A novel EEG motor imagery classification model using feature fusion of temporal convolution and attention

Motor imagery classification using electroencephalography (EEG) signals is a fundamental component of Brain-Computer Interface (BCI) systems. It enables individuals with physical disabilities to control robotic limbs and perform various movements. However, the inherently noisy nature of EEG signals poses significant challenges for their effective utilization in this domain. In this study, we propose a novel end-to-end deep learning model based on feature fusion of multiple deep learning blocks, including a Convolutional Neural Network (CNN), Temporal Convolutional Network (TCN), and Squeeze and Excitation (SE) attention mechanism, enabling the model to learn discriminative features for classifying raw motor imagery signals without any preprocessing. The proposed architecture employs novel feature fusion strategies to maximize classification performance and computational efficiency. The CNN extracts initial spatial features, the TCN captures temporal dependencies, and the SE attention mechanism emphasizes the most informative features from the CNN output. The model was evaluated on the BCI Competition IV 2a and 2b datasets. Training was conducted for 500 epochs (2a dataset) and 200 epochs (2b dataset), using only the first session of each subject for training and validation. The average classification accuracies on the completely isolated test sets (second session) were 78.12 % and 85.72 % for the 2a and 2b datasets, respectively. These results demonstrate that the proposed model effectively classifies multi-class motor imagery signals.