By Marcus Bennett

A French automaker ordered a second automated weld-seam inspection cell in April 2024. The first began production use four months before that, in early 2024, at the same body-in-white (the assembled but unpainted car shell) plant.

Valladolid, Spain-based integrator DGH Group built the system. It handles both MIG and laser welding joints — the two most common body-in-white joining methods in European passenger-car plants.

DGH — an automation specialist now part of GROUPE ADF — took the work on commission. The integrator calls its customer only "a large French automotive group." No OEM name, no vehicle programs. Both cells run MVTec HALCON machine-vision software and MVTec's Deep Learning Tool.

The setup targets cracks, incomplete seams, and irregular bead patterns. MIG welds leave wider, textured beads; laser welds produce narrow, smoother joints. HALCON holds detection rates steady across both, DGH says. The two-cell program marks DGH's first commercial scale-up of an AI-based weld-inspection setup inside GROUPE ADF's customer base.

How does the cell inspect a weld?

When a body reaches the station, a programmable logic controller (PLC) triggers a sequence of 2D cameras. Images stream into HALCON over the GigE Vision protocol, a standardised machine-vision camera interface. The software runs two deep-learning models in sequence:

• Instance Segmentation localises the weld-seam zone on each image with pixel-level precision.
• Anomaly Detection flags any deviation from the trained "good weld" baseline.

Pass or fail results return to the PLC and render on a line-side display.

Why train the model on good welds?

Anomaly detection needs predominantly defect-free images to learn what "normal" looks like. "Anomaly detection had two decisive advantages for us: On the one hand, the detection rates are very high and robust," said Guillermo Martín, Innovation & Technology Director at DGH. A small library of defect images still sets the threshold separating good from bad.

Engineers reviewed each captured photo before it entered the training set, weeding out anything ambiguous. DGH labels images in MVTec's free Deep Learning Tool, where a Smart Label feature lets an engineer click inside the seam. The tool frames the segment automatically.

After labelling, images split 50% training, 25% validation, 25% testing. The trained model then loads back into HALCON for runtime inference.

What does this change for the line?

Two practical effects stand out. Detection rates hold steady across reflective bare steel and fluctuating booth lighting that defeat rule-based vision checks. The system also cuts the line's dependence on skilled visual inspectors, a workforce that has thinned across European body shops for years.

For line operators, the inspection station now handles the visual judgement that previously sat with senior quality staff. The deployment also standardises quality output across shifts, removing the variability that comes from human inspectors working different patterns on different days. DGH says defects were "demonstrably minimized" and the system now catches every defect class the customer specified.

What to watch next

DGH has not disclosed the OEM's name, vehicle programs, or the annual volume covered by the two cells. Watch for:

• A third cell order, which would move the program from pilot to standardised equipment across that OEM's body shops.
• Similar deployments at other European carmakers supplied by GROUPE ADF, now DGH's parent.