نوشين مومني شهركي

Brain-computer interfaces (BCI), as one of the most advanced technologies for human-machine interaction, play a critical role in assisting individuals with motor impairments. However, achieving high accuracy in motor imagery (MI) classification within these systems faces challenges such as the complexity of EEG signals, limited access to sufficient training data, and the high cost of data acquisition. This study proposes an effective solution to improve the performance of BCI systems by utilizing transfer learning techniques and pre-trained deep neural networks, alongside innovative approaches in model architecture design. In this research, three architectures—ResNet50, InceptionV3, and InceptionResNetV2—were employed, which had been pre-trained on large datasets. These networks demonstrated high performance in EEG data analysis due to their ability to extract rich features. Specifically, attention mechanisms and LSTM were integrated with the InceptionResNetV2 architecture to optimally capture spatiotemporal patterns of the signals. Furthermore, the InceptionResNetV2 model was eva‎luated both with and without transfer learning, achieving accuracies of 98.88% and 91.50%, respectively, highlighting the importance of transfer learning in improving classification accuracy. A comparison of results showed that the InceptionResNetV2-LSTM-Attention architecture outperformed other models, including InceptionV3 (91.33%) and ResNet50 (90.09%). The primary innovation of this research lies in the creative combination of attention mechanisms and LSTM with the advanced InceptionResNetV2 architecture, as well as the investigation of transfer learning's role in enhancing the accuracy of BCI systems. This method not only improved classification accuracy but also reduced the need for complex configurations.

رابط مغز-رايانه , تصور حركتي , انتقال يادگيري

Brain-Computer Interfaces (BCI) , Motor Imagery (MI) , Transfer Learning (TL)

noosheen momeni shahraki

Dr.Abbas Erfanian Omidvar