A torque estimation method based on dynamic nucleus DMP

Abstract

The application discloses a torque estimation method based on a dynamic kernel DMP, and relates to the technical field of rehabilitation robots and intelligent control. The method first generates an exoskeleton joint angle and a torque signal, and then constructs a dynamics model to solve a theoretical torque; dynamics residual is calculated and smoothed; a dynamic kernel DMP model is then constructed and the weight is trained; finally, the torque of the exoskeleton itself is estimated, and the human-machine interaction torque is calculated through difference operation. Compared with the prior art, the application can effectively adapt to unmodeled dynamic working conditions and multi-source disturbances, improve the accuracy of torque estimation and the robustness under different working conditions.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation robots and intelligent control technology, specifically to a torque estimation method based on dynamic kernel DMP. Background Technology

[0002] Lower limb motor dysfunction is a common sequela of stroke, trauma, and age-related degenerative diseases. Traditional manual rehabilitation training suffers from numerous drawbacks, including uneven training intensity, low standardization, high labor costs, and insufficient repetition, making it difficult to meet the needs of large-scale, refined, and routine rehabilitation training. Lower limb rehabilitation exoskeletons, relying on biomimetic motion mechanisms and intelligent control technology, can provide stable limb support, standardized gait guidance, and precise movement assistance for patients with lower limb dysfunction. They are currently a core intelligent device for assisting patients in rebuilding lower limb motor function and improving rehabilitation efficiency.

[0003] Human-machine interaction torque is a core state parameter for achieving compliant assistance in lower limb rehabilitation exoskeletons. Exoskeleton systems require precise calculation of human-machine interaction torque to match personalized assistance torques and control strategies. Dynamic Motion Elements (DMPs) can fit complex motion torque curves using kernel function weighting, making them suitable for modeling and estimating periodic gait movements. Traditional fixed-kernel DMPs use a constant kernel width parameter for torque fitting, which cannot adjust the fitting strategy in real time based on model errors. They suffer from insufficient generalization ability in scenarios with large dynamic model deviations and strong operational disturbances, and struggle to achieve high-precision local fitting under stable conditions. Summary of the Invention

[0004] This invention proposes a torque estimation method based on a dynamic kernel DMP (Dynamic Dynamic Management Platform), aiming to solve the problems of poor fitting accuracy and weak adaptability of traditional fixed kernel DMPs, and improve the estimation accuracy and robustness of human-machine interaction torque under complex working conditions. First, exoskeleton joint angle and torque signals are generated, and the joint kinematic parameters are solved. The theoretical torque is calculated by combining the robot's dynamics model. The dynamic residual is solved by comparing the theoretical torque with the actual torque. Then, the DMP kernel width is dynamically adjusted based on the residual, and finally, the human-machine interaction torque is obtained. The method is implemented through the following technical solutions:

[0005] Step 1: Generate exoskeleton joint angle and torque signals.

[0006] Step 2: Construct a dynamic model and solve for the theoretical torque.

[0007] Step 2.1: Based on the principles of rigid body dynamics, establish an ideal dynamic model of the lower limb exoskeleton that includes inertial, damping, and gravity terms. Substitute the kinematic parameters to calculate the theoretical torque without disturbance or nonlinear error. The dynamic model expression is:

[0008] ,

[0009] in, For rotational inertia, It is the damping coefficient. Indicates the gravity coefficient. , , These represent joint angle, angular velocity, and angular acceleration, respectively.

[0010] Step 3: Calculate and smooth the dynamic residuals.

[0011] Step 3.1: Compare the actual torque with the theoretical torque in real time, calculate the absolute difference, and obtain the original dynamic residual. The expression is:

[0012] ,

[0013] in, This represents the actual torque of the robot's joints.

[0014] Step 3.2: The original residual sequence is smoothed using a five-point moving average filtering algorithm to suppress high-frequency residual jitter caused by sensor noise and random disturbances.

[0015] Step 4: Construct a dynamic kernel DMP model, with the kernel width adaptively adjusted based on the dynamic residuals, and the kernel function... The expression is:

[0016] ,

[0017] in, For phase variables, Indicates the first The center of each kernel function, It is the first Time of the first Time-varying kernel width.

[0018] Step 4.1: Initialize the basic parameters of the DMP model and calculate the initial baseline kernel width. : ,

[0019] in, The number of kernel functions. This represents the baseline width coefficient.

[0020] Step 4.2: Real-time calculation of the first... Each joint dynamic residuals at time step .

[0021] Step 4.3: Calculate the time-varying kernel width based on the real-time dynamic residual. :

[0022] ,

[0023] in, This is the residual adjustment coefficient.

[0024] Step 4.4: Substitute the time-varying kernel width into the kernel function to dynamically adjust the effective fitting interval and weight coverage of the kernel function.

[0025] Step 5: Train the weights of the dynamic kernel DMP model. The calculation formula is: ,

[0026] in, For the robot in the Time of the first The actual torque of each joint It refers to the length of time. Indicates the first The first joint The weights of each kernel function.

[0027] Step 6: Estimate the robot's own torque online using the trained dynamic kernel DMP model. The specific expression is:

[0028] ,

[0029] in, Indicates the number of kernel functions.

[0030] Step 7: Construct the measured torque containing small-amplitude random noise Calculate the torque of human-computer interaction :

[0031] .

[0032] Step 8: Use the mean absolute error to quantify the accuracy of torque estimation.

