3D Convolutional Neural Network for EEG Signal Processing

Abstract

In this thesis, we propose a new approach for Electroencephalography (EEG) based Motor Imagery (MI) feature extraction and classification using a Deep learning (DL) technique. Recently, a few efforts on applying DL techniques for EEG time-series data analysis have proved the advantages for Brain Computer Interface (BCI). Herein, we introduce a method to transform EEG data into a 3D structure of image to robust the frequency and spatial features related to MI. Then, those 3D EEG images are fed into a 3D Convolutional Neural Network (3D CNN) architecture to robust representations for learning and classification. To our best knowledge, it is the first study attempting to develop a 3D CNN configuration for the topic of single-subject EEG-based MI BCI. Our empirical results on BCI Competition IV Dataset 1 (real human data) demonstrates that our new approach outperforms the baseline method - Common Spatial Pattern (CSP) - with higher performance for all subjects. In addition, based on our observation, the 3D CNN model can bring the benefits in capturing and learning essential frequency and spatial representations of MI from 3D EEG images; thus, it leads to better classification accuracy.