Skip to Main Content
This paper aims at dealing with a critical issue for electromyography (EMG) recognition. The issue is related to the stability of an EMG-based prosthesis control. Traditional EMG recognition systems receive EMG patterns and send them into classifiers directly, which generally results in unstable situations if the classes of some of the input EMG patterns are not included in the training of the classifiers. The EMG patterns whose class labels are not defined in the training phase are called non-target patterns. There should be a filter and this filter should be able to reject all non-target EMG patterns. As such, only target EMG patterns are fed into classifier, thus achieving a high-accuracy EMG classification. To this end, we propose in this paper a one-class classification-based non-target EMG pattern filtering scheme. By introducing a novel one-class classifier, called support vector data description (SVDD), into the filtering scheme, the goal mentioned above can easily be achieved. SVDD is a powerful machine learning technique. It can be built on a single class and find a flexible boundary to enclose the target class by using the so-called kernel trick. In experiments, we will show that if the filtering scheme is not performed, the traditional EMG classification system suffers from unstable situations. Contrarily, the whole classification system will achieve satisfactory and stable performance no matter what the input EMG patterns are target or non-target ones, if the proposed filtering scheme is embedded.