Efficient Prestress Wedge Flaw Detection Using a Lightweight Computational Framework.

Prestressing wedges are critical in bridge and road construction, but flaws in wedge threads lead to severe safety hazards, construction delays, and costly maintenance. Traditional manual inspection remains labor-intensive and inconsistent, particularly under variable illumination and complex surface conditions. However, few studies have investigated improving the inspection effectiveness. Therefore, this study aims to propose a lightweight FasterNET-YOLOv5 framework for accurate and robust prestress wedge flaw detection in industrial applications. The framework achieves a detection precision of 96.3%, recall of 96.2, and mAP@0.5 of 96.5 with 18% faster end-to-end inference speed, enabling deployable system configuration on portable or embedded devices, making the approach suitable for real-time industrial inspection. Further practical guidance for workshop inspection illumination conditions was confirmed by robustness evaluations, as white lighting background provides the most balanced performance for incomplete thread and scratch defects. Moreover, a mechanical model-based inverse method was exploited to link the detections from machine vision. The results also demonstrate the potential for broader 3D surface inspection tasks in threaded, machined, and curved components of intelligent, automated, and cost-effective quality control. In general, this research contributes to computational inspection systems by bridging deep learning-based flaw detection with engineering-grade reliability and deployment feasibility.