نوع مقاله : مقاله پژوهشی
نویسندگان
1 گروه مهندسی برق، دانشگاه صنعتی سیرجان، سیرجان،کرمان
2 دانشکده مهندسی برق و کامپیوتر، دانشگاه سمنان، سمنان، ایران
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
The identification of electroencephalography (EEG) signals related to motor imagery plays a key role in the analysis and evaluation of neural functions in brain-computer interface (BCI) systems. However, considerable individual differences in EEG patterns pose a significant challenge for designing accurate and generalizable models. Moreover, the ability to successfully recognize motor imagery from shorter signal durations has a direct impact on improving the efficiency and practical usability of these technologies. In this study, an innovative hybrid framework is proposed for classifying motor imagery EEG signals, introducing a two-stage architecture based on the combination of an enhanced Informer and the EEGNet model. In this architecture, the EEG signals, after initial frequency feature extraction, are first fed into the enhanced Informer module. This module, leveraging sparse attention mechanisms and adaptive frequency filters (FAA), effectively captures long-term temporal dependencies within the EEG data. The output of the Informer is then passed to the EEGNet model, which, through its specialized convolutional layers (spatial convolution, depthwise convolution, and separable temporal convolution), purposefully extracts spatial-temporal features from the EEG signals and generates a compact and discriminative representation for final classification. Experimental results demonstrate that the proposed model achieves 85.20% accuracy in cross-subject evaluation on the standard PhysioNet dataset with short 2-second trial durations. Comparative analyses with state-of-the-art models indicate that the proposed approach offers competitive and improved performance, particularly in handling shorter signal durations and participant diversity.
کلیدواژهها [English]
)