A compound learning control method based on integral of error sign

Abstract

The application discloses a compound learning control method based on error sign integral, and relates to the fields of rehabilitation robots and intelligent control technologies. The method first constructs a dynamic model of a robot, and decouples the model into a slow subsystem and a fast subsystem by using singular perturbation theory. An error sign integral term is introduced into a parameter updating law of the slow subsystem, an improved compound learning adaptive law is constructed, and finally, a moment is output to a driver. Compared with the prior art, the application does not depend on acceleration feedback, can effectively inhibit time-varying non-repetitive disturbances such as sudden spasm of a patient, and improves the safety of human-machine interaction.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation robots and intelligent control technology, specifically to a composite learning control method based on error symbol integration. Background Technology

[0002] With the advent of an aging population, the number of people with limb disabilities caused by diseases such as stroke is rapidly increasing. Upper limb rehabilitation robots can assist patients in repetitive training and promote neural function remodeling. Existing robot control methods based on singular perturbation compound learning can achieve online parameter estimation and have the advantage of low cost. However, in actual rehabilitation training, this method is difficult to adapt to sudden situations such as muscle spasms and drastic fluctuations in active torque.

[0003] The disturbances caused by the aforementioned unforeseen circumstances to the robot system are highly time-varying. Existing composite learning control methods only include proportional error terms in their parameter update laws, lacking an integral compensation mechanism for continuous time-varying disturbances. This easily leads to increased transient tracking errors and makes it difficult to quickly adjust the control output when the patient's actions change abruptly. Furthermore, because patient intentions change frequently and irregularly, directly introducing integral terms into traditional adaptive control can easily cause integral saturation and control overshoot, further deteriorating transient response performance and even endangering patient safety. Therefore, how to construct a composite learning control strategy with time-varying disturbance memory and rapid compensation capabilities without increasing acceleration feedback has become a key technical problem for improving the safety of upper limb rehabilitation robots. Summary of the Invention

[0004] This invention proposes a composite learning control method based on error symbol integration, aiming to suppress time-varying disturbances in patients, improve the robot system's ability to compensate for sudden disturbances, and ensure the safety of rehabilitation training. First, a dynamic model of the robot is constructed. Then, singular perturbation theory is used to decouple the dynamic model into two subsystems: a fast-varying subsystem and a slow-varying subsystem. Combining this with a composite learning adaptive law, an error symbol integration term is introduced into the parameter update law. Finally, the torque of the inner-loop motor is calculated, and the torque is output to the driver. The method is implemented through the following technical solution:

[0005] Step 1: Measure the position, angular velocity, and torque signals of the robot's shoulder joint.

[0006] Step 2: Construct a dynamic model of the robot to provide a theoretical basis for subsequent singular perturbation decoupling and controller design.

[0007] ,

[0008] ,

[0009] ,

[0010] ,

[0011] in, , These are the inertia matrix and the Coriolis matrix, respectively. These are the terms of gravity and friction, respectively. Indicates driving torque. For spring torque, It is the rate of change of torque. These represent the joint position, angular velocity, and angular acceleration, respectively. Let be the angular acceleration of the motor. These are the inertia, damping, and stiffness matrices of the motor, respectively. This indicates the output torque of the motor.

[0012] Step 3: Using singular perturbation theory, the model is decoupled into two subsystems: a fast-varying one and a slow-varying one. The flexible joint system with dual time-scale characteristics is decomposed into slow-varying rigid body dynamics and fast-varying actuator elastic dynamics, simplifying the controller design.

[0013] Step 3.1: Introduce small parameters The equations for the fast variable subsystem are obtained as follows:

[0014] ,

[0015] in, , This is a diagonal matrix containing constant terms related to the actuator stiffness, damping, and small parameters. yes The second derivative of .

[0016] Step 3.2: When the small parameter equals 0, the equations of the slow-varying subsystem are obtained:

[0017] ,

[0018] ,

[0019] in, For the nonlinear terms matrix of the robot, It is the filtering error. The first derivative, Indicates the reference angular acceleration. This is a slowly varying control component.

[0020] Step 4: Introduce an error symbol integral term, construct a composite learning parameter update law, and use historical error symbol information to form memory compensation for time-varying disturbances, thereby enhancing the system's ability to suppress sudden disturbances.

