Fixed Views Beat a Circling Robot on Engine EOL Checks
Diesel and gasoline engine assemblies and transmissions carry points on the top, sides, end faces and sometimes the underside. The unit is large, and a single robot has limited reach and approach angles. Circling the whole assembly means a long path, many repositions, a settling pause at every stop, and an imaging pose that changes with position. The constraint is not a lack of cameras; it is that one camera cannot get there and cannot hold still.
The robot spends its cycle travelling, not imaging
Covering several faces with one robot costs path time. Depart, arrive, settle, shoot, move on: the exposure itself may be a small fraction of the cycle. With four faces, that path repeats four times and the takt collapses. Worse, residual vibration after each stop has to decay before the shutter fires, and those waits add up to more than the imaging.
Repeatability is what lets a threshold stay tight
If distance, angle and lighting differ at the same inspection point from cycle to cycle, the decision threshold has to be loosened. A loose threshold raises false calls and escapes together. The value of a fixed viewpoint is not the extra robot; it is that the same point is imaged under nearly identical conditions every time, so parameters can be preset and locked instead of adjusted on the floor.
Four faces, four owners
The cell splits the top, sides, end faces and underside into groups, each robot owning one group of fixed views with a 20 MP industrial camera and a combination of very large high-brightness area light and high-brightness bar light. Paths shorten, imaging stabilizes, and responsibility becomes traceable: if one face develops a problem, the record points to a robot and a view.
There is a quieter benefit. Each group can optimise its own camera and light settings without compromising a difficult angle elsewhere.
Two lights for two kinds of features
A large assembly outgrows ordinary illumination. Brightness falls from the centre toward the edges, and bolts at the edge are judged on an image that is already fading. The oversized area light lifts the whole surface; the bar light directs light onto small features such as bolts and pipework. The division is practical: area light is even but flattens detail, bar light reveals detail but cannot cover a large face. Use only one and some class of point will always be judged badly.
The underside belongs in the plan
The bottom face is uneven, close to the floor and easy to omit because it is awkward for people and was never assigned to a machine. This cell writes the underside into the robot view map, with a robot carrying a camera into that region rather than leaving it for an inspector to crawl under. Putting the hardest face on the allocation table is more reliable than reminding people not to forget it.
Seventy seconds is a parallel result
The cycle is about 70 seconds and accuracy is 99 percent, subject to site acceptance results. The takt comes from converting serial walk-and-shoot into parallel work: four faces proceed at once and total time approaches the slowest branch. That makes balance mandatory. If one robot owns far more points than the others, it alone sets the line rate.
Code reading and whole-unit photo retention fit the same window; judgment, evidence and release are three actions in one takt.
Questions raised before the layout is frozen
Why not one wide-angle camera for everything?
A wide view drops pixel density at every assembly point, so small parts and barcodes become unreliable. It also cannot see the back and underside of a multi-face assembly.
Will the robots interfere with each other?
They can, and that is a layout task. Working envelopes and shutter timing must be separated at the programming stage so two robots never enter the same space together.
Does changing the assembly model require new hardware?
Usually not. Inspection items and thresholds switch with the assembly type while robot positions and camera configuration stay as they are.
Can the 99 percent figure be guaranteed?
It is project data. Escape and false-call rates should be read from running records, subject to site acceptance results.