Robust Time Series Contrastive Representation Learning Via Explicit Seasonal-Trend Disentanglement

We address universal self-supervised representation learning for time series under label scarcity and distribution shifts. A key challenge is that prevalent encode-then-decompose pipelines only implicitly separate seasonal and trend signals, which mixes noise across components and degrades robustness, especially with sudden jumps or stochastic movements. In this work, we propose an explicit, decomposition-aware contrastive framework: original time series are split via a Fourier transform into seasonal (high-frequency) and trend (low-frequency) components. Seasonality is modeled by a frequency-enhanced attention encoder with amplitude-induced augmentations; trend is captured by a tree-structured causal convolution encoder with random perturbations to handle non-stationary drift. The two views are trained contrastively and then fused with lightweight fine-tuning, delivering state-of-the-art accuracy, improved robustness to jumps, and interpretable attributions.