<p>Deep learning has been increasingly applied in electroencephalogram (EEG) artifact research. However, existing research is generally limited to detecting specific types of artifacts and suffers from a shortage of labeled data. To address these issues, a multi-artifact detection method based on channel masking and a semi-supervised network across multiple feature domains was proposed in this paper. Firstly, the channel correlations of five different types of artifacts were analyzed, and their features were utilized to re-label and mask the channel tags. Then, a semi-supervised network with an attention mechanism was used to learn features from various feature domains in the data. Finally, a custom multi-class loss function was applied to further enhance the model’s focus on artifacts, aiming to reduce the impact of noisy labels. The EEG data used in this study were collected from patients diagnosed with benign epilepsy with centro-temporal spikes (BECTS) and sourced from the dataset provided by the Children’s Hospital of Zhejiang University School of Medicine (CHZU). Compared with recent multi-artifact detection approaches and commonly used classification loss functions, the proposed method achieved the shortest training time and an average inference latency of 35.52 ms per 1-second EEG segment, enabling real-time processing in continuous EEG monitoring scenarios. Moreover, it obtained an F1 score of 90.99% and a recall of 86.32%.</p>