Upper limb exoskeleton adaptive control method based on multi-modal data fusion

Abstract

The invention discloses an upper limb exoskeleton adaptive control method based on multi-modal data fusion, and relates to the technical field of upper limb exoskeleton control. According to the upper limb exoskeleton self-adaptive control method based on multi-modal data fusion, multi-modal data such as user electromyographic signals, joint movement signals and motor output torque time sequence data are obtained; a human-computer interaction torque observation set is generated through an upper limb exoskeleton reverse power model in combination with joint movement signals and motor output torque time sequence data; based on the observation set, carrying out adaptive fusion on the electromyographic signals and the joint movement signals to obtain a fusion weight set, and predicting a user expectation instruction; performing multi-target collaborative optimization by combining the observation set and the expected instruction to generate a control parameter set; according to the upper limb exoskeleton driving control method and system, the upper limb exoskeleton is subjected to driving control processing through the control instruction, so that the exoskeleton control precision is improved, and the upper limb fine assistance requirement is met.

Description

Technical Field

[0001] This invention relates to the field of upper limb exoskeleton control technology, specifically to an adaptive control method for upper limb exoskeleton based on multimodal data fusion. Background Technology

[0002] With the continuous advancement of robotics, sensor technology, and human-computer interaction technology, upper limb exoskeleton systems have shown broad application prospects in fields such as medical rehabilitation, strength enhancement, and occupational assistance. In order to achieve natural, efficient, and ergonomic human-machine collaborative work, the core key lies in enabling the exoskeleton system to accurately understand the user's intentions and provide compliant assistance accordingly.

[0003] Human upper limb movements are highly flexible and complex. The accurate perception of movement intentions and the adaptive control of exoskeletons are the core of improving the practicality of devices and user experience. In recent years, multimodal data fusion technology has gradually become a research hotspot in this field. Through multi-source information, it is possible to more comprehensively reflect human movement intentions and provide richer input dimensions for exoskeleton control.

[0004] Existing technology, such as the intelligent wearable exoskeleton adaptive auxiliary control method and exoskeleton system disclosed in patent application CN120382467A, involves the following steps: acquiring human walking motion data samples and constructing a human motion model based on a human-machine closed-chain motion model; collecting multimodal data of the target user while wearing the exoskeleton in real time; calling the human motion model and the human-machine closed-chain motion model to generate human-machine motion coordinated exoskeleton control commands based on the multimodal data; and dynamically adjusting the exoskeleton's motion control parameters according to the control commands. This method constructs a precise human motion model based on a human-machine closed-chain motion model, generates human-machine motion coordinated exoskeleton control commands based on the model, and dynamically adjusts the exoskeleton's motion control parameters according to the commands. This can accurately match the differences in body shape, muscle strength, and exercise habits of different users, effectively improving wearing comfort and safety.

[0005] Based on the above findings, the limitations of existing technologies include at least the following issues: Existing technologies do not construct dedicated dynamic models for upper limb movement characteristics, and they do not perform deep fusion and adaptive weighting of multi-source heterogeneous data such as electromyography, joint motion, and motor torque. This results in difficulty in accurately observing human-computer interaction torque, lag and bias in user movement intention recognition, and a disconnect between control commands and the user's actual movement expectations. Furthermore, existing technologies do not generate impedance control sets and auxiliary ratio sets adapted to upper limb movements through multi-objective collaborative optimization, nor do they combine admittance control models to achieve accurate torque mapping. This leads to a mismatch between exoskeleton auxiliary torque and the user's real-time interaction torque and movement needs, resulting in insufficient smoothness of human-computer collaboration. Moreover, they lack refined processing such as loss compensation and hierarchical torque calculation for upper limb movements, resulting in poor control accuracy and individual adaptability, making it difficult to meet the adaptive assistance needs in complex and delicate upper limb movement scenarios. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an adaptive control method for upper limb exoskeleton based on multimodal data fusion. This method solves the problems of existing technologies, such as the lack of a dedicated upper limb model and deep fusion of multiple data, resulting in low control accuracy and poor adaptability, which makes it difficult to meet the needs of fine upper limb assistance.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an adaptive control method for an upper limb exoskeleton based on multimodal data fusion, comprising the following steps: acquiring multimodal data of the upper limb exoskeleton, wherein the multimodal data includes user electromyographic signals, joint motion signals, and time-series data of motor output torque; based on the joint motion signals and the time-series data of motor output torque, performing inverse dynamic recursive processing through a pre-established upper limb exoskeleton inverse dynamic model to generate a human-computer interaction torque observation set for the upper limb exoskeleton; and based on the human-computer interaction torque observation set, controlling the user electromyographic signals... The signal and the joint motion signal are adaptively fused to generate a fusion weight set for the upper limb exoskeleton, and predictive analysis is performed to obtain the user's expected commands for the upper limb exoskeleton. Multi-objective collaborative optimization processing is then performed on the human-computer interaction torque observation set and the user's expected commands to generate a control parameter set for the upper limb exoskeleton, including an impedance control set and an auxiliary ratio set. Based on the control parameter set, the user's expected commands are torque-mapped using an admittance control model to generate joint torque control commands for the upper limb exoskeleton. Finally, the upper limb exoskeleton is driven and controlled based on these joint torque control commands.

[0008] Further, the specific steps for generating the human-computer interaction torque observation set of the upper limb exoskeleton are as follows: Based on the joint motion signal, extract the joint angular acceleration time series set of the upper limb exoskeleton; input the joint motion signal and the joint angular acceleration time series set, along with the structural parameter set stored in the database, into the upper limb exoskeleton inverse dynamic model to analyze the reference driving joint torque time series set of the upper limb exoskeleton; perform loss compensation processing on the motor output torque time series data to generate the actual output torque time series set of the upper limb exoskeleton; perform collaborative integration processing based on the reference driving joint torque time series set and the actual output torque time series set to generate the human-computer interaction torque observation set of the upper limb exoskeleton.

[0009] Further, the specific steps for analyzing the reference driving joint torque time series set of the upper limb exoskeleton are as follows: The joint motion signal, the joint angular acceleration time series set, and the structural parameter set are processed for time synchronization and marked as joint inputs; based on the upper limb exoskeleton inverse dynamic model, the joint inputs are processed for hierarchical torque calculation and collaborative synthesis to generate the joint component torque time series set of the upper limb exoskeleton; the joint component torque time series set is processed for synchronous superposition to generate the reference driving joint torque time series set of the upper limb exoskeleton.

[0010] Further, the specific steps for generating the fusion weight set of the upper limb exoskeleton are as follows: Based on the human-computer interaction torque observation set, construct an interaction torque feature subset of the upper limb exoskeleton; Based on the user's electromyography (EMG) signal and the joint motion signal, extract the EMG activation confidence and joint motion confidence of the upper limb exoskeleton respectively; Perform adaptive allocation processing on the interaction torque feature subset, the EMG activation confidence, and the joint motion confidence to generate an initial weight set of the upper limb exoskeleton; Perform optimization and suppression processing on the initial weight set to generate the fusion weight set of the upper limb exoskeleton.

[0011] Further, the specific steps for generating the initial weight set of the upper limb exoskeleton are as follows: input the subset of interactive torque features into a preset nonlinear mapping model to extract the torque regulation factor set of the upper limb exoskeleton; based on the torque regulation factors, perform collaborative mapping and constraint processing on the electromyographic activation confidence and the joint motion confidence to generate the initial weight set of the upper limb exoskeleton.

[0012] Furthermore, the nonlinear mapping model includes an input layer, a feature fusion layer, and an activation output layer. The specific steps for extracting the torque regulation factor set of the upper limb exoskeleton are as follows: In the input layer, the interactive torque feature subset is received and preprocessed; in the feature fusion layer, attention association processing is performed on the preprocessed interactive torque feature subset to generate a torque aggregation feature vector of the upper limb exoskeleton; in the activation output layer, nonlinear transformation processing is performed on the torque aggregation feature vector to output the torque regulation factor set of the upper limb exoskeleton.

