An adaptive impedance control method for exoskeleton based on multi-point human-computer interaction parameter optimization

Abstract

The present invention belongs to the field of robot system control technology, specifically, an exoskeleton adaptive impedance control method based on multi-point human-machine interaction parameter optimization. The method is based on a single-point expectation model and a single-point interaction environment dynamics model, and designs an impedance index function and optimization problem for multi-point human-machine interaction to obtain the optimal impedance model of the exoskeleton joint space. By introducing process variables, the original impedance control problem is converted into a nonlinear system optimal control problem. By designing an intelligent learning control method based on an evaluation neural network, the optimal control performance of the exoskeleton system is guaranteed. By designing a weight parameter update algorithm based on dynamic regression extension and hybrid technology, the transient performance of the neural network parameter learning process is improved, and the requirements for traditional continuous excitation conditions are relaxed, and finally the adaptability and flexibility to the wearer in the human-machine interaction process are achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of robot system control, and in particular is an exoskeleton adaptive impedance control method based on multi-point human-machine interaction parameter optimization. Background Art

[0002] As a wearable device, exoskeleton robots can provide strength support and motion assistance to the wearer. They are widely used in limb motor function rehabilitation training scenarios and play an important role in the field of intelligent rehabilitation.

[0003] When the wearer uses the exoskeleton for rehabilitation training, there is a high degree of tight coupling between the patient and the robot's multiple joints, which leads to multi-point human-machine physical interaction problems, posing challenges to the coordinated control of human-machine motion and the effectiveness of rehabilitation training. Existing studies on impedance control methods usually only focus on the single interaction point of the robot's end effector, without modeling multi-point human-machine interaction and corresponding collaborative control design. At the same time, the impedance parameters of some traditional exoskeleton impedance control are usually set manually, ignoring the differences in the motor abilities of different patients, and failing to meet the personalized adaptation of patient needs.

[0004] In response to the above situation, in recent years, some learning-based intelligent control technologies have been widely studied to improve the adaptability of controllers to the environment. However, the weight update strategy of neural networks mostly uses the gradient descent method, which relies on harsh persistent excitation (PE) conditions and is not conducive to its implementation in engineering applications. Summary of the invention

[0005] In view of this, the present invention provides an exoskeleton adaptive impedance control method based on multi-point human-computer interaction parameter optimization to solve the technical problems existing in the above-mentioned prior art.

[0006] In order to solve the above problems, the present invention adopts the following technical solutions:

[0007] An exoskeleton adaptive impedance control method based on multi-point human-computer interaction parameter optimization comprises the following steps:

[0008] Step 1, establishing a single-point interactive environment dynamics model for the robot based on a dynamics model, and a single-point expected impedance model; the dynamics model is a flexible exoskeleton system dynamics model established based on the Euler-Lagrange model, and the single-point interactive environment dynamics model is an environmental dynamics model that uses a linear time-invariant (LTI) system to describe the external environment;

[0009] Step 2: Based on the single-point interaction environment dynamics model and the expectation model, the optimal impedance parameter problem of the original robot system is converted into a linear quadratic optimal regulator (LQR) problem and solved to obtain the optimal impedance model for the multi-point human-machine interaction problem;

[0010] Step 3: Based on the optimal impedance model obtained in step 2 and the actual motion control information of the robot, process variables are defined, and the process variables are introduced into the robot dynamics model to convert the impedance control problem of the original robot system into an optimal control problem of the nonlinear system;

[0011] Step 4: Set a cost function for system performance evaluation, solve the cost function using Hamilton function and Bellman optimality principle, and obtain the Hamilton-Jacobi-Bellman equation (HJB) for solving the optimal control problem of nonlinear systems;

[0012] Step 5. Design an intelligent learning controller based on the critic neural network, use the dynamic regression extension and hybrid (DREM) method as the weight convergence strategy of the critic neural network, solve the Hamilton-Jacobi-Bellman (HJB) equation proposed in step 3, and thus complete the adaptive impedance control of the multi-point interactive exoskeleton.

