Tier-1 Automotive Supplier Hits 99.7% Defect Detection with Custom AI Vision System

Our Warsaw R&D team started with the physics problem: how do you reliably image a polished aluminum surface in a dusty, hot foundry with variable ambient lighting?

The answer was controlled lighting. We designed a dome illumination system with polarized filters that creates uniform, diffuse light across the caliper surface while eliminating the specular reflections that had defeated the previous system. The entire optical assembly sits inside a custom IP67 enclosure built to withstand foundry conditions.

For sensors, we moved away from standard color cameras. Monochrome sensors paired with telecentric lenses provided the resolution and distortion-free imaging needed to detect micro-fractures at the edges of complex caliper geometries.

Generic object detection models couldn’t distinguish between acceptable surface variations and actual defects. We needed semantic segmentation – a model that could precisely outline the boundaries of pores, scratches, and hairline fractures.

We customized a U-Net architecture and trained it on thousands of labeled images from the client’s production line. But rare defect types posed a challenge – some critical fracture patterns appeared only a handful of times in historical data.

Our solution: synthetic data generation. We created photorealistic renders of rare defect types to augment the training set, ensuring the model could catch failures it had never seen in real production. The final model achieved 99.7% detection accuracy, surpassing the 99% KPI.

A conveyor moving at 1.2 meters per second doesn’t wait for cloud inference. We deployed the model on NVIDIA Jetson Orin modules positioned directly at the inspection station, achieving inference times under 45ms per frame.

The final integration challenge was connecting the vision system to the production line. We wrote custom firmware to interface with the client’s PLCs via OPC UA. When a defect is detected, the system triggers a pneumatic divert arm to reject the part instantly – no operator intervention, no line stoppage.

The result: 100% automated inspection at full production speed, with every defective part caught and rejected before it leaves the station.