[0013] Further, the specific steps for obtaining the user's expected instructions for the upper limb exoskeleton are as follows: The user's electromyographic signals and joint motion signals are processed to generate a user intention motion feature set for the upper limb exoskeleton; based on the fusion weight set, the user intention motion feature set and the joint motion signals are processed to obtain a fused motion feature set for the upper limb exoskeleton; the fused motion feature set is then subjected to temporal trend fitting and prediction smoothing to obtain the user's expected instructions for the upper limb exoskeleton.

[0014] Further, the specific steps for generating the control parameter set of the upper limb exoskeleton are as follows: read the joint motion signal, and combine the human-machine interaction torque observation set with the user's expected command to extract the human-machine collaboration feature set of the upper limb exoskeleton; perform adaptive analytical optimization and verification processing on the human-machine collaboration feature set to generate the impedance control set of the upper limb exoskeleton; and perform demand adaptation processing on the human-machine collaboration feature set to generate the auxiliary ratio set of the upper limb exoskeleton.

[0015] Further, the specific steps for generating the impedance control set of the upper limb exoskeleton are as follows: set an optimization variable set and construct a multi-objective optimization function in combination with the human-machine collaboration feature set; perform single-objective transformation analysis on the multi-objective optimization function to generate the initial impedance control set of the upper limb exoskeleton; and perform calibration on the initial impedance control set based on preset constraints to generate the impedance control set of the upper limb exoskeleton.

[0016] Further, the specific steps for generating joint torque control commands for the upper limb exoskeleton are as follows: input the user-expected commands and the impedance control set into the admittance control model to extract the initial joint torque control commands for the upper limb exoskeleton; read the human-computer interaction torque observation set and, in conjunction with the auxiliary ratio set, modify the initial joint torque control commands to obtain the joint torque control commands for the upper limb exoskeleton.

[0017] The present invention has the following beneficial effects: (1) The adaptive control method of upper limb exoskeleton based on multimodal data fusion can accurately obtain the human-computer interaction torque observation set by collaborative acquisition and processing of electromyographic multimodal data and combined with the upper limb exoskeleton inverse dynamic model. This improves the recognition and observation ability of human-computer interaction state. When generating the human-computer interaction torque observation set, the joint angular acceleration time series set is extracted from the joint motion signal. Combined with the exoskeleton structural parameters, the layered torque calculation and synchronous superposition are completed through the inverse dynamic model to obtain the theoretical driving joint torque. At the same time, the loss compensation processing of the motor output torque is performed to obtain the actual output torque. The two are then integrated in a coordinated manner, which not only considers the dynamic characteristics of the exoskeleton itself, but also corrects the torque loss caused by motor transmission, friction, etc., so that the observed human-computer interaction torque is closer to the real interaction state. This can effectively avoid the control deviation caused by inaccurate torque observation, make the exoskeleton more sensitive to the user's upper limb movement resistance and assistance needs, and thus quickly capture the real state of human-computer interaction, making the movement of the exoskeleton and the user's upper limb more closely matched.

[0018] (2) The adaptive control method of upper limb exoskeleton based on multimodal data fusion adaptively fuses electromyographic signals and joint motion signals through human-computer interaction torque observation set. By constructing an interactive torque feature subset, extracting electromyographic activation confidence and joint motion confidence, and then extracting torque regulation factors through a nonlinear mapping model, the adaptive allocation and optimization suppression of weights are completed to form a fusion weight set. This effectively avoids the inaccurate intention recognition caused by simple splicing of multi-source data. When the user's expected instruction is obtained, the intention is first calculated on the electromyographic and joint motion signals, and then the fusion weight set is combined to complete modal collaboration and generate a fusion motion feature set. Finally, through time-series trend fitting and prediction smoothing, the expected instruction that fits the user's real motion intention is obtained, so that the intention recognition is more in line with the real-time interaction scenario, and the auxiliary movements of the exoskeleton are kept in sync with the user's motion ideas, thereby improving the smoothness of human-computer collaboration.

[0019] (3) The adaptive control method of upper limb exoskeleton based on multimodal data fusion generates an impedance control set and an auxiliary ratio set through multi-objective collaborative optimization, and then completes torque mapping and correction by combining the admittance control model to form a precise joint torque control command. This achieves fine-grained control of upper limb exoskeleton control parameters. When generating the control parameter set, the human-machine collaboration feature set is extracted first. By constructing a multi-objective optimization function, single-objective transformation analysis and constraint calibration, an impedance control set adapted to the current interaction state is obtained. At the same time, the corresponding auxiliary ratio set is generated in combination with the user's movement needs. This takes into account both the stability of human-machine interaction and the adaptation of auxiliary force. When generating joint torque control commands, the user's expected command is first combined with the impedance control set through the admittance control model to obtain the initial command. Then, it is corrected based on the human-machine interaction torque observation set and the auxiliary ratio set, so that the final control command not only meets the user's movement expectations, but also matches the real-time interaction torque and auxiliary needs. This allows the exoskeleton to not hinder the user's active movement, but also provide sufficient support, improve control accuracy and individual adaptability, and meet the usage needs of complex and fine upper limb movement scenarios.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0021] Figure 1 This is a flowchart of an adaptive control method for an upper limb exoskeleton based on multimodal data fusion according to the present invention.

[0022] Figure 2 This is a flowchart illustrating the specific steps involved in generating a human-computer interaction torque observation set for an upper limb exoskeleton in an adaptive control method for an upper limb exoskeleton based on multimodal data fusion, as described in this invention.

[0023] Figure 3 This is a flowchart illustrating the specific steps involved in generating the fusion weight set of the upper limb exoskeleton in the adaptive control method for upper limb exoskeleton based on multimodal data fusion of the present invention. Detailed Implementation

[0024] Please see Figure 1This invention provides a technical solution: an adaptive control method for an upper limb exoskeleton based on multimodal data fusion, comprising the following steps: acquiring multimodal data of the upper limb exoskeleton (in the current cycle), the multimodal data including user electromyography signals (collected from the surface of the target muscle group of the user's upper limb), joint motion signals, and time-series data of motor output torque (of the exoskeleton joints); based on the joint motion signals and the time-series data of motor output torque, performing reverse dynamic recursive processing through a pre-established upper limb exoskeleton inverse dynamic model to generate a human-computer interaction torque observation set for the upper limb exoskeleton (in the current cycle) (characterizing the net interaction torque of the user's active action on the exoskeleton drive shaft); based on the human-computer interaction torque observation set, performing adaptive fusion processing on the user electromyography signals and joint motion signals to generate a fusion weight set for the upper limb exoskeleton (in the current cycle), and performing predictive analysis to obtain the user's expected command for the upper limb exoskeleton (in the next cycle); Multi-objective collaborative optimization processing is performed on the human-computer interaction torque observation set and user-expected commands to generate a control parameter set for the upper limb exoskeleton (in the next cycle), including an impedance control set and auxiliary ratios. Based on the control parameter set, the user-expected commands are torque-mapped using an admittance control model to generate joint torque control commands for the upper limb exoskeleton (in the next cycle). Based on the joint torque control commands, the upper limb exoskeleton (in the next cycle) is driven and controlled by sending the torque control commands corresponding to each driven joint directly to the servo driver or motor drive module of the corresponding joint according to the joint number. The servo driver outputs the corresponding drive current to the corresponding drive motor according to the received torque control commands, so that the motor outputs a drive torque that matches the torque control commands, thereby driving the exoskeleton linkage to complete the corresponding movement.

