Exercise function-based self-adaptive impedance control method and system for outer limb robot

Abstract

The invention discloses an outer limb robot self-adaptive impedance control method and system based on a motor function. According to the method, motion information and EMG / EEG multi-mode signals are collected, a motion intention is predicted through weighted stacking and nonlinear compensation, the weight is updated based on an admittance model, and a planning trajectory is learned iteratively; the motor function index is obtained by decomposing the electromyographic signal through a non-negative matrix, the damping and rigidity are dynamically adjusted through feed-forward-impedance double-loop control, and the system stability is verified through the passivity theory. According to the method, the self-adaptive response of the outer limb robot to the human body motor function change is realized, the accuracy and safety of man-machine interaction are improved, and the method is suitable for outer limb robot application scenes such as rehabilitation training and industrial assistance.

Description

Technical Field

[0001] This invention relates to the field of exolimb robot control technology, specifically to an adaptive impedance control method and system for exolimb robots based on kinematics. Background Technology

[0002] With the aging population and increasing number of patients with neuromuscular injuries, the demand for extremity robots in rehabilitation training and assistive devices is growing. Traditional extremity impedance control has many shortcomings, such as impedance parameters being typically set manually, ignoring the differences in users' motor abilities, and failing to meet personalized adaptation requirements. To address these issues, adaptive impedance controllers for extremities have made progress in several aspects in recent years. However, in the practical application of adaptive impedance controllers, the following technical problems still exist: human body impedance characteristics exhibit significant individual differences and dynamic changes, making it difficult to quickly establish accurate personalized mapping relationships; manual parameter adjustment is difficult; learning-based control methods have application limitations, with some algorithms relying on stringent conditions, making engineering implementation challenging.

[0003] Chinese patent CN111904795A discloses a variable impedance control method for rehabilitation robots that combines trajectory planning. This invention dynamically adjusts the damping and stiffness values ​​of the controller in real time based on the tracking error of the interactive task and the difference between the user's perceived position and the endpoint, while introducing trajectory planning that conforms to the principles of human movement. However, the tuning of the damping and stiffness parameters of this invention relies on multiple preset constants, which are not iterated in actual practice, potentially leading to deviations in the motion results. Furthermore, the motion path for healthy individuals cannot be adapted to all disabled patients, resulting in low personalization.

[0004] Chinese patent CN120791801A discloses a robot variable impedance control method, system, and control device based on model predictive control. This invention constructs an iterative vector based on the current state vector and the control vector, and uses an optimization model to iterate until preset requirements are met under constraints such as the upper and lower bounds of impedance and stiffness parameters, the rate of change, and passivity. However, the objective function of this patent does not consider human motion characteristics and physiological states, which can easily lead to a disconnect between robot motion and human intention in human-robot collaborative scenarios; the iterative optimization process is complex, requiring multiple rounds of calculation per cycle, resulting in insufficient real-time performance and difficulty in adapting to high-frequency response scenarios.

[0005] In summary, in practical applications of industrial production, adaptive impedance controllers suffer from insufficient personalization and complex iterative optimization processes leading to a lack of real-time performance. These shortcomings result in a disconnect between the robot's movement intentions and the human body's movements, insufficient flexibility in adapting to different patients, and difficulty in meeting the demands of high-frequency response scenarios, thus limiting the engineering implementation and widespread adoption of the technology. Summary of the Invention

[0006] To address the technical problems in existing technologies, such as the inability to adjust impedance parameters in real time according to changes in human motor function, the lack of full utilization of historical human motion information leading to a disconnect between robot behavior and human intent, and the lack of a mechanism for real-time detection of functional changes and feedback adjustment of impedance parameters, this invention provides an adaptive impedance control method for extremity robots based on motor function. This invention constructs a motor function index based on muscle co-analysis using non-negative matrix factorization and introduces this index into a feedforward-impedance dual-loop control structure for real-time adjustment of damping and stiffness parameters. Simultaneously, by combining an intent prediction model based on historical motion information with an iterative learning trajectory planning mechanism, the invention achieves an adaptive response of the extremity robot to changes in human motor function while ensuring the passive stability of the system.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides an adaptive impedance control method for an extremity robot based on kinematic function, comprising the following steps: S1: Based on historical motion information, establish a human motion intention perception and prediction model; S2: Based on the admittance model, establish a mapping function between human historical motion information and interaction force, and update the weight vector in real time; S3: Iterative learning is used to plan the movement trajectory of the external limbs to achieve complete tracking within a limited time. S4: Based on nonnegative matrix decomposition, the activation matrix of muscle synergy during the operation is analyzed by electromyography signals to obtain the characteristic index of muscle synergy, namely the motor function index. S5: Based on a dual-loop structure of feedforward control and impedance control, the damping and stiffness parameters are adjusted in real time according to the kinematics index. S6: Stability analysis and verification of closed-loop systems based on passivity theory.

[0008] In step S1, current and historical motion information of the human body is collected to construct a multi-dimensional motion information vector, which is represented as follows: L(t) = [1, X(t), X(t-τ), …, X(t-(h-1)τ)] T ; Where X(t) is the position, velocity and acceleration of the arm end at the current moment, X(t-τ) to X(t-(h-1)τ) are historical motion information, τ represents the time delay parameter, i.e. the sampling interval, and h is the length of the historical window.

[0009] Human movement intention is represented as: x' human (t) = A T L(t) + φ(X(t-τ)); Where, x' human (t) is the derivative of the predicted human movement intention, A T = [a0, a1, …, a h ] are the weighting coefficients for motion information at different times, and φ(X(t-τ)) is a nonlinear compensation term used to correct the prediction error of the linear model.

[0010] In step S2, the admittance model is expressed as: M d (ẍ - ẍ d ) + B d (ẋ - ẋ d ) + K d (x - x d ) = F; Where ẍ, ẋ, and x are the acceleration, velocity, and position output by the exolimb robot, respectively. d , ẋ d x d It refers to the reference acceleration, velocity, and position; F is the externally applied force; M is the reference acceleration, velocity, and position. d B d K d These are the mass matrix, damping matrix, and stiffness matrix, respectively.

