Large Auto Parts: 3D Plus Deep Learning in 98 s
Automotive parts are large and carry many inspection points. In the past, full dimensional inspection relied on CMM sampling, and one part took long enough that not every unit coming off the line could be covered. Missing or wrong parts depended on visual inspection, with both escape risk and workload. This cell uses a robot carrying a 3D measurement camera and a 2D industrial camera, adds deep learning for assembly judgment, and completes full dimensional measurement and error proofing in one load. Correlation with the CMM is 0.15 mm and the cycle is 98 seconds, subject to site acceptance results.
0.15 mm is a consistency figure, not an accuracy figure
The number describes the level of agreement between vision measurement and CMM measurement: the smaller it is, the closer the two are. It answers not how accurate the device is, but whether vision data can sit on the same trend chart as CMM data.
The robot solves reach, not resolution
With large parts and many measurement faces, a fixed camera position struggles to combine field of view and working distance across every point. The robot carries the camera to each face along a planned path and covers all points from multiple poses. Camera count can therefore be low, but path planning and positioning accuracy are demanded more heavily.
3D measures, 2D reads detail, deep learning judges assembly
The 3D camera supplies spatial dimension data, the 2D camera handles local detail and missing or wrong features, and deep learning moves assembly judgment from rule comparison to feature learning. The division is clear: the robot solves whether the camera can get there, 3D solves whether the value is right, and 2D plus deep learning solve whether the part is fitted correctly.
One load, two verdicts, 98 seconds against a 100 second cap
Measurement and error proofing share calibration and imaging, and deep learning identifies missing or wrong features on the same image batch used for dimensioning, so one inspection outputs both values and assembly verdicts without a second load.
The requirement is a cycle no longer than 100 seconds, and the measured result is 98 seconds, a 2 second margin. That margin is thin, so path, exposure and computation cannot fluctuate noticeably. Faster lines need more cameras capturing in parallel or a better path.
Reflection and changeover are the two walls
Machined surfaces reflect and disturb imaging, so light configuration and exposure are optimised by zone. When several variants share the line, inspection programs are managed as recipes and switched with one action. A recipe holds capture points, inspection items, algorithm parameters and acceptance criteria; miss one and changeover falls back to trial and error.
Where the boundary to metrology sits
The system targets inline full inspection for interception and trend monitoring. High-accuracy metrology and arbitration still belong to the CMM. Treating inline measurement as a replacement for the metrology room misreads what the number means.
Questions from the project review
What does 0.15 mm correlation with the CMM mean?
It is the agreement level between inline vision results and CMM results, used to judge whether vision can replace CMM sampling. Smaller means closer agreement.
How is full dimensional inspection completed in one pass?
The robot carries 3D and 2D cameras along a planned path across every measurement face, then feature extraction and dimensional computation run together, producing all values from one load.
Why can error proofing run at the same time?
Deep learning identifies missing or wrong features on the same image batch, and both judgments share calibration and imaging, so one inspection outputs two result types.
Is there margin in 98 seconds?
The cap is 100 seconds and the measured result is 98, so margin is limited. Faster lines need parallel cameras or better paths, subject to site acceptance results.
What role does SPC play?
Data enters the statistical system in real time, monitoring process capability on key dimensions, so drift can be corrected early and batch defects prevented at source.