[0025] Specifically, such as Figure 2 As shown, the joint motion signals are the joint angles and angular velocities at each time point collected by the inertial measurement units and joint encoders of each drive joint (such as the shoulder, elbow, and wrist joints) of the upper limb exoskeleton. The motor output torque timing data are the output torques at each time point collected by the torque sensors built into the servo motors of each drive joint of the upper limb exoskeleton. The specific steps for generating the human-computer interaction torque observation set of the upper limb exoskeleton are as follows: Based on joint motion signals, a time series set of joint angular accelerations of the upper limb exoskeleton is extracted. Specifically, the acquired joint motion signals are subjected to low-pass filtering for noise reduction. A second-order Butterworth low-pass filter with a cutoff frequency of 10Hz is selected to eliminate environmental electromagnetic interference, sensor zero-drift errors, and high-frequency noise, retaining the effective time series data of joint angles and joint angular velocities. Based on the discrete time difference method, the joint angular acceleration at each time point of each joint is calculated to form a time series set of joint angular accelerations. The calculation formula is as follows: ; in, For the first The driving joint in the first Joint angular acceleration at each time point For the first The driving joint in the first Joint angular velocity at each time point For the first The driving joint in the first Joint angular velocity at each time point The unified sampling period for multimodal data (i.e., the time interval between two adjacent time points). , To drive the total number of joints, , Given the total number of time points, the joint angular acceleration at each time point of all driven joints is integrated to form the joint angular acceleration time series set; The joint motion signals and joint angular acceleration time series, along with the structural parameter set stored in the database, are input into the upper limb exoskeleton inverse dynamic model to analyze the baseline driving joint torque time series of the upper limb exoskeleton (i.e., the theoretical driving joint torque of each driving joint at each time point). Loss compensation processing is performed on the motor output torque timing data to generate the actual output torque timing set of the upper limb exoskeleton (i.e., the actual output torque of each driven joint at each time point), which is as follows: The compensation items include motor friction loss torque and transmission mechanism clearance loss. The motor friction loss torque is calculated using a combination of Coulomb friction and viscous friction, and the compensation formula is as follows: ; in, For the first The driving joint in the first The total frictional loss torque at each time point The coulomb friction torque of the motor (pre-calibrated and stored in the database). The sign function is 1 when the angular velocity is positive, -1 when it is negative, and 0 when it is zero. The motor viscous friction coefficient is preset (calibrated in advance and stored in the database); the transmission mechanism clearance loss uses a fixed compensation value. (Based on the exoskeleton transmission mechanism model, pre-calibrated), the final formula for calculating the actual output torque at each time point is: ; in, Let be the actual output torque of the i-th drive joint at time k. Let be the output torque of the i-th drive joint at the k-th time point; The actual output torque at each time point of all drive joints is integrated to form the actual output torque timing set; It should be noted that, The calibration process is as follows: The upper limb exoskeleton is placed in a no-load, free-moving state. Position / velocity closed-loop control is turned off, and the drive motor is put into torque output mode. The drive motor is controlled to rotate forward at a very low angular velocity (e.g., 0.1 rad / s) at a constant speed, and the output torque required to maintain the constant speed motion is recorded. After the motion stabilizes, the motor is switched to rotate in reverse at the same angular velocity at a constant speed, and the output torque required to maintain the constant speed motion is recorded again. The absolute values ​​of the torque data recorded during forward and reverse rotation are taken, and the average value is calculated. This average value is the Coulomb friction torque of the motor in the drive joint. , will be calibrated Store in the database; The calibration process is as follows: With the exoskeleton unloaded, control the drive motor to rotate forward and backward sequentially at several different constant angular velocities (e.g., 0.2 rad / s, 0.5 rad / s, 0.8 rad / s, 1.0 rad / s). After the motion stabilizes at each speed range, record the corresponding motor output torque. Plot the angular velocity as the x-axis, and the corresponding motor output torque and Coulomb friction torque. The difference is used as the ordinate. Multiple sets of test data are linearly fitted, and the slope of the fitted line is the motor viscous friction coefficient of the drive joint. The calibration results Store in the database; Based on the collaborative integration of the benchmark drive joint torque time series set and the actual output torque time series set, a human-computer interaction torque observation set for the upper limb exoskeleton is generated. Specifically, the difference between the actual output torque and the theoretical drive joint torque at each time point for each drive joint is calculated to obtain the net user interaction torque of a single drive joint at a single time point. That is, the human-computer interaction torque observation set.

[0026] The specific steps for analyzing the reference drive joint torque timing set of the upper limb exoskeleton are as follows: The joint motion signals, joint angular acceleration time series, and structural parameter set are time-synchronized and marked as joint inputs, specifically as follows: The database accesses the upper limb exoskeleton structural parameter set, which includes the link mass corresponding to each drive joint. The coordinates of the center of mass in the link's own coordinate system (Fixed and unchanging, pre-calibrated and stored, representing the fixed coordinates from the rotation axis to the center of mass in the link's own coordinate system), the moment of inertia of the link about the joint's rotation axis. The data can be directly exported from the 3D design model, design drawings, or manufacturer's technical specifications of the exoskeleton mechanical structure and stored in the database. Subsequently, the joint angles of each driving joint at each time point were extracted from the joint motion signals. Joint angular velocity Extract the joint angular acceleration of each driven joint at each time point from the joint angular acceleration time sequence set. As the exoskeleton links rotate around the drive joint, the link orientation changes with the joint angle. Changes in the position coordinates of the link's center of mass in the global coordinate system It will vary with the joint angle Real-time changes require coordinate rotation transformation to change the coordinates of the fixed centroid in the linkage's own coordinate system. The coordinates are converted to real-time centroid coordinates that vary with the joint angle in the global coordinate system. The transformation formula is derived based on the joint rotation matrix. Considering that the joints driven by the upper limb exoskeleton rotate around a single axis (such as the shoulder joint around the coronal axis and the elbow joint around the sagittal axis), the rotation transformation formula is (taking rotation around the z-axis as an example, the rotation axis can be adjusted according to the actual joint rotation direction): ; After the coordinate transformation is completed, the joint motion parameters are converted to the timestamp. , , Transformed real-time centroid position coordinates and structural parameters , Perform timing synchronization alignment to ensure that all parameters of the i-th drive joint match at time point k. After synchronization, integrate the synchronization parameters of all drive joints and all time points and mark them as joint input. ; Based on the inverse dynamics model of the upper limb exoskeleton, layered torque calculation and collaborative synthesis processing are performed on the joint inputs to generate a time series set of joint torque components of the upper limb exoskeleton. Specifically, the inverse dynamics model of the upper limb exoskeleton is a multi-joint inverse dynamics calculation model established based on rigid body dynamics theory. Each link of the exoskeleton is regarded as a rigid body (ignoring link deformation to conform to actual mechanical design). The joint inputs are then processed... Layered torque calculation is performed, calculating the inertial torque, Coriolis force and centrifugal torque, and gravitational torque of each driving joint at each time point. The three-layer calculation process is carried out independently and synchronously in time. Inertial moment calculation: Based on rotational inertia in the combined input and joint angular acceleration The formula for the moment of inertia of a rigid body rotation is used to solve it. The formula is as follows: ; in, For the first The driving joint in the first The inertial torque at a given time point reflects the inertial resistance torque generated by angular acceleration when the connecting rod rotates; Solution of Coriolis force and centrifugal torque: Based on the moment of inertia in the joint input and joint angular velocity A combined calculation model of centrifugal torque and Coriolis torque is adopted, and the formula is as follows: ; in, For the first The driving joint in the first The sum of the Coriolis force and the centrifugal torque at each time point For the first The Coriolis force coefficients of each driven joint (derived from the real-time center of mass position coordinates) and moment of inertia It is derived that (the position of the center of mass is updated synchronously with the position of the center of mass at each time point), the first term is the centrifugal torque, and the second term is the Coriolis torque, which together reflect the additional torque generated by the angular velocity when the connecting rod rotates; It should be noted that, The derivation steps are as follows: Using the rotation axis of the i-th driving joint as a reference, a self-coordinate system is established that moves synchronously with the link. The magnitude of the Coriolis torque is equal to the product of the link mass, the joint angular velocity, and the radial velocity of the center of mass, multiplied by 2. To match the linear form of the torque calculation model, we extract the part of the torque expression that is linearly related to the angular velocity as a coefficient, which is the Coriolis force coefficient. This coefficient is determined by the mass of the connecting rod, the real-time distance from the center of mass to the axis of rotation, and the radial velocity of the center of mass. These parameters can all be calculated from the real-time position coordinates of the center of mass. Gravitational moment calculation: Based on the link mass in the combined input Real-time centroid position coordinates and joint angle Combined with the gravitational acceleration g (taken as 9.8 m / s²), 2 The formula for calculating the gravitational torque is: ; in, For the first The driving joint in the first Gravitational torque at each point in time The distance from the rotation axis of the i-th drive joint to the center of mass of the link (based on the coordinates of the fixed center of mass in the link's own coordinate system). Calculations show that (This is a fixed value and is independent of the joint angle). It reflects the influence of joint angles on the direction and magnitude of gravitational torque, while the real-time center of mass position coordinates Used to assist in calibrating the direction of the gravitational torque to ensure consistency with the real-time attitude of the link; After the layered calculation is completed, the inertial torque, Coriolis force, centrifugal torque, and gravitational torque of each driving joint at each time point are synthesized in a coordinated manner and integrated into a time series set of joint component torques. , ; The time series set of joint component torques is synchronously superimposed to generate the benchmark driving joint torque time series set of the upper limb exoskeleton. Specifically, the three types of component torques at each time point of each driving joint are synchronously superimposed to obtain the theoretical driving joint torque of a single driving joint at a single time point. The superposition formula is as follows: ; in, For the first The driving joint in the first The theoretical driving joint torque at each time point reflects the theoretical driving force required to drive the joint when the exoskeleton moves on its own (without considering the user's force). Finally, the theoretical driving joint torques at each time point of all driving joints are integrated to form the benchmark driving joint torque time series set of the upper limb exoskeleton.