[0011] A function J(x) is established based on the admittance model to map the dynamic influence of human historical motion and interaction forces on robot motion. human (t-τ), F), this function consists of the energy of the system under the action of interaction forces and position / velocity errors and the accumulated terms of historical errors: J(x human (t-τ), F) = ½(F - F des ) T M d -1 (F - F des ) + ½B d ||Δẋ|| 2 + ½K d ||Δx|| 2 + γΣ k=0 h-1 ||x human (t-kτ) - x des (t-kτ)|| 2 ; Where, Δẋ = ẋ - ẋ d Δx = x - x d F des γ represents the expected compliance level, and γ represents the historical weight, ranging from 0.01 to 0.1. human(t-kτ) represents the actual motion state of the human body at the historical moment t-kτ.

[0012] To minimize the aforementioned cost function and achieve adaptive convergence in motion intent prediction, the cost function J(x) is adjusted accordingly. human (t-τ), F) Find the weight vector The partial derivatives yield the gradient vector, and an online update law for the weight vector is constructed based on the gradient descent method: ; in, This is the updated weighted coefficient vector at the current time. This is the weighted coefficient vector from the previous historical moment. The adaptive learning rate step size is greater than 0. Let be the gradient of the partial derivative of the cost function with respect to the weighted coefficient vector.

[0013] In step S3, the input signal and output error of the controlled object are used as control signals, and the trajectory of the external limb movement is planned through iterative learning. The reference trajectory acceleration is expressed as: ẍ r (t) = x' human (t) + f(x human (t-τ), F); Where, f(x) human (t-τ), F) represents a function of the influence of human historical information and interaction forces on the trajectory of external limbs, expressed as: f(x human (t-τ), F) = (F - b·ẋ human (t-τ) - k s ·x human (t-τ)) / m; Where F is the external interaction force, b is the damping coefficient, and k is the damping coefficient. s It is the stiffness coefficient, m is the mass, and x is the stiffness coefficient. human (t-τ) represents the historical position, ẋ human (t-τ) is the historical velocity.

[0014] The dynamic equations of the external limbs are: M D q̈ k (t) + B D q̇ k (t) = u k (t) + F k (t) - f ds,k (t); Among them, M D B is the positive definite inertial matrix of the external limbs. D For the positive definite damping matrix of the external limbs, uk (t) is the control input for the k-th iteration, F k (t) represents the human-computer interaction force in the k-th iteration, f ds,k (t) represents the external disturbance force in the k-th iteration.

[0015] The learning law of iterative learning control is expressed as: u k+1 (t) = u k (t) + Γ1e k (t) + Γ2ė k (t) + Γ3ΔF k (t) + Γ4Δx his (t); Where Γ1>0 is the error proportional gain, Γ2>0 is the error differential gain, Γ3>0 is the interaction force error coupling gain, Γ4>0 is the historical motion error coupling gain, and e k (t) = y d (t) - y k (t) represents the output error, ė k (t) represents the rate of change of error, ΔF k (t) represents the interaction force error, Δx his (t) = x k (t) - x k (t-τ) represents the difference between the current position and the historical position.

[0016] In step S4, multi-channel surface electromyography (EMG) signals are acquired during the procedure, and an EMG signal matrix is ​​constructed, which is represented by nonnegative matrix decomposition as follows: V = W c×p C p×t ; Among them, W c×p Let C be the muscle synergy activation matrix, describing the weights of each muscle in p synergistic modules, with dimensions c×p; p×t The corresponding time coefficient matrix describes the activation intensity of each collaborative module over time, with a dimension of p×t.

[0017] Define the motor function index η(t) as the similarity between the current coordination matrix and the baseline coordination matrix: η(t) = (1 / p)Σ i=1 p max j ((W base,i ·W curr,j ) / (||W base,i ||·||W curr,j ||)); Among them, W baseW is the baseline co-weight matrix obtained by NMF decomposition when the patient is in a healthy / fatigue-free state. curr η(t) is the current collaborative weight matrix obtained by real-time acquisition of electromyographic signals and online NMF decomposition. η(t)∈[0,1], the closer it is to 1, the more consistent the current muscle exertion pattern is with the non-fatigue state. When η(t) decreases, it indicates a decline in function.

[0018] In step S5, the adaptive impedance controller employs a dual-loop structure combining feedforward control and impedance control. The joint space dynamic equations of the exolimb robot are: M(q)q̈ + C(q,q̇)q̇ + G(q) + F f (q̇) = τ total + τ h ; Where q, q̇, and q̈ are the joint angle, angular velocity, and angular acceleration vectors, respectively; M(q) is the inertia matrix; C(q,q̇)q̇ represents the Coriolis force and centrifugal force terms; G(q) represents the gravity term; and F... f (q̇) represents the friction term, τ h For the active torque of the human body, τ total The total torque is output to the controller.

[0019] The feedforward control law is: τ ff = M(q)q̈ d + C(q,q̇)q̇ d + G(q) + F f (q̇ d ) + α(1 - η(t))τ̂ h ; Among them, q̇ d ,q̈ d The desired joint trajectory is generated by the intent recognition module, α>0 is an adjustable gain, which amplifies the assist when the function is weaker, τ̂ h This is the estimated torque for human intention.

[0020] Human Intentional Torque τ̂ h The task space intention force F was obtained by collecting EEG signals and using support vector regression. EEG Then, it is mapped to joint space via Jacobi transpose: τ̂ h = J T (q)·F EEG ; Impedance control establishes a dynamic relationship between human-computer interaction force and trajectory deviation in the task space: Among them, ex = x d - x represents the end position error, M d For a fixed target inertia matrix, B d (t), K d (t) represents the adaptive damping and stiffness, and F represents the interaction force measured by the six-dimensional force sensor.

