A vision system built around a convolutional neural network has classified spot weld defects on a body-in-white line with 95.8% accuracy and delivered a maintenance-prognosis accuracy of 95.87%, according to researchers at the Autonomous University of Queretaro in Mexico.

The paper, co-authored by Alfonso Alejo-Ramirez, Rogelio Cedeño-Moreno (Ph.D.), Luis A. Morales Hernandez (Ph.D.) and Juan C. Jauregui-Correa (Ph.D.), uses two 12-megapixel monochrome cameras mounted on each side of a cabin at 1.3 meters. It locates individual welds, measures nugget and heat-affected zone (HAZ) geometry, and flags when a specific robot needs service.

What does the system inspect?

The work targets quality at the cabin-assembly station, where each car body carries thousands of resistance spot welds laid down by robotic cells. Inspection runs immediately after cabin assembly, before the body moves downstream.

Under normal conditions, the researchers define a "good" weld by inner and outer elliptical contours, a fusion zone, a smooth rounded surface, and a diameter of 2 to 3 millimeters. Tracked defects include cold spots, poor welds, deformed metal, ejection, and pitting.

The pipeline extracts nugget diameter and HAZ size from camera images, then routes the data to a CNN that learns each robot's propensity to produce defective welds.

How was the model trained?

The training database covers three consecutive days of uninterrupted production: 8,092 images from 674 cabin assemblies, each at 4,096 by 3,000 pixels. From these, the team extracted 63,073 individual spot weld images at 30 by 30 pixels.

For classification, the team capped each class at 1,000 images, with 800 for training and 200 for validation. For nugget and HAZ sizing, separate models used 100 images for training and 30 for validation. Boundaries were labeled manually.

Measured against ground truth, the sizing models returned average errors of about 5% for nugget diameter and 3% for HAZ diameter.

What performance did the model deliver?

The classification network hit 95.8% accuracy, 95.6% precision, 96% recall, and a 95.8% F1-score. Each cabin classification runs at roughly 0.5 second.

The maintenance-prognosis network, trained on cumulative 30-hour windows of defect data per robot, hit 95.87% accuracy, 94.72% precision, 93.97% F1-score, and 94.42% recall. It also ran at about 0.5 second per evaluation.

What triggers a maintenance call?

The team mapped defect rates to four service categories aligned with ISO 18278-1:2022, which holds vehicle-chassis failures below 1% as the efficiency benchmark:

• "No maintenance" for defect rates below 0.1%
• "Preventive maintenance" for 0.2% to 0.4%
• "Predictive maintenance" for 0.6% to 0.9%
• "Corrective maintenance" at 1% or above

Those labels trained the prognostic CNN to flag when a robot crosses the 1% threshold versus drifting below it.

What are the stated limitations?

Three issues stand out. First, the database sorts welds only into "good" or "bad," with no separation by defect type. Second, no destructive testing confirmed that visually "good" welds meet mechanical strength requirements. Third, the lighting and fixture tolerances baked into one body-in-white station may not transfer to a second plant or OEM.

What to watch next

Two validation steps will decide whether this research turns into a procurement order. The team plans destructive tests on representative welds from each of the four quality categories to correlate visual classification with mechanical strength.

A second milestone: whether the 0.5-second-per-cabin cycle time and the 95.8% accuracy hold when the model is retrained on a different OEM's body-in-white station and production mix.