[0027] In this implementation scheme, a second-order Butterworth filter is used to denoise the joint motion signal, and then the discrete time-difference method is used to calculate the angular acceleration. This eliminates sensor drift and electromagnetic interference while ensuring the integrity and synchronization of the time-series data for angle, angular velocity, and angular acceleration, avoiding subsequent calculation deviations due to data distortion. In the theoretical torque calculation, a coordinate rotation transformation is first performed to transfer the link's center of mass from its own coordinate system to the global coordinate system. Then, based on rigid body dynamics, the inertial torque, Coriolis and centrifugal torque, and gravitational torque are calculated in layers. Finally, these are superimposed to obtain the theoretical driving joint torque, completely reproducing the exoskeleton's own motion. The required driving force is calculated without omitting any key torque component, ensuring that the theoretical calculation results are more closely aligned with the actual stress conditions of the mechanical structure. Simultaneously, friction and transmission clearance loss compensation is applied to the motor output torque. Both Coulomb friction and viscous friction are calibrated on actual machines. Combined with fixed clearance compensation values, the original torque collected by the sensor is corrected to the true effective torque. Finally, by calculating the difference between the theoretical torque and the actual output torque, the obtained human-computer interaction torque can accurately reflect the real interaction force between the user and the exoskeleton. This eliminates interference from the exoskeleton's own movement and mechanical losses, making the interaction force observation results more reliable.

[0028] Specifically, such as Figure 3As shown, the user's electromyogram signal is the surface electromyogram amplitude at each time point corresponding to each driving joint of the upper limb exoskeleton collected by an electromyogram sensor attached to the surface of the target muscle groups (deltoid muscle, biceps brachii, triceps brachii, wrist flexor muscle, wrist extensor muscle) of the user's upper limb. The specific steps to generate the fusion weight set of the upper limb exoskeleton are as follows: Based on the human-machine interaction torque observation set, construct an interaction torque feature subset of the upper limb exoskeleton, specifically: Take the absolute value of the user's net interaction torque at each time point of each driving joint as the interaction torque amplitude ; Based on the difference between the user's net interaction torque at the current time point k and the previous time point k - 1, divide it by the unified sampling period Calculate point by point to obtain the torque change rate of each driving joint at each time point , representing the instantaneous dynamic change rate of the torque at each time point. When k = 1 and there is no data for the previous time point, this value is taken as 0; Take the kth time point as the window end point, select a time series sliding window with a length of W (the window range is n = k - W + 1 to n = k), and calculate the information entropy of the torque time series distribution within the window point by point to represent the stability of the torque within the window corresponding to each time point (the larger the entropy value, the more stable the torque; the smaller the entropy value, the more fluctuating the torque), and obtain the torque fluctuation entropy of each driving joint at each time point , and the calculation formula is: ; Among them, is the proportion of the absolute value of the user's net interaction torque at the nth time point within the window, and W is the preset length of the time series sliding window (such as 5 - 10 time points); for the initial time points where k < W, calculate using the existing time points (n = 1 to n = k) to ensure that each time point corresponds to a unique entropy value, is the user's net interaction torque of the ith driving joint at the kth time point, is the user's net interaction torque of the ith driving joint at the nth time point within the window; Read the actual output torque of the motor at each time point of each driving joint and perform a ratio processing with the user's net interaction torque at the corresponding time point, that is, user's net interaction torque / actual output torque of the motor, to obtain the proportion of the user's active torque​​​​​​​​​​​​​​​Based on user electromyography (EMG) signals and joint motion signals, the confidence scores for EMG activation and joint motion of the upper limb exoskeleton are extracted, specifically as follows: Electromyography (EMG) activation confidence extraction: The surface EMG amplitude of the EMG signal corresponding to the i-th driving joint at each time point is subjected to temporal low-pass filtering to remove environmental electromagnetic noise, baseline drift, and high-frequency interference, while retaining the effective EMG activation components, resulting in the filtered EMG amplitude time series. Based on the filtered EMG amplitude time series, the EMG activation intensity is extracted point by point using the sliding window root mean square method to characterize the true activation level of the muscle corresponding to the i-th driving joint at the current time point. The sliding window length is a fixed preset parameter (e.g., the length of 5 time points). For time points in the initial stage where the window length has not been fully accumulated, the activation intensity is directly calculated using the current time point and all previously collected effective time point data to ensure that an effective activation intensity value is generated at each time point. Electromyographic activation intensity of each actuating joint at each time point Substitute the preset S-shaped nonlinear function and calculate the confidence level of electromyographic activation point by point. Normalized to the [0, 1] interval, the confidence level of the electromyographic signal at the current time point is represented (the closer the value is to 1, the more effective the electromyographic activation and the lower the noise ratio; the closer it is to 0, the closer the signal is to noise and the less effective the activation). The confidence level is calculated using the following formula: ; in, This is the confidence level adjustment coefficient. The preset electromyographic noise threshold (used to distinguish effective activation from background noise). The acquisition steps are as follows: Subjects were instructed to perform standard upper limb contraction movements (such as elbow flexion and extension). Multiple sets of complete electromyographic (EMG) signal sequences, including resting, mildly activated, moderately activated, and maximally activated states, were collected, based on a pre-defined EMG noise threshold. The collected electromyographic signals were divided into noise regions (intensity less than...). ) and effective activation interval (intensity greater than With the goal of achieving a confidence level close to 0 in the noise interval and a confidence level close to 1 in the effective activation interval, the coefficients are adjusted iteratively. The value of is chosen to maximize the match between the sigmoid function output and the manually labeled activation validity tags. Ultimately, the function converges. The value is the confidence level adjustment coefficient; The preset steps are as follows: Obtain multiple sets of sample data, that is, keep the subject's upper limbs completely relaxed, collect raw electromyography (EMG) signals for a duration of not less than 30 seconds, calculate the root mean square (RMS) value of the collected resting EMG signals, and use this as the baseline intensity of the background noise. Set the EMG noise threshold to 1.5 to 2 times the average value of the background noise RMS value. ; Based on the joint angular velocity of the i-th driving joint at time point k in the joint motion signal Calculate the confidence level of joint motion point by point. And normalized to the interval [0, 1], it represents the effectiveness of the joint movement at the current moment, and the calculation formula is: ; in, The preset joint motion hysteresis threshold is determined as follows: The upper limb exoskeleton is controlled to complete multiple full joint reciprocating movements (such as elbow flexion / extension and shoulder rotation) at a constant angular velocity under no-load conditions. Joint angular velocity signals are simultaneously acquired. From the acquired angular velocity signals, angular velocity data during joint motion direction switching is selected. The fluctuation amplitude and stability deviation of the angular velocity during the switching phase are extracted as the baseline feature of hysteresis noise. The joint motion hysteresis threshold is set to 1.2 to 1.5 times the baseline feature of hysteresis noise. ; An adaptive allocation process is performed on the interaction torque feature subset, electromyographic activation confidence, and joint motion confidence to generate the initial weight set of the upper limb exoskeleton. The initial weight set is optimized and suppressed to generate a fusion weight set for the upper limb exoskeleton, specifically as follows: Based on the initial electromyographic weights of k-1 at the previous time step Initial weights of joint movements Initial weights of electromyography at the current time point k Initial weights of joint movements By performing point-by-point iterative correction, the corrected initial electromyographic weights are obtained. Correcting the initial weights of joint movements The optimized formula is: ; in, The iterative learning rate (value 0 < <1), the acquisition steps are as follows: recruit subjects to complete multiple sets of complete action sequences containing different upper limb movement patterns, simultaneously collect electromyographic signals and joint movement signals, and record the corresponding initial weight sequence as the basic data for subsequent optimization; Select multiple conditions that satisfy 0 < Candidate values ​​for the <1 constraint, such as 0.1, 0.3, 0.5, 0.7, and 0.9, cover different learning rate ranges. Set a convergence threshold (e.g., the absolute value of the weight change at 3 consecutive time points is less than 10-3). For each candidate value, record the number of time points from the initial state of the weight sequence to the point where the condition is met (i.e., the number of convergence steps). The fewer the number of convergence steps, the faster the convergence speed. Fluctuation calculation: Extract the weight values ​​at 10 consecutive time points after convergence from the converged weight sequence, and calculate the standard deviation of this set of values. The smaller the standard deviation, the smaller the fluctuation of the converged weight and the higher the stability. Convergence speed and volatility are used as two evaluation dimensions, with weights of 60% and 40% respectively, to highlight the priority of convergence speed. The convergence steps for each candidate value are converted into a normalized convergence speed score, and the standard deviation of fluctuation is converted into a normalized stability score. The convergence speed score is multiplied by 60% of the weight, and the stability score is multiplied by 40% of the weight to obtain the comprehensive score for each candidate value. The candidate value with the highest overall score is selected as the final iterative learning rate. ; It should be noted that when k=1, the corrected initial weights for electromyography and the corrected initial weights for joint motion are directly taken from the initial weights for electromyography and the initial weights for joint motion. And corrected the initial weights of electromyography. Correcting the initial weights of joint movements Normalization was performed to obtain the electromyographic weights. Joint motion weights That is, the fusion weight set.

