Checking 4,000 Body Welds in 0.4 Seconds
A passenger car body carries roughly 4,000 to 5,000 welds, and weld quality decides body strength and vehicle safety. The welding line runs fast and the weld count is high, so manual sampling cannot cover every spot, and a missed weld that leaves the plant is a vehicle-level quality risk. Traceability also requires the weld verdict to travel with the body, so the inspection system has to interlock with the line's information system.
The blind area left by sampling is settled at vehicle level
With four or five thousand welds, even a high sampling ratio is still sampling. The cost of a missed weld does not show up at the welding line; it shows up in final assembly, at delivery or in the field. The requirement is therefore not a bigger sample but full inspection fitted inside the online cycle, which is a different engineering problem from raising the ratio.
Balancing single-frame coverage against pixel density
A 5472 x 3648 high-resolution industrial camera covers a large weld area in one frame, and stations are laid out so every weld on the body is imaged, with denser imaging where welds cluster. Each weld then has enough pixels to be judged. On a station with sparse welds across a wide field, the pixel count per weld must be checked so resolution is not diluted by coverage.
Experts draw the ROI, the CNN learns the features
Process experts define ROIs on the image in advance so the algorithm concentrates on critical areas and ignores background. The CNN learns weld texture and morphology and builds a standard weld feature library rather than depending on hand-written rules. When the process changes, the ROI can be redrawn without retraining the whole model, which is the practical difference between a learned feature set and a rule set that has to be rewritten by hand.
The confidence threshold is the knob between false and missed calls
At run time image features are compared with the standard library and a confidence is output: above the threshold passes, below is abnormal. Too high and false calls multiply, too low and misses multiply, so the setting comes from the library comparison and is balanced during site validation. Verdicts bind to weld positions, and a missed weld is flagged immediately.
Mixed models, one library per model
Weld layouts differ widely between models, so the standard library is organised per model and the matching baseline is loaded for the current body. A new model needs samples accumulated before it can go live; an old model's library cannot be forced onto it. The library keeps growing with production as new weld forms are added.
Interference, cycle and the limits
Surface condition around a weld varies and affects imaging; the lighting and the library comparison together suppress it. The inference unit sits close to the line and returns a verdict within 0.4 seconds, with a 0.5-second work cycle, and inference and capture run as parallel pipelines so the next body is imaged while the previous one is judged.
The project reaches 99.9 percent accuracy at 0.4 seconds per part, saving about RMB 210,000 per station per year, subject to site acceptance results. Before go-live, historical missed-weld samples should be used to validate the threshold and the completeness of the library, and camera resolution should be set from weld density and station field of view.
Questions from the line
How are 4,000 to 5,000 welds fully covered?
The high-resolution camera covers several welds per frame and station layout covers the whole body, with expert ROIs ensuring each weld is inside an inspection area.
Can 0.4 seconds preserve judgment quality?
Through multi-weld frames, parallel ROI inference and pipelined capture. Capability is set by pixel density and library completeness, and should be verified on missed-weld samples before go-live.
Does a model change need retraining?
Not wholesale. Libraries are organised per model, new models accumulate samples first, and ROI can be redrawn when the process changes.