An Onset Based Approach for Feature Extraction and Classification of EMG Signals

Electromyography (EMG) signals and their characteristics have been extensively studied and used in the last decade for rehabilitation engineering and robotics applications. In these applications, an EMG signal-based control system performs different steps, from signal acquisition and processing to classification and real-time control. Among these, feature extraction is one of the important steps. It is calculated from the segments of the pre-processed signals. Hence, accuracy and processing time are critical performance indicators of different classifiers. Therefore, in this work, an onset based approach is proposed for identifying active EMG signals, where a norm-based threshold method for segmentation and feature calculation is employed, resulting in fewer samples. The Multi-Layer Perceptron (MLP) and Linear Discriminant Analysis (LDA) classifiers are trained and tested for the same. The results show an accuracy of 81% with LDA and 87% with MLP, respectively. It has also been observed that few channels show enhanced accuracy while others result in comparable accuracy with significantly reduced processing time.