[0029] The specific steps for generating the initial weight set of the upper limb exoskeleton are as follows: Input the subset of interactive torque features into a pre-defined nonlinear mapping model to extract the set of torque regulation factors for the upper limb exoskeleton; Based on the torque regulation factor, the confidence scores of electromyographic activation and joint motion are co-mapped and constrained to generate the initial weight set of the upper limb exoskeleton, specifically as follows: Torque control factor generated point by point Using the allocation criterion, confidence levels of electromyographic activation were respectively... Joint motion confidence Perform point-by-point weighted mapping to obtain the initial weights for electromyography. Initial weights of joint movements The mapping formula is: ; Initial weights for electromyography Initial weights of joint movements Normalization is performed by applying point-by-point normalization constraints to the initial weights of the electromyography (EMG) signals and the initial weights of the joint motion signals at the same driving joint and time point. This involves first calculating the sum of the two types of initial weights at the same driving joint and time point, then dividing the initial weights of the EMG signals and the initial weights of the joint motion signals by this sum to obtain the normalized initial weights for a single joint, ensuring that both types of weights are non-negative and that their sum is always 1. Normalized initial weights of electromyography Initial weights of joint movements As the initial weight set.

[0030] The nonlinear mapping model includes an input layer, a feature fusion layer, and an activation output layer. The specific steps for extracting the torque regulation factor set of the upper limb exoskeleton are as follows: In the input layer, a subset of interaction torque features is received and preprocessed, specifically by: processing the interaction torque amplitude of each driving joint at each time point. Torque change rate Torque fluctuation entropy User-initiated torque ratio Perform max-min normalization to map each feature value to the interval [0, 1]; In the feature fusion layer, attention association processing is performed on the preprocessed subset of interactive torque features to generate a torque aggregation feature vector for the upper limb exoskeleton, specifically as follows: Based on pre-trained attention weight matrix For each drive joint at each time point, the normalized interaction torque amplitude is calculated. Torque change rate Torque fluctuation entropy User-initiated torque ratio Calculate the attention weights for each feature class to compute the normalized interaction torque magnitude. Attention weights For example: ; in, For the first Attention calculation of weight vectors for torque-like features. Furthermore, torque features 1, 2, 3, and 4 correspond to the normalized interactive torque amplitude, torque change rate, torque fluctuation entropy, and user-initiated torque proportion attention calculation weight vector, respectively. Calculate the weight vector for the attention of the first type of torque feature (i.e., the normalized interactive torque amplitude); Normalized rate of change of torque Torque fluctuation entropy User-initiated torque ratio Attention weights , , and Logical consistency; Normalized interaction torque amplitude Torque change rate Torque fluctuation entropy User-initiated torque ratio Each with its corresponding attention weight , , , Weighted processing is performed to obtain the aggregated torque value of each driving joint at each time point. To form a torque aggregation feature vector; In the activated output layer, the torque aggregation feature vector is subjected to a nonlinear transformation to output the torque control factor set of the upper limb exoskeleton, which is as follows: Combined with the pre-trained output layer weight matrix (A 4×1 matrix) and the bias vector (A 4×1 matrix) Aggregate the torque values ​​of each driving joint at each time point. After linear mapping, the LeakyReLU activation function is used to perform nonlinear transformation, obtaining the torque control factor of each driving joint at each time point. (torque control factor) To calculate the results point-by-point, each point corresponds to a specific time point and driving joint in the subset of interactive torque features. A larger value indicates a higher weighting priority for the electromyographic signal under the current interactive torque state, and vice versa. This forms the torque regulation factor set, i.e.: ; It should be noted that the attention weight matrix Output layer weight matrix (A 4×1 matrix) and the bias vector (For a 4×1 matrix) The pre-training steps are as follows: During the model training phase, multiple datasets of upper limb interactive torques from different users and under different movement patterns were collected. A subset of interactive torque features (torque amplitude, torque change rate, torque fluctuation entropy, and the proportion of user-initiated torque) were used as input, and a standard torque control factor matching the user's movement intention was used as the label to construct a supervised learning sample set. The mean squared error between the network's output torque control factor and the standard label was used as the loss function, and an adaptive moment estimation algorithm was employed to iteratively optimize the network parameters until the loss function converged, resulting in the trained attention weight matrix. Output layer weight matrix and bias vector .

[0031] Among them, the output layer weight matrix It is a matrix of dimension n×1, where n is the number of drive joints of the upper limb exoskeleton. The i-th row of the output layer weight matrix represents the weighted coefficients of the torque aggregation features corresponding to the i-th driving joint obtained during training; the bias vector... For a vector of dimension n×1, This represents the i-th element of the output layer bias vector, which is the bias compensation value for the i-th driving joint obtained during training.

[0032] The pre-training steps for the nonlinear mapping model are as follows: Training dataset construction: Subjects of different ages and upper limb motor abilities were recruited, and multiple sets of interactive torque data containing different upper limb movement patterns (such as elbow flexion and extension, shoulder rotation, and multi-joint coordinated movement) were collected. The interactive torque amplitude, torque change rate, torque fluctuation entropy, and user active torque ratio of each driving joint at each time point were extracted as input feature subsets. At the same time, standard torque regulation factors were generated as labels through expert annotation and movement intention matching algorithms to construct a supervised learning sample set. Model initialization: attention weight matrix, output layer weight matrix and bias vector. The attention weight matrix contains attention calculation weight vectors corresponding to 4 types of torque features. The dimension of the output layer weight matrix is ​​the number of driven joints × 1. The dimension of the bias vector is the same as that of the output layer weight matrix. The initial values ​​are generated using a random normal distribution. Iterative optimization training: The sample set is divided into a training set and a validation set. The mean square error between the torque control factor output by the network and the standard label is used as the loss function. The adaptive moment estimation algorithm (Adam) is used to perform iterative parameter optimization. In each iteration, a subset of input features is input into the model. The predicted torque control factor is obtained through input layer preprocessing, attention weighting of the feature fusion layer and nonlinear transformation of the activation output layer. The loss value is calculated and backpropagation is used to update the attention weight matrix, output layer weight matrix and bias vector. Convergence determination and parameter storage: Continue iterating until the loss function value on the validation set does not decrease significantly for 10 consecutive iterations. The model is then determined to have converged. The trained attention weight matrix, output layer weight matrix, and bias vector are stored in the model parameter library as fixed parameters of the nonlinear mapping model for subsequent online calculation of the torque control factor set.

