EMG Based Binary Finger Movement Classification by Tensor Decomposition

The Electromyography (EMG) signal refers to the electrical activity generated by skeletal muscles. For example, while performing different hand motions, the hand muscles contract and relax which produce EMG signals. Thus EMG signals can be a good representative of those motions. Traditionally, time domain features or frequency domain features, or both timefrequency domain frequency features are used to perform the classification of different hand gestures. However, these feature extraction techniques were only used in two-dimension/matrix format. Normally EMG signals are multidimensional as they consist of channels, time, frequency, subjects, repetition, and movements which makes EMG signals higher-order arrays or tensors. Therefore, higher-order feature extraction methods need to be performed on tensors to extract underlying patterns. In this paper, we propose a feature extraction method that takes two 3rd-order EMG data tensors of index finger flexion and index finger extension motion from 21 subjects and applies canonical polyadic decomposition (CP) to extract features. Finally, different types of classifiers were used to classify both finger motions. The average accuracy achieved was greater than 94% for all classifiers with highest accuracy of 98.99% for Fine KNN classifer.