Training Data Augmentation Method for Brain-machine Interface using Error-related Potentials

In the development of brain-machine interfaces (BMIs), studies have been conducted to improve the accuracy of action using an electroencephalogram (EEG) to detect the neural activity of users when the BMI outputs an error against the intention of users and to provide feedback to the BMI. In this case, only a small number of EEG datasets can be prepared at the time of the error output. In addition, a large bias exists in the number of datasets between the error and no error outputs that affects the discrimination accuracy of machine learning. In this study, four methods for augmenting training data for one-dimensional convolutional neural networks were proposed, and their discrimination rates were compared. The method that eliminated the bias in the number of datasets between the labels of error and no error by data expansion by merging two datasets at the time of error output, obtained the highest discrimination rate of 0.728 compared with the other three methods.