[0021] The impedance parameter is adjusted in real time according to η(t): B d (t) = B0[1 + β(1 - η(t))]; K d (t) = K min + (K max - K min )·η(t); Where B0 is the reference damping value, β is the damping adjustment coefficient, and K min K max These are the lower and upper bounds for stiffness adjustment, respectively.

[0022] The resistance force is mapped to the joint space through the Jacobian transpose, generating a feedback control torque: τ fb = J T (q)·F imp ; The controller ultimately outputs the total control torque: τ total = τ ff + τ fb ; In step S6, the stability of the closed-loop system is verified using energy and passivity theory. The total energy function V(t) of the system is defined as the sum of the kinetic energy of the robotic arm and the potential energy of the virtual impedance: V(t) = ½q̇ T M(q)q̇ + ½e q T K d (t)e q ; Among them, ½q̇ T M(q)q̇ represents the physical kinetic energy of the robotic arm in the joint space, ½e q T K d (t)e q The virtual spring potential energy, e, is established for the impedance controller. q = q d - q represents the tracking error.

[0023] Differentiating the energy function V(t) and simplifying it using the robot's oblique symmetry, the upper bound of the system's rate of energy change is: V̇(t) ≤ -ė q T B d (t)ė q + q̇ T (τ ext + Δ model ) + ½ e q T K̇ d (t) e q ; Because of B d The matrix V̇(t) = B0[1 + β(1 - η(t))] is always positive definite, and the system always provides positive damping, continuously consuming system energy. By selecting a sufficiently large β, it is ensured that the damping dissipation power is always greater than the power term introduced by the stiffness change, thus ensuring that V̇(t) ≤ 0 always holds. The closed-loop system satisfies the passivity condition, guaranteeing the global asymptotic stability and safety in the human-machine collaboration process.

[0024] Secondly, this invention provides an adaptive impedance control system for exoskeletal robots based on kinematics, aiming to solve the technical problem of insufficient dynamic matching between exoskeletal robots and human kinematics, and to improve the accuracy, coordination, and system stability of human-computer interaction. This system is implemented using the aforementioned control method, and through the collaborative work of multiple modules, it completes human motion intention perception, trajectory planning, dynamic control, and stability verification. Specifically, it includes the following functional modules: The motion information acquisition module is used to collect current and historical motion information of the human body and construct a multi-dimensional motion information vector L(t) containing parameters such as position, velocity, and acceleration, providing data support for motion intention prediction. The motion intention prediction module is connected to the motion information acquisition module. It establishes a prediction model x' of the derivative of human motion intention by combining weighted superposition and nonlinear compensation terms. human (t) = A T L(t) + φ(X(t-τ)) enables accurate perception and prediction of human movement intentions; The weight update module is signal-connected to the motion intention prediction module, and constructs a mapping function J(x) between human historical motion information and interaction force based on the admittance model. human (t-τ), F), the gradient vector is obtained by taking the partial derivative of the cost function with respect to the weighted coefficient vector, and the weighted coefficient vector is updated in real time online based on the gradient descent method; The iterative learning trajectory planning module is connected to the motion intention prediction module and the weight update module respectively. It establishes a discrete dynamics system model of the external limbs, maps the predicted human intention and interaction force to the desired trajectory, and updates the control input u according to the learning law, which includes trajectory error, error change rate, interaction force error, and historical motion error. k+1 (t), to achieve precise planning of external limb movement trajectories; The electromyography signal analysis module is used to acquire multi-channel surface electromyography (EMG) signals, analyze the muscle synergy activation matrix during the operation through a non-negative matrix factorization algorithm, calculate the matching similarity between the current synergy matrix and the benchmark synergy matrix, and obtain the motor function index η(t) that characterizes the human movement state. The dual-loop control module, connected to the iterative learning trajectory planning module and the electromyography signal analysis module respectively, includes a feedforward control unit and an impedance control unit. The feedforward control unit predicts the human movement intention based on EMG / EEG multimodal signals and generates a compensation torque τ. ff The impedance control unit dynamically adjusts the damping parameter Bd(t) and stiffness parameter Kd(t) in real time based on the kinematic performance index η(t), generating a feedback torque τ. fb The dual-loop control module outputs a total control torque τ through torque fusion. total = τ ff + τ fb To the external limb robot actuator, to achieve human-machine collaborative control; The stability verification module is used to construct the energy function V(t) of the closed-loop system based on the passive property theory. Through positive definite damping dissipation and stiffness change compensation mechanism, it verifies that the closed-loop system meets the passive property condition and ensures the global asymptotic stability of the system.

[0025] This system achieves adaptive response of extremity robots to changes in human motor function through the organic collaboration of various modules, effectively improving the coordination and control accuracy of human-computer interaction, while ensuring the safety and stability of system operation. It can be widely used in extremity robot-related application scenarios such as rehabilitation training and industrial assistance.

[0026] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) By introducing historical motion information and nonlinear compensation terms, this invention can predict motion trends in a short time. It responds faster than traditional methods based on instantaneous data, significantly reduces prediction errors, and makes human intention prediction more accurate and real-time.

[0027] (2) This invention utilizes muscle synergy characteristics to quantify motor function, so that damping and stiffness no longer depend on fixed parameters, but can dynamically change with fatigue, functional decline and other states. The impedance parameters can be adaptively adjusted in real time with changes in motor function, thereby improving the safety and comfort of rehabilitation training.

[0028] (3) This invention makes full use of historical errors by iteratively learning the control law, improves the convergence speed of repeated actions, can achieve complete tracking within a finite number of iterations, and has low computational cost, improves trajectory tracking accuracy, and meets the requirements of high-frequency control.

[0029] (4) The present invention adopts a double-ring structure combined with feedforward intention torque and impedance feedback, which enables the system to actively compensate for the user's intention, while avoiding excessive assistance or mechanical stiffness, and has stronger human-computer interaction compliance and stability.

[0030] (5) This invention does not require solving large-scale optimization problems, the update law is simple in form, it is suitable for embedded controller implementation, has good engineering implementation potential, and can be widely used in actual rehabilitation training and occupational assistance systems. Attached Figure Description

[0031] Figure 1 This is the overall flowchart of the present invention.

