Classification of Lower Limb Electromyographical Signals Based on Autoencoder Deep Neural Network Transfer Learning

In this paper we aim to propose an Artificial Neural Network (ANN) model in order to classify lower limb surface Electromyographical signals, without applying any pre-processing to the inputs. To this end, first three different autoencoder models have been proposed in order to choose the autoencoder architecture with the best performance. Then encoder layers of the best autoencoder model transferred to a classification network. This network aims to classify raw input sEMG signals into three categories, namely; sitting, standing and walking. After training the classification network, we were able to successfully classify input sEMG signals and reach to 81.51% accuracy for all the classes and 91% and 94% accuracy for the first two classes, with the unseen subjects. The major difference of this research with previous ones lies in fact that evaluation of our network occurs in a realistic scenario, where data from test and validation subjects is not seen by the network before and also our transferred network uses unlabeled raw sEMG signals for extracting features from signal, which gives us the ability to transfer this model in situations where labeled data is not available.