LHR-STD: Learning Heterogeneous Representations through Spatiotemporal Decoupling

Multivariate time series analysis faces the challenge of modeling heterogeneous temporal correlations. This paper proposes to decouple the learning of heterogeneous representations from spatiotemporal perspectives, introducing a time-varying learning and spatial iteration framework. The methodology comprises two core components: Temporal Correlation Learning (TCL) addresses granularity imbalance in long-term dependency modeling through multi-scale convolutions and multi-head attention fusion, while Association Iteration Modeling (AIM) dynamically updates node association paths via graph attention with sparse pruning to suppress weak correlation propagation, thereby reducing model complexity. Experiments across five industrial datasets demonstrate an average 4.87% F1-score improvement over nine baseline methods, validating the method’s effectiveness in parsing complex correlation systems.