Inline Weld Seam Inspection Cell

Background

A tier-two automotive fabricator inspected robotic MIG welds by sampling — an operator pulled roughly one part in forty to a light booth. Defects that escaped sampling surfaced at the customer, and each containment event cost more than a week of production margin.

The problem

What made it hard.

Inspection had to finish inside the robot cell's 45 ms index window or it would slow the line, which ruled out sending frames to a server. Weld appearance shifted constantly with spatter, tip wear and shielding gas flow, so a fixed-threshold vision recipe drifted within a shift. The plant also refused any inspection system that required opening the OT network.

Approach

What we built.

We built a self-contained cell: two GigE cameras on a rigid frame with controlled ring lighting, feeding a Jetson Orin NX running a segmentation model quantised to INT8 through TensorRT. The model outputs a seam mask and a per-millimetre defect class — porosity, undercut, burn-through, insufficient fill. Inference lands at 38 ms end to end, inside the index window, with the verdict pushed to the cell PLC over OPC UA as a simple pass/divert bit. Nothing leaves the cell. A local ring buffer keeps the last 5,000 part images so quality engineers can retrain against real escapes; retraining runs offline and ships back as a signed model bundle on removable media.