[0033] In this implementation scheme, multiple features are extracted from the human-computer interaction torque observation set to comprehensively capture the dynamic state of human-computer interaction and avoid weight allocation bias caused by single features. When extracting confidence, the activation intensity is extracted using the sliding window method after denoising the electromyographic signal. Combined with the sigmoid function and the calibrated noise threshold and adjustment coefficient, the validity of the electromyographic signal is accurately judged. The joint motion confidence is calculated based on angular velocity and hysteresis threshold, which fits the actual joint motion. The nonlinear mapping model fuses torque features through an attention mechanism. The extracted regulatory factors can dynamically match different interaction states. Combined with iterative correction and normalization processing, the initial weight set is optimized. Throughout the process, all parameters are calibrated on actual machines and trained on samples. Iterative learning rate optimization ensures fast weight convergence and small fluctuations, avoiding interference from invalid signals and dynamically adjusting weights according to the real-time interaction state. This allows for higher weights when the electromyographic signal is valid and more weights when the joint motion is more reliable. The final fused weight set fits the actual motion requirements and improves the accuracy of subsequent command generation.

[0034] Specifically, the steps to obtain the user's desired commands for the upper limb exoskeleton are as follows: Intention decoding is performed on the user's electromyographic signals and joint motion signals to generate a user intention motion feature set for the upper limb exoskeleton, which specifically includes: Extracting electromyographic activation intensity based on user electromyographic signals And combined with the preset maximum safe angular velocity of each drive joint (The preset process is as follows: obtain the nominal maximum safe angular velocity of each drive joint from the design technical specifications of the exoskeleton mechanical structure and the joint component specifications provided by the manufacturer, and multiply the nominal value by a safety redundancy factor of 0.7~0.8 as the final preset maximum safe angular velocity), and the motion direction symbol. , Increase the intensity of electromyography activation Mapped to expected joint angular velocity via electromyography ,Right now: ; Electromyographic prediction of joint angular velocity By sampling period By integrating point by point, the expected joint angles of electromyography can be obtained. ,Right now: ; Simultaneously, the expected joint angles based on electromyography... Applying physiological range-of-motion constraints to the upper limbs: For each driven joint, based on the preset upper and lower limits of the corresponding physiological motion angles of the upper limb joints (i.e., based on the standard range of motion of the upper limb anatomical joints, determining the universal upper and lower limits of the degrees of freedom of each joint (shoulder, elbow, wrist); before wearing the exoskeleton, performing static joint motion calibration on the user, collecting the user's own maximum / minimum joint motion angles, and correcting the universal standard to adapt to individual differences; within the calibrated individual range of motion, reserving a safety redundancy of 5°~10°, further narrowing the upper and lower limits to prevent the user's joints from approaching their physiological limits, and finally obtaining the preset upper and lower limits of joint angles), the expected joint angles obtained by point-by-point integration are constrained by the electromyographic values; if the calculated value exceeds the physiological upper limit, it is clamped to the physiological upper limit value; if it is lower than the physiological lower limit, it is clamped to the physiological lower limit value. The same drive joint, at the same time point and By integrating these features, the user intention motion characteristics of the corresponding driven joints at the corresponding time points are obtained, thus obtaining the user intention motion feature set. Based on the fusion weight set, modal co-processing is performed on the user intent motion feature set and joint motion signals to obtain the fusion motion feature set of the upper limb exoskeleton. Specifically, based on the joint motion signals, the joint angle of each driven joint is read at each time point. Joint angular velocity And the expected joint angles of electromyography corresponding to the user's intentional movement feature set. and expected joint angular velocity via electromyography By fusing weights, the corresponding electromyographic weights are concentrated. Joint motion weights By performing comprehensive processing, the fused joint angle of each driven joint at each time point is obtained. , fusion joint angular velocity That is, the fusion of motion feature sets, the specific formula of which is: ; By performing temporal trend fitting and prediction smoothing on the fused motion feature set, the user-expected commands for the upper limb exoskeleton are obtained, specifically as follows: Extract the fusion joint angles and fusion joint angular velocities at all time points within the current cycle to construct the global fusion motion feature sequence for the current cycle; Based on this global feature sequence, a time-series trend fitting method is adopted, which involves sequentially traversing all time points in the current cycle's global fusion motion feature sequence, statistically analyzing the first and last time point feature values ​​of the fusion joint angle and fusion joint angular velocity, calculating the feature difference between the first and last time points, and then dividing the feature difference by the total number of time points in the current cycle to obtain the overall average slope of the fusion joint angle and fusion joint angular velocity over time. Based on the overall average slope, the fusion joint angle and fusion joint angular velocity features at each time point in the current cycle are corrected point by point to make the features at each time point fit the overall trend of change, eliminate local abrupt changes and noise interference, and obtain the trend features of joint angle and joint angular velocity after global fitting in the current cycle. Based on the overall average change slope and the global fitting characteristics at the end of the current cycle, the initial expected joint angle and initial expected joint angular velocity of each drive joint in the next control cycle are obtained by calculating along the time sequence direction. Using the average change amplitude of adjacent time points in the global fitting feature sequence of the current cycle as the constraint benchmark (the average change amplitude is calculated by arithmetic mean, that is, the sum of the feature change amplitudes of all adjacent time points in the current cycle divided by the number of adjacent groups), the initial expected joint angle and initial expected joint angular velocity of the next control cycle are subject to temporal smoothing constraints. That is, the feature change amplitude of every two adjacent time points in the global fitting feature sequence of the current cycle is calculated, and the average value of all adjacent change amplitudes is taken as the constraint threshold. If the difference between the initial expected feature of the next control cycle and the fitting feature at the end of the current cycle exceeds the constraint threshold, the difference is adjusted according to the constraint threshold to make the feature change amplitude of the next cycle consistent with the natural change amplitude in the current cycle and suppress parameter mutations. For the first control cycle, the global fitting features at the end of the current cycle are directly used as the initial expected features for the next control cycle. Then, smoothing is performed according to the aforementioned adjacent change amplitude constraints to obtain the expected joint angles of each drive joint (denoted as...). ), desired joint angular velocity (denoted as ), as the user's expected instruction.

[0035] The specific steps for generating the control parameter set of the upper limb exoskeleton are as follows: Read joint motion signals and, in conjunction with the human-computer interaction torque observation set and the user's expected commands, extract the human-computer collaboration feature set of the upper limb exoskeleton, specifically: Based on the human-computer interaction torque observation set, extract the net user interaction torque of each drive joint at each time point within the current cycle. The average net user interaction torque of each drive joint is obtained by statistical averaging. ; Based on joint motion signals, extract the joint angle at the end of the current cycle. Joint angular velocity , and the desired joint angle Desired joint angular velocity The difference is then processed to obtain the angle tracking deviation of each drive joint. Deviation from angular velocity tracking ,by For example: ; And the average net interactive torque of each drive joint for users. Angle tracking deviation Deviation from angular velocity tracking Normalization (such as min-max normalization) will normalize the average net interaction torque of users. Angle tracking deviation Deviation from angular velocity tracking As a human-machine collaboration feature set; Adaptive analytical optimization and verification processing is performed on the human-machine collaboration feature set to generate the impedance control set of the upper limb exoskeleton; The human-machine collaboration feature set is processed for demand adaptation to generate an auxiliary ratio set for the upper limb exoskeleton, specifically: the average net user interaction torque for each driven joint. Angle tracking deviation Deviation from angular velocity tracking The weighted values ​​are then processed and standardized to map them to a range of 0-1, yielding the auxiliary ratio (i.e., demand intensity, denoted as ) of each driving joint of the upper limb exoskeleton. ), to form an auxiliary ratio set; It should be noted that the steps for obtaining the weight coefficients for each parameter in this weighted processing are as follows: the average net interaction torque, angle tracking deviation, and angular velocity tracking deviation of each drive joint are averaged, the average net interaction torque, the average angle tracking deviation, and the average angular velocity tracking deviation are extracted, and then summed to obtain the required sum value. The above average value is then compared with the required sum value, and the corresponding result is used as the weight coefficient for each parameter.

