Probabilistic damage quantification via the integration of non-parametric time-series and Gaussian process regression models

One of the major challenges the structural health monitoring (SHM) community is facing today is related to enabling accurate and robust damage detection and quantification under varying operating and environmental conditions with a limited number of sensing data sets. In particular, state-of-the-art active-sensing SHM technologies face significant difficulties in accurately detecting and quantifying damage within uncertain environments due to the lack of appropriate statistical learning and inference methods. As such, it is critical for the community to have available tools that accurately detect and quantify damage, in a probabilistic sense, using the –oftentimes limited– available information. In this study, a novel probabilistic damage detection and quantification approach is proposed based on the integration of non-parametric statistical time series representations and Gaussian Process Regression Models (GPRMs). Initially, non-parametric models based on Short-Time Fourier Transform (STFT) power spectral density (PSD) estimates are used in order to detect damage and statistically determine signal (wave) paths that carry the most information about damage size. Next, GPRMs are trained using Damage Index (DI) values from selected wave propagation paths and used to estimate the damage size based on the current DI sets. The experimental assessment is presented for data recorded from a notched aluminum plate. It is shown that, using wave propagation paths selected by non-parametric statistical models, estimation accuracy increases and damage size estimation confidence levels become narrower, thus providing a robust and efficient damage detection and quantification approach.