An investigation of feature combinations of time-domain power spectral descriptors feature extraction for myoelectric control of hand prostheses

Abstract

Upper limb prostheses controlled with Pattern Recognition (PR) and myoelectric signals have great promise for amputees who lost an upper limb since it can control large number of movements intuitively. One of the existing challenges with such PR systems include the need to develop new feature extraction techniques to facilitate clinical implementation of PR systems to satisfy amputees' needs. In this paper, the features of newly proposed Time Domain Power-Spectral Descriptors (TD-PSD) feature extraction method will be investigated in order to find the best feature set to classify eight hand and finger movement. Two congenital female transradial amputees were recruited and the myoelectric signals which are also known as Electromyography (EMG) signals, were collected from different surface EMG sensors when the two participants performed eight finger and hand movements. Results showed that a subset of four TD-PSD features achieved similar performance to that of the full set of TD-PSD features, with average error rates of the classification being equal to approximately 7% which is within the acceptable error rates of a usable PR system where it should be below 10%.