CETRNet: Convolutional-EMA Temporal Residual Network for Motor Imagery Decoding
<p>Hongyu Cai<sup>1</sup></p> · Academic Journal of Computing & Information Science · 2025
Brain-Computer Interfaces (BCIs) utilize wearable electroencephalography (EEG) coupled with artificial intelligence (AI) to interpret neural activity. It finds significant applications in healthcare and robotics. And Motor imagery (MI) decoding is the basis of external device control via EEG. However, achieving high decoding accuracy remains a significant challenge, hindering the widespread adoption and advancement of BCI systems. To address this challenge, the present study proposes the Convolutional-EMA Temporal Residual Network (CETRNet), a novel deep learning architecture designed to improve the classification of MI signals from EEG data. The network consists of several key modules that enhance classification performance while maintaining parameter efficiency to reduce computational requirements. The initial processing stage includes dedicated temporal and spatial blocks that capture essential spatio-temporal features, followed by channel attention mechanisms that prioritize relevant spatial information. An Exponential Moving Average (EMA) module is integrated to capture long-range temporal dependencies and detect inherent periodic patterns in the EEG data. Subsequently, higher-level temporal abstractions are derived through a temporal convolutional residual block, which also implements data augmentation using a convolutional sliding window technique. Evaluation on the BCI Competition IV-2a benchmark dataset demonstrated that CETRNet achieved a subject-specific accuracy of 83.33%, highlighting its potential for reliable classification of MI-EEG signals.