Assessing the effect of transfer learning on myoelectric control systems with three electrode positions

Abstract

A motion recognition module integrated into the myoelectric control system learns the parameters of the classifier using training data recorded during multiple days. However, such module will experience troubles (decreases in recognition accuracy) in daily life because the electrode position of the operation day is slightly different from past days. This paper aims to adjust model parameters using transfer learning and assess its performance. Twenty-two kinds of motions in free space were observed from one subject along isometric force of the right forearm by using a wearable device over a one-month period. In the assessment, myoelectric signals were recorded with three different electrode position sets. Furthermore, five models were trained to recognize each degree-of-freedom motion. A linear discriminant analysis (LDA) classifier with covariate shift adaptation showed better results (Mean ± SE: 3.18 ± 0.308%) than a classical LDA classifier (4.34 ± 0.387%) in overall accuracy. These results suggest that transfer learning improves the robustness and usability of myoelectric control systems with wearable devices in daily life.