Abstract
Convolutional neural networks (CNN) provide an interesting model to automatically identify patterns on signals. This study presents the end-to-end deep learning derived from time-frequency representations of EMG signals to identify physical activity. End-to-end learning of CNN allows the network to automatically learn features from time-frequency representations, without requiring the design of hand-crafted expert features. This type of learning eliminates the requirement for complex multi-step machine learning processing methods. The purpose of this article is to present the framework of the end-to-end learning used to classify physical activity on EMG signals into normal and aggressive classes. This paper proposes the novel approach of using the time-frequency representations produced from EMG signals as the inputs of the CNN to identify activity patterns. The importance of selecting the optimal time-frequency analysis method to represent EMG data is investigated. Three convolutional neural networks were evaluated for two time-frequency representations: the spectrogram and the scalogram. From the analysis, it can be proven that EMG signal representation affects the performance of CNNs. Using the scalogram images to train CNNs echived higher accuracy compared to the spectrogram images. Simple CNN obtained the highest classification accuracy with 94.61%.