Engine EOL in 65 s: Pose, OCR and Missing-Part Checks
An engine, tens to over a hundred kilograms, travels on a slat conveyor through a gantry station straddling the line. Two jobs happen at once: confirm no component is wrongly or doubly fitted, and confirm the stamped number, nameplate and scanned barcode identify one unit. One looks at appearance, the other reads characters, and both must finish inside 65 seconds.
Two questions, two decision logics
A missing-part question asks whether the right component is present and fitted, and its evidence is a difference between image regions. A three-code question asks whether three strings identify one machine, and its evidence is what OCR returned. Blend them and either the appearance channel is tuned so hot that the character channel inherits false calls, or the reverse. This cell judges the channels separately and merges at release: failure in either blocks the unit until a person reviews it.
A locating camera fixes the unit before anything else aligns
A fixed camera establishes the pose of the engine in the station, and the inspection camera then visits each area in turn. This step cannot be skipped. The engine is hoisted and lands slightly differently each time; a camera on fixed coordinates may capture the flywheel edge or half the housing. With the unit pose known, each inspection area can be computed, and geometric matching has a correct starting point.
Why a collaborative robot carries the camera
Inspection points sit on several faces. Fixed cameras multiply hardware and debugging, and a model change rebuilds a bank of brackets. A collaborative robot moves one camera along a planned path, covering many points with less hardware, and a new model only needs a new path. The price is that every point waits for the arm to arrive, so path planning and shutter timing must line up with the judgment flow or the takt will not hold.
After alignment, judge the strength of the difference
Geometric matching aligns a standard template to the live image, and a missing part becomes a clear difference region. Gray-level comparison finds the difference, and a binary variance measure quantifies its strength and extent; above threshold, the unit is judged to have a missing or wrong part. Variance replaces single-pixel gray because a missing part removes a boundary, not a few pixels, and point comparison would confuse noise with absence.
Stamp and nameplate cannot share a light setup
Stamped characters vary in depth and their edges wear, so direct recognition misreads them. The nameplate surface reflects, which eats the contrast between characters and background. This cell frames the stamp area by geometric matching, normalizes gray levels to strengthen the characters, then runs OCR; the nameplate gets its own lighting and exposure settings. The two areas cannot share illumination, a step often skipped in commissioning.
Three identifiers cross-check each other
The stamp, the nameplate and the scanned code are read independently and then compared; any mismatch is an anomaly. Cross-checking is more reliable than trusting one read, because it catches low-confidence misreads. The cycle is 65 seconds, synchronized with the conveyor without adding a stop; confidence is above 98 percent, missed reports are 0 and false reports are at most 5 percent, subject to site acceptance results.
A light curtain pauses the robot on intrusion, the basic condition for human-machine collaboration. Results and images are archived by engine number for batch traceability in plants producing tens of thousands of units, and a review lane rechecks anomalies while passed units are lifted away.
How blind spots and takt are weighed
Fixed multi-camera or collaborative robot?
Concentrated points, loose takt and a single variant favour fixed cameras. Points spread across faces, a takt that allows the camera to move and frequent model changes favour the collaborative robot.
Can a badly worn stamp still be read?
Light wear is handled by normalization and cross-checking. Once character structure breaks, OCR confidence drops and the unit should go to manual confirmation, not a lower threshold.
How much output does 65 seconds support?
It is the inspection cycle for one engine and runs with the line without adding stop time. Line output also depends on line speed, shift pattern and station count, subject to site acceptance results.