Aircraft Fuselage Corrosion Detection Using Artificial Intelligence

Abstract

Corrosion identification and repair is a vital task in aircraft maintenance to ensure continued structural integrity. Regarding fuselage lap joints, typically, visual inspections are followed by non-destructive methodologies, which are time-consuming. The visual inspection of large areas suffers not only from subjectivity but also from the variable probability of corrosion detection, which is aggravated by the multiple layers used in fuselage construction. In this paper, we propose a methodology for automatic image-based corrosion detection of aircraft structures using deep neural networks. For machine learning, we use a dataset that consists of D-Sight Aircraft Inspection System (DAIS) images from different lap joints of Boeing and Airbus aircrafts. We also employ transfer learning to overcome the shortage of aircraft corrosion images. With precision of over 93%, we demonstrate that our approach detects corrosion with a precision comparable to that of trained operators, aiding to reduce the uncertainties related to operator fatigue or inadequate training. Our results indicate that our methodology can support specialists and engineers in corrosion monitoring in the aerospace industry, potentially contributing to the automation of condition-based maintenance protocols.

Keywords: aircraft corrosion inspection; automatic corrosion detection; aviation maintenance; corrosion science; deep learning; material fatigue; rust detection.

Figure 1

An example showing pillowing corrosion and surface micro deformations. (a) A specialist using a flashlight performs fuselage visual inspection; (b) the fuselage photo scanned showing surface micro perturbations caused by corrosion.

Figure 2

D-Sight Aircraft Inspection System (DAIS). (a) DAIS 250c, and (b) schematic of DAIS functioning.

Figure 3

Two samples taken with the DAIS 250c device of a simple shear lap joint. (a) An image with no corrosion; and (b) the corrosion pillowing affecting rivets in red color.

Figure 4

Wing inspection images created with DAIS 250C. Images are composed as a mosaic to cover the entire wing. Each image sample is enumerated in order to control the inspection. A specialist then perform the inspection of each sample.

Figure 5

Learning curves for training and testing using the DenseNet architectecture, whose accuracy achieved the top results. Each plot corresponds to a five-fold split.

Figure 5

Learning curves for training and testing using the DenseNet architectecture, whose accuracy achieved the top results. Each plot corresponds to a five-fold split.

Figure 6

Learning curves for training and testing using the SqueezeNet architectecture. Although it is considered a light weight architecture, SqueezeNet achieved the second best results. Each plot corresponds to a five-fold split.

Figure 7

Loss curves with error bars depicting the loss for the epochs for the different architectures evaluated: DenseNet (a), ResNet (b), SqueezeNet (c), and InceptionV3 (d).

Figure 8

Examples of images that the Deep Learning architecture DenseNet-201 did not predict well. Column (a) corresponds to images misclassified with corrosion or false positives (FP), while column (b) includes images with misclassified corrosion detected or false negatives (FN). In (c), we use Grad-CAM for the visual interpretability of the DenseNet-201’s gradients over the image in (b).

Figure 9

Visualization results through class-activated maps overlaid on input true positive images together with the raw images. The maps were extracted from the last convolution layer using the algorithm GradCam for DenseNet-201 (first two rows) and SqueezeNet (last two rows). The heat maps stand for the larger heights learned during the training.