End-Of-Line Poka-Yoke For Engine And Transmission Assemblies: What Vision Can And Cannot Catch

Overview

An end-of-line station is the last chance to stop a defect before a vehicle. It is also the place where expectations are set highest and audited least: a station that claims to check everything will be trusted to check everything, and the gap between the claim and the capability is where recalls come from.

Big Bird Industrial (formerly Shimada Big Bird Industrial) has built assembly inspection stations for engine and transmission lines since 2001. This article separates what a vision system can genuinely verify at end of line from what it cannot, and describes how to write a check list that a quality engineer can defend in an audit.

The distinction matters more than the technology. Most failed poka-yoke projects did not fail technically; they failed because the specification promised something the physics does not allow.

1. The Four Things Vision Actually Verifies

End-of-line vision is good at four categories of check, and almost everything useful falls into one of them.

Presence and absence

A bolt, a cap, a clip, a cover, a sensor. This is the bread and butter of poka-yoke and the category with the highest return, because missing parts are common and the consequence is often immediate.

Orientation and sequence

A bracket fitted backwards, two similar parts swapped, a bearing cap in the wrong position. These are harder than presence checks because the difference can be small, and they reward good lighting far more than they reward a bigger model.

Text and codes

Nameplate, barcode, laser-marked serial, date code, and the consistency between them. Optical character recognition on a cast or painted surface is a solved problem in good light and an unsolved one in bad light, which is why the illumination design decides the recognition rate.

Routable routing

Hoses, harnesses, pipes and their clips: whether the routing matches the drawing and whether a clamp is present at the right point. This is the category where a robot-held camera earns its cost, because the views needed are all around the assembly.

2. What Vision Cannot Verify

The honest list is shorter than the marketing list, and writing it down is what keeps a station credible.

Vision cannot verify torque. A correctly installed bolt and an under-torqued bolt look identical. Torque is a process signal from the tool, and the right architecture is to bind the tool result to the part serial rather than to photograph the joint.

Vision cannot verify internal assembly. Whether a piston ring is correctly seated, whether a bearing shell is in place inside a closed housing, whether a seal is installed in a bore that is capped afterwards — none of these are visible, and no amount of resolution changes that.

Vision cannot verify fluid fill, pressure integrity or function. Those need a test stand, and a test stand is a different instrument with a different acceptance logic.

Vision cannot prove absence of a defect it was never trained on. A station validated on a fixed defect set has measured performance on that set, not on the universe of possible failures.

3. Designing The Check List

A useful end-of-line station is built from a list of checks, each with an owner, a method and a failure action. The list is not a description of the camera; it is a description of the process gate.

Writing the method column forces the honest answer. A row that says vision for torque is a row that will be argued about at the first audit.

4. Recognition Rate Is A Property Of The Set

A recognition rate of 99.5 percent is a statement about a specific population, in specific light, with a specific camera. It is not a property of the algorithm. Two stations quoting the same number can have very different behaviour on the parts that actually cause escapes: the dirty cap, the partially occluded clip, the label applied at an angle.

The practical way to qualify a station is a seeded defect run: known defects inserted at known positions, repeated across shifts, with the result reported per check rather than overall. A station that catches 99 percent of seeded missing caps and 80 percent of reversed brackets has two different problems, and an overall number hides both.

It also matters who seeds the defects. If the supplier builds the defect set, the set will reflect what the supplier's system already does well. The set should include defects collected from the field, even if that means shipping parts.

5. Cycle Time And The Robot Question

Most end-of-line assemblies need views from several directions, which is why a robot-held camera is common. The alternative is a multi-camera fixed rig, which is faster but only covers the views it was built for. The trade is straightforward: a robot covers new positions by teaching, a fixed rig covers them by adding hardware.

Cycle time is dominated by how many stops the robot makes, and by how long each stop takes to settle. A typical engine assembly station with a dozen checks runs in the range of eighty to ninety seconds, and the largest improvement usually comes from reducing settle time rather than from accelerating the motion.

Path duplication matters for mixed lines: a station that can copy a taught path to a new variant in an hour is a station that survives a model change. One that requires reprogramming from scratch will be bypassed the first time production is under pressure.

6. Making It Auditable

The last design question is what the station leaves behind. Every check should produce an image, a verdict, a timestamp and a part serial, because the audit question is never did the system pass this part but why did it pass this part.

Retention policy deserves a decision rather than a default. Keeping every image for a year is expensive; keeping none is indefensible. A workable rule is to keep all failures, all first-off parts, and a sample of passes, with a retention period agreed with quality.

Finally, the station should record its own health: light output, camera focus, calibration date and the last verification result. A poka-yoke station that has quietly drifted is more dangerous than no station at all, because the line believes it is protected.

7. The Escalation Path

A station that can stop the line needs an escalation path, and the path needs to be designed before the station is switched on. In practice three outcomes are possible for a failed check: rework at the station, route to a repair loop, or scrap. Each needs a defined owner and a defined record.

The most common failure in service is not a missed defect; it is an operator overriding a check because the escalation path is unclear and production is under pressure. Once an override is routine, the station is decorative.

The countermeasure is to make overrides visible and attributable: every override logged with a reason code and a name, reviewed daily. A station with fifty overrides a week is telling you something about the specification, not about the operators.

Frequently Asked Questions

Can vision replace the torque station?

No. Torque is a process value, not a visual one. The correct integration is to bind the tool result to the part serial and let the vision station block release when the result is missing.

How many checks can one station handle?

The limit is cycle time and view count, not software. Each additional view costs a robot stop, and each stop costs seconds.

Is deep learning required for presence checks?

Often not. Classical detection is more predictable on a fixed part in a fixed fixture. Deep learning earns its place on orientation, texture and cluttered scenes.

What is a realistic recognition target?

Above 99 percent on a well-defined check in controlled light, with the number stated per check and measured on a seeded set that includes field returns.

Conclusion

End-of-line vision is a narrow instrument with a wide reputation. It is excellent at presence, orientation, text and routing; it is blind to torque, internals, fluid and function. A station that says so in its specification is trusted for the checks it does, and that trust is worth more than an inflated claim.

Design the check list first, choose the views second, and qualify with seeded defects from the field. In that order the station becomes a gate; in any other order it becomes a photo booth.