Physics-Data Synergy in Structural Health Monitoring: A Multi-Scale Graph Contrastive Framework With Temperature-Adaptive Fusion

Abstract

Operational Structural Health Monitoring (SHM) currently faces a tripartite dilemma: the scarcity of labeled damage data from real-world infrastructure, the masking effects of environmental variability, and the prohibitive computational costs associated with high-fidelity physics simulations. To resolve these conflicts, this article presents a physics-data dual-driven framework that harmonizes finite element analysis (FEA) with self-supervised contrastive learning. The architecture employs a novel dual-matrix strategy: a static, physics-based matrix encoding mechanical constraints derived from modal analysis, and a dynamic monitoring matrix learned via hierarchical InfoNCE objectives across node, subgraph, and global levels. To distinguish genuine structural degradation from benign environmental shifts, we introduce an attention-based temperature-adaptive fusion mechanism. This module dynamically modulates the weighting of physics-driven versus data-driven features based on observed thermal gradient patterns. Furthermore, by leveraging parametric damage simulations as negative training samples, the framework achieves open-set classification capabilities, enabling the detection of unforeseen failure modes. Extensive validation on three benchmark datasets (IASC-ASCE, Z24 Bridge, and Guangzhou Tower) demonstrates that this approach outperforms state-of-the-art methods by margins of 4.4-6.1%, achieving 94.2% accuracy on the IASC-ASCE benchmark with only 20% labeled data. Notably, the model exhibits exceptional label efficiency, retaining 89.7% accuracy with merely 5% labeled samples—effectively matching supervised baselines that require sixfold more annotation. This work establishes a robust methodology for unifying structural mechanics with deep representation learning, ensuring reliability under limited supervision and varying thermal conditions.