Modular Bayesian damage detection for civil structures. A proof of concept case-study

In the last decades a significant portion of research in Structural Health Monitoring has been developed upon the principles of vibration-based methods, where monitored modal properties have been treated as a set of features for damage detection. Unfortunately, factors external to the structural system, such as environmental/operational effects, have been shown to mask relevant anomalous patterns. This paper presents a modular Bayesian damage identification framework which considers external effects explicitly, along with other sources of uncertainty. A calibrated finite element model of the Tamar bridge is used in order to identify anomalous features in the bridge main/stay cables and its bearings. Displacements, natural frequencies, temperature and traffic monitored throughout a year are used to form a reference baseline, which is compared against a posterior identification with one month of monitored data. The proposed framework allows to account for observation errors, estimation of damage and model discrepancy of the predictive model. Multiple response Gaussian processes emulate the model response surface and its discrepancy enhancing the identification task while minimising costly computations. Results indicate that the main cables and bearings have no structural defects. Although the stay cables strain increased considerably, this could be attributed to the inherent variability of their properties. Finally, a considerable amount of uncertainty is propagated when estimating the proposed damage metric, and some options to reduce it are discussed.