Bolt-Loosening Monitoring Framework Using an Image-Based Deep Learning and Graphical Model.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Structural Engineering

Background:

• Traditional bolt-loosening detection methods are often labor-intensive and costly.
• High-quality training datasets are crucial for developing effective deep learning models.
• Acquiring diverse real-world data for bolt inspection can be challenging and expensive.

Purpose of the Study:

• To propose and validate a novel framework for detecting loosened bolts using synthetic images generated from a graphical model.
• To demonstrate the feasibility of training a deep learning model with synthetic data for bolt-loosening detection.
• To assess the practical applicability of this methodology on real-scale structures.

Main Methods:

• Development of a framework integrating computer graphics for synthetic image generation and deep learning for analysis.
• Training a deep learning model using synthesized bolt images.
• Validation of the model on a lab-scaled bolted joint and subsequently on a real-scale truss bridge.

Main Results:

• The deep learning model trained on synthetic images accurately recognized bolts and detected looseness.
• The framework proved feasible in lab settings and practical for real-world applications.
• Successful evaluation on a historical truss bridge in Danang, Vietnam.

Conclusions:

• Synthetic data generated via graphical models can effectively train deep learning models for loosened bolt detection.
• This methodology significantly reduces the time and cost of acquiring training data.
• The approach accelerates the adoption of vision-based deep learning models in practical structural health monitoring.