[0013] Furthermore, the implementation method of step 1 is:

[0014] Step 1.1, establishing a single-point expected impedance model based on the dynamic model and the single-point interactive environment dynamic model for the robot;

[0015] The dynamic model of the robot system is:

[0016]

[0017] Where q∈R n represents the angle of the robot in the joint space, represents the angular velocity of the robot in the joint space, The angular acceleration of the robot in the joint space, n represents the robot's degree of freedom, M(q)∈R n×n represents the inertia matrix, represents the centrifugal Coriolis force matrix, G(q)∈R n×n represents the gravity matrix, τ∈R n is the control input torque, τ d ∈R n is the external torque;

[0018] The single-point interactive environment dynamics model for the robot is:

[0019]

[0020] In Cartesian space, i represents the i-th point: where M e,i ∈R n×n represents the inertia of the environment, Ce,i ∈R n×n Represents damping, G e,i ∈R n×n Represents the stiffness impedance value, x i ∈R n is the actual position of the point, is the velocity at that point, is the acceleration of the point, and M e,i ∈R n×n ,C e,i ∈R n×n ,G e,i ∈R n×n are the inertia, damping and stiffness impedance values ​​of the human-computer interaction environment, f i is the interaction force between the point and the environment.

[0021] The single-point expected impedance model is established as:

[0022]

[0023] Among them, M d,i ∈R n×n represents the expected inertia, C d,i ∈R n×n represents the expected damping, G d,i ∈R n×n Represents the desired stiffness impedance value.

[0024] Furthermore, the implementation method of step 2 is:

[0025] Step 2.1: Convert the optimal impedance parameter problem of the original robot system into a linear quadratic optimal regulator (LQR) problem:

[0026] Assume M d,i It has been given, is the state vector containing the position and velocity information of the point, For i Derivatives of, non-inertial forces According to the single-point expected impedance model and the single-point interactive environment dynamics model, the following formula is obtained:

[0027]

[0028] in, M d,i It is a parameter set by humans; I n It represents the n-th order identity matrix;

[0029] definition The indicators that need to be optimized are as follows:

[0030]

[0031] Among them, t0 is the current time, R i , Q i are constant coefficients set and are all positive definite matrices

[0032] f i ' as the "control input" of the single-point interactive environmental dynamics model to transform the optimal impedance parameter problem into a linear quadratic optimal regulator (LQR) problem.

[0033] Step 2.2: Solve the LQR problem:

[0034] The simplified form of the Riccati equation is:

[0035] The optimal solution of the LQR problem is obtained by using the Riccati equation as follows:

[0036]

[0037] Among them, P i For the equation The solution; P i,3 With P i,4 P i A portion of a block matrix;

[0038] Will Substitution In, we get:

[0039]

[0040] According to the non-inertial force formula Get the optimal solution C for each impedance parameter of a single point d,i :

[0041]

[0042] So far, the single-point optimal impedance model of each equivalent point of contact between the robot and the human is obtained.

[0043] Step 2.3: Convert the single-point optimal impedance model of each equivalent point to the joint space and solve the over-coupling problem:

[0044] The optimal impedance of a single point in the joint space is solved using the inverse kinematics formula:

[0045]

[0046] Among them, φ i (q) is the kinematic solution of the robot point i, and J i (q) is the Jacobian matrix;

[0047] In order to solve the problem of multi-point impedance over-coupling in the joint space, the optimization index of the multi-interaction point system is defined as:

[0048]

[0049] Among them, J(M), J(C), and J(G) are the optimization indicators of inertia, damping, and stiffness values ​​for solving the over-coupling problem in the joint space, respectively; a i is the importance of the i-th point, if a i The larger it is, the closer the optimal impedance of the joint space is to the ideal impedance of the i-th point in the joint space;

[0050] According to the optimal impedance of a single point in the joint space, solve J(M), J(C) and J(G) to obtain the optimal impedance parameters in the joint space. Thus, the ideal impedance model of the joint space point of the robot system is obtained That is, the optimal impedance model for multi-point human-computer interaction problems.

