SWTT: A Hybrid Stationary Wavelet Transform and TCN-Transformer Network for EEG Artifact Removal

Electroencephalography (EEG) is widely used in clinical diagnosis, cognitive neuroscience, and brain–computer interfaces. During acquisition, it is inevitably affected by artifacts from electrooculography (EOG) and electromyography (EMG), severely degrading signal quality and impacting subsequent analysis. In practice, EEG is often contaminated by multiple artifacts simultaneously, and the degree of contamination can vary among different artifacts. Therefore, artifact removal is an important task for improving the quality of EEG and ensuring the reliability of subsequent analysis. However, most existing methods only consider single artifact removal scenarios. Even when considering mixed artifact scenarios, these methods often fail to effectively utilize the frequency characteristics of different artifacts, thus limiting their ability to accurately separate and suppress artifacts in the frequency domain. This study aims to establish an effective and robust EEG artifact removal method that can suppress both EOG and EMG artifacts while preserving EEG. To this end, we develop the stationary wavelet TCN-Transformer network (SWTT), a deep neural network based on SWT, TCN, and Transformer, designed for removing EEG artifacts. SWTT first uses SWT to decompose the noisy signal into time-aligned multi-scale time-frequency coefficients, which can separate different types of noise to some extent and make it easier to extract effective features. Then, a two-branch network is used for signal recovery in the wavelet domain. Specifically, the TCN branch uses dilated convolution to extract local temporal patterns, while the Transformer branch models cross-band global dependencies through a self-attention mechanism. Finally, the processed coefficients are used to reconstruct the time-domain signal using inverse SWT and then refined using a lightweight time-domain module to obtain the final result. Experiments on benchmark datasets demonstrate that SWTT achieves state-of-the-art performance in both single and mixed artifact removal tasks. Furthermore, because EOG and EMG artifacts exhibit distinct frequency-domain distributions, the network, by learning from the multi-scale time-frequency coefficients provided by SWT, can automatically capture these different time-frequency characteristics, and we observe that the model demonstrates distinct recovery behaviors for different artifact types.