ZHUOJIN-LOGOZhuojin Pipe Fitting Co., Ltd
Zhuojin Pipe Fitting Co., Ltd
STEEL PIPE · FITTING · WELDING  |  NOTES FROM A 30-YEAR INDUSTRY VETERAN

Image Based Defect Detection for Malleable Iron Pipe Fittings

Why fitting surface defects are hard to catch

The paper addresses surface defect detection on malleable iron pipe fittings using a convolutional feature network with attention based feature fusion. The practical problem is familiar: fittings have complex geometry, dark casting surfaces, mold parting lines, flash, sand residue, and small defects that look like normal casting texture. A shrinkage cavity near a thread runout, a cold shut on a fillet, or a fine crack at a bend can be missed by visual inspection and over rejected by a crude camera system. The paper builds a dataset and reports average precision at IoU 50 of 91.3 percent, with the biggest gain on small defects, about 5.6 percent. That is the kind of improvement that matters on a production line, because small defects are the ones that pass inspection and later cause leakage or fracture in the field.

What the improved detection method changes

Item Paper result Inspection practice meaning
Baseline method HRNetV2p style network Good feature extraction but weak scale balance
Improved method Attention based feature fusion Better balance between small and large defects
Average precision at IoU 50 91.3 percent Useful screening level, not absolute proof
Small defect gain 5.6 percent Helps catch early cracks and fine shrinkage
Medium and large gain about 2.7 percent Confirms benefit is not only for tiny defects

The key lesson is not that a camera replaces the inspector. The lesson is that lighting, camera angle, scale normalization, and defect labeling decide whether the system sees a real defect or a shadow. In malleable iron fittings, surface roughness and casting skin can mimic cracks. If the training data are biased toward clean parts, the system will miss dirty castings. If the data are too noisy, it will reject good parts and create a false reject backlog. I have seen quality systems collapse because false rejects were treated as normal variation. A good deployment needs a PDCA cycle, defect review by metallurgy, and a clear rule for when the camera decision is advisory and when it is final.

Inspection line practice and judgement

For a real fitting plant, the inspection station must control part orientation, surface cleaning, illumination, and image resolution. The camera system should be paired with magnetic particle or dye penetrant checks for critical fittings, because surface imaging alone cannot confirm depth. The paper’s 91.3 percent precision is encouraging, but the plant must define acceptable escape rate and false reject rate before installation. If a fitting is used in a threaded water line, a small surface crack may be tolerable until it opens under pressure. If the same fitting is used in a critical industrial line, the same crack may be a reject. The method is valuable because it improves scale balance, but the final standard must be written by the inspector, the foundry, and the customer together, not by a single image score.