Finger Movements Classification for Myoelectric Control Based on Artificial Neural Network

A human-computer interface (HCI) based on electromyographic (EMG) signals represents one of the communication options between humans and a terminal device. This paper presents a classification of 13 finger movements based on an artificial neural network (ANN) for this purpose. EMG signals corresponding to movements were acquired from five healthy participants. Data acquisition using four bipolar channels was executed from the anterior forearm hand muscles. Four features in the time domain (TD) were extracted from pre-processed records. ANN models with different hyperparameters were investigated for classification purposes. The model with 36 hidden neurons in one hidden layer and sigmoid activation function achieved the overall highest accuracy of 96.8 %. Finally, two subjects performed a real-time classification experiment (one of them was not involved in the training process). The average accuracy of 89.23 % was achieved in this scenario.

Paper

Full text

PDF

Finger Movements Classification for Myoelectric Control Based on Artificial Neural Network

Semantic Scholar · Engineering · 2021

Abstract

A human-computer interface (HCI) based on electromyographic (EMG) signals represents one of the communication options between humans and a terminal device. This paper presents a classification of 13 finger movements based on an artificial neural network (ANN) for this purpose. EMG signals corresponding to movements were acquired from five healthy participants. Data acquisition using four bipolar channels was executed from the anterior forearm hand muscles. Four features in the time domain (TD) were extracted from pre-processed records. ANN models with different hyperparameters were investigated for classification purposes. The model with 36 hidden neurons in one hidden layer and sigmoid activation function achieved the overall highest accuracy of 96.8 %. Finally, two subjects performed a real-time classification experiment (one of them was not involved in the training process). The average accuracy of 89.23 % was achieved in this scenario.

Similar papers

© 2026 NYSGPT2525 LLC