Robot control with a neural network and a muscle-machine interface

Namen magistrskega dela je implementacija krmiljenja robota z vmesnikom mišice-stroj. Področje dela obsega predobdelavo in analizo površinskih elektromiogramov z računalniškimi algoritmi z namenom klasifikacije gibov zapestja in področje robotike. Za klasifikacijo gibov uporabljamo umetno nevronsko mrežo. Končni izdelek je robot, sestavljen iz kock LEGO, ki ga je preko računalnika z uporabo protokola Bluetooth možno krmiliti z zapestnico Myo. Preučili smo vplive različnih velikosti okna za izračun korenjene srednje kvadratne vrednosti površinskih elektromiogramov, topologij nevronske mreže, eksperimentalnih protokolov za zajemanje elektromiogramov, števila učnih epoh nevronske mreže ter podali analizo pravilnosti razpoznave gibov z dokončno izbrano nevronsko mrežo na več merjencih.The goal of this thesis is the implementation of robot control with a muscle-machine interface. The field of work spans the preprocessing and analysis of surface electromyograms using computer algorithms with the purpose of wrist gesture classification in the field of robotics. For gesture classification an Artificial Neural Network is used. The final product is a robot, made out of LEGO bricks, which can be controlled with the Myo Armband via a computer and the Bluetooth protocol. We studied the impact of different sizes of windows for calculating the Root Mean Square value of surface electromyograms, neural network topologies, experimental protocols for capturing electromyograms, number of training epochs of the selected neural network and presented an analysis of the gesture recognition correctness using the final chosen neural network on multiple human subjects