Towards probabilistic data-driven damage detection in SHM using sparse Bayesian learning scheme

Despite continuous evolution and development of structural health monitoring (SHM) technology, interpreting a huge amount of sensed data from a sophisticated SHM system to extract useful information about structural health condition remains a challenge. Aiming to resolve this problem, a novel application of probabilistic data-driven damage detection method was proposed in the context of Sparse Bayesian Learning (SBL) scheme. The framework involves constructing a new structural damage index and establishing SBL regression model as reference base only using the data acquired in health state. The construction of the structural damage index is based on damage-sensitive frequency band, which is determined by NExT using vibration monitoring data. The structure will be classified to be damaged as the structural damage index based on new data deviates from the index predicted by SBL regression reference model, and further, the Bayes factor is adopted to quantify the damage degree. In addition, the relationship between the Bayes factors and the resonance frequency change rate is investigated in detail. The proposed methodology features the following merits: (i) It is probabilistic data-driven method exempting from physical model of the structure, excitation/loading information, and (ii) it belongs to the unsupervised model in need for structural damage detection, which can be formulated using only monitoring data from health state in the absence of monitoring data from damaged state. Damage detection and discrimination capabilities of the proposed methodology are verified using field monitoring data acquired from a cable-stayed bridge. Finally, a discussion of the SBL-based approach is made and further challenges pertaining to damage detection processes in the context of SHM are identified.