Bacterial Memetic Algorithm based Feature Selection for Surface EMG based Hand Motion Recognition in Long-term Use

Pattern recognition algorithms have been applied in the surface electromyography (sEMG) based hand motion recognition for their promising accuracy. Research on proposing new features, improving classifiers and their combinations has been extensively conducted in the past decade. Meanwhile, the feature projection methodology, has been routinely exploited between the phases of feature extraction and classification. However, limited publications have been seen addressing the feature selection, which is a vital alternative in dimensionality reduction for pattern recognition. Recent development of sEMG acquisition devices have contributed to more signal capturing sites or even detection arrays of high density in the application. In this paper, the memetic evolutionary method named bacterial memetic algorithm (BMA) has been adopted as the feature selection strategy for sEMG based hand motion recognition. A case study of 4 subjects in long-term use has been conducted to demonstrate the feasibility of the proposed strategy, that comparable recognition accuracy with reduced computation cost has been achieved. A further discussion on the feature redundancy and inter-subject use has also been demonstrated based on the experimental results derived from BMA based feature selection.