Wavelet-Based Electromyographic Feature Selection Method for Real-Time Ankle Movement Recognition

This paper proposes a feature selection method aiming to determine the optimal feature subset for online recognition of four ankle movements, namely abduction, adduction, dorsiflexion, and plantarflexion. Sequential forward selection means is utilized to select from an original feature set comprising information from the mean absolute value (MAV) of wavelet decomposition coefficients, of the wavelet multi-level reconstructions, and of the signal itself. The impacts of different window lengths and overlap percentage on online classification accuracy are also evaluated. Support vector machines algorithm is adopted to recognize the ankle movements. The highest real-time classification accuracy of 97.32% is achieved by means of nine features, windows of 0.1 s length, and 80% overlap between windows. It shows that the proposed method is capable of selecting features for real-time ankle movement recognition with high accuracy and low latency effect.