Corroded Bolt Identification Using Mask Region-Based Deep Learning Trained on Synthesized Data.

Area of Science:

• Structural Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning models require extensive datasets for optimal performance.
• Limited datasets hinder accuracy and generalization in neural networks.
• Existing corroded bolt detection methods struggle with partially rusted components.

Purpose of the Study:

• To develop a robust deep learning method for identifying corroded bolts in steel structures.
• To address the limitations of existing methods in detecting partially corroded bolts.
• To improve the accuracy and generalization of corrosion detection using synthesized data.

Main Methods:

• Utilized a Mask Region-based Convolutional Neural Network (Mask-RCNN) with a Resnet50 backbone and Feature Pyramid Network.
• Developed a four-step procedure for autonomously synthesizing diverse training datasets of corroded bolts.
• Trained the Mask-RCNN detector on the synthesized datasets to identify corrosion severity.

Main Results:

• The proposed method achieved high accuracy in detecting corroded bolts across varying distances and perspectives.
• Achieved an overall accuracy rate of 96.3% at a 1.0 m capturing distance.
• Demonstrated 97.5% accuracy for a 15° perspective angle, confirming robustness.

Conclusions:

• The deep learning approach effectively detects corroded bolts and classifies their severity using synthesized data.
• The method offers a reliable solution for monitoring the structural integrity of steel components.
• Synthesized data generation is a viable strategy to enhance deep learning model performance in corrosion detection.