The electroencephalogram (EEG) provides essential data for analyzing brain activities. However, artifacts such as electrooculography (EOG) and electromyography (EMG) often interleave with the EEG signals, significantly affecting the quality of EEG signal analysis. The heterogeneous distribution of these artifacts in the time-frequency domain makes it challenging to remove multiple artifacts using a unified model. In this paper, we propose an artifact-aware EEG denoising model, referred to as \(A^2\) DM, to effectively remove various types of artifacts in a unified manner. We first obtain an artifact representation that indicates the type of artifact from a pre-trained artifact classification model. This artifact representation is then used as prior knowledge, which is fused into the denoising model in the time-frequency domain. This enables the model to become aware of the artifact type and precisely remove artifacts based on their type. Due to the heterogeneous distributions of artifacts in the frequency domain, we introduce a frequency enhancement module that can identify specific types of artifacts based on their representation and remove them using a hard attention mechanism. Additionally, we design a time-domain compensation module to enhance the denoising capability of \(A^2\) DM by compensating for potential losses of global information. Comprehensive experiments demonstrate that \(A^2\) DM significantly outperforms the novel CNN in denoising EEG signals, showing a notable 12% improvement in correlation coefficient (CC) metrics. This work demonstrates that artifact representation can be used in artifact removal models to effectively remove multiple types of artifacts.
