Answers are grounded in this paper and cite their sources.
This paper develops and validates an industrially deployable vision framework for online resistance spot-weld (RSW) inspection, integrating weld localization, three-level appearance-quality classification, and shop-floor deployment. The framework combines controlled imaging for reflective sheet metal, a HALCON deep-learning detector with a compact SqueezeNet backbone, a three-class Good/Medium/Bad labeling strategy to account for ambiguous weld appearances, and a C#-HALCON deployment program supporting confidence-threshold adjustment, ROI filtering, alarms, and image traceability. The Medium class is further intended as a warning category for potential process drift; this early-warning function is formulated as a hypothesis and has not yet been validated with time-resolved production data. On an in-house dataset of 500 line-acquired images containing 2061 labeled welds, the optimized model achieves 97.6 ± 0.3% recall and 93.3 ± 0.4% precision (mean ± standard deviation over three independent training runs with different random seeds), with a network inference time of 28.1 ms per image. An independent field test on 2312 welds demonstrates a controllable precision–miss-rate trade-off, achieving 93.4–98.3% precision and a 2.3–4.7% miss rate under different confidence thresholds. Comparisons with YOLOv8, RT-DETR, and lightweight detectors further demonstrate the practical effectiveness of the proposed framework for online industrial inspection.
No takes yet. Share an insight, caveat, or question.
Xue et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: