Hybrid structural health monitoring using data-driven modal analysis and model-based Bayesian inference.

Abstract

Civil infrastructures that are valuable assets for the public and owners must be adequately and periodically maintained to guarantee safety, continuous service, and avoid economic losses. Vibration-based structural health monitoring (VBSHM) has been a significant tool to assess the structural performance of civil infrastructures over the last decades. Challenges in VBSHM exist in two aspects: operational modal analysis (OMA) and Finite element model updating (FEMU). The former aims to extract natural frequency, damping ratio, and mode shapes using vibrational data under normal operation; the latter focuses on minimizing the discrepancies between measurements and model prediction. The main impediments to real-world application of VBSHM include 1) uncertainties are inevitably involved due to measurement noise and modeling error; 2) computational burden in analyzing massive data and high-fidelity model; 3) updating structural coupled parameters, e.g., mass and stiffness. Bayesian model updating approach (BMUA) is an advanced FEMU technique to update structural parameters using modal data and account for underlying uncertainties. However, traditional BMUA generally assumes mass is precisely known and only updating stiffness to circumvent the coupling effect of mass and stiffness. Simultaneously updating mass and stiffness is necessary to fully understand the structural integrity, especially when the mass has a relatively large variation. To tackle these challenges, this dissertation proposed a hybrid framework using data-driven and model-based approaches in two sequential phases: automated OMA and a BMUA with added mass/stiffness. Automated stochastic subspace identification (SSI) and Bayesian modal identification are firstly developed to acquire modal properties. Following by a novel BMUA, new eigen-equations based on two sets of modal data from the original and modified system with added mass or stiffness are derived to address the coupling effect of structural parameters, e.g., mass and stiffness. To avoid multi-dimensional integrals, an asymptotic optimization method and Differential Evolutionary Adaptive Metropolis (DREAM) sampling algorithm are employed for Bayesian inference. To alleviate computational burden, variance-based global sensitivity analysis to reduce model dimensionality and Kriging model to substitute time-consuming FEM are integrated into BMUA. The proposed VBSHM are verified and illustrated using numerical, laboratory and field test data, achieving following goals: 1) properly treating parameter uncertainties; 2) substantially reducing the computational cost; 3) simultaneously updating structural parameters with addressing the coupling effect; 4) performing the probabilistic damage identification at an accurate level.