Automotive Industry

The hard-to-catch defect problem

In automotive parts production the difficult cases are not the obvious defects but the ambiguous ones — flaws sitting on the specification boundary, scratches that appear and disappear depending on lighting and angle.

Rule-based vision cameras struggle here. Tighten the threshold and good parts get scrapped; loosen it and defects slip through. People end up re-inspecting, and inspection headcount becomes the bottleneck.

Two directions

Detection — catching what gets missed

Deep learning vision inspection (V²) detects fine defects that rule-based systems miss. Using anomaly detection trained only on images of good parts means it can be applied early, before enough defect samples have accumulated. The models are light enough to run on edge devices, so line speed is unaffected.

Recording — knowing what was made, and when

The general manufacturing MES core tracks output and LOTs by line in real time and aggregates equipment utilization. When a defect appears, being able to trace back to the LOT, the machine and the conditions is what makes root cause analysis and prevention possible. Both detection and prevention run on top of this record.

Prevention — reducing the conditions that create defects

Detection alone does not lower the defect rate. Linking process variables such as molding conditions, mold temperature and material lot to actual quality results shows which conditions raise defect rates. Adjusting the process to avoid those conditions is the more fundamental fix.

Rollout

Validate on one line, then extend to other lines running the same process. Because equipment and conditions differ between lines, models rarely transfer unchanged, so a retraining procedure is designed alongside the rollout.