Physics-Informed Sparse Structure Learning for Generalizable Pattern Recognition with Distributed Acoustic Sensing System

Distributed acoustic sensing (DAS) systems in complex physical environments often exhibit poor generalization and limited interpretability, primarily due to strong noise, environmental variability, and dynamic system uncertainties. To overcome these challenges, we propose PISS (Physics-Informed Sparse Structure learning), a physics-informed framework that improves robustness and generalization for multi-class event recognition in DAS. Instead of relying on explicit governing equations, PISS leverages sparse symbolic regression to extract compact, physically meaningful dynamical expressions from a limited number of clean samples. These class-specific expressions act as soft structural priors, guiding model outputs to remain within a physically consistent and interpretable domain. To enhance robustness against real-world disturbances, we adopt a noise-augmented training strategy that simulates diverse environmental perturbations. This allows the model to learn invariant physical patterns while suppressing spurious noise artifacts. A Transformer-based architecture with hyperbolic embedding captures spatiotemporal dependencies while enforcing physical constraints. Extensive experiments on diverse industrial datasets demonstrate that PISS maintains over 90% classification accuracy under both cross-domain conditions and severe noise perturbations. In addition, SHAP-based interpretability analysis reveals that our model produces more consistent and semantically aligned attributions across events, indicating stronger decision transparency. Compared to leading industrial intrusion detection baselines, PISS delivers superior generalization, physical consistency, and interpretability-making it a reliable and deployable solution for real-world sensing systems operating under noisy and uncertain physical conditions.