[0032] Figure 2 This is a flowchart of the motion intention prediction method of the present invention.

[0033] Figure 3 This is a flowchart of the weight vector update method of the present invention.

[0034] Figure 4 This is a flowchart of the iterative learning method for motion planning according to the present invention.

[0035] Figure 5 This is a flowchart of the motion energy characterization method of the present invention.

[0036] Figure 6 This is a flowchart of the feedforward control and impedance control method of the present invention.

[0037] Figure 7 This is a flowchart of the stability analysis of the present invention. Detailed Implementation

[0038] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Contents not described in detail in the embodiments of the present invention are prior art known to those skilled in the art. Those skilled in the art can make equivalent substitutions or improvements to the following embodiments without departing from the principles of the present invention, and all such substitutions or improvements should fall within the protection scope of the present invention.

[0039] Example 1 like Figure 1 As shown, an adaptive impedance control method for an exolimb robot based on kinematics includes the following steps: Step S1: Based on historical motion information, establish a human motion intention perception and prediction model.

[0040] like Figure 2 As shown, during external limb manipulation, human movement is continuous and periodic, possessing a certain velocity and acceleration. This step involves measuring current and historical motion information to form a motion information vector L(t), and then weighting it using coefficients A. T The product of these terms is used to represent the intention of human movement.

[0041] Specifically, the three-dimensional position, three-dimensional velocity, and three-dimensional acceleration data of the arm's end effector are acquired using an inertial measurement unit (IMU) or an optical motion capture system. The sampling frequency is set to 100Hz~500Hz, the time delay parameter τ is set to 10ms~50ms, and the history window length h is set to 5~20.

[0042] The motion information vector is represented as: L(t) = [1, X(t), X(t-τ), …, X(t-(h-1)τ)] T ; Where X(t) represents the current position, velocity, and acceleration of the arm's end effector, X(t-τ) to X(t-(h-1)τ) represent historical motion information, τ represents the time delay parameter, i.e., the sampling interval, and h is the length of the historical window. In this embodiment, X(t) is a 9-dimensional vector containing three-axis position (x, y, z) and three-axis velocity (v). x , v y , v z ) and triaxial acceleration (a x , a y , a z ).

[0043] Considering the influence of historical motion information on intention perception, the human motion intention is represented as: x' human (t) = A T L(t) + φ(X(t-τ)); Where, x' human (t) is the derivative of the predicted human movement intention, representing the expected human movement state estimated by the model. A T = [a0, a1, …, a h ] represents the weighting coefficients for motion information at different times, with initial values ​​set to a uniform distribution, i.e., a i = 1 / (h+1). φ(X(t-τ)) is a nonlinear compensation term used to correct the prediction error of the linear model, reflecting the nonlinear influence of historical motion information. The nonlinear compensation term is implemented using radial basis functions (RBF), and its expression is: φ(X(t-τ)) = Σ j=1N w j exp(-||x(t-τ) - c j || 2 / (2σ j 2 )); Where N is the number of RBF centers, ranging from 10 to 50; w j c represents the weight of the j-th basis function; j σ is the center vector of the j-th basis function; j is the width parameter of the j-th basis function, with a value ranging from 0.1 to 1.0.

[0044] Step S2: Based on the admittance model, establish a function of human historical motion information and interaction force, and update the weight vector in real time.

[0045] like Figure 3 As shown, the admittance model describes the dynamic response characteristics of an exolimb robot under external forces, expressed as: M d (ẍ - ẍ d ) + B d (ẋ - ẋ d ) + K d (x - x d ) = F; Where ẍ, ẋ, and x are the acceleration, velocity, and position output by the exolimb robot, respectively. d , ẋ d x d The reference acceleration, velocity, and position are F, which is the externally applied force as input, and M is M. d B d K d These are the mass matrix, damping matrix, and stiffness matrix, respectively. In this embodiment, M... d The values ​​are taken as a diagonal matrix of 5kg to 15kg, B d A diagonal matrix with values ​​ranging from 50 N·s / m to 200 N·s / m, K d The values ​​range from 100N / m to 500N / m for a diagonal matrix.

[0046] A function J(x) is established based on the admittance model to map the dynamic influence of human historical motion and interaction forces on robot motion. human (t-τ), F), mainly consists of the energy of the system under the action of interactive forces and position / velocity errors and the accumulated term of historical errors: J(x human (t-τ), F) = ½(F - F des ) T M d -1(F - F des ) + ½B d ||Δẋ|| 2 + ½K d ||Δx|| 2 + γΣ k=0 h-1 ||x human (t-kτ) - x des (t-kτ)|| 2 ; The physical meanings of each item are as follows: ½(F - F des ) T M d -1 (F - F des The interaction force deviation energy term quantifies the actual interaction force F and the expected compliance force F. des The deviation energy reflects the degree of human-machine conflict; ½B d ||Δẋ|| 2 This is a velocity deviation damping term, used to suppress velocity oscillations during trajectory tracking; ½K d ||Δx|| 2 This is the positional deviation stiffness term, used to maintain trajectory tracking accuracy; γΣ k=0 h-1 ||x human (t-kτ) - x des (t-kτ)|| 2 The historical trajectory error term integrates the cumulative deviation of historical movement intentions to improve robustness. γ is the historical weight, with a value ranging from 0.01 to 0.1.

[0047] Since the reference trajectory in the cost function is obtained by integrating the derivative of the motion intention containing the weighted coefficient vector A, the cost function is essentially an implicit function of the weighted coefficient vector A. For the cost function J(x) human The gradient vector is obtained by taking the partial derivative of (t-τ),F with respect to the weighting coefficient vector A. An online update law is then constructed based on the gradient descent method to achieve adaptive convergence of the human intent model weights. ; in, This is the updated weighted coefficient vector at the current time. This is the weighted coefficient vector from the previous historical moment. The adaptive learning rate step size is greater than 0.

[0048] Step S3: Use iterative learning to plan the movement trajectory of the external limbs to achieve complete tracking within a limited time.

