Performance Evaluation of Temporal and Frequential Analysis Approaches of Electromyographic Signals for Gestures Recognition using Neural Networks

Abstract

Listen

Translate article

Now-a-days, human-machine interfaces are increasingly intuitive and straightforward to design, but there is difficulty capturing electromyographic signal data using the least amount of hardware. This work takes the signals of a human forearm as input parameters describing a series of five gestures, using a dataset of 8 channels of electromyographic signals, using as a capture device a Thalmic Labs Inc. handle called Myo armband. The aim is to compare the performance of the artificial neural network using data in the time domain as input to the learning system. The same data are pre-processed to the frequency domain, looking for an improvement in the neural network's performance since transforming the input signals of the system to the frequency domain minimizes the problems inherent to this type of signal. This transformation is achieved using the fast Fourier transform. Consequently, it seeks to reach a neural network architecture that recognizes the gestures captured with the Myo armband in a high percentage of performance to be used in stand-alone applications, using the TensorFlow libraries of Python for its design. As a result, a comparison of the neural network trained with data in time versus the same data expressed in the frequency domain is obtained, seen from the increase in performance and the percentage of gesture detection.