Pattern Classification of EMG Signals Using Neural Networks during a Series of Motions

This paper proposes a pattern classification method of time series EMG signals using neural networks in order to estimate a motion intended by a human operator. To achieve successful classification for non-stationary EMG signals, a new network structure that combines a common back-propagation neural network with recurrent neural filters is used. This network is suitable to express time-varying characteristics of time-series EMG signals. Also dynamics of a terminal attractor is incorporated in the learning rule in order to regulate convergence time. The convergence time is always less than the upper limit of a specified time, so that mental stress of the operator for waitting the convergence of learning can be reduced. In the experiments, the EMG signals measured from four subjects during a series of six motions are used. It is shown from the results that the proposed network can achieve a relatively high classification performance and the learning converges within a specified time.