Efficient Prestress Wedge Flaw Detection Using a Lightweight Computational Framework | JoVE Visualize

Area of Science:

Computational engineering
Artificial intelligence in manufacturing
Non-destructive testing methods

Background:

Prestressing wedges are vital components in infrastructure, but thread flaws pose significant safety risks and increase maintenance costs.
Current manual inspection methods for wedge flaws are inefficient, inconsistent, and struggle with varying conditions.
Limited research exists on enhancing the effectiveness of automated inspection for these critical components.

Purpose of the Study:

To develop a lightweight and efficient deep learning framework for accurate flaw detection in prestressing wedges.
To improve the robustness and real-time applicability of automated inspection systems in industrial settings.
To bridge the gap between AI-based flaw detection and engineering-grade reliability for quality control.

Main Methods:

Implementation of a lightweight FasterNET-YOLOv5 deep learning framework tailored for flaw detection.
Evaluation of detection performance using precision, recall, and mAP@0.5 metrics.
Robustness testing under various illumination conditions, specifically evaluating white lighting backgrounds.
Integration with a mechanical model-based inverse method for linking machine vision detections.

Main Results:

The FasterNET-YOLOv5 framework achieved high detection accuracy with 96.3% precision, 96.2% recall, and 96.5 mAP@0.5.
Achieved an 18% faster end-to-end inference speed compared to existing methods.
Identified white lighting as optimal for detecting incomplete threads and scratches.
Demonstrated successful linkage of machine vision detections via a mechanical model-based inverse method.

Conclusions:

The proposed FasterNET-YOLOv5 framework provides an accurate, robust, and efficient solution for prestressing wedge flaw detection.
The system's speed and accuracy enable deployment on portable devices for real-time industrial inspection.
The research highlights the potential for broader applications in automated quality control of threaded and machined components.
This work advances computational inspection by combining deep learning with practical engineering requirements.