Exploration on Surrogate Models for Inverse Identification of Delamination Cracks in CFRP Composites using Electrical Resistance Tomography

摘要

Non-destructive evaluation (NDE) techniques to detect and measure internal inter-ply delamination and intra-ply matrix cracking damage are needed for Carbon Fiber Reinforced Polymer (CFRP) materials used in aerospace structures. The electrical resistance tomography (ERT) is a NDE technique that uses the inherent changes in conductive properties of the composite to characterize damage. Identification of damage requires solving the inverse problem that minimizes the difference between the model predicted and the measured change in resistance at specified electrode locations. Using numerical finite element models of the laminate directly in the inverse identification is computationally expensive and requires the development of an accurate surrogate model. This research investigates efficient numerical modeling techniques for inverse identification of delamination damage location and size in composite laminates using ERT based NDE. Traditional ERT approaches are focused on damage detection. For structural health prognosis, in addition to detection, the inverse identification also has to accurately quantify the damage. This paper explores the use of Response Surfaces and Kriging approximations for surrogate models applied on ERT. Since the electrical resistance change measurements across the different electrode pairs for given damage state could be correlated, Singular Value Decomposition (SVD) is used to identify the principal components for surrogate model fitting and dimension reduction. Different options for fitting surrogates, such as direct fit to individual resistance measurements versus fit to the contributions of the principal components to the resistance changes are compared. It was found that the surrogate models with best prediction accuracy and inverse identification results were obtained using Kriging. It was also found that surrogate models created by fitting the reduced coefficient matrix obtained from SVD exhibit similar performance than surrogate models created by directly fitting resistance/resistance change values.