Improving myoelectric pattern recognition using invariant feature extraction

The existing algorithms of myoelectric pattern recognition (MPR) are far from enough to satisfy the criteria which an ideal control system for upper extremity prostheses should fulfill. This study focuses on the criterion of short training, or possibly zero training. Due to the non-stationarity inhered in surface electromyography (sEMG) signals, the system may need to be re-trained day by day in the extended usage of myoelectric protheses. However, as the subjects perform the same motion types in different days, we hypothesize there still exists some invariant characteristics in the sEMG signals. Therefore, give a set of training data from several days, it is possible to find an invariant component in them. To this end, an invariant feature space analysis (IFSA) framework based on kernel feature extraction is proposed in this paper. A desired transformation, which minimizes the dissimilarity between sEMG feature distributions of different days and maximizes the dependence between the training data and their corresponding labels, is found. The results show that the generalization ability of the classifier trained on previous days to the unseen testing days can be improved by using IFSA. More specifically, IFSA significantly outperforms Baseline (original input feature) with average classification rate of 1.11% to 1.69% (p <; 0.0001) in task including 9 motion classes or 13 motion classes. This implies that the promising approach can help for achieving the zero-training of MPR.