Ultrasonic testing of rivet in multilayer structure using a convolutional neural network on edge device.

Area of Science:

• Aerospace Engineering
• Materials Science
• Non-Destructive Testing

Background:

• Aircraft rivets are critical structural components in air intakes, fuselages, and wings.
• Extreme operational conditions can lead to pitting corrosion in rivets, compromising aircraft integrity and safety.
• Early detection of rivet corrosion is essential for maintaining airworthiness.

Purpose of the Study:

• To develop and evaluate an ultrasonic testing (UT) method integrated with a convolutional neural network (CNN) for detecting pitting corrosion in aircraft rivets.
• To design a lightweight CNN model suitable for real-time deployment on edge devices.
• To assess the performance of the proposed method with limited training data.

Main Methods:

• Utilized ultrasonic testing to acquire data from rivets, including those with artificial pitting corrosion.
• Developed a lightweight convolutional neural network (CNN) model for corrosion detection.
• Trained the CNN model using a small dataset of 3 to 9 artificial pitting corrosive rivets.
• Implemented and tested the CNN model on an edge device (Jetson Nano) for real-time performance evaluation.

Main Results:

• The integrated UT-CNN approach demonstrated high detection rates for pitting corrosion.
• With only three training rivets, the system achieved 95.2% detection accuracy.
• Increasing the training data to nine rivets improved detection accuracy to 99%.
• The CNN model operated in real-time on the Jetson Nano with a low latency of 1.65 ms.

Conclusions:

• The proposed ultrasonic testing method combined with a lightweight CNN is effective for detecting aircraft rivet corrosion.
• The system demonstrates robust performance even with minimal training samples, making it practical for real-world applications.
• The real-time capability and efficiency on edge devices enable potential integration into aircraft structural health monitoring systems.