| Περίληψη: | Electroencephalography (EEG)-based Brain Machine Interfaces (also known as Brain Computer Interfaces BCIs), particularly those using motor-imagery (MI) data, have the potential to become groundbreaking technologies in both clinical and entertainment settings. That is attributed to their ability to create a direct communication pathway between the human brain and the external environment, by translating recorded brain activity into messages or commands. MI data describes the mental performance of movements not accompanied by any kind of peripheral (muscular) activity. However, despite the extensive research on EEG in recent years, it is still challenging to interpret EEG signals effectively, mainly because they are non-stationary, they can suffer from external noise and are prone to unique signal artifacts. As a result, the majority of research on this field employs computationally demanding preprocessing techniques or complicated artifact removal stages that pose limitations regarding the robustness and the propriety of the models for real-world applications, neglecting at the same time the spatio-temporal information provided by the EEG signals. In this thesis, we introduce a 2D Convolutional Recurrent Neural Network (2D CNN-LSTM) for precisely identifying human intended movements by exploiting both the spatial features and the temporal dynamics of 'raw' EEG data. In order to achieve this, we introduce a transformation that is able to convert the 'raw' EEG signals into 3D structures based on the spatio-temporal analysis of electrode space on the scalp region. In addition, we design a 3D CNN that uses the same input data with the 2D CNN-LSTM model and an 1D CNN model that utilizes only the 'raw' EEG signals, without any kind of transformation, and we evaluate them under similar experimental conditions. Both the 3D CNN model and the 2D CNN-LSTM model outperform the 1D CNN model by more that 30% in terms of accuracy, with the 2D CNN-LSTM model demonstrating the best performing capabilities of all (95.52%). Finally, for the sake of comparison, we test and evaluate two of the most common approaches in the field of BCIs, Common Spatial Patterns algorithm and Wavelet Packet Decomposition, relatively to the Deep Learning models in terms of performance.
|
|---|