[0021] Step 4.1: Linearize the equations of the slowly varying subsystem to provide a computable mathematical expression for adaptive parameter identification:

[0022] ,

[0023] ,

[0024] ,

[0025] in, To expand the inertia matrix, It is a regression matrix transpose, Indicates the reference angular velocity. Let be the vector of parameters to be estimated. It is the extended inertia matrix of the motor.

[0026] Step 4.2: Calculate the outer ring reference torque This enables the slow-varying subsystem to simultaneously achieve trajectory tracking error convergence and parameter adaptive adjustment:

[0027] ,

[0028] ,

[0029] ,

[0030] in, The tracking error between the desired trajectory and the actual trajectory. It is filtering error. This represents the desired trajectory of the joint. For tracking error The first derivative, It is a positive definite diagonal matrix for controlling the gain. Indicates position control gain. These are the estimated values ​​of the parameter vector to be estimated.

[0031] Step 4.3: Introduce an error sign integral term into the parameter update law to construct the parameter update equation with disturbance memory:

[0032] ,

[0033] ,

[0034] ,

[0035] ,

[0036] ,

[0037] in, It is a specific form of regression matrix. yes The regression matrix after low-pass filtering, yes transpose, , Let these represent the desired angular velocity and desired angular acceleration of the joint, respectively. For torque prediction error, It is the torque obtained after low-pass filtering of the outer loop reference torque. This represents the positive definite learning rate matrix. for transpose, It is the integral generalized prediction error based on historical data windows. , , Indicates the weighting factor. For the error sign integral term, It is a symbolic function. It represents the Hadamardi (or Hadama) stack.

[0038] Step 5: Calculate the inner loop control torque based on the outer loop reference torque, and convert the outer loop reference torque into actual motor drive commands through the fast-change PD control law to achieve rapid tracking of the reference torque by the fast-change subsystem.

[0039] ,

[0040] in, For the outer ring reference torque, It is a positive definite diagonal matrix for controlling the gain. It serves as both the inner loop control torque and the motor output torque.

[0041] Step 6: Output torque to the driver.

[0042] This invention introduces an error sign integral term into the composite learning parameter update law, which can accumulate the sign information of historical errors, provide fast and smooth compensation for sudden disturbances such as patient spasms, and enhance the robot system's ability to suppress time-varying disturbances. Attached Figure Description

[0043] Figure 1 The overall flowchart of an embodiment of the present invention. Detailed Implementation

[0044] Figure 1This is a flowchart of an embodiment. This embodiment provides a composite learning control method based on error symbol integration. The specific process includes: measuring the position, angular velocity, and torque of the robot's shoulder joint; constructing a dynamic model of the robot system; decoupling the model into two subsystems, a fast-varying one and a slow-varying one, using singular perturbation theory; introducing an error symbol integration term to construct a composite learning parameter update law; calculating the inner loop control torque and outputting the torque to the driver; and evaluating the algorithm's anti-interference performance by simulating sudden disturbances.

[0045] Step 1: Measure the position, angular velocity and torque of the shoulder joint of the upper limb rehabilitation robot, and set the training task as shoulder abduction and adduction. Introduce a 15 Nm interference torque during the model training process to simulate shoulder muscle spasm for 1 second.

[0046] Step 2: Construct the robot's dynamic model:

[0047] ,

[0048] ,

[0049] ,

[0050] ,

[0051] in, , These are the inertia matrix and the Coriolis matrix, respectively. These are the terms of gravity and friction, respectively. Indicates driving torque. For spring torque, It is the rate of change of torque. These represent the joint position, angular velocity, and angular acceleration, respectively. Let be the angular acceleration of the motor. These are the inertia matrix and damping matrix of the motor, respectively. This indicates the output torque of the motor. Here is the stiffness matrix.

[0052] Step 3: Decouple the model into two subsystems, fast-varying and slow-varying, using singular perturbation theory.

[0053] Step 3.1: Introduce small parameters The equations for the fast variable subsystem are obtained as follows:

[0054] ,

[0055] in, , This is a diagonal matrix of constant terms related to the actuator stiffness, damping, and small parameters. yes The second derivative of .

[0056] Step 3.2: When the small parameter equals 0, the equations of the slow-varying subsystem are obtained:

[0057] ,

[0058] ,

[0059] in, For the nonlinear terms matrix of the robot, It is the filtering error. The first derivative, Indicates the reference angular acceleration. This is a slowly varying control component.

