A Comparative Study of Visual Identification Methods for Highly Similar Engine Tubes in Aircraft Maintenance, Repair...

Area of Science:

• Aerospace Engineering
• Computer Vision
• Machine Learning

Background:

• Manual identification of aerospace components is error-prone and inefficient, especially for visually similar parts.
• Existing automated identification methods struggle with aerospace components due to high visual similarity and lack of distinct features.
• The maintenance, repair, and overhaul (MRO) sector requires robust automated identification for critical parts like engine tubes.

Purpose of the Study:

• To investigate object-inherent property identification methods for aerospace components.
• To address the challenge of identifying visually similar engine tubes lacking distinctive markings.
• To evaluate the effectiveness of 2D image and 3D point cloud data approaches using digital image processing and deep learning.

Main Methods:

• Development and implementation of various identification methods using 2D image and 3D point cloud data.
• Application of digital image processing and deep learning techniques, including PointNet.
• Creation of a benchmark dataset with visually similar demonstrator tubes for performance evaluation.

Main Results:

• 3D approaches demonstrated clear superiority over 2D image analysis methods.
• PointNet and point cloud alignment achieved the highest performance in the benchmark.
• The proposed methods were evaluated for applicability to the aerospace component identification problem.

Conclusions:

• 3D data processing offers a more effective solution for identifying visually similar aerospace components.
• Deep learning models like PointNet are highly promising for automated aerospace part identification.
• The study provides a benchmark and validates 3D methods for critical MRO processes.