Optimization of Torque-Control Model for Quasi-Direct-Drive Knee Exoskeleton Robots Based on Regression Forecasting.

Area of Science:

• Robotics
• Biomechanics
• Machine Learning

Background:

• Controlling lower-limb exoskeleton robots, especially in human-machine coupled systems, presents challenges due to complex, non-linear sensor-to-torque mapping.
• Accurate modeling of joint torque is difficult with traditional mathematical tools.

Purpose of the Study:

• To develop a nonlinear regression forecasting model for predicting knee joint torque in exoskeleton robots.
• To create a knee joint torque-control model that compensates for mechanical and control system delays.
• To evaluate the effectiveness of the developed control model in assisting stair climbing using surface electromyography (sEMG).

Main Methods:

• Collected knee torque and inertial measurement unit (IMU) data from an exoskeleton robot during stair climbing.
• Employed a multivariate network model combining Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) for nonlinear approximation.
• Implemented time-shifting techniques during model training to compensate for system delays and generated multiple control models.

Main Results:

• The CNN-LSTM model successfully learned nonlinear approximations for knee joint torque prediction.
• Testing on a lightweight knee exoskeleton showed reduced rectus femoris (RF) sEMG activity by 20.87% and increased vastus medialis (VM) sEMG activity by 17.45%.
• Different time-shifting parameters influenced the exoskeleton's performance, with optimized shifts enhancing assistance.

Conclusions:

• The developed machine learning control model effectively assists knee joints during stair climbing in exoskeleton robots.
• The approach of using CNN-LSTM with time-shift compensation addresses the challenges of nonlinear mapping and system delays.
• Experimental results validate the model's ability to reduce user muscle effort and improve exoskeleton performance.