Adaptive learning of multi-finger motion recognition based on support vector machine

A common source for controlling hand prosthesis is the myoelectric signal (MES, also termed electromyography, EMG) that are collected from human body. For a pattern recognition-based EMG control scheme, research has found that the classification accuracy obtained offline may deteriorate owing to signal instinct or changed environment, which results in a reduced system stability. Based on support vector machine (SVM), this paper proposed an adaptive learning procedure intending to keep the classification accuracy. The general idea was to rearrange the training samples of the classifier in real-time by measuring their Kuhn-Tucker (KT) conditions. To regulating the learning effectiveness and system complexity, a forgetting factor was applied to each EMG sample considering its life period. The proposed learning algorithm was validated on a multi-session MES dataset collected from a series of control scenarios, within which the myoelectric signals were collected from two healthy subjects while performing a large variety of finger motions. The experimental results showed that the accuracy of the classifiers could be effectively maintained. In addition, the introduced forgetting factor can effectively confine the classifier's complexity in the long run.