Classification of waist motion by neural networks — Toward power assist suit for caregiver

This paper presents a signal processing for discrimination of waist motions including forward/backward bending and right/left twist. The system is planned to implement in a waist power assist suit for caregiver which our research group is currently developing. To this end, the surface electromyogram (SEMG) signals on the lower back are analyzed in the first step, and the discrimination method is then proposed using four sets of feedforward neural networks (NNs) in which each network is a binary classifier for each of four motions. The source for discrimination is SEMG signals on right and left erector spinae muscles. With a peripheral FFT-based prefilter, the motion start point is detected, and the feature vectors, which are inputs of NNs, are calculated with SEMG signals just behind the start point. It is shown that a multi-class classifier based on combination use of four sets of NNs appropriately discriminates each motion.