[0051] Furthermore, the implementation method of step 3 is:

[0052] Step 3.1. Error formula between the robot system and the ideal impedance model for:

[0053]

[0054] Define process variables z and q r ;

[0055]

[0056] Among them, Λ and Γ are both set auxiliary matrices, and Λ and Γ satisfy the conditions: τ l is the filtered torque, τ l The calculation formula is:

[0057] Step 3.2: Treat process variables as errors Introduced into the robot dynamics model, the following model formula is obtained:

[0058]

[0059] where q r is the introduced process variable, and its derivative is The second-order derivative is By excess variables z and q r Calculated;

[0060] Step 3.3, design the controller τ, which is expressed as follows:

[0061]

[0062] Among them, K en ∈R n×n is a positive definite matrix, u l ∈R n is the feedforward learning control quantity;

[0063] The controller τ is introduced into the model obtained in step 2.2, and an adaptive dynamic programming method is used to change u l (t) makes z track 0, and obtains the nonlinear system for the process variable

[0064]

[0065] in, is the derivative of z, g(z)=-M -1 (q).

[0066] Furthermore, the implementation method of step 4 is:

[0067] Step 4.1, construct the cost function for evaluation as shown below:

[0068]

[0069] in is the utility function, A is a positive definite matrix, and T represents the transpose.

[0070] Step 4.2: Construct the HJB equation:

[0071] definition is the optimal value function;

[0072] Then the Hamilton function is:

[0073]

[0074] in, represents the partial derivative of the cost function with respect to state z, is the symbol of partial derivative;

[0075] set up and

[0076] The Hamilton function is solved by using the Bellman optimality principle, and the Hamilton-Jacobi-Bellman equation as a nonlinear optimal controller is obtained:

[0077]

[0078] in, is the optimal control feedforward learning amount defined.

[0079] Furthermore, the implementation method of step 5 is as follows:

[0080] Step 51, design an evaluation neural network, and use the neural network to approximate the optimal cost function online to obtain an approximate optimal control strategy:

[0081] The expression for evaluating the neural network is:

[0082]

[0083] Among them, W c is the ideal weight of the evaluation neural network, φ(z) is the nonlinear part of the designed evaluation neural network, and ε(z) is the error of the evaluation neural network;

[0084] Since the ideal weight W c It is impossible to calculate accurately, so the following formula is used to fit V * (z);

[0085]

[0086] in, is the approximate value of the optimal value function, is the true weight of the neural network; the approximate optimal control strategy is defined as follows:

[0087]

[0088] Step 5.2: Based on the approximate optimal control strategy, a dynamic regression extension and hybrid (DREM) method is constructed as a weight convergence strategy for evaluating the neural network, thereby solving the HJB equation and completing the multi-point interactive exoskeleton adaptive impedance control. The steps include:

[0089] Transform the Hamilton function to: Θ = -Π T W c -ε HJB ;

[0090] in and The fitting error is defined as is the partial derivative of φ with respect to Z;

[0091] Introducing the Hamiltonian operator, The formula is transformed into the following formula:

[0092] Y c =-Φ c W c-Ξ c

[0093] Among them, Y c satisfy And Φ c satisfy where β c is a positive constant, Ξ c By an ε HJB The filter transformation term and an exponential decay term;

[0094] exist Multiply both sides by the adjoint matrix adj{Φ c}, transformed into:

[0095]

[0096] in, For c The filtering transformation, y c = adj{Φ c}Y c ∈R l , M c = adj{Φ c}Φ c = det{Φ c}I l =m c I l ∈R l×l For decoupling operation, where m c = det{Φ c}∈R,I l It is an l-order identity matrix; since M c The diagonal property of , so we get the scalar regression formula as shown below:

[0097]

[0098] where y c(i) , W c(i) , Respectively represent y c , W c and The elements of the i-th row;

[0099] According to the scalar regression formula, the weight iteration strategy is as follows:

[0100]

[0101] Among them, γ1 is the weight learning rate, γ 1(i) is the i-th row number of γ1. and They are and The number of rows i;

[0102] According to the weight updating strategy, the HJB equation can be solved, thereby realizing adaptive impedance control of multi-point interactive exoskeleton.

