Estimation of finger joint angles from sEMG using a recurrent neural network with time-delayed input vectors

This paper reports a new technique for estimating continuous finger joint angles from surface electromyogram (sEMG). Using an artificial neural network including a feedback stream (recurrent structure) and a time-delay factor for input, continuous angles scaled to range between 0 and 1 are able to estimated with network from feature vectors. Feature vectors extracted from sEMG are scaled to range between 0 and 10. Target hand motions are free state, fist with five fingers, grip with four fingers except thumb, and thumb flexion only. In this paper, two types of estimation networks are compared. The type 1 network is an older system that cannot be used to train the dynamics of estimation system. The type 2 network is a newer system that can train the dynamics with a recurrent structure by a feedback stream and time-delay factor for input. A comparing of the two types networks show that estimations of finger joint angles with type 2 are better than those with type 1. In particular, the results from type 2 are better than those from type 1 at the transition from one motion to another motion.