EMGTTL: Transformers-Based Transfer Learning for Classification of ADL using Raw Surface EMG Signals

TL;DR

This paper introduces EMGTTL, a transformer-based transfer learning framework that classifies Activities of Daily Living from raw surface EMG signals without explicit feature extraction, improving accuracy across datasets.

Contribution

The study presents a novel transformer-based transfer learning approach for ADL classification using raw sEMG signals, eliminating the need for feature extraction.

Findings

Transformer architecture achieved 64.47% accuracy on DB1 and 68.82% on DB4.

Transfer learning improved accuracy to 66.75% (pre-trained on DB4) and 71.04% (pre-trained on DB1).

Method reduces reliance on time-consuming feature extraction in sEMG analysis.

Abstract

Surface Electromyography (sEMG) is widely studied for its applications in rehabilitation, prosthetics, robotic arm control, and human-machine interaction. However, classifying Activities of Daily Living (ADL) using sEMG signals often requires extensive feature extraction, which can be time-consuming and energy-intensive. The objective of this study is stated as follows. Given sEMG datasets, such as electromyography analysis of human activity databases (DB1 and DB4), with multi-channel signals corresponding to ADL, is it possible to determine the ADL categories without explicit feature extraction from sEMG signals. Further is it possible to learn across the datasets to improve the classification performances. A classification framework, named EMGTTL, is developed that uses transformers for classification of ADL and the performance is enhanced by cross-data transfer learning. The methodology is implemented on EMAHA-DB1 and EMAHA-DB4. Experiments have shown that the transformer architecture achieved 64.47% accuracy for DB1 and 68.82% for DB4. Further, using transfer learning, the accuracy improved to 66.75% for DB1 (pre-trained on DB4) and 71.04% for DB4 (pre-trained on DB1).