Real-Time Load Estimation for Load-lifting Exoskeletons Using Insole Pressure Sensors and Machine Learning

원문 초록

This paper presents a novel method for real-time lifting-load estimation to enhance the control strategies of upper-limb assistive exoskeletons. By leveraging cost-effective insole pressure sensors, the proposed system extracts differential pressure data that minimizes disturbances from variations in body weight and sensor placement. Two modeling approaches are explored: a channel-based method that employs traditional regression techniques-Elastic Net, Support Vector Regression (SVR), and Multi-Layer Perceptron (MLP)-and a map-based method that utilizes transfer learning with a pre-trained MobileNetV2 model. The experiment is in the preliminary test stage, covering load ranges from 2 kg to 10 kg in increments of 0.5 kg, and collecting data from three subjects to test the approach. In the Channel-based method, the average Weighted Mean Absolute Percentage Error(WMAPE) for three subjects showed that the SVR achieved 13.46%, with the MLP performing similarly. In the Map-based method, using data from one subject, the Fully Fine-Tuned MobileNetV2 model reached a WMAPE of 9.74%. The results indicate that the integration of insole sensor technology with advanced machine learning models provides an effective solution for dynamic load estimation, potentially reducing the risks of over- and under-compensation in exoskeleton control.