An interpretable physics-consistent neural framework for wave-based system identification of buildings

Structural health monitoring (SHM) is important for rapid post-earthquake condition assessment and resilience-oriented management of civil structures. Among system identification methods, wave-based approaches are attractive because they are sensitive to localized stiffness changes and may reduce some limitations of global modal indicators, including potential influences associated with soil-structure interaction. However, many artificial-intelligence (Al)-based identification methods remain difficult to interpret physically, which limits their reliability in engineering applications. This study proposes an interpretable physics-consistent neural framework (PCNF) for wave-based system identification of buildings. The PCNF is derived directly from the layered Timoshenko beam formulation, in which the state transition of each structural layer is mapped onto a neural computational graph with physically meaningful trainable parameters. Structural parameter identification is therefore reformulated as a physics-guided gradient-based learning problem. The PCNF is applied to a 54-story office building in Los Angeles using records from nine earthquakes. The identified layer-wise shear-wave velocities exhibit coefficients of variation not exceeding 5% across the nine earthquakes and are generally more stable than those reported by established wave-based techniques. These results suggest that the proposed PCNF provides a stable, physically interpretable, and AI-integrated framework for wave-based structural system identification and monitoring of instrumented high-rise buildings.