Convolutional Neural Networks for Hole Inspection in Aerospace Systems

Area of Science:

• Aerospace Engineering
• Computer Vision
• Manufacturing Technology

Background:

• Manual inspection for foreign object debris (FOd) in aerospace manufacturing is time-consuming, fatiguing, and error-prone.
• Current methods using visual checks with flashlights and mirrors lack reliability and efficiency.

Purpose of the Study:

• To introduce HANNDI, a novel handheld automated optical inspection system.
• To enable rapid, reliable, and automated inspection of FOd in rivet holes, machined holes, and fastener sites on the factory floor.

Main Methods:

• Development of a compact handheld device integrating controlled optics, illumination, and onboard deep learning (YOLO-based CNNs).
• Implementation of focal sweeps, image alignment, and fusion for an all-in-focus representation.
• Utilizing a dual CNN pipeline for hole detection/localization and debris classification.

Main Results:

• HANNDI achieved near 95% per-class precision and recall on a large, proprietary dataset of aerospace assets.
• Demonstrated an effective task time of 13.6 seconds per hole in end-to-end tests on aircraft parts.
• All training data was collected with the prototype, ensuring consistent imaging conditions.

Conclusions:

• HANNDI represents the first handheld automated optical inspection system with mechanical geometry enforcement, controlled illumination, and embedded CNN inference.
• The system offers a practical solution for robust factory floor deployment, enhancing aerospace manufacturing quality control.
• This technology significantly improves the speed and accuracy of foreign object debris inspection.