[0103] After adopting the above technical solution, the present invention has the following advantages:

[0104] 1. Unlike many traditional impedance control studies that only focus on the end effector of the robot arm, the present invention takes into account the problem of multi-point human-machine interaction, and incorporates the state information of multiple human-machine interaction points into the adaptive control process to solve the optimal solution of impedance parameters, which is more suitable for lower limb exoskeleton robots with highly human-machine coupling.

[0105] 2. Taking into account the different motor abilities of different patients, the present invention describes the dynamic model of the human-machine interaction environment using the LTI system, so that it can adaptively adjust the impedance parameters of the robot to meet the personalized needs of different patients.

[0106] 3. Since the continuous excitation condition is difficult to meet in actual engineering, in the process of evaluating the learning of neural network parameters, the present invention designs a weight convergence criterion based on the DREM method, relaxes the continuous excitation condition, and improves the transient performance of the process, which is suitable for engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] Figure 1 This is a control block diagram of the exoskeleton adaptive impedance control method based on multi-point human-machine interaction parameter optimization;

[0108] Figure 2 A specific flow chart for implementing the exoskeleton adaptive impedance control method based on multi-point human-computer interaction parameter optimization. DETAILED DESCRIPTION

[0109] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0110] Example 1

[0111] like Figure 1-Figure 2 As shown, an exoskeleton adaptive impedance control method based on multi-point human-computer interaction parameter optimization includes the following steps:

[0112] Step 1: Establish a robot dynamics model, a single-point interactive environment dynamics model for the robot, and a single-point expected impedance model. The robot dynamics model of this embodiment is a flexible exoskeleton system dynamics model based on the Euler-Lagrangian model, and the single-point interactive environment dynamics model for the robot is an environmental dynamics model that uses a linear time-invariant (LTI) system to describe the external environment.

[0113] The robot dynamics model is:

[0114]

[0115] Where q∈R n represents the angle of the robot in the joint space, represents the angular velocity of the robot in the joint space, The angular acceleration of the robot in the joint space, n represents the robot's degree of freedom, M(q)∈R n×n represents the inertia matrix, represents the centrifugal Coriolis force matrix, G(q)∈R n×n represents the gravity matrix, τ∈R n is the control input torque, τ d ∈R n is the external torque.

[0116] The single-point interactive environment dynamics model for the robot is:

[0117]

[0118] In Cartesian space, i represents the i-th point: where M e,i ∈R n×n represents the inertia of the environment, C e,i ∈R n×n Represents damping, G e,i ∈R n×n Represents the stiffness impedance value, x i ∈R n is the actual position of the point, is the velocity at that point, is the acceleration of the point, and M e,i ∈R n×n ,C e,i ∈R n×n ,G e,i ∈R n×n are the inertia, damping and stiffness impedance values ​​of the human-computer interaction environment, f i is the interaction force between the point and the environment.

[0119] The single point expected impedance model is:

[0120]

[0121] M d,i ∈R n×n represents the expected inertia, C d,i ∈R n×n represents the expected damping, G d,i ∈R n×n Represents the desired stiffness impedance value.

