EEG Verilerinden Gürültü Giderme

Abstract

Electroencephalography (EEG) is a vital tool for non-invasive brain activity monitoring, widely used in clinical and research settings, but often contaminated by noise from muscle movements, eye blinks, and electrical interference, which can obscure neural information. This thesis explores advanced machine learning techniques, focusing on autoencoders with Neural Ordinary Differential Equations (NODEs) and Residual Networks (ResNet), to enhance EEG denoising. While traditional methods like Independent Component Analysis (ICA) have been effective in separating EEG signals from artifacts by leveraging statistical independence, they struggle with the dynamic and nonlinear nature of EEG data. To overcome these limitations, this research integrates autoencoders with NODEs and ResNet, combining autoencoders' dimensionality reduction with NODEs' continuous-time dynamics and ResNet's skip connections to handle the complexity of multivariate EEG signals. The proposed hybrid framework significantly improves denoising accuracy, computational efficiency, and adaptability to different noise levels in bio-signals, outperforming traditional methods. Results, evaluated through metrics like Mean Squared Error (MSE), Relative Root Mean Squared Error (RRMSE), and correlation coefficients, show substantial improvements in noise removal for both synthetic and real EEG datasets, marking a significant advancement in EEG signal processing. Keywords: Electroencephalography (EEG), Denoising, Machine Learning, Independent Component Analysis (ICA), Neural Ordinary Differential Equations (ODEs), Residual Network, Autoencoders, Signal Processing, Brain Waves, Noise Removal