[0060] Step 4: Introduce the error sign integral term to construct the composite learning parameter update law.

[0061] Step 4.1: Linearize the equations of the slowly varying subsystem:

[0062] ,

[0063] ,

[0064] ,

[0065] in, To expand the inertia matrix, It is a regression matrix transpose, Indicates the reference angular velocity. Let be the parameter vector to be estimated. It is the extended-side inertia matrix of the motor. It is an identity matrix.

[0066] Step 4.2: Calculate the outer ring reference torque :

[0067] ,

[0068] ,

[0069] ,

[0070] in, The tracking error between the desired trajectory and the actual trajectory. It is filtering error. This represents the desired trajectory of the joint. For tracking error The first derivative, It is a positive definite diagonal matrix for controlling the gain. Indicates position control gain. These are the estimated values ​​of the parameter vector to be estimated.

[0071] Step 4.3: Introduce an error sign integral term into the parameter update law to construct the parameter update equation with disturbance memory:

[0072] ,

[0073] ,

[0074] ,

[0075] ,

[0076] ,

[0077] in, It is a specific form of regression matrix. yes The regression matrix after low-pass filtering, yes transpose, , Let these represent the desired angular velocity and desired angular acceleration of the joint, respectively. For torque prediction error, It is the torque obtained after low-pass filtering of the outer loop reference torque. This represents the positive definite learning rate matrix. for transpose, For the error sign integral term, Indicates the weighting factor. It is the integral generalized prediction error based on historical data windows. It is a symbolic function. For Hadama accumulation.

[0078] Step 5: Calculate the inner loop control torque based on the outer loop reference torque:

[0079] ,

[0080] in, , It is a positive definite diagonal matrix for controlling the gain. For the outer ring reference torque, It is the inner loop control torque.

[0081] Step 6: Output torque to the actuator to drive the robot to move.

Claims

1. A composite learning control method based on error sign integral, characterized in that, Includes the following steps: Step 1: Measure the position, angular velocity, and torque of the robot's shoulder joint; Step 2: Construct the robot's dynamic model; Step 3: Decouple the model into two subsystems, a fast-varying one and a slow-varying one, using singular perturbation theory: ， ， in, For the small parameters introduced, It is the inertia matrix of the motor. This represents a diagonal matrix containing constant terms related to the actuator stiffness, damping, and minor parameters. For spring torque, It is the rate of change of torque. express The second derivative, For the nonlinear terms matrix of the robot, It is the filtering error. The first derivative, Indicates the reference angular acceleration. For slowly varying control components, It is the inertia matrix. These represent the joint position, angular velocity, and angular acceleration, respectively. This provides the output torque for the motor. Step 4: Introduce the error sign integral term and construct the composite learning parameter update law, which is implemented in the following steps: Step 4.1: Linearize the dynamic equations of the slowly varying subsystem: ， ， ， in, These are the inertia matrix and the Coriolis matrix, respectively. These are the terms of gravity and friction, respectively. Represents the extended inertia matrix. For the regression matrix transpose, It is the reference angular velocity. This represents the parameter vector to be estimated. This is the extended inertia matrix of the motor. It is the identity matrix; Step 4.2: Calculate the outer ring reference torque : ， ， ， in, The tracking error between the desired trajectory and the actual trajectory. It is filtering error. Indicates the location of the joint. It is the desired trajectory of the joint. For tracking error The first derivative, The positive definite diagonal matrix representing the control gain. For position control gain, It is the estimated value of the parameter vector to be estimated; Step 4.3: Introduce an error sign integral term into the parameter update law to construct the parameter update equation with disturbance memory: ， ， ， ， ， in, It is a specific form of regression matrix. yes The regression matrix after low-pass filtering, yes transpose, , Let these represent the desired angular velocity and desired angular acceleration of the joint, respectively. For torque prediction error, It is the torque obtained after low-pass filtering of the outer loop reference torque. This represents the positive definite learning rate matrix. for transpose, It is the integral generalized prediction error based on historical data windows. , , Indicates the weighting factor. For the error sign integral term, It is a symbolic function. It represents the Hadamardi (or Hadama) stack; Step 5: Calculate the inner loop control torque based on the outer loop reference torque: ， in, For the outer ring reference torque, It is a positive definite diagonal matrix for controlling the gain. This indicates the torque prediction error. It serves as both the inner loop control torque and the motor output torque; Step 6: Output torque to the driver.

Images