[0049] like Figure 4 As shown, after effectively perceiving the human's movement intention, the system will consider the influence of the person's historical movement information and interactive forces on the trajectory of the external limbs to estimate the person's next movement intention and adjust the trajectory in real time. The reference trajectory acceleration is expressed as: ẍ r (t) = x' human (t) + f(x human (t-τ), F); Where, f(x) human (t-τ), F) represents a function of the influence of human historical information and interaction forces on the trajectory of external limbs: f(x human (t-τ), F) = (F - b·ẋ human (t-τ) - k s ·x human (t-τ)) / m; To simplify the analysis, the effects of gravity and Coriolis force are ignored, and the external limbs are simplified as a second-order linear system with the following dynamic equations: M D q̈ k (t) + B D q̇ k (t) = u k (t) + F k (t) - f ds,k (t); Among them, M D B is the positive definite inertia matrix for the external limbs, with values ​​ranging from 2kg to 10kg; D The external limb positive definite damping matrix has a value ranging from 20 N·s / m to 100 N·s / m; u k (t) is the control input for the kth iteration; F k (t) represents the human-computer interaction force in the k-th iteration; f ds,k (t) represents the external disturbance force in the k-th iteration.

[0050] Combining the state vector x of the kth iteration in iterative learning k (t) = [q k (t), q̇ k (t)] T , where q k (t) represents the position of the external limbs, q̇ k Given velocity (t), the state equation function f is: f(x k (t), uk (t), t) = [q̇ k (t), (1 / M D (-B) D q̇ k (t) + u k (t) + F k (t) - f ds,k (t))] T ; The output equation function g is: g(x) k (t), u k (t), t) = y k (t) = q k (t); The system's output error during the k-th run is: e k (t) = y d (t) - y k (t), where y d (t) represents the desired output trajectory.

[0051] The learning law for iterative learning control is: u k+1 (t) = u k (t) + Γ1e k (t) + Γ2ė k (t) + Γ3ΔF k (t) + Γ4Δx his (t); The physical meaning and typical values ​​of each gain parameter are as follows: Γ1>0 represents the error proportional gain, which controls the strength of the error correction on the input, typically ranging from 0.5 to 2.0. Γ2>0 represents the error differential gain, which suppresses overshoot of the external limb trajectory; a typical value is 0.1~0.5. Γ3>0 represents the interaction force error coupling gain, which quantifies the impact of the interaction force on the input update; its typical value is 0.05~0.2. Γ4>0 represents the historical motion error coupling gain, which is compensated using the changing trend of historical trajectory data, with a typical value of 0.02~0.1.

[0052] ė k (t) = de k (t) / dt is the time derivative of the output error, reflecting the rate of change of the error. ΔF k (t) represents the interaction force error. Δx his (t) = x k (t) - x k(t-τ) represents the difference between the current position and the position at a past time.

[0053] Step S4: Based on nonnegative matrix decomposition, analyze the muscle synergy activation matrix during the operation using electromyographic signals to obtain the muscle synergy characteristic index, i.e., the motor function index.

[0054] like Figure 5 As shown, multi-channel surface electromyography (EMG) signals are acquired during the acquisition process. In this embodiment, an 8- to 16-channel surface EMG sensor is used, with a sampling frequency of 1000 Hz to 2000 Hz. After bandpass filtering (20 Hz to 450 Hz), noise reduction, and full-wave rectification of the raw signal, an EMG signal matrix V is constructed.

[0055] The electromyography signal matrix is ​​decomposed using nonnegative matrix decomposition: V = W c×p C p×t ; Among them, W c×p The muscle synergy activation matrix describes the weight of each muscle in p synergistic modules, with dimensions c×p (number of muscles × number of synergies); C p×t The corresponding time coefficient matrix describes the activation intensity of each collaborative module over time, with dimensions p×t (number of collaborative modules × time point). In this embodiment, the number of collaborative modules p ranges from 3 to 6.

[0056] To quantify the human motor function index η(t) in real time, standard movements are first performed with the patient (user) in a healthy / fatigue-free state. The baseline co-weight matrix W is then obtained through NMF decomposition. base Then, during the actual use of the external limb robot by the patient, electromyography (EMG) signals were collected in real time and online NMF decomposition was performed to obtain the current collaborative weight matrix W. curr (t).

[0057] When the body is fatigued or engages in compensatory exercise, muscle activation patterns change, leading to W curr (t) and W base The structures produce differences. The kinetic performance index η(t) is defined as the similarity between the current coordination matrix and the baseline coordination matrix: η(t) = (1 / p)Σ i=1 p max j ((W base,i ·W curr,j ) / (||W base,i ||·||W curr,j ||)); For each co-operation vector W in the benchmark matrix base,iFind the vector W that is most similar to it in the current matrix. curr,j Calculate the cosine similarity between the two; average the similarity of all collaborative modules to obtain the overall functional index η(t). η(t)∈[0,1], the closer it is to 1, the more consistent the current muscle exertion pattern is with the non-fatigue state; when η(t) decreases, it indicates a decline in function, requiring increased auxiliary force and reduced stiffness to protect the user.

[0058] Step S5: Based on the dual-loop structure of feedforward control and impedance control, the damping and stiffness parameters are adjusted in real time according to the kinematic performance index.

[0059] like Figure 6 As shown, the adaptive impedance controller of this invention adopts a dual-loop structure of feedforward control and impedance control. The feedforward loop predicts human motion intention based on EMG / EEG multimodal signals and generates a compensation torque; the impedance loop dynamically adjusts the target damping B according to the kinetic energy index η(t). d (t) and stiffness K d (t) enables compliant human-computer interaction. The two work together to form a closed-loop control system of perception, decision-making, and execution.

