Improving motor imagery EEG classification by CNN with data augmentation

Brain Computer Interface (BCI) system enables human brain to communicate with the external world without the involvement of muscle and peripheral nerves. Motor Imagery(MI) Electroencephalogram (EEG) is one of brain signals commonly used in the BCI system. Recently, deep learning models such as Convolutional Neural Network (CNN) have received widespread attention and provided better classification performance in MI EEG classification compared to other state of art approaches because they can learn the features that are most relevant to the task at hand. However, the performance of CNN largely depends on its architecture as well as the quality and quantity of training data. Current MI EEG data are scarce because the data collection is relatively expensive and therefore effective data augmentation methods are particularly important to improve the MI classification performance. In this paper, we first propose a shallow CNN architecture as well as a new and effective data augmentation method to compensate the shortcoming of data insufficiency, then we apply the method of superposing and normalizing the signals of the same labels across subjects and time to generate additional EEG data. The proposed superimposed data augmentation method can enable the signals preserve the intrinsic characteristics and reduce the signals drift over time and subjects. We evaluate the proposed architecture and method on the PhysioNet dataset, the experimental results show that the proposed CNN architecture performs better than the previous architectures and can achieve an average accuracy of 91.06% in two-class classification tasks. In addition, the proposed data augmentation method can improve the average accuracy from 73.46% to 76.78% in four-class classification tasks for all 109 subjects, which proves the effectiveness of the proposed method.