USE OF DEEP MACHINE LEARNING METHODS OF ARTIFICIAL NEURAL NETWORKS FOR DESIGNING ALGORITHMS OF ELECTROMYOGRAPHY SIGNAL RECOGNITION IN BIONIC PROSTHESIS

Abstract

For sterling application of scientific and engineered achievements in field of bionic prosthesis it’s required to provide comfortable  and natural human‑prosthesis interface for an end‑user. In this article we are looking into ways and methods of analysis of the  signal collected through electromyography activity of muscles on the skin surface. Such signal is nonstationary and unstable  by  its  nature,  dependent  on  various  factors.  sEMG  based  interface  has  several  unsolved  problem  at  the  moment,  such  as  insufficient accuracy of recognition and noticeable delay caused by signal recognition and processing. Article is dedicated to  application of deep machine learning required to provide decent recognition of electromyography signals. In the course of the  research hardware was developed to register muscle activity. Data collecting system and algorithms of gesture recognition have  been designed as well. In conclusion decent results were achieved by using convolutional neural network, with two‑dimensional input, since data stream has obvious translational orientation. In the future, modification of neural network architecture, learning  algorithms and experiments with structure of data are planned.