[0060] The joint space dynamic equation of an exolimb robot is: M(q)q̈ + C(q,q̇)q̇ + G(q) + F f (q̇) = τ total + τ h ; Where q, q̇, and q̈ are the joint angle, angular velocity, and angular acceleration vectors, respectively; M(q) is the positive definite symmetric inertia matrix; C(q,q̇)q̇ represents the Coriolis force and centrifugal force terms; G(q) represents the gravity term; F f (q̇) represents the friction term, using the Coulomb friction plus viscous friction model; τ h The active torque of the human body; τ total The total torque is output to the controller.

[0061] (1) Feedforward control The feedforward control law is: τ ff = M(q)q̈ d + C(q,q̇)q̇ d + G(q) + F f (q̇ d ) + α(1 - η(t))τ̂ h ; Among them, q̇ d ,q̈ dThe desired joint trajectory is generated by the intent recognition module; α>0 is an adjustable gain, typically ranging from 0.5 to 2.0, amplifying the assist when the function is weaker; τ̂ h This is the estimated torque for human intention.

[0062] Human Intentional Torque τ̂ h The calculation steps are as follows: acquire EEG signals and use support vector regression (SVR) to obtain the task space intention force F. EEG (Unit: N), the SVR output is normalized intention force intensity or direction information, the amplitude of which is mapped to the task space equivalent through system calibration, and then mapped to the joint space through Jacobian transpose: τ̂ h = J T (q)·F EEG ; Feedforward control can compensate before the patient makes a noticeable movement, significantly reducing system lag.

[0063] (2) Impedance control Establish a dynamic relationship between human-computer interaction force and trajectory deviation in the task space: Among them, e x = x d - x represents the end position error; M d The inertial matrix of the fixed target is taken as 5kg~15kg; B d (t), K d (t) represents adaptive damping and stiffness; F represents the interaction force measured by the six-dimensional force sensor.

[0064] The impedance parameter is adjusted in real time according to η(t): B d (t) = B0[1 + β(1 - η(t))]; K d (t) = K min + (K max - K min )·η(t); Where B0 is the reference damping value, typically ranging from 50 N·s / m to 100 N·s / m; β is the damping adjustment coefficient, typically ranging from 0.5 to 2.0; K min K max These are the lower and upper limits for stiffness adjustment, with typical values ​​of 50 N / m and 500 N / m, respectively.

[0065] The resistance force can be solved from the dynamic relationship between human-computer interaction force and trajectory deviation: This resistance force is mapped to the joint space via the Jacobian transpose, generating a feedback control torque: τ fb = J T (q)·F imp ; (3) General control law The final output total control torque of the controller is the sum of the feedforward torque and the feedback torque: τ total = τ ff + τ fb ; Step S6: Perform stability analysis and verification of the closed-loop system based on the passivity theory.

[0066] like Figure 7 As shown, to verify the effectiveness and safety of the dual-loop control system proposed in this invention, the stability of the closed-loop system is proven using the energy and passivity theory.

[0067] (1) Construction of energy function of closed-loop system The exoskeleton robot and the human body are considered as a coupled dynamic system. The total energy function V(t) of the system is defined as the sum of the kinetic energy of the robotic arm and the virtual impedance potential energy: V(t) = ½q̇ T M(q)q̇ + ½e q T K d (t)e q ; Among them, ½q̇ T M(q)q̇ represents the physical kinetic energy of the robotic arm in the joint space; ½e q T K d (t)e q The virtual spring potential energy, e, is established for the impedance controller. q = q d - q represents the tracking error; K d (t) is the positive definite stiffness matrix that is adjusted in real time according to the functional index η(t).

[0068] (2) Proof of system passivity Taking the derivative of the energy function V(t), we get: V̇(t) = q̇ T M(q)q̈ + ½q̇ T Ṁ(q)q̇ + e q T K d (t)ė q + ½ e q TK̇ d (t)e q ; The dynamic equation M(q)q̈ + C(q,q̇)q̇ + G(q) = τ total + τ h Substitute the values ​​and utilize the robot's oblique symmetry q̇ T Simplify by using (Ṁ(q) - 2C(q,q̇))q̇ = 0.

[0069] Considering the adaptive damping term τ fb,damping = -B d (t)ė q The upper bound of the system's rate of energy change can be derived as follows: V̇(t) ≤ -ė q T B d (t)ė q + q̇ T (τ ext + Δ model ) + ½ e q T K̇ d (t) e q ; (3) Core arguments for stability From the energy derivative formula, we know that the term -ė q T B d (t)ė q This represents the damping power dissipation of the system. According to the adaptive damping control law B... d (t) = B0[1 + β(1 - η(t))], where B0 > 0, β > 0, and the functional index η(t)∈[0,1], which guarantees that B d (t) is always a positive definite matrix, which means that no matter how the human body’s functional state changes, the controller always provides positive damping to the system, continuously consuming the system’s energy, thereby suppressing oscillations and improving the system’s stability.

[0070] Due to stiffness parameter K d (t) As time changes, the energy derivative will include a stiffness rate of change term ½e. q T K̇ d (t)e q When stiffness increases, this factor may inject energy into the system, thus adversely affecting system stability. To counteract this potential instability, this invention introduces a fatigue compensation coefficient β into the damping adjustment law. Since the fatigue process of human muscles has a slow-varying characteristic, i.e., the rate of change of η(t) is relatively small, by selecting a sufficiently large β, the damping dissipation power ė can be guaranteed.q T B d (t)ė q It is always greater than ½e, the power term introduced by the change in stiffness. q T K̇ d (t)e q This ensures, in a physical sense, that the total energy change rate of the system always satisfies V̇(t) ≤ 0, thus guaranteeing the passivity of the closed-loop system.

[0071] The iterative learning control (ILC) introduced in step S3 gradually compensates for the model's uncertainty error Δ by continuously learning from repetitive periodic actions. model As the number of iterations k increases, the perturbation term q̇ T Δ model The value gradually approaches zero, enabling the system to converge precisely to the equilibrium point under steady-state conditions.

[0072] In summary, the control system proposed in this invention satisfies the passivity condition through a positive definite damping dissipation and stiffness variation compensation mechanism. This means that during human-machine physical interaction, the system always exhibits energy dissipation characteristics, fundamentally ensuring the system's global asymptotic stability and safety.

