Electrophysiological time series (ETS) are characterized by high dimensionality, strong temporal and spatial dependencies, inter-channel variability, and non-stationary signal patterns, all of which pose significant challenges for effective feature learning. Data augmentation is essential for improving model robustness and generalization. However, existing augmentation strategies often fall short in terms of generalizability and domain-specific sensitivity, limiting their ability to capture the complex, multimodal dynamics intrinsic to ETS signals. To address these challenges, we propose NeuroCEA, a novel contrastive representation learning framework tailored for ETS, which introduces Cross-domain Embedding Augmentation (CEA) to improve representation quality. Specifically, CEA utilizes a Cross-domain Spatial Structure (CSS) to project ETS into a unified space that captures dependencies across time, channel, numerical, and frequency domains—all essential for interpreting electrophysiological patterns. To ensure computational efficiency, we propose a compact 2D CSS matrix construction mechanism via dimensionality reduction indexing, which maintains the cross-domain structure while enabling efficient local feature extraction. Additionally, we introduce Contrastive Knowledge Distillation (CKD), which integrates CEA with a Structured Space Transformer (SST) in a teacher-student paradigm to construct high-quality contrastive sample pairs. This enhances both feature discrimination and generalization, while preserving the physiological integrity of the signals. Extensive experiments on ETS classiuficatio n benchmarks show that NeuroCEA outperforms state-of-the-art methods. To validate its generalization capability beyond ETS, we further evaluate NeuroCEA on general multivariate time series (MTS) tasks, including both classification and forecasting, demonstrating robust performance and scalability across diverse temporal domains.