Sensitivity and Robustness of Neural Networks for Defect-Depth Estimation in CFRP Composites

Abstract

Carbon fiber reinforced composites are lightweight materials that possess a desirable mechanical performance manifested in high strength to weight ratio. However, due to their anisotropic and low thermal conductive nature; the detection of delamination and impact damages using thermography is still a challenging task. This research paper aims at conducting a comprehensive analysis of the neural network when used as a post-processor to quantify defects depths from thermograms of CFRP; a pulsed thermography inspection routine was used to generate the thermograms. A thorough study is conducted to compare and analyze the effect of the different parameters on the accuracy and the robustness of the neural network outcome. The optimized neural network architecture and hyper-parameters are realized based on the accuracy yields (defect depths estimate). This research explores the (i) network related parameters and the (ii) thermography-data specific factors such as frame rate, number of considered points per defect location, etc.; showing that the robustness of the neural network is found more sensitive to: the network architecture, noise, training hyperparameters and the type, amount and diversity of the training data.