[0036] The specific steps for generating the impedance control set of the upper limb exoskeleton are as follows: Set up an optimization variable set and, in conjunction with the human-machine collaborative feature set, construct a multi-objective optimization function, specifically: using the stiffness coefficients of each driven joint... Damping coefficient To optimize variables, an optimization variable set is constructed, which is then combined with the average net user interaction torque. Angle tracking deviation Deviation from angular velocity tracking Construct matching error functions respectively (Characterizing the deviation between the theoretical output impedance torque of the exoskeleton and the actual interaction torque of the user), Intent adaptation error function (Deviation between the stiffness coefficient and damping coefficient, which are adaptively adjusted according to the user's force intensity, and the ideal human body reference), force adaptation error function. (Characterizing the deviation between the actual stiffness-damping ratio and the ideal ratio benchmark for human joints), that is: ; in, For the preset first The ideal stiffness-damping ratio for each drive joint is determined by the following preset steps: Multiple subjects were recruited, and biomechanical equipment such as joint stiffness testers and multi-channel electromyography acquisition systems were used to measure the inherent stiffness and damping coefficient of human joints under different movement modes such as slow flexion and extension and fast swing. After collecting multiple sets of data, the average value was taken to obtain the physiological stiffness-damping ratio benchmark value of human joints. Obtain the inherent stiffness and damping parameters of the mechanical structure from the exoskeleton's technical specifications or design documents. Match and correct the exoskeleton's mechanical characteristics with the human physiological baseline values. If the exoskeleton's mechanical damping is too high, appropriately reduce the stiffness ratio to make the overall proportion closer to the natural state. Motion smoothness quantification: By collecting joint angular velocity sequences through a motion tracking system, the root mean square value of the third derivative of the angular velocity (jerk) is calculated. The smaller this value, the smoother the motion.

[0037] Torque tracking accuracy quantification: The actual interactive torque is collected by the interactive torque sensor, and the mean square error between the actual torque and the theoretical output impedance torque is calculated. The smaller the value, the higher the tracking accuracy. Human comfort measurement: The visual analog scale (VAS) was used to have subjects rate their comfort level on a scale of 0-10 for each ratio. At the same time, muscle tension in electromyography was collected as an objective indicator. The higher the score and the lower the muscle tension, the higher the comfort level. The quantitative results of motion smoothness, torque tracking accuracy, and human comfort are assigned weights of 0.3, 0.4, and 0.3 respectively. The comprehensive score of each ratio is calculated, and the ratio with the highest comprehensive score is selected as the ideal stiffness-damping ratio of the drive joint. Based on the above matching error function Intended to adapt error function Force adaptation error function Construct a multi-objective optimization function (It aims to minimize each error), that is: ; in, , , The matching error adjustment coefficient, intent adaptation error adjustment coefficient, and force adaptation error adjustment coefficient stored in the database are obtained in the following steps: Multiple subjects were recruited to complete multiple sets of movement sequences containing different upper limb movement patterns. Complete data such as interaction torque, joint movement trajectory, and electromyography signals were collected simultaneously. Based on these data, matching error, intention adaptation error, and force adaptation error were calculated to construct a benchmark dataset containing the three types of errors. Select multiple groups that satisfy Candidate coefficient combinations, such as (0.4, 0.3, 0.3), (0.3, 0.4, 0.3), (0.3, 0.3, 0.4), etc., cover different error weight allocation scenarios; For each set of candidate coefficients, substitute them into the multi-objective optimization function, calculate the corresponding optimization result, and perform quantitative evaluation from three dimensions: Torque tracking accuracy: The mean square error between the actual interactive torque and the theoretical output impedance torque is calculated. The smaller the value, the higher the accuracy.

[0038] Intent response speed: Calculates the time delay from the activation of electromyographic signals to the corresponding movement generated by the exoskeleton; the smaller the value, the faster the response.

[0039] Stability of stiffness-damping ratio: The standard deviation of the fluctuation of the stiffness-damping ratio is calculated. The smaller the value, the higher the stability.

[0040] The quantification results of torque tracking accuracy, intention response speed, and force adaptation stability are assigned weights of 0.4, 0.3, and 0.3 respectively. The comprehensive score of each candidate coefficient combination is calculated, and the combination with the highest comprehensive score is selected as the final error adjustment coefficient. A single-objective transformation analytical process is performed on the multi-objective optimization function to generate the initial impedance control set for the upper limb exoskeleton, specifically as follows: The preset stiffness-damping ratio is as follows: Multi-objective optimization function Simplify to a single-variable function (soon) Substitute into the original multi-objective optimization function In the middle, all functions that originally contained Replace all items with After the replacement is completed, you will get ), by solving Minimum optimal damping coefficient : ; The optimal stiffness coefficient can be obtained based on the stiffness-damping ratio. ,Right now ; Integrating all drive joints , This forms the initial impedance control set; Based on preset constraints, the initial impedance control set is calibrated to generate the impedance control set for the upper limb exoskeleton, specifically as follows: The preset constraints are: , ; in, , , , The preset minimum stiffness coefficient, maximum stiffness coefficient, damping stiffness coefficient, and damping stiffness coefficient for each drive joint are, in order. These can be preset through the equipment's technical specification or publicly available technical documents such as the EksoGT upper limb exoskeleton and ReWalk Restore. Optimal stiffness coefficients of each drive joint in the initial impedance control set Optimal damping coefficient Each value is compared with a preset constraint. If the value exceeds the constraint range, the boundary value is used; otherwise, the original value is retained. After calibration, the optimal stiffness coefficient of each drive joint is obtained. Optimal damping coefficient This forms an impedance control set.

[0041] In this implementation scheme, when generating the user's desired commands, the expected joint motion parameters are first calculated using electromyography (EMG) signals. Combined with physiological range-of-activity constraints, this prevents exceeding the safe range of joint movement, protecting the user's body. Then, an integrated weight set combines EMG and joint motion signals, undergoes time-series trend fitting and smoothing to eliminate abrupt interference, and can predict motion parameters for the next cycle, ensuring smooth command transitions and avoiding stuttering. Furthermore, when generating the control parameter set, a human-machine collaboration feature set is extracted, comprehensively considering user interaction torque, angle, and angular velocity tracking deviations to fully reflect the human-machine interaction state. A multi-objective optimization function is constructed, combining human physiological stiffness damping ratios and exoskeleton mechanical characteristics. Through single-objective transformation and constraint calibration, a suitable impedance control set is generated. Simultaneously, an auxiliary ratio set is dynamically generated based on the user's force application and tracking deviation. All parameters have undergone real-world testing, calibration, and optimization verification, ensuring control accuracy while adapting to different motion scenarios, allowing the exoskeleton's auxiliary force and impedance characteristics to precisely match the user's real-time needs.

