Informações do Trabalho
Analyzing Data Augmentation Methods for Convolutional Neural Network-based Brain-Computer Interfaces
GABRIEL OLIVEIRA MOREIRA FARIA
HEDER SOARES BERNARDINO
Brain-computer interfaces (BCIs) are systems that use brain signals to interact with external devices, with a wide variety of applications spanning from entertainment to healthcare. Within BCIs, motor imagery brain signals are one of the most commonly used, enabling individuals to control external devices by mentally simulating body movements. In this context, convolutional neural networks (CNNs) can be used to classify motor imagery electroencephalogram (EEG) signals and translate them into specific commands. However, these signals are usually noisy and can vary significantly over time and among people, making it frequently necessary to collect a large amount of data to calibrate these models. This process can be time-consuming and fatiguing for BCI users, a particularly significant issue in healthcare settings. In order to address this problem, this work investigates the effectiveness of data augmentation in reducing the need for data and improving classification accuracy. Five data augmentation methods previously documented in the literature are evaluated across two motor imagery benchmark datasets utilizing a few-parameter CNN. The evaluations encompass a variety of configurations of EEG electrodes, motor imagery tasks, and training sample sizes. Each method is then compared against one another in terms of both accuracy and data distribution effects. The overall findings indicate that data augmentation can alleviate the need for original data, resulting in higher accuracy even with fewer training samples. The findings also indicate that combining different data augmentation methods can lead to further improvements in accuracy.
brain-computer interface, motor imagery, convolutional neural network, data augmentation
Obter arquivos extras