Orthogonal Features Based EEG Signals Denoising Using Fractional and Compressed One-Dimensional CNN Autoencoder.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Electroencephalogram (EEG) signals are crucial for brain activity monitoring but are susceptible to noise, primarily muscle artifacts (MA).
• Traditional denoising techniques like decomposition, thresholding, and filtering have limitations in effectively removing noise while preserving signal integrity.
• The increasing use of portable, low-energy devices necessitates efficient deep learning architectures for real-time signal processing.

Purpose of the Study:

• To develop an advanced, fractional one-dimensional convolutional neural network (CNN) autoencoder for robust EEG signal denoising.
• To introduce and evaluate a novel hyper-parameter (α) controlling the fractional order for improved gradient-based learning.
• To compress the deep learning model using randomized singular value decomposition (RSVD) for efficient deployment on resource-constrained devices.

Main Methods:

• EEG signals were transformed into an orthogonal domain using Tchebichef moments prior to input into the CNN autoencoder.
• A fractional calculus approach was integrated into the CNN architecture, introducing a hyper-parameter (α) for fractional gradient calculation.
• The model's trainable parameters were compressed via the randomized singular value decomposition (RSVD) algorithm.

Main Results:

• Significant improvements in restored EEG signal quality were observed through tuning the fractional order hyper-parameter (α).
• The proposed fractional and compressed CNN autoencoder demonstrated superior performance compared to existing state-of-the-art denoising methods.
• Experiments conducted on standard EEG datasets (Mendeley and Bonn) validated the effectiveness of the proposed approach.

Conclusions:

• The fractional CNN autoencoder offers a promising and effective solution for denoising EEG signals contaminated by muscle artifacts.
• Model compression using RSVD enables efficient implementation on portable, low-energy devices without compromising denoising performance.
• The introduced fractional order hyper-parameter provides a valuable tool for optimizing EEG signal restoration quality.