Efficient Contaminant Identification in sEMG Signals using Machine Learning

Surface electromyography (sEMG) signal classification applications, such as upper limb prosthesis, have increased in recent years. Eliminating unwanted noise signal for the exact and controlled motion of prosthesis is imperative. Noise in signal make a model biased and lead to mis-classifications. During the usage of a prosthetic arm, mis-classifications of hand gestures can cause it to behave erratically. The choice of a proper filter depends on the type of noise. Therefore, it is necessary to correctly identify the noise and filter it without increasing its hardware complexity. In this article, a study on three approaches to identify four commonly occurring noise in sEMG signal is performed, which is useful for filter selection. The three approaches include the tabular feature, sequence feature, and image feature based classification. Six different classifiers are used to compare the three approaches on the basis of accuracy, time consumption, and memory consumption. Models are also tested on varying noise levels for robustness analysis. On all the levels of noise, the sequence feature based classification approach gives a reasonably good accuracy while consuming the least time and memory.