Multi-length Windowed Feature Selection for Surface EMG Based Hand Motion Recognition

Abstract

Feature selection for surface electromyography based hand motion recognition has been seen to retrieve an optimal or quasi-optimal feature subset for classification. This work aims to consider the influence of channel, feature and window length simultaneously with an emphasis on the multiple segmentation. The bacterial memetic algorithm is applied to select the feature candidates from time domain and autoregressive coefficients, which is measured by the inter-day hand motion recognition accuracy. The evaluation is conducted on a case study of 3 able-bodied subjects performing 9 hand motions in consecutive 7 days with 4 different window lengths adopted for the electromyographic data segmentation. Classification in combination with the multi-length windowed feature selection achieved an improved recognition accuracy in comparison with using solely the single-length windowed features in inter-day scenarios and indicated that complementary information to full length segmentation resides in the sub-windows, thus providing feasible feature combinations for conventional pattern recognition based solutions to prosthetic control.