[0073] Example 2 This embodiment provides a specific implementation method in a rehabilitation training application scenario.

[0074] For upper limb rehabilitation training scenarios, the exolimb robot is a 7-DOF collaborative robotic arm equipped with a 6-DOF force / torque sensor (range ±100N / ±10N·m, accuracy 0.1N / 0.01N·m). It employs a 16-channel surface electromyography sensor (Delsys Trigno, sampling rate 2000Hz) and an 8-channel electroencephalogram (EEG) acquisition system (Emotiv EPOC+, sampling rate 256Hz).

[0075] The control system runs on an embedded real-time controller with a control cycle of 1ms. Specific parameter settings are as follows: Motion intention prediction module: sampling frequency 200Hz, time delay τ=20ms, historical window length h=10, number of RBF centers N=20; Admittance model parameters: M d =diag(10,10,10)kg, initial B d =diag(100,100,100) N·s / m, initial K d =diag(200,200,200)N / m; Iterative learning control of gain: Γ1=1.0, Γ2=0.3, Γ3=0.1, Γ4=0.05; Muscle synergy analysis: Number of synergistic modules p=4; Impedance parameter adjustment: B0 = 80 N·s / m, β = 1.5, K min =80N / m, K max =400N / m, α=1.2.

[0076] Experimental results show that, compared with traditional fixed impedance control, the trajectory tracking error is reduced by about 65% after 5 iterations using the method of this invention; when the user experiences fatigue (η(t) decreases from 1.0 to 0.6), the damping automatically increases from 80 N·s / m to 128 N·s / m, and the stiffness decreases from 400 N / m to 272 N / m, effectively protecting the user's safety; the lead time for predicting human intentions reaches about 150 ms, significantly improving the system response speed.

[0077] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the inventive concept of the present invention and the contents of the specification and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An adaptive impedance control method for an exolimb robot based on kinematics, characterized in that: Includes the following steps: S1: Based on historical motion information, establish a human motion intention perception and prediction model; collect current and historical human motion information, construct a multi-dimensional motion information vector L(t), and obtain a prediction model for the derivative of human motion intention through weighted superposition and nonlinear compensation terms: x' human (t) = A T L(t) + φ(X(t-τ)); Where, x' human (t) is the derivative of the predicted human movement intention, A T Let X(t) be the weighted coefficient vector, L(t) be the motion information vector, φ(X(t-τ)) be the nonlinear compensation term, and τ be the time delay parameter. S2: Based on the admittance model, establish a mapping function J(x) between human historical motion information and interaction force. human (t-τ), F), the gradient is obtained by taking the partial derivative of the cost function with respect to the weighted coefficient vector, and the weighted coefficient vector is updated online in real time based on the gradient descent method; S3: Iterative learning is used to plan the movement trajectory of the external limbs; a discrete dynamic system model of the external limbs is established, mapping the predicted human intention and interaction force to the desired trajectory. The control input u is updated through iterative learning control based on the learning law composed of trajectory error, error change rate, interaction force error, and historical motion error. k+1 (t); S4: Based on nonnegative matrix decomposition, the activation matrix V of muscle synergy during the operation is analyzed by electromyography signals to obtain the motor function index η(t); S5: Employs a dual-loop structure combining feedforward control and impedance control; the feedforward loop predicts human motion intent based on EMG / EEG multimodal signals and generates a compensating torque τ. ff The impedance loop adjusts the damping B in real time according to the kinematic performance index η(t). d (t) and stiffness K d (t), generating feedback torque τ fb The controller outputs a total torque τ total = τ ff + τ fb ; S6: Construct the energy function V(t) of the closed-loop system based on the passive theory, and perform stability analysis and verification of the closed-loop system.

2. The method according to claim 1, characterized in that: In step S1, the motion information vector L(t) is represented as: L(t) = [1, X(t), X(t-τ), …, X(t-(h-1)τ)] T ; Where X(t) represents the position, velocity, and acceleration of the arm's end at the current moment, X(t-τ) to X(t-(h-1)τ) represent historical motion information, τ is the time delay parameter, i.e., the sampling interval, and h is the length of the historical window; The weighting coefficient vector A T = [a0, a1, …, a h ] represents the weighting coefficients for motion information at different times; the nonlinear compensation term φ(X(t-τ)) is used to correct the prediction error of the linear model and reflects the nonlinear influence of historical motion information.

3. The method according to claim 2, characterized in that: In step S2, the admittance model is expressed as: M d (ẍ - ẍ d ) + B d (ẋ - ẋ d ) + K d (x - x d ) = F; Where ẍ, ẋ, and x represent the acceleration, velocity, and position output by the exolimb robot, respectively. d , ẋ d x d For reference acceleration, velocity, and position, F is the externally applied force, and M is the reference value. d B d K d These are the mass matrix, damping matrix, and stiffness matrix, respectively. The mapping function J(x) human (t-τ), F) consists of the interaction force deviation energy term, velocity deviation damping term, position deviation stiffness term, and historical trajectory error term: J(x human (t-τ), F) = ½(F - F des ) T M d -1 (F - F des ) + ½B d ||Dẋ|| 2 + ½K d ||Δx|| 2 + gS k=0 h-1 ||x human (t-kτ) - x des (t-kτ)|| 2 ; Among them, F des γ represents the expected compliance force, and γ represents the historical weight, with a value ranging from 0.01 to 0.

1. The weighted coefficient vector is updated in real time based on the gradient descent method, and its update law is: ； in, This is the updated weighted coefficient vector at the current time. This is the weighted coefficient vector from the previous historical moment. An adaptive learning rate step size greater than 0, Let be the gradient of the partial derivative of the cost function with respect to the weighted coefficient vector.

