Enhancing motor-imagery BCI robustness for edge computing applications

This paper presents a robustness analysis for an embedded motor-imagery brain–computer interface. The analyzed model is based on the motor imagery algorithm EEGNet, adjusted with channel and time window reduction and temporal down-sampling to address the computational and memory limitations of microcontroller units. By addressing these limitations, battery operated autonomous, wearable mechanical devices can be enhanced to meet requirements for motor imagery and hence, transmission to external computing engines can be avoided. Considering the safety requirements of the application, the algorithm must be ensured to be robust to any variations in the sensors. To address this issue, this work examines the performance of the introduced model regarding data augmentation techniques. Tests conducted on the Physionet EEG Motor Movement/Imagery Dataset indicate strong natural resilience to noise, but reduced accuracy when the data is augmented with patch perturbations. Furthermore, this work proposes training the model on additional augmented data to achieve higher robustness. With additional training data augmented by noise, the models achieve up to 3.26% higher classification for 4- class MI tasks with a remaining high robustness to noise and low robustness to patch perturbation. For the augmentation by patch perturbation, the model achieves similar performances for 4-class MI tasks and robustness to noise, however, a significantly improved robustness to patch perturbation by over 15%.