Developing abnormal detection models based on digitalization images of 3D aircraft for maintenance inspections

Abstract The aviation industry is rapidly growing, and the increasing density of operational aircraft has led to higher demands for aircraft maintenance, resulting in a significant workload for technicians. The application of machine learning technologies, especially Convolutional Neural Networks (CNNs), can assist engineers in timely finding problems on the aircraft’s surface, aiming to reduce the time needed for visual inspections and increase the evaluation process’s accuracy. Inheriting from the previous 3D aircraft modeling dataset, we developed user-friendly and completely free open-source You Only Look Once (YOLO) and transformer-based object detection (RF-DETR) models through various versions to create a promising approach for this challenge. Desiring high reliability, ease of use, and a short training time, this paper provides an overview of the workflow for building and in-depth evaluating the anomaly detection models. The article also delivers the capability of those models in this context and includes a case study of airplane passenger windows, building on an integrated image dataset of 3D aircraft digitalization with surface change detection, all at a low cost.