Lesson 14 of 20 · Chapter 14
Quality Maintenance through Computer Vision and AI
A line can run well and still lose at the last metre: a missing seal, a crooked label, a weak weld. Cameras can now check every unit at full speed. But a camera that only sorts good from bad is an expensive filter that lets the process make scrap faster. This chapter is about using vision to keep the process inside normal, so the defect is prevented, not just caught.
Choose how you want to learn
You can change this any time. Every level does the same hands-on steps. Press Listen on any part to hear it read aloud.
- 1 · Every unit, at speed
- 2 · Detect or prevent
- 3 · Truth comes first
- 4 · When the camera errs
- 5 · Close the loop
- 6 · Try it at work
- 7 · Your card
- 8 · Check yourself
Part 1
From checking a sample to checking every unit
Old quality checks take a sample, because checking every unit would slow the line. A camera can check every unit without stopping, the same way on every shift.
The point is not more policing. It is to move people from policing to prevention.
Where the camera goes matters most. Put it where the decision still has leverage: before the step that makes a defect expensive or permanent.
Many early pilots fail because the camera is mounted where it is easiest, not where it is economically decisive: catch a misapplied label before packaging, an incomplete weld before coating, a missing seal before filling. The camera must also know what it is looking at (a product-mode handshake), or a model trained on one variant will misclassify another and drown the line in false rejects.
A camera station is not a gadget on a bracket. Optics, lighting, triggering and algorithms work together: shadows and glare create fake defects, motion blur creates false edges. It needs routine lens cleaning, lighting checks and air purge checks, or the plant will blame AI for what is actually dirt, vibration and drift.
“It is about shifting people from policing to prevention.”
“The camera should be placed where the decision still has leverage.”
“So a practical in-line inspection design begins with a product-mode handshake: the vision system must know what it is currently inspecting.”
Read the whole chapter free in the book reader.
Three lines, two possible camera positions each. Choose where the decision still has leverage.
Part 2
Detection alone is not quality maintenance
A camera that only rejects bad units can let the process run out of control and make waste faster.
Quality maintenance asks the next question: why are the defects happening, and what keeps them from starting?
In the chapter’s example, label rejects burst in the first twenty minutes after changeover. The cause: each shift set the label guide by feel. A fixed reference point ended it.
The team joined the reject data with state transitions and found micro-stops labelled “manual adjust” in the same window. A short gemba observation showed the applicator guide set by feel, with each shift using slightly different reference points. The fix went into the changeover standard, and the camera became a teacher of normal, not just a judge of output.
Vision also makes “done” measurable after maintenance. When a station is adjusted, rebuilt or cleaned, the camera shows at once whether output returned to normal at speed, not just whether the machine runs: the quality counterpart to “prove it on the trend.”
“If the camera simply sorts good and bad, it can become an expensive filter that still allows the process to run out of control, producing waste faster.”
“Now the camera becomes a teacher of normal, not just a judge of output.”
“That is quality maintenance, because the defect mechanism was reduced, not merely detected.”
Label rejects per hour, the week before and the week after. Choose what the plant did. The counts are an illustration.
Part 3
Machine learning does not create truth
Machine learning learns from labelled examples. If three people label the same picture three different ways, the model learns the disagreement.
So first, write a short defect dictionary: each defect, with pictures of what counts and what does not.
And teach the model the defects that cost money. A model that rejects every tiny cosmetic mark makes false rejects, and people start bypassing it.
Defects cluster during start-up, after changeovers and cleaning, during warm-up, or when a component nears the end of its life, so training data must cover those states. Ambiguous cases need a review path, or the model gets retrained on shifting opinions and drift returns in digital form.
The model’s confidence score becomes action only through thresholds that are set deliberately and governed. A low-confidence result should trigger a secondary check, especially early on, not an automatic reject. The goal is stable capability that protects customers and protects flow.
“To keep continuity with the discipline we have repeated since Chapter 9, the first rule is simple: machine learning does not create truth.”
“If three people label the same image differently, the model will learn disagreement.”
“The goal is stable capability that protects customers and protects flow.”
