Implementation of a neural network of low computational cost for its application in arm prostheses

Abstract

A prostheses implementation represents a design challenge in its different stages. The control systems and the total system cost play a very important role. In this work, a control proposal is presented using artificial neural networks (ANN) for pattern recognition using electromyographic (EMG) signals, which are obtained from the arm muscle (biceps). A single channel EMG surface sensor is used to acquire the EMG signals and by means of adjacent windows the feature extraction is carried out in order to reduce the input values to the neural network. The neural network is trained with the features extracted from the EMG signals, using a method of muscle tension thresholds for activation and a labeling technique for the output called One Hot Encode. The resulting ANN was embedded in a low-cost microcontroller and an accuracy of approximately 93% was achieved.