[0033] This invention constructs a dynamic model based on a lower limb exoskeleton to reproduce unmodeled dynamic conditions such as friction, coupling, deformation, and noise, thus solving the simplification bias problem of traditional dynamic models. It proposes a time-varying kernel DMP algorithm driven by dynamic residuals, which achieves adaptive adjustment through time-varying kernel width, balancing estimation accuracy and generalization ability. It does not require manual parameter tuning, has strong anti-interference ability, and high estimation accuracy, making it suitable for human-machine collaborative control problems of exoskeletons. Attached Figure Description

[0034] Figure 1 This is an overall flowchart of an embodiment of the present invention. Detailed Implementation

[0035] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0036] Figure 1 This is a flowchart of an embodiment. This embodiment provides a torque estimation method based on dynamic kernel DMP. The specific process includes: generating exoskeleton joint angle and torque signals; constructing a dynamic model to solve for theoretical torque; calculating and smoothing dynamic residuals; constructing a dynamic kernel DMP model and training weights; estimating the exoskeleton's own torque; differentially solving for human-computer interaction torque; and quantitatively evaluating the model's accuracy and adaptive performance.

[0037] The specific implementation steps of a torque estimation method based on dynamic kernel DMP are as follows:

[0038] Step 1: Generate exoskeleton joint angles and actual torque signals.

[0039] Step 2: Construct a dynamic model and solve for the theoretical torque.

[0040] Step 2.1: Based on the principles of rigid body dynamics, establish an ideal dynamic model of the lower limb exoskeleton that includes inertial, damping, and gravity terms. Substitute the kinematic parameters to calculate the theoretical torque without disturbance or nonlinear error. The dynamic model expression is:

[0041] ,

[0042] in, The moment of inertia is 0.5 Nm for the hip joint and 0.2 Nm for the knee joint. These are the damping coefficients: 0.5 Nm for the hip joint and 0.2 Nm for the knee joint. The value represents the gravitational coefficient, with 10 Nm for the hip joint and 5 Nm for the knee joint. , , These represent joint angle, angular velocity, and angular acceleration, respectively.

[0043] Step 3: Calculate and smooth the dynamic residuals to quantify the degree of model mismatch.

[0044] Step 3.1: Compare the actual torque with the theoretical torque in real time, calculate the absolute difference, and obtain the original dynamic residual. The expression is:

[0045] ,

[0046] in, This represents the actual torque of the robot's joints.

[0047] Step 3.2: The original residual sequence is smoothed using a five-point moving average filtering algorithm to suppress high-frequency jitter in the residual caused by sensor noise and random disturbances.

[0048] Step 4: Construct a dynamic kernel DMP model. The kernel width is adaptively adjusted based on the dynamic residuals. The kernel function expression is:

[0049] ,

[0050] in, For phase variables, Indicates the first The center of each kernel function, It is the first Time of the first Time-varying kernel width, For the first Time of the first One kernel function.

[0051] Step 4.1: Initialize the basic parameters of the DMP model and calculate the initial baseline kernel width. : ,

[0052] in, The number of kernel functions. This represents the baseline width coefficient.

[0053] Step 4.2: Calculate the dynamic residuals of each joint in real time to obtain the first... Each joint dynamic residuals at time step .

[0054] Step 4.3: Calculate the time-varying kernel width based on the real-time dynamic residual. :

[0055] ,

[0056] in, This is the residual adjustment coefficient.

[0057] Step 4.4: Substitute the time-varying kernel width into the kernel function to dynamically adjust the effective fitting interval and weight coverage of the kernel function.

[0058] Step 5: Train the weights of the dynamic kernel DMP model. The calculation formula is: ,

[0059] in, For the robot in the Time of the first The actual torque of each joint It refers to the length of time. Indicates the first The first joint The weights of each kernel function.

[0060] Step 6: Estimate the robot's own torque online using the trained dynamic kernel DMP model. The specific expression is:

[0061] ,

[0062] in, Indicates the number of kernel functions.

[0063] Step 7: Construct the measured torque signal containing small-amplitude random noise Calculate the torque of human-computer interaction :

[0064] .

[0065] Step 8: Use the mean absolute error to quantify the accuracy of torque estimation.

Claims

1. A torque estimation method based on dynamic kernel DMP, characterized in that, Includes the following steps: Step 1: Generate robot joint angle and torque signals; Step 2: Calculate the theoretical torque based on the robot's dynamics model. The expression for the dynamics model is: ， in, For rotational inertia, It is the damping coefficient. Indicates the gravity coefficient. , , These represent joint angle, angular velocity, and angular acceleration, respectively. Step 3: Calculate the dynamic residuals The expression is: ， in, This refers to the actual torque of the robot's joints; Step 4: Construct a dynamic kernel DMP model, with the kernel width adaptively adjusted based on the dynamic residuals, and the kernel function... The expression is: ， in, For phase variables, Indicates the first The center of each kernel function, It is the first Time of the first Each time-varying kernel width; Step 5: Train the weights of the dynamic kernel DMP model. The calculation formula is: ， in, For the robot in the Time of the first The actual torque of each joint It refers to the length of time. Indicates the first The first joint The weights of each kernel function; Step 6: Use the trained dynamic kernel DMP model to estimate the robot's own torque online. The specific expression is: ， in, Indicates the number of kernel functions; Step 7: Calculate the human-computer interaction torque : ， in, This represents the measured total torque; Step 8: Quantify the accuracy of torque estimation using mean absolute error; The time-varying kernel width mentioned in step 4 is implemented in the following steps: Step 4.1: Initialize the basic parameters of the DMP model and calculate the initial baseline kernel width. : , in, The number of kernel functions. Indicates the base width coefficient; Step 4.2: Calculate the dynamic residuals of each joint in real time to obtain the first... Each joint Dynamic residual at time step ; Step 4.3: Calculate the time-varying kernel width based on the real-time dynamic residual. : ， in, This is the residual adjustment coefficient; Step 4.4: Substitute the time-varying kernel width into the kernel function to dynamically adjust the effective fitting interval and weight coverage of the kernel function.

Images