CD3Net: A Contrastive Diffusion Model with Domain Adaptive Data Synthetic Network for Motor Imagery and Emotion classification

Brain-computer interfaces (BCI) has experienced considerable progress in recent years, highlighting the central role of electroencephalograms (EEG) in this technology. However, raw EEG data contains significant noise, which can be divided into subject-specific and task-specific characteristics. To adaptively reduce noise, this paper proposes the Contrastive Diffusion Model with Domain Adaptive Data Synthetic Network(CD 3 Net), a contrastive diffusion model for synthesizing domain adaptation EEG data. Firstly, we separated the clean module and the noise module in the reverse process of diffusion to achieve denoising , then obtained the synthesized data. Secondly, in the clean module, a Cell-ProUNet based Clean Signal Extraction module is proposed to effectively capture subtle feature changes in EEG signals. In the noise module, to learn individual noise based on different tasks, a novel module called the Domain Noise Adaptive Extraction module is introduced. At last, we obtained the new synthesized data through CD 3 Net. Experimental results achieve better performance in EEG classification tasks when compared to other existing models, highlighting the efficiency of CD 3 Net in improving the reliability of data synthetic. The code is available at https://github.com/xiyuanssxx/CD3NET . • The proposed CD 3 Net can remove noise and synthesize clean EEG signals. • The Cell-ProUNet can expand the feature space and address the issue of gradient vanishing during model training to extract clean signals. • The Domain Noise Adaptive Extraction can adaptively extract task noise and subject noise simultaneously. • The Cross Domain Attention mechanism enhances the focus on different domains under different tasks. • Contrastive Denoise Module ensures further separation of clean signals from noise.