Corrosion Detection on Aircraft Fuselage with Multi-teacher Knowledge Distillation

The procedures of non-destructive inspection (NDI) are employed by the aerospace industry to reduce operational costs and the risk of catastrophe. The success of deep learning (DL) in numerous engineering applications encouraged us to check the usefulness of autonomous DL models also in this field. Particularly, in the inspection of the fuselage surface and search for corrosion defects. Herein, we present the tests of employing convolutional neural network (CNN) architectures in detecting small spots of corrosion on the fuselage surface and rivets. We use a unique and difficult dataset consisting of \(1.3\times 10^4\) images (\(640\times 480\)) of various fuselage parts from several aircraft types, brands, and service life. The images come from the non-invasive DAIS (D-Sight Aircraft Inspection System) inspection system, which can be treated as an analog image enhancement device. We demonstrate that our novel DL ensembling scheme, i.e., multi-teacher/single-student knowledge distillation architecture, allows for 100% detection of the images representing the “moderate corrosion” class on the test set. Simultaneously, we show that the proposed ensemble classifier, when used for the whole dataset with images representing various stages of corrosion, yields significant improvement in the classification accuracy in comparison to the baseline single ResNet50 neural network. Our work is the contribution to a relatively new discussion of deep learning applications in the fast inspection of the full surface of an aircraft fuselage but not only its fragments.