Consistency-Weighted Moving Average and Representation alignment for Domain Generalization in EEG Motor Imagery Decoding

Zihao Chen^*Jiaming CaiBohao ZhangShuo ZhangJiawei WangJun Li

• Hebei University, Baoding, China

Cross-subject domain variability in EEG signals causes instability in feature extraction and a decline in classification performance. To overcome these limitations, we propose CW-MARA, a domain generalization framework that mitigates cross-subject variability by stabilizing model updates and enforcing consistent representation alignment, enabling the reliable extraction of domain-invariant EEG features even in the absence of target-domain data.} The proposed method employs a Consistency-Weighted Moving Average (CWMA) framework that adaptively updates the teacher model parameters, using the cosine similarity between student and teacher model parameters to control the updated strength. To enhance feature robustness and generalization, we design a Representation alignment block (RAB) to directly regularize feature representations, thereby improving cross-subject generalization. It includes two feature-space contrastive objectives: the Contrastive Large-Margin Loss (CLML), which enhances intra-class compactness and inter-class separability using Mixup, and the Multi-Grain Contrastive Loss (MGCL), which captures cross-subject consistency across temporal and spectral domains. textcolor{red}{Our extensive evaluation on BCI-IV 2a, 2b, OpenBMI, and self-collected datasets shows that CW-MARA consistently surpasses existing methods, improving robustness and cross-subject generalization, and thus contributing to more reliable MI-BCI systems.