[0122] Step 2: Based on the single-point expectation model and the single-point interaction environment dynamics model, the optimal impedance problem of the original robot system is converted into a linear quadratic optimal regulator (LQR) problem and solved to obtain the optimal impedance model for the multi-point human-machine interaction problem. The implementation method is as follows:

[0123] Step 2.1: Convert the optimal impedance parameter problem of the original robot system into a linear quadratic optimal regulator (LQR) problem:

[0124] Assume M d,i It has been given, is the state vector containing the position and velocity information of the point, For i Derivatives of, non-inertial forces According to equation (2) and equation (3), we can get equation (4):

[0125]

[0126] in, M d,i It is a parameter set by humans; I n It represents the n-th order identity matrix;

[0127] definition The indicators that need to be optimized are as follows:

[0128]

[0129] Among them, t0 is the current time, R i , Q i are all positive definite matrices, and Q i ≥0, R i ≥0;

[0130] f i 'As the "control input" of the single-point interactive environmental dynamics model, the original optimal impedance parameter problem can be converted into an LQR problem.

[0131] Step 2.2: Solve the LQR problem:

[0132] The simplified form of the Riccati equation is:

[0133] The optimal solution of the LQR problem is obtained by using the Riccati equation as follows:

[0134]

[0135] Among them, P i For the equation The solution;

[0136] P i Divided into the following four parts:

[0137]

[0138] Substituting (8) into (7), we obtain:

[0139]

[0140] According to the non-inertial force formula Get the optimal solution C for each impedance parameter of a single point d,i :

[0141]

[0142] So far, the optimal impedance model of each equivalent contact point between the flexible exoskeleton system and the human body is obtained.

[0143] Step 2.3: Convert the single-point optimal impedance model of each equivalent point to the joint space and solve the over-coupling problem:

[0144] The optimal impedance of a single point in the joint space is solved by using the inverse kinematics formula. That is, if the i-th point is to satisfy the optimal impedance obtained in equation (6), the impedance parameter that the single point needs to satisfy in the joint space should be:

[0145]

[0146] In formula (11), φ i (q) is the kinematic solution of the robot point i, and J i (q) is the Jacobian matrix.

[0147] In order to solve the problem of multi-point impedance over-coupling in the joint space, the optimization index of the multi-interaction point system is defined as:

[0148]

[0149] In formula (12), J(M), J(C), and J(G) are the optimization indicators of inertia, damping, and stiffness values ​​for solving the over-coupling problem in the joint space, respectively; a i is the importance of the i-th point, if a i The larger it is, the closer the optimal impedance of the joint space is to the ideal impedance of the i-th point at the joint;

[0150] According to the optimal impedance of a single point in the joint space, the optimal impedance parameter of the point in the joint space is obtained by solving equation (12):

[0151]

[0152] According to step 2.3, the optimal impedance parameters of each equivalent point in the joint space are calculated one by one, and the ideal impedance model of the entire robot system can be obtained:

[0153]

[0154] Step 3: After defining the process variables and introducing them into the original exoskeleton system, the impedance control problem of the original flexible exoskeleton system is converted into an optimal control problem of a nonlinear system. The implementation method is as follows:

[0155] Step 3.1. Error formula between the robot system and the ideal impedance model for:

[0156]

[0157] Step 3.2, define process variables:

[0158] Select two positive definite matrices Λ and Γ and satisfy the following conditions:

[0159]

[0160] In formula (16), Λ and Γ are both set auxiliary matrices;

[0161] For τ d The filtering is performed as follows:

[0162]

[0163] In formula (17), τ l is the filter torque;

[0164] The final process variables z and q r The conditions for satisfying are:

[0165]

[0166] Step 3.3: Set process variables z and q r As error Introduced into the robot dynamics model, the following model formula is obtained:

[0167]

[0168] In formula (19), for q r The first derivative of for q r The second-order derivative of; From formula (18), we can get

[0169] Step 3.4, design the following controller τ that approaches the ideal impedance system:

[0170]

[0171] Among them, K en ∈R n×n is a positive definite matrix, and u l ∈R n is the feedforward learning control quantity;

[0172] Step 3.5: Substitute equation (20) into equation (19) and use the adaptive dynamic programming method to change u l (t) makes z track 0, and obtains a nonlinear system for the process variable, which is as follows:

[0173]

[0174] in is the derivative of z, g(z)=-M -1 (q).

