Metodologia para classificação de sinais EMG no controle de membros artificiais

Abstract

One of the major challenges for prosthesis development is to produce devices which mimic their natural counterparts. In general, artificial limbs don’t have proper feedback by which the user can assess the status of the prosthesis and the control is very unnatural. Preferably, a subconscious control is desired. Myoelectric control has been widely used as an alternative strategy designed for easier control. However, there is still a lot do be done in order to achieve artificial limbs as dextrous as human limbs. In an attem pt to contribute to the researches towards better artifical limbs, it has been developed an EMG processing system, capable of generate input control to a four degrees of freedom prosthesis. Two major muscle groups (biceps and triceps) were used as source of electromyografic signals, which were discriminated into four different classes: elbow flexion, elbow extension, wrist pronation and wrist supination. Those patterns were classified by an artificial neural network, which received as inputs the EMG signal features extracted by a,n autoregressive model. The minimum number of pairs of electrodes and their best positioning for detection, Processing and classification were also investigated. To do so, five pairs of electrodes (two on the biceps - long head (Bl) and short head (B2) - and three on the triceps -long head (T l), médium head (T2) and lateral head(T3)) and one pair of electrodes (on plexo brachial) configuration were considered. Isometric and isotonic contractions were analyzed for each one of those two configurations. The EMG signals were studied in several combinations for each type of contraction. The results show that the configurations using two pairs of electrodes (positioned on B2 and T l) and three pairs of electrodes (positioned on B2, T l and T2 or B2, T l and T3), provided accuracy as good as 100%, for the EMG pattern recognition process.