Classificação de sinais eletromiográficos utilizando redes neurais artificiais, análise discriminante linear e floresta aleatória

Abstract

There is a gap in the upper limb prosthesis market in relation to lower limb prostheses, this is due to the sum of a smaller market, after all, only approximately 20% of amputations performed are of upper limbs, added to a greater difficulty in developing these prostheses (ZIEGLER-GRAHAM, 2008), and to make it all worse, there is still a high acquisition cost in buying one. Modern advances in artificial intelligence and access to data processing, along with the rising of startups and scientists interested in using the best that data processing can offer, ensured a great technological advance in prosthesis models and a considerable reduction in costs. In this work, based on the public database Ninapro (Non-Invasive Adaptive Hand Prosthetics), three different artificial intelligence techniques were used, seeking to discover which one is the most promising in the classification of myoelectric signals. The algorithms used are Artificial Neural Networks, Linear Discriminant Analysis and Random Forest, with all its development, parameter adjustment, validation and testing being presented during the work. Based on the tests carried out, Random Forest was identified as the most promising of the three approaches, reaching an accuracy that ranged from 92% in sets of 5 movements to 84% in sets with all 52 movements.