[0175] Step 3: Set a cost function for system performance evaluation, solve the cost function using Hamilton function and Bellman optimality principle, and obtain the Hamilton-Jacobi-Bellman equation (HJB) for solving the optimal control problem of the nonlinear system. The implementation method includes:

[0176] Step 3.1, construct the cost function for evaluation as shown below:

[0177]

[0178] Among them, Q is a non-negative local cost function, is the utility function, A∈R n×n is a positive definite constant matrix and T represents the transpose.

[0179] Step 3.2 Construct the HJB equation for solving the nonlinear optimal control problem:

[0180] The optimal value function is defined as:

[0181]

[0182] The Hamilton function is defined as:

[0183]

[0184] In formula (24), And V * (0)=0, is the partial derivative sign; then its optimal condition should be:

[0185]

[0186] Step 3.2: Use the Bellman optimality principle to solve the Hamilton function and obtain the Hamilton-Jacobi-Bellman equation (HJB) for solving the nonlinear optimal control problem as follows:

[0187]

[0188] Step 5: Design an intelligent learning controller based on the critic neural network technology, use the dynamic regression extension and hybrid (DREM) method as the weight convergence strategy of the critic neural network, solve the Hamilton-Jacobi-Bellman equation (HJB) proposed in step 3, and complete the adaptive impedance control of the multi-point interactive exoskeleton. The implementation method is as follows:

[0189] Step 5.1: Design an evaluation neural network and use the neural network to approximate the optimal cost function online, thereby obtaining an approximate optimal control strategy. Specifically:

[0190] The expression for evaluating the neural network is:

[0191]

[0192] In formula (27), W c is the ideal weight of the evaluation neural network, φ(z) is the nonlinear part of the designed evaluation neural network, and ε(z) is the error of the evaluation neural network.

[0193] Since the ideal weight W c It is impossible to calculate accurately, so formula (28) is used to fit V * (z);

[0194]

[0195] In formula (28), is the approximate value of the optimal value function, is the real weight of the neural network, and the approximate optimal control strategy is defined as follows:

[0196]

[0197] Step 5.2: construct a dynamic regression extension and hybrid (DREM) method as a weight convergence strategy for evaluating the neural network, thereby solving the HJB equation and completing the multi-point interactive exoskeleton adaptive impedance control. The steps include:

[0198] Rewrite the Hamilton function:

[0199] Θ=-Π T Wc -ε HJB (30)

[0200] in and The fitting error is defined as and That is, the partial derivative of φ with respect to z;

[0201] Introducing the Hamiltonian operator, equation (27) is transformed into:

[0202] Y c =-Φ c W c -Ξ c (31)

[0203] In formula (30), Y c satisfy And Φ c satisfy where β c is a positive constant, and Ξ c By an ε HJB It consists of a filtered transformation term and an exponential decay term.

[0204] Multiply both sides of equation (28) by the adjoint matrix adj{Φ c}, transformed into:

[0205]

[0206] Where: y c = adj{Φ c}Y c ∈R l , M c = adj{Φ c}Φ c = det{Φ c}I l =m c I l ∈R l×l For decoupling operation, where m c = det{Φ c}∈R,I l It is an l-order identity matrix; since M c The diagonal property of , so we get the scalar regression formula as shown in formula (33):

[0207]

[0208] where y c(i) , W c(i) , Respectively represent yc , W c and The elements of the i-th row;

[0209] The weight iteration strategy is defined as follows:

[0210]

[0211] In formula (34), γ1 is the weight learning rate, γ 1(i) is the i-th row number of γ1. and They are and The i-th row number of .

[0212] According to this strategy, the HJB equation can be solved and the adaptive impedance control of the multi-point interactive exoskeleton can be completed.

