Direct comparison of SVM and LR classifier for SEMG signal classification using TFD features

Abstract

The performance of any bio-medical signal based robotic device depends upon the classification accuracy of the system. In this context, surface electromyography (SEMG) signals acquisition was done from 15 right hand dominated healthy subjects. After the data acquisition, some pre-processing steps like smoothing, rectification and normalization was performed followed by Daubechies 4 (db 4 ) discrete wavelet transform (DWT) de-noising. DWT was also utilized for SEMG signals decomposition up to four levels. DWT decomposed the SEMG signal in approximation and detailed coefficients and finally fourth level approximation and detailed coefficients were exploited for time-frequency domain features extraction. The performance of individual time-frequency domain features was critically compared with support vector machine (SVM) classifier and linear regression (LR) model. The results indicated that the performance of SVM classifier was found better than the LR classifier. SVM classifier achieved the 95.8% classification accuracy with all combined time-frequency domain features in the form of the feature vector.