Comparação entre as redes LVQ e MLP no controle de próteses virtuais para membros superiores

Abstract

During the rehabilitation process, individuals who have experienced a total or partial loss of upper limbs are exposed to many risks. Besides this, a great mental effort is required during the training phase to adapt to a real prosthesis. In many cases, the use of Virtual Reality in Medicine has proven to be an excellent tool for evaluation and support as well as mitigates risk and reduces the mental effort required. In order to be useful, virtual prosthesis must have a great similarity with the real world. For this reason, artificial neural networks have been explored to be applied in the training phase in order to provide real time response. The objective of this study is to compare the performance of the LVQ and MLP neural networks in EMG pattern recognition. To achieve this, different feature extraction techniques for simulation and control of virtual prostheses for upper limbs are investigated. Using the LVQ neural network, autoregressive model as a feature extraction technique and an average of 10% of all training patterns, achieved up 99% of efficiency for the hand movements and 97% of efficiency the arm movements.