Assembly Error Proofing: Camera Count Follows Blind Spots
Automotive assembly parts are large and feature-rich. They arrive on the line conveyor, passing units are released and anomalies intercepted. Three problems must be prevented: a missed machining operation, a missing component, and a wrong component or wrong position. Manual visual inspection under a tight cycle looks first at the areas most likely to fail and samples the rest from experience, so the error-proofing duty is not really fulfilled.
The camera count is set by occlusion
The cell uses several fixed industrial cameras for full multi-view coverage. More cameras are not there to make images sharper; they are there to remove blind spots. An assembly has a complex spatial structure, and one camera looking from one direction cannot see the back and the recesses, which is exactly where a missing feature may hide. Drawing the occlusion relationships before fixing camera positions is easier than fixing accuracy first and fitting positions later; the latter often ends with brackets installed and the critical feature outside every camera view.
A feature should appear in two or three views
With multiple views, the system cross-validates the same assembly feature across them. If one view shows a part and another does not, the judgment condition is unstable and the item needs review. Single-view misjudgment has many sources: reflection, shadow and local dirt can all make a fitted part look missing. Multiple views let those incidental effects cancel each other out.
Measurement and error proofing share one station
Measurement and recognition happen at the same station, so the assembly does not travel offline. Measurement correlation with a CMM reaches 0.15 mm, subject to site acceptance results. The figure needs careful wording: it describes the level of agreement between vision measurement and high-accuracy measurement, showing that vision can support position judgment of assembly features. It does not mean vision replaces a CMM for final metrology. Treating correlation as accuracy is a common misunderstanding in this class of project.
117 seconds is a layout result
The cycle is 117 seconds, held under 120 seconds to match the inline rhythm. It is set by camera count, capture and judgment time per camera, and the degree of parallelism across views. Once a fixed multi-camera layout is locked, compressing the cycle later means adding cameras or raising per-camera processing speed, not optimizing an algorithm. That is why camera planning has to be right at proposal stage; later changes involve brackets, cabling and calibration.
The model must learn states, not just presence
Assembly features vary widely in type and form, and the model has to learn the state distribution of each, not merely whether something is there. Wrong assembly depends on this: a part is present, slightly displaced or reversed, and a presence-only check passes all of them. Recognition accuracy is above 99.9 percent, subject to site acceptance results. That level comes from multi-view feature extraction and combined judgment, provided the samples cover the anomaly states; a wrong fit never seen in training cannot be judged.
Data tied to the batch
Inspection data covers every view and every verdict, archived against the assembly batch for repair traceability and process analysis. When an anomaly is intercepted, the record shows which view and which feature produced the call, and that chain determines whether the problem can be located to a specific process.
Where the positions come from and how to read the number
Fixed multi-camera or robot-mounted camera?
Stable points, few variants and a loose cycle favour fixed multi-camera, which is simpler and more reliable. Scattered points, changing layouts and spatial state judgment favour a robot. The former is stable after one commissioning; the latter changes over faster.
Is 0.15 mm correlation enough?
Enough for assembly feature position judgment, not for tighter dimensional tolerance. Define the tolerance band of each measurement item first, then check whether vision covers it; do not compare a correlation number directly with a tolerance.
Why is wrong assembly harder than missing?
A missing part has a clear boundary in the image. A wrong part is present but incorrect, perhaps only slightly offset or reversed, so the model must understand the normal range of states rather than run a presence check.