[0042] Specifically, the steps for generating joint torque control commands for the upper limb exoskeleton are as follows: Input the user-expected commands and impedance control set into the admittance control model, and extract the initial joint torque control commands for the upper limb exoskeleton, which are specifically: Based on user-expected commands, extract angle tracking deviation. Deviation from angular velocity tracking The optimal stiffness coefficient of each drive joint in the impedance control set. With optimal damping coefficient The input is fed into the admittance control model to obtain the initial joint torque values ​​of each driven joint. To generate initial joint torque control commands, the admittance control model is specifically represented as follows: ; The initial joint torque control commands are modified by reading the human-computer interaction torque observation set and combining it with the auxiliary ratio set to obtain the joint torque control commands for the upper limb exoskeleton. Specifically: Based on the human-computer interaction torque observation set, the average net interaction torque of each drive joint is extracted. And in combination with the corresponding joint's assist ratio and initial joint torque value Generate joint torque control values ​​for each drive joint of the upper limb exoskeleton. This is to generate joint torque control commands, namely: .

[0043] In this implementation scheme, the user's expected commands and impedance control set are input into the admittance control model. Combined with angle and angular velocity tracking deviations, as well as optimal stiffness and damping coefficients, the initial torque control commands for each drive joint are accurately calculated. This ensures that the initial torque matches the user's expected movement and meets impedance control requirements, avoiding a disconnect between torque output and user intent. Secondly, the initial torque commands are corrected by combining the user's average net interaction torque from the human-computer interaction torque observation set and the adaptation parameters from the auxiliary ratio set. This further calibrates the torque output. The correction process fully considers the user's actual interaction force exertion and real-time assistance needs. For example, when the user exerts strong force, the exoskeleton torque output is reduced by adjusting the auxiliary ratio; when the force exertion is weak, it is appropriately increased, making the torque output more closely match the user's real-time state. This ensures that the torque output of each joint of the exoskeleton is accurate and stable, thereby adapting to various upper limb movement scenarios and improving the user experience.

[0044] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0045] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An adaptive control method for an upper limb exoskeleton based on multimodal data fusion, characterized in that, Includes the following steps: Acquire multimodal data of the upper limb exoskeleton, including user electromyography signals, joint motion signals, and timing data of motor output torque; Based on the joint motion signals and the timing data of the motor output torque, the reverse dynamic recursion processing is performed through the pre-established upper limb exoskeleton inverse dynamic model to generate a human-computer interaction torque observation set of the upper limb exoskeleton; Based on the human-computer interaction torque observation set, the user's electromyography signal and the joint motion signal are adaptively fused to generate a fusion weight set for the upper limb exoskeleton, and predictive analysis is performed to obtain the user's expected commands for the upper limb exoskeleton. The human-computer interaction torque observation set and the user's expected command are subjected to multi-objective collaborative optimization processing to generate a control parameter set for the upper limb exoskeleton, including an impedance control set and an auxiliary ratio set. Based on the control parameter set, the user's expected command is processed by torque mapping through the admittance control model to generate joint torque control command for the upper limb exoskeleton. Based on the joint torque control command, the upper limb exoskeleton is driven and controlled.

2. The adaptive control method for upper limb exoskeleton based on multimodal data fusion according to claim 1, characterized in that, The specific steps for generating the human-computer interaction torque observation set of the upper limb exoskeleton are as follows: Based on the joint motion signals, the time series set of joint angular acceleration of the upper limb exoskeleton is extracted; The joint motion signals and the time series set of joint angular acceleration, along with the set of structural parameters stored in the database, are input into the upper limb exoskeleton inverse dynamic model to analyze the reference driving joint torque time series set of the upper limb exoskeleton. The timing data of the motor output torque is processed for loss compensation to generate the actual output torque timing set of the upper limb exoskeleton; Based on the reference drive joint torque timing set and the actual output torque timing set, a human-computer interaction torque observation set for the upper limb exoskeleton is generated through collaborative integration processing.

3. The adaptive control method for upper limb exoskeleton based on multimodal data fusion according to claim 2, characterized in that, The specific steps for analyzing the reference drive joint torque timing set of the upper limb exoskeleton are as follows: The joint motion signals, the joint angular acceleration time series, and the structural parameter set are subjected to time synchronization processing and marked as joint inputs; Based on the inverse dynamic model of the upper limb exoskeleton, the joint input is subjected to hierarchical torque calculation and collaborative synthesis processing to generate a time series set of joint torque components of the upper limb exoskeleton; The joint component torque timing set is synchronously superimposed to generate the reference drive joint torque timing set of the upper limb exoskeleton.

4. The adaptive control method for upper limb exoskeleton based on multimodal data fusion according to claim 1, characterized in that, The specific steps for generating the fusion weight set of the upper limb exoskeleton are as follows: Based on the aforementioned human-computer interaction torque observation set, a subset of interaction torque features for the upper limb exoskeleton is constructed; Based on the user's electromyography (EMG) signals and the joint motion signals, the EMG activation confidence and joint motion confidence of the upper limb exoskeleton are extracted respectively. The interaction torque feature subset, the electromyographic activation confidence, and the joint motion confidence are adaptively assigned to generate an initial weight set for the upper limb exoskeleton. The initial weight set is optimized and suppressed to generate a fusion weight set for the upper limb exoskeleton.

5. The adaptive control method for upper limb exoskeleton based on multimodal data fusion according to claim 4, characterized in that, The specific steps for generating the initial weight set of the upper limb exoskeleton are as follows: The interaction torque feature subset is input into a preset nonlinear mapping model to extract the torque regulation factor set of the upper limb exoskeleton; Based on the torque control factor, the electromyographic activation confidence and the joint motion confidence are co-mapped and constrained to generate an initial weight set for the upper limb exoskeleton.

6. The adaptive control method for upper limb exoskeleton based on multimodal data fusion according to claim 5, characterized in that, The nonlinear mapping model includes an input layer, a feature fusion layer, and an activation output layer. The specific steps for extracting the torque regulation factor set of the upper limb exoskeleton are as follows: In the input layer, the subset of interactive torque features is received and preprocessed. In the feature fusion layer, attention association processing is performed on the preprocessed interactive torque feature subset to generate a torque aggregation feature vector of the upper limb exoskeleton. In the activated output layer, the torque aggregation feature vector is subjected to nonlinear transformation processing to output the torque control factor set of the upper limb exoskeleton.

7. The adaptive control method for upper limb exoskeleton based on multimodal data fusion according to claim 1, characterized in that, The specific steps to obtain the user's desired instructions for the upper limb exoskeleton are as follows: The user's electromyography signals and joint motion signals are processed to generate a user intention motion feature set for the upper limb exoskeleton. Based on the fusion weight set, modal co-processing is performed on the user intention motion feature set and the joint motion signal to obtain the fusion motion feature set of the upper limb exoskeleton; The fused motion feature set is subjected to time-series trend fitting and prediction smoothing to obtain the user's expected commands for the upper limb exoskeleton.

8. The adaptive control method for upper limb exoskeleton based on multimodal data fusion according to claim 1, characterized in that, The specific steps for generating the control parameter set for the upper limb exoskeleton are as follows: Read the joint motion signals and combine them with the human-computer interaction torque observation set and the user's expected command to extract the human-computer collaboration feature set of the upper limb exoskeleton; Adaptive analytical optimization and verification processing is performed on the human-machine collaboration feature set to generate the impedance control set of the upper limb exoskeleton; The human-machine collaboration feature set is subjected to demand adaptation processing to generate an auxiliary ratio set for the upper limb exoskeleton.

9. The adaptive control method for upper limb exoskeleton based on multimodal data fusion according to claim 8, characterized in that, The specific steps for generating the impedance control set of the upper limb exoskeleton are as follows: Set an optimization variable set and combine it with the human-machine collaboration feature set to construct a multi-objective optimization function; The multi-objective optimization function is subjected to single-objective transformation analysis to generate the initial impedance control set of the upper limb exoskeleton; Based on preset constraints, the initial impedance control set is calibrated to generate the impedance control set of the upper limb exoskeleton.

10. The adaptive control method for an upper limb exoskeleton based on multimodal data fusion according to claim 1, characterized in that, The specific steps for generating joint torque control commands for the upper limb exoskeleton are as follows: The user-expected command and the impedance control set are input into the admittance control model to extract the initial joint torque control command of the upper limb exoskeleton. The initial joint torque control command is modified by reading the human-computer interaction torque observation set and combining it with the auxiliary ratio set to obtain the joint torque control command of the upper limb exoskeleton.

Images