Deep Learning Based Crack Detection in Inhomogeneous X-Ray Images for High Pressure Turbine Blades in Aviation

To ensure safety in aviation, there are stringent requirements for aircraft components, not only in design and manufacturing but also their maintenance. For economic reasons, the maintenance process is intended to be as efficient as possible while still meeting required safety standards. The increase in automation level contributes to enhancing the economic efficiency of the maintenance process. A highly time-consuming step involves the manual assessment of X-ray images. Advances in deep learning algorithms suggest a promising prospect for employing these algorithms to automatically inspect X-ray images. This paper introduces a deep learning-based approach to enhance the level of automation in the inspection of X-ray images of high-pressure turbine blades by automatically detecting the presence of internal axial cracks. We encountered challenges posed by the heterogeneous X-ray images with very small and variable damage patterns, along with the issue of limited available data. To address these challenges, we applied damage mapping to define a region of interest and used the concept of transfer learning. Three different model architectures of deep neural networks (Resnet18, Inception-V3 and Densenet161) were compared, and it is demonstrated that all three are well-suited (average AUPRC > 0.95 and average F2-Score > 0.93) for our task. We are therefore able to help increase the degree of automation of the HPT blade X-rays’ inspection process and thus make the maintenance process more efficient.