Artificial Intelligence For Classification And Control To Improve Upper Prosthetic Limb Functionality

Abstract

This paper describes an electromagnetic (EMG) signal classification method utilizing CNNs (convolution neural networks) aimed at idealizing a simple way of using a prosthetic limb. By employing an EMG database of 200 EMG samples taken from subject forearm muscle during (1) gripping hand and (2) releasing hand actions; a one-dimensional convolutional neural network model (1D CNN) was developed to learn the distinctive features of grip/release EMG patterns and categorize those actions into corresponding classes. Eighty percent (80%) of the database was used to train the network; twenty percent (20%) was used for evaluation of model performance. The neural network architecture consisted of convolutional and pooling layers to extract features from the EMG patterns, and dense layers to classify the grip/release EMG samples as binary classifications. The CNN model classified both categories with 100% accuracy on the test set and accurately distinguished between the grip and release hand positions. The weights from the trained CNN model were stored on an ESP32 microcontroller for EMG signal classification.