Bayesian Information Fusion for Fatigue Crack Growth Diagnosis using Ultrasonic Guided Wave Pitch-catch in a Piezoelectric Actuator-sensor Network

Abstract

Diagnosis of fatigue crack growth in thin metallic structures using guided ultrasonic waves has been shown to be a promising methodology for aerospace and structural engineering applications. Typically, the non-destructive evaluation (NDE) tests are performed using the pitch-catch mode, where Lamb waves are induced in the metallic structure using a PZT actuator, and the wave field at a suitable location on the structure is measured using a PZT sensor. The NDE methods employ a model-based, baselinereference approach, where any changes in the sensed signal are assumed to be caused by fatigue crack growth. These NDE methods, however, are prone to inaccuracies due to aleatory uncertainty in system properties as well as physics model discrepancy. Probabilistic treatment of the diagnosis faces two key challenges: a large parameter space, and computationally expensive numerical models of the governing physics. In this work, we discuss a probabilistic crack diagnosis framework to overcome the aforementioned challenges; and to tackle the aleatory and epistemic uncertainties in the process. We build a Bayesian network that describes a Lamb-wave pitch-catch NDE method using a low-fidelity, physics-based model of the same. We perform global sensitivity analysis to quantify the contribution of various parameters to the variance of the damage-sensitive output signal feature(s) using this model. We retain the parameters with higher contribution, and build a medium-fidelity, one-way coupled, multi-physics finite element model to simulate the piezoelectric effect and Lamb wave propagation. We use the finite element model to generate training data to train a Gaussian process (GP) surrogate model that can output the damage-sensitive signal feature for given values of pertinent system parameters. We use the GP surrogate to perform Bayesian diagnosis of crack size considering data corrupted by measurement and process noise. We fuse the information obtained from various actuator-sensor paths in a PZT network to sequentially update the crack growth estimate. The proposed Bayesian estimation and fusion approach can potentially improve the performance of Lamb-wave pitch-catch NDE methods for metallic structures.