Active-sensing structural health monitoring via statistical learning: an experimental study under varying damage and loading states

Active-sensing acousto-ultrasound structural health monitoring (SHM) constitutes an important family of methods for both metallic and composite structures. However, the presence of varying operational and environmental conditions in the real world can significantly affect their accuracy and robustness in the face of uncertainty. In this context, statistical learning methods that can be based on Gaussian Process regression models (GPRMs) and statistical time-series models can be incorporated in the damage diagnostic process to account for and properly represent such uncertainties. Towards this end, the main objective of this paper is the postulation and experimental assessment of two statistical learning approaches, based on GPRMs and time-varying time series models, for active sensing SHM under varying structural and loading states under uncertainty. The proposed methods involve GPRM representation of the non-linear mapping between the actual states with (i) traditional damage indices (DIs) and (ii) parameters of time-dependent autoregressive (TAR) models. The experimental validation and comparative assessment is based on a series of experiments on an aluminum coupon outfitted with a network of piezoelectric actuators/sensors subjected to different static loads under increasing damage size.