EMG based finger movement estimation

Abstract

The advancements in electronics these past years enabled the scientific community to read the smaller voltages created by the muscles called Electromyography Signals (EMG). The aim of this dissertation was to capture EMG signals from the forearm and investigate their relationship with the finger joint angles of the index and middle finger Metacarpophalangeal Joint (MCP). The EMG signals were captured through six muscles located in the forearm: Flexor Pollicis Longus, Extensor Pollicis Longus, Flexor Digitorum Superficialis, Extensor Indicis, Flexor Carpi Radialis, Extensor Carpi Ulnaris. The joint angles were measured using a goniometer for six discrete angles 15o, 30o, 45o, 60o, 75o and 90o. The index or middle finger flexed and extended for 10 times each time stopping for 2 seconds each time it reached the angle. This was repeated 15 times for the index and middle finger. Only the steady state parts were the finger reached the angle were considered for this dissertation. Four features were then extracted from each the EMG data muscle channels: Mean Absolute Value, Variance, Waveform Length and Willison Amplitude. A three layered Multi-Layer Perceptron (MLP) was used to estimate the joint angles from the EMG features. Four different MLPs were trained for four different scenarios. Four different scenarios were considered: Two individually trained neural networks for index and middle finger respectively, a neural network trained with the index and middle finger data together and a neural network with index and middle fingers moving together at the same time in same direction. The best results were obtained for the individually trained neural networks with R2 of 0.75 for the index and R2 of 0.59 for the middle finger. Furthermore the index and middle robotic hand was actuated with the Arduino MATLAB toolbox, and the voltage outputs measured to ensure the angle was set the right voltage levels at input.