Three moments in setting up a vision model. What would the chapter do?
Part 4
Most errors are not fixed by retraining
False rejects and missed defects are not just technical errors. They are cultural events: the line stops to argue with the camera.
When it errs, ask why first. A dirty lens? Fading lights? Still set for the last product? Most of these are fixed by restoring basic conditions, not by retraining the model.
So the camera station gets the same care as any machine: operators keep the lens clean, maintenance keeps the lights and calibration in shape.
Autonomous maintenance protects the station’s basic conditions: clean optics, stable lighting, secure mounts, correct air purge, and reference-image verification. Planned maintenance protects calibration and replaces components before they drift into instability. Only a genuinely new defect, or a real change in the product, calls for updating the dictionary and retraining.
Defect bursts can also be joined to equipment signals. If rejects rise with temperature drift at a sealing station, or current variance on a conveyor drive, the plant has a hypothesis that crosses quality and maintenance without argument, and in time can predict defects before the reject boundary is crossed.
“False rejects and missed defects are not just technical errors; they are cultural events.”
“Each of these has a different countermeasure, and most of them are not “retrain the model” first.”
“Autonomous maintenance protects the camera station’s basic conditions: clean optics, stable lighting, secure mounts, correct air purge, and reference image verification.”
Five reasons a camera rejected good parts. Which fix does each one need?
Part 5
Correct, restore, or stop
Closed-loop means the camera does not just report; it triggers a response. There are three tiers.
Tier 1: a small automatic correction, inside limits engineering and quality approved. Tier 2: a short controlled stop to clean or re-seat, with a work order. Tier 3: stop and escalate, because the process is out of control.
Compensation is not restoration: if corrections keep going one way or hit their limit, restore the cause. After any stop, prove it: the next 20 units must pass at speed. And no permanent changes on the fly.
Which tier applies is decided in advance, per defect family and per station: a missing component is not the same as a small print contrast drift. The chapter’s example rule: three misalignment rejects in a rolling window of 20 units signals drift worth correcting.
Letting the system “auto-tune” without a boundary is the first dangerous mistake. Thresholds, correction limits and stop rules are part of normal: reviewed, documented and released under change control. The bench below lets you set what the camera sees and choose the response.
“In practice, a closed-loop system uses at least three response tiers.”
“For example: three misalignment rejects in a rolling window of 20 units, or a steady decline in print contrast over 60 seconds.”
“Compensation is not restoration.”
“A simple rule protects culture: no permanent changes on the fly.”
The Closed-Loop Station. Set what the camera sees at a labelling station, then choose the response and the rules around it.
Part 6
Try it where you work
Do these on a real machine: at work if you work with equipment, or at home on a car, washing machine, furnace or lawn mower if you do not. Tick each one when it is done. Safety first: look and clean only what you are allowed to touch, never reach into moving equipment, and follow your site’s lockout rules.
Part 7
Your “My Camera Station” card
One in-line camera, set up to prevent defects rather than just catch them. Nothing you type here leaves this device.
My Camera Station
Global Sovereign University · Modern TPM, lesson 14
Vision that prevents
- Place the camera where the decision has leverage.
- Tell it what product it is seeing.
- Write the defect dictionary first.
- Restore the basics before retraining.
- Correct, restore, or stop, by rule.
- Prove recovery: 20 units at speed.
A camera that only sorts is an expensive filter. No permanent changes on the fly.
Part 8
Check yourself
Six real situations. Pick what you would do. You will see why after each one.
Lesson 14 complete.
You can now place a camera where the decision has leverage, turn detection into prevention, keep the model honest with a defect dictionary, fix false rejects at the root, and close the loop with tiered responses. Next: Lesson 15, early equipment management and virtual prototyping.
Modern Total Productive Maintenance (TPM) by Dr. Gene A Constant. This lesson comes from chapter fourteen. Read the whole book free in the reader, or keep a copy.
Buy a copy — it funds a free seat
GSU is a 501(c)(3) nonprofit and book sales are its only income. A purchase is not a tax-deductible donation. This lesson is free and always will be.