Enhanced Real-Time Raw sEMG Signal Classification Through Bypass of Manual Feature Engineering and Extraction

Abstract

AbstractThe use of surface electromyographic (sEMG) signals obtained from the residual stump of an arm has been embraced for the purpose of developing prosthetics and bionic limbs. These sEMG signals are processed and classified into gestures. As the signals are stochastic and non-stationary, classification of these signals is a complex task. Conventional methods of feature recognition require handcrafted, manual feature extraction followed by classification. This paper explores the bypass of this preliminary step of feature extraction in order to reduce lag and computing complexity. Through data collected with the aid of an 8-channel Myoband developed by Thalmic Labs, our group trained and implemented three main raw data classifier models for real-time sEMG signal classification of nine gesture classes. The three methods presented are a raw naïve Bayes model implementation (81.55%), a novel raw sEMG ConvNet (96.88%), and finally, a novel ConvXGB model implementation (90.62%). The ConvXGB implemented is an adaptation of the new deep learning model developed in 2020. The significance of these novel models developed is the reduction in computational resources required for real-time classification of signals, leading to a reduction in lag time in real-time prosthetic devices, and the increased viability of embedded-systems classifications. The findings have the potential to also pioneer a new generation of more responsive prosthetic devices that are also easy to control.KeywordsMulti-channel time series classificationsEMG signalsProsthetic devicesBypassing feature extractionConvXGB modelGRU-FCN modelNaïve Bayes model