Enhancing CFRP delamination detection: A Bayesian neural network approach Integrating spatial and temporal information

Non-destructive testing (NDT) plays a pivotal role in ensuring the structural integrity of Carbon Fiber Reinforced Polymers (CFRP). We propose an innovative methodology that leverages spatial and temporal dimensions of thermal imaging data to enhance the precision and reliability of delamination detection. The approach involves training a Multiple Input Multiple Output Convolutional Bayesian Neural Network (MIMO-CBNN) using as an input direct model that generates two sequences of temporal samples from top and bottom thermal cameras. We systematically connect corresponding samples from the top and bottom cameras. Each paired sample captures the temporal evolution. To further exploit temporal dynamics, we introduce the concept of linking residual images between the two input datasets. This enhances the network's ability to discern nuanced changes in temperature patterns over time, contributing to a more comprehensive understanding of defect evolution. Integrating spatial and temporal information not only refines the network's feature extraction [2] but also significantly reduces output result variance, resulting in more accurate and reliable predictions. The proposed MIMO-CBNN is to address classification and regression tasks simultaneously. The classification output spatially maps defects by identifying whether a pixel belongs to the delamination or not. The regression output provides precise coordinates and dimensions of the delamination, offering quantitative insights into the defect's location and extent. Extensive numerical simulations illustrate the above procedure.