4. The method according to claim 3, characterized in that: In step S3, the dynamic equation of the external limb is: M D q̈ k (t) + B D q̇ k (t) = u k (t) + F k (t) - f ds,k (t); Among them, M D B is the positive definite inertial matrix of the external limbs. D For the positive definite damping matrix of the external limbs, u k (t) is the control input for the k-th iteration, F k (t) represents the human-computer interaction force in the k-th iteration, f ds,k (t) represents the external disturbance force during the k-th iteration; The learning law of the iterative learning control is: u k+1 (t) = u k (t) + Γ1e k (t) + Γ2ė k (t) + Γ3ΔF k (t) + Γ4Δx his (t); Where Γ1>0 is the error proportional gain, Γ2>0 is the error differential gain, Γ3>0 is the interaction force error coupling gain, Γ4>0 is the historical motion error coupling gain, and e k (t) represents the output error, ė k (t) represents the rate of change of error, ΔF k (t) represents the interaction force error, Δx his (t) represents the difference between the current position and the historical position.

5. The method according to claim 4, characterized in that: In step S4, nonnegative matrix decomposition is used to decompose the multi-channel surface electromyography (EMG) signal: V = W c×p C p×t ; Among them, W c×p Let C be the muscle synergy activation matrix, describing the weights of each muscle in the p synergistic modules. p×t This is the corresponding time coefficient matrix; The motor function index η(t) is the current coordination matrix W. curr (t) and the benchmark co-operation matrix W base Matching similarity: η(t) = (1 / p)Σ i=1 p max j ((W base,i ·W curr,j ) / (||W base,i ||·||W curr,j ||)); Where η(t)∈[0,1], the benchmark coordination matrix W base The collaborative weight matrix is ​​obtained through non-negative matrix decomposition when the user is in a healthy / fatigue-free state.

6. The method according to claim 5, characterized in that: In step S5, the joint space dynamic equation of the exophoric robot is: M(q)q̈ + C(q,q̇)q̇ + G(q) + F f (q̇) = τ total + τ h ; Where q, q̇, and q̈ are the joint angle, angular velocity, and angular acceleration vectors, respectively; M(q) is the inertia matrix; C(q,q̇)q̇ represents the Coriolis force and centrifugal force terms; G(q) represents the gravity term; and F... f (q̇) represents the friction term, τ h The active torque of the human body; The feedforward control law is: τ ff = M(q)q̈ d + C(q,q̇)q̇ d + G(q) + F f (q̇ d ) + α(1 - η(t))τ̂ h ; Among them, q̇ d ,q̈ d For the desired joint trajectory, α>0 is an adjustable gain, τ̂ h This is the estimated torque for human intention.

7. The method according to claim 6, characterized in that: The estimated human intention torque τ̂ h Obtained through the following methods: EEG signals were collected and support vector regression was used to obtain the task space intention force F. EEG Mapped to joint space via Jacobi transpose: τ̂ h = J T (q)·F EEG ; Among them, J T (q) is the transpose of the Jacobian matrix.

8. The method according to claim 7, characterized in that: The impedance control establishes a dynamic relationship between human-computer interaction force and trajectory deviation in the task space: in, e x = x d - x represents the end position error, M d The inertia matrix of the fixed target is given, and F represents the interaction force. The damping B d (t) and stiffness K d (t) is adjusted in real time according to the kinetic performance index η(t): B d (t) = B0[1 + β(1 - η(t))]; K d (t) = K min + (K max - K min )·η(t); Where B0 is the reference damping value, β is the damping adjustment coefficient, and K min K max These are the lower and upper bounds for stiffness adjustment, respectively. The feedback torque τ fb Through resistance force F imp Obtained by mapping to joint space via Jacobi transpose: τ fb = J T (q)·F imp .

9. The method according to claim 8, characterized in that: In step S6, the total energy function V(t) of the system is the sum of the kinetic energy of the robotic arm and the virtual impedance potential energy: V(t) = ½q̇ T M(q)q̇ + ½ e q T K d (t) e q ; Among them, ½q̇ T M(q)q̇ is the physical kinetic energy of the robotic arm in the joint space, ½ e q T K d (t) e q The virtual spring potential energy, e, is established for the impedance controller. q = q d - q represents the tracking error; By employing a positive definite damping dissipation and stiffness change compensation mechanism, the system energy change rate V̇(t) ≤ 0 is guaranteed to always hold, satisfying the passivity condition and ensuring the global asymptotic stability of the closed-loop system.

10. An adaptive impedance control system for an exolimb robot based on kinematics, characterized in that: The system is implemented using any one of the control methods described in claims 1-9, including: The motion information acquisition module is used to collect current and historical motion information of the human body and construct a multidimensional motion information vector L(t); The motion intention prediction module, connected to the motion information acquisition module, is used to obtain a prediction model x' of the derivative of human motion intention by combining weighted superposition and nonlinear compensation terms. human (t) = A T L(t) + φ(X(t-τ)); The weight update module, connected to the motion intention prediction module, is used to establish a mapping function J(x) based on the admittance model. human (t-τ), F), the gradient vector is obtained by taking the partial derivative of the cost function with respect to the weighted coefficient vector, and the weighted coefficient vector is updated in real time online based on the gradient descent method; The iterative learning trajectory planning module, connected to the motion intention prediction module and the weight update module, is used to update the control input u according to the learning law of iterative learning control. k+1 (t), to realize the trajectory planning of external limb movements; The electromyography signal analysis module is used to acquire multi-channel surface electromyography (EMG) signals and obtain the motor function index η(t) through non-negative matrix decomposition. The dual-loop control module, connected to the iterative learning trajectory planning module and the electromyography signal analysis module, includes a feedforward control unit and an impedance control unit. The feedforward control unit is used to predict human movement intentions based on EMG / EEG multimodal signals and generate a compensation torque τ. ff The impedance control unit is used to adjust the damping B in real time according to the kinematic performance index η(t). d (t) and stiffness K d (t), generating feedback torque τ fb The dual-loop control module outputs a total control torque τ. total = τ ff + τ fb External limb robot actuator; The stability verification module is used to construct the energy function V(t) of the closed-loop system based on the passivity theory and verify that the closed-loop system satisfies the passivity condition.

Images