[0213] In summary, it can be seen that, unlike the traditional impedance control method that only focuses on the end effector, the exoskeleton adaptive impedance control method provided by the present invention takes into account multiple human-machine interaction points and is more suitable for flexible lower limb exoskeleton robots with high human-machine coupling. At the same time, the differences in motor abilities of different patients are also taken into account. According to the single-point interaction environment dynamics model and the robot's state information, the impedance parameters can be adaptively adjusted to meet the personalized needs of patients. In addition, the DREM method is introduced into the parameter learning of the evaluation neural network, which relaxes the continuous excitation conditions, improves the transient performance of the parameter learning process, and effectively improves the performance of the exoskeleton.

[0214] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. An exoskeleton adaptive impedance control method based on multi-point human-computer interaction parameter optimization, comprising the following steps: Step 1: Establish a single-point interactive environment dynamics model for the robot based on a dynamics model, and a single-point expected impedance model; the dynamics model is a flexible exoskeleton system dynamics model established based on the Euler-Lagrange model, and the single-point interactive environment dynamics model is an environmental dynamics model that uses a linear time-invariant system to describe the external environment; Step 2: Based on the single-point interaction environment dynamics model and the expectation model, the optimal impedance parameter problem of the original robot system is converted into a linear quadratic optimal regulator problem and solved to obtain the optimal impedance model for the multi-point human-machine interaction problem; the specific method includes: Step 2.1: Convert the optimal impedance parameter problem of the original robot system into a linear quadratic optimal regulator problem: Assume M d,i It has been given, is the state vector containing the position and velocity information of the point, For i Derivatives of, non-inertial forces According to the single-point expected impedance model and the single-point interactive environment dynamics model, the following formula is obtained: in, It is a parameter set by humans; I n It represents the n-th order identity matrix; definition The indicators that need to be optimized are as follows: Among them, t0 is the current time, R i , Q i are set constant coefficients and are all positive definite matrices; f i 'As the "control input" of the single-point interactive environmental dynamics model, the optimal impedance parameter problem can be converted into a linear quadratic optimal regulator problem; Step 2.2, using the Riccati equation to solve the linear quadratic optimal regulator problem, and obtain the single-point optimal impedance model for each equivalent point where the robot contacts the human; Step 2.3: Use the kinematic inverse solution formula to solve the optimal impedance of each single point in the joint space, convert the single-point optimal impedance model of each equivalent point to the joint space, and solve the over-coupling problem by defining the multi-intersection system optimization index to obtain the optimal impedance model for the multi-point human-computer interaction problem. in represents the optimal impedance parameter of point i in the joint space; a i is the importance of the i-th point, if a i The larger it is, the closer the optimal impedance of the joint space is to the ideal impedance of the i-th point in the joint space; Step 3: Based on the optimal impedance model for the multi-point human-machine interaction problem and the actual motion control information of the robot, the process variables are defined, and the process variables are introduced into the robot dynamics model to convert the impedance control problem for the original robot system into the optimal control problem for the nonlinear system. The specific method includes: Step 3.

