Physics-Guided Machine Learning for Condition Assessment of Building Structures in Operational Environments

Abstract

Structural condition assessment is critical for ensuring the safety and functionality of building structures, yet existing methods face significant challenges, including data scarcity, noise contamination, and limited generalisation across diverse operational environments. Traditional machine learning-based approaches often rely on extensive labelled datasets and assume consistent data distributions, which are impractical in real-world scenarios. Furthermore, these methods frequently lack interpretability, limiting their adaptation to practical applications. To address these issues, advanced frameworks are required to enhance accuracy, robustness, and scalability in structural damage detection and condition assessment.

A series of physics-guided machine learning frameworks are developed in this research to overcome these above-mentioned challenges, mainly including transfer learning and physics-informed machine learning. Transfer learning methods leverage simulated frequency response function (FRF) data to pre-train deep convolutional neural networks (CNNs) and fine-tune them using limited real-world measurements, significantly improving damage localisation and severity identification. Additionally, a Joint Maximum Discrepancy and Adversarial Discriminative Domain Adaptation (JMDAD) framework is developed to eliminate the need for labelled target data. By aligning feature distributions at both domain and class levels and leveraging transmissibility functions, this approach enhances robustness against noise and environmental variations while effectively detecting damage in real structures.

Physics-informed machine learning methods further embed physical constraints into machine learning models to improve interpretability and reliability. The Parallel Neural Ordinary Differential Equations (PNODEs) framework integrates state-space equations to provide physical constraints, enabling accurate damage quantification and enhanced model reliability. Additionally, the Temporal-Spatial Neural Operator (PhySTN) framework combines a spatial feature mapping encoder with a physics-informed time operator to enable structural parameter identification and response reconstruction from sparse sensor data, addressing challenges in data insufficiency.

The proposed frameworks are validated through extensive numerical simulations and experimental studies, including nonlinear numerical models, experimental structures, benchmark frames, and real-world applications. These methods demonstrate significant improvements in damage detection accuracy, scalability, and interpretability, offering reliable and efficient solutions for structural health monitoring. By addressing the challenges of insufficient data and enhancing the explainability of machine learning-based condition assessment, this research contributes valuable advancements to the field.