D-STNet: Dynamic Spectrum-Time Fusion With Adaptive Covariance Alignment for Motor Imagery EEG Decoding

Abstract

Motor imagery electroencephalography (MI-EEG) decoding is limited by intersession nonstationarity and covariance mismatch. Existing alignment and deep decoding methods often use fixed preprocessing or unified feature learning, which limits adaptation to evolving covariance statistics and heterogeneous spectral–temporal patterns. We propose D-STNet, a dynamic spectrum-time fusion network that combines offline batch-level adaptive covariance alignment with dual-branch representation learning. Its static-dynamic Euclidean alignment (SD-EA) initializes a nontrainable covariance reference matrix from offline training data, updates it from unlabeled evaluation batches, and constructs a batch-specific whitening transform for covariance realignment while the trainable network parameters remain fixed. The backbone combines geometry-aware spectral learning (GASL) for covariance-aligned spectral–spatial representation, temporal gated multihead attention (TGMA) for waveform-level temporal dynamics, and dual-branch projection fusion (DBPF) for integration in a shared latent space. On brain–computer interface (BCI) Competition IV-2a, the High-Gamma Dataset, and BCI Competition IV-2b, D-STNet achieves accuracies of 83.00%, 94.36%, and 85.23%, outperforming the strongest baselines by 7.54, 1.63, and 1.47 percentage points, respectively. These results support offline covariance-statistics-level adaptive alignment as a compact strategy for MI-EEG decoding under covariance mismatch.