Supervised Single-channel EEG Decomposition using Detector-kernel Networks for Noise Reduction

Signal decomposition techniques are ubiquitously utilized for the analysis, noise reduction, and classification of electroencephalogram (EEG) signals. Predominantly, these techniques harness the spatial features of EEG signals, thereby targeting multi-channel signals. However, decompositions employing spatial features often prove inefficient when applied to scenarios involving limited channel recording, such as with consumer-use equipment. This paper presents a novel approach for single-channel EEG signal decomposition. Rather than relying on a multitude of electrodes, we leverage training samples to cultivate a decomposition model based on convolutional neural networks. Our model hypothesizes that an observed EEG signal has essential components called kernels and comprises the time-shifted and amplitude-modulated kernels. Accordingly, the convolutional neural network-based signal decomposition model estimates these kernels and identifies their occurrence positions, thereby pinpointing time-shift and amplitude modulation. We validated our proposed decomposition method in the context of a noise reduction problem within a BCI application. Our experiment substantiated that our method effectively decomposes noise-related and non-noise-related components.