A Hierarchical Learning-Based Approach for the Automatic Defect Detection and Classification of AFP Process Using Thermography

This paper presents a novel learning-based approach for automatic detection and classification of production defects in the automated fiber placement (AFP) process using thermography as a solution to the error-prone and time consuming manual inspection problem. We introduce a hierarchical framework designed to achieve reliable performance, optimize computational resources, and address challenges such as inherent data imbalance. First, a high-level lay-up status classifier, that utilize traditional vision and classical machine learning algorithms, is fetched to decide whether a lay-up region is healthy or defective. Then, a low-level model based on a proposed deep learning architecture classifies the defective instance into specific defect classes. A comprehensive thermal image database, including both natural and synthetically induced defect experiments, is built and used to train, test, and evaluate the models. The performance of each classification level is analyzed individually, yielding promising results with accuracy rates exceeding 95%. Moreover, the proposed approach demonstrates real-time operation capability.