Motion Discrimination Method from EMG Signals Using Statistically Structured Neural Networks
Toshio Tsuji, Daiichiro Mori, Koji Ito · IEEJ Transactions on Electronics Information and Systems · 1992
The present paper proposes a method to estimate the motion intended by a human operator from his EMG signals using the statistically structured neural network. EMG signals gradually appear at the beginning of the motion and begin to vanish at the end of the motion. Therefore they should be regarded as non-stationary signals and only weak assumptions are made about the probability density functions. In order to classify such non-stationary signals, the neural network presented here is statistically structured using a Gaussian mixture model which can approximate an unknown probability density function by a finite mixture of multivariate Gaussian component densities. The experimental results for non-stationary EMG signals show that the network can learn the unknown densities of the EMG signals and it can discriminate six motions of forearm and hand based on Bayesian rule from unlearned EMG patterns with the accuracy above 90%.