 milliseconds, and the upper limb swing angle and trunk tilt angle are collected by a six-axis inertial measurement unit sensor with a sampling period of 10 milliseconds. The joint interaction force, upper limb swing angle and trunk tilt angle are transmitted to the central computing and control unit.

10. The system according to claim 8, characterized in that, Statistical processing of the joint interaction forces within the sliding time window yields quantitative indicators of muscle fatigue, including: Set the sliding time window length to 30 seconds, and perform statistical calculations on the joint interaction forces within the window to obtain the average interaction force, force variance, and force decay rate. Based on the average interaction force, force variance, force decay rate and the initial maximum active force benchmark value, the muscle fatigue quantification index is calculated by a weighted fusion formula. The upper limb swing angle and trunk tilt angle are processed by Kalman filtering to obtain the filtered upper limb swing angle and the filtered trunk tilt angle. The upper limb swing abnormality is calculated based on the filtered upper limb swing angle, and the trunk tilt abnormality is calculated based on the filtered trunk tilt angle. The posture imbalance index is obtained by weighted fusion of the upper limb swing abnormality and the trunk tilt abnormality, where the weight coefficients in the weighted fusion are 0.4 and 0.6, respectively.

Images