Feature weighting concatenated multi-head self-attention for amputee EMG classification

Abstract

ReliefF and neighborhood component analysis (NCA) concatenated multi-head self-attention (MSA) based multi-channel amputee EMG signals classification model is proposed in this paper. It is inspired by the Transformer and Vision Transformer models, and designed to be lightweight for prosthetic applications. The ReliefF and NCA layers are integrated to the MSA for class separability concatenation of 8-channel EMG signals. The contribution as weight concatenation is performed on publicly available amputee dataset, and the effects of ReliefF and NCA are compared to the conventional MSA architecture against varying contraction levels. Six hand gestures with three contraction levels are recognized using the popular features of waveform length (WL) and root mean square (RMS) depending on three evaluation schemes (within the same force level, unseen level and all levels). The proposed class separability concatenation yields up to 2.08% increase rates when compared to the conventional MSA model.