چكيده لاتين
Brain–computer interfaces (BCIs) enable direct communication between the human brain and computers by recording and analyzing the activity of the central nervous system. One of the most important and challenging applications of BCIs is the classification of motor imagery (MI) based on electroencephalography (EEG) signals. Although MI classification from EEG offers significant advantages, the non-stationary and noisy nature of these signals imposes considerable limitations. To address these issues, recent studies have employed deep learning methods—particularly convolutional neural networks (CNNs)—which have demonstrated strong capability in extracting temporal, spatial, and spectral features from EEG. However, since CNNs were originally designed for image processing, their performance in handling non-stationary signals such as EEG remains challenging.
In this thesis, we propose a CNN-based model tailored for EEG signal processing. Specifically, we introduce a new convolutional layer, termed NewCWConv, which resembles traditional convolutional layers but employs filters based on the continuous wavelet transform (CWT). To evaluate its effectiveness, we design EEGWaveletNet, a model structurally similar to the well-known EEGNet, with the key difference that its first layer is replaced with the proposed NewCWConv to extract time–frequency features. The model is evaluated using the BCI Competition IV 2a dataset, achieving an average accuracy of 75.71% across nine subjects in a four-class classification task.
Finally, we compare the performance of EEGWaveletNet with benchmark deep learning models, including EEGNet, ShallowConvNet, and DeepConvNet, which are widely used in MI classification. The results show that our proposed model achieves a meaningful improvement in accuracy compared to EEGNet, despite having only 920 more parameters. Moreover, while EEGWaveletNet (with 2,636 parameters) achieves performance comparable to ShallowConvNet (40,644 parameters) and DeepConvNet (152,119 parameters), it requires significantly fewer parameters. These findings demonstrate that incorporating the NewCWConv layer leads to a more compact model, better suited for extracting informative features from EEG signals.