1. Error formula between the robot system and the ideal impedance model for: Define process variables z and q r : Among them, Λ and Γ are both set auxiliary matrices, and Λ and Γ satisfy the condition that Λ+Γ=M d -1 C d , τ l is the filtered torque, τ l The calculation formula is: Step 3.2: Introduce process variables into the robot dynamics model and obtain the following model formula: Among them, q r is the introduced process variable, and its derivative is The second-order derivative is By excess variables z and q r Calculated; Step 3.3, design the controller τ, which is expressed as follows: Among them, K en ∈R n×n is a positive definite matrix, u l ∈R n is the feedforward learning control quantity; The controller τ is introduced into the model obtained in step 3.2, and an adaptive dynamic programming method is used to change u l (t) makes z track 0, and obtains the nonlinear system z for the process variable: in, is the derivative of z, g(z)=-M -1 (q); Step 4: Set a cost function for system performance evaluation, solve the cost function using Hamilton function and Bellman optimality principle, and obtain the Hamilton-Jacobi-Bellman equation for solving the optimal control problem of nonlinear systems; specifically, including: Step 4.1, construct the cost function for evaluation as shown below: in is the utility function, A is a positive definite matrix, and T represents the transpose; Step 4.2: Construct the HJB equation: definition is the optimal value function; Then the Hamilton function is: in, represents the partial derivative of the cost function with respect to the state z, is the symbol of partial derivative; set up The Hamilton function is solved using the Bellman optimality principle, and the Hamilton-Jacobi-Bellman equation as a nonlinear optimal controller is obtained: in, is the optimal control feedforward learning amount defined; Step 5: Design an intelligent learning controller based on the evaluation neural network, use the dynamic regression expansion and hybrid method as the weight convergence strategy of the evaluation neural network, solve the Hamilton-Jacobi-Bellman equation proposed in step 4, and complete the adaptive impedance control of the multi-point interactive exoskeleton; specifically include: Step 5.1, design an evaluation neural network, and use the neural network to approximate the optimal cost function online to obtain the approximate optimal control strategy: The expression for evaluating the neural network is: Among them, W c is the ideal weight of the evaluation neural network, φ(z) is the nonlinear part of the designed evaluation neural network, and ε(z) is the error of the evaluation neural network; Since the ideal weight W c It is impossible to calculate accurately, so the following formula is used to fit V * (z); in, is the approximate value of the optimal value function, is the true weight of the neural network; the approximate optimal control strategy is defined as follows: Step 5.2: Based on the approximate optimal control strategy, a dynamic regression extension and hybrid method is constructed as a weight convergence strategy for evaluating the neural network, thereby solving the Hamilton-Jacobi-Bellman equation and completing the multi-point interactive exoskeleton adaptive impedance control. The steps include: Transform the Hamilton function to: Θ = -Π T W c -ε HJB ; in and The fitting error is defined as is the partial derivative of φ with respect to Z; Introducing the Hamiltonian operator, The formula is transformed into the following formula: Y c =-Φ c W c -X c Among them, Y c satisfy And Φ c satisfy where β c is a positive constant, Ξ c By an ε HJB The filter transformation term and an exponential decay term; exist Multiply both sides by the adjoint matrix adj{Φ c }, transformed into: in, For c The filtering transformation, y c = adj{Φ c }Y c ∈R l , M c = adj{Φ c }Φ c = det{Φ c }I l =m c I l ∈R l×l For decoupling operation, where m c = det{Φ c }∈R,I l It is an l-order identity matrix; since M c The diagonal property of , so we get the scalar regression formula as shown below: where y c(i) , W c(i) , Respectively represent y c , W c and The elements of row i; According to the scalar regression formula, the weight iteration strategy is as follows: Among them, γ1 is the weight learning rate, γ 1(i) is the number of the i-th row of γ1; and They are and The i-th row number of .

2. According to claim 1, an exoskeleton adaptive impedance control method based on multi-point human-computer interaction parameter optimization is characterized in that: The dynamic model of the robot system is: Where q∈R n represents the angle of the robot in the joint space, represents the angular velocity of the robot in the joint space, The angular acceleration of the robot in the joint space, n represents the robot's degree of freedom, M(q)∈R n×n represents the inertia matrix, represents the centrifugal Coriolis force matrix, G(q)∈R n×n represents the gravity matrix, τ∈R n is the control input torque, τ d ∈R n is the external torque; The single-point interactive environment dynamics model for the robot is: In Cartesian space, i represents the i-th point: where M e,i ∈R n×n represents the inertia of the environment, C e,i ∈R n×n Represents damping, G e,i ∈R n×n Represents the stiffness impedance value, x i ∈R n is the actual position of the point, is the velocity at that point, is the acceleration of the point, and M e,i ∈R n×n ,C e,i ∈R n×n ,G e,i ∈R n×n are the inertia, damping and stiffness impedance values ​​of the human-computer interaction environment, f i is the interaction force between the point and the environment; The single-point expected impedance model is established as: Among them, M d,i ∈R n×n represents the expected inertia, C d,i ∈R n×n represents the expected damping, G d,i ∈R n×n Represents the desired stiffness impedance value.

Images