Investigation of damage size and orientation effects in active-sensing SHM via statistical non-parametric time series methods

In today’s world, engineering structures are subject to many sources of uncertainty, from varying operational and environmental conditions to complex damage evolution patterns. As such, in the context of active-sensing guided-wave Structural Health Monitoring (SHM), these uncertainties must be initially accounted for, where probabilistic SHM methods are developed for robustly analyzing complex systems, accurately reaching a decision, and properly defining confidence intervals on that decision. To this end, stochastic time series methods have attracted the interest of the vibration-based SHM community in order to overcome the above challenges. Advantages such as a high accuracy in modelling system behavior and dynamics, as well as robustness against uncertainties make these methods attractive for SHM applications. The goal of this work is the introduction and critical assessment of a probabilistic active-sensing guided-wave SHM framework for damage detection, localization and quantification based on the parametrization of damage state-that is, damage size and orientation-via the use of integrated nonparametric time series representations and Bayesian Gaussian Process (GP) models. The main novel contribution of this work involves the parameterization of wave propagation patterns and corresponding damage-sensitive features under different damage states via the integration of statistical models and Bayesian learning. The proposed framework is validated using a series of Spectral Element Model (SEM) based simulations and experimental data from a stiffened Aluminum (Al) panel and a carbon fiber-reinforced polymer (CFRP) coupon. Results are compared with those of two state-of-the-art Damage Indices (DIs). Stochastic, nonparametric time series representations are proposed for identifying the wave propagation paths within a “hotspot” configuration that are most sensitive to damage. Bayesian Gaussian Processes are used to represent the nonlinear functional mapping between the time series models/traditional DIs and damage size/ orientation