Automated Defect Classification in Aerospace X-ray Images Using CNNs

Non-destructive testing (NDT), through radiography, is used to reveal internal structures and defects without causing damage to inspected components [1]. Analysing X-ray images helps determine the presence and characterise the extent of damage or anomalies in various composite parts. In particular, these tests are crucial in manufacturing composite materials [2], such as helicopter rotor blades [3]. Traditional NDT methods rely on manual inspection by human experts, which can be time-consuming and inaccurate. By leveraging automatic learning-based NDT approaches, the testing process can be faster and more accurate than a human operator: reported human detection accuracies range from 70% to 80% [4]. By automatically extracting complex multi-scale features, deep-learning-based approaches have enabled significant progress in analysing industrial (aircraft) composite parts [5]. This work leverages deep learning, particularly CNN-based approaches, to detect defects in helicopter composite parts through a binary classification approach of X-ray images. The potential of deep learning-based approaches is investigated in the specific context of defective helicopter rotor hubs (suffering from the ripple defect). Different CNN models are studied, and an end-to-end image processing and classification methodology is proposed. The impact of generalization methods is also investigated. Finally, we show that the proposed pipeline enables generalization to images acquired by new sensors. To the best of our knowledge, the generalization of the defect classification model to the images acquired using new sensor parameters is lacking in the literature. Additionally, the utilization of defect data from helicopter rotor hubs for CNN-based applications with high accuracy is limited.