A model trained on six months of production data predicts scrap well in testing and poorly on the line. The reason is simple: during those six months, each shift ran the machines its own way. The model learned the habits of three teams, not the physics of the process.
AI learns from the past. If the past is chaos, the model learns chaos.
What lean gives AI
Standard work creates one way of working. Without it, variation between shifts and setups looks like a pattern and the model chases it.
5S and visual order fix the positions of parts and keep camera views and sensors clean. A vision system that sees a different background every morning will never be stable.
TPM basics bring machines into their basic condition. Otherwise a maintenance model predicts the effects of poor care, which a cleaning routine would remove at no cost.
Problem solving removes the obvious causes. AI is for what remains when simple causes are fixed.
Daily management creates the routine in which a warning is acted upon within a defined time. A prediction nobody reacts to is noise.
Fix the simple causes first
In one plant the escape rate of molded housings was 610 parts per million. Before any camera was bought, the team installed better lighting at the manual inspection, wrote a defect catalog with limit samples and photos, and trained all three shifts on it. Escapes fell to 420 ppm.
The vision pilot that followed started from a cleaner base: labels were consistent because the catalog existed, and the light was already controlled. The improvement before AI made the AI better.
Four checks before a pilot
Ask four questions about the process behind a use case:
- Is there a current standard, and is it followed?
- Is the problem defined with a baseline in numbers?
- Have the obvious causes been fixed with simple means?
- Is it clear who acts on a result, how and how fast?
A “no” to any of them is not a reason to give up. It is the first work package of the project, and often the cheapest one.
The lean test
If a checklist, a poka-yoke or a better light solves the problem, use that. It costs less, breaks less and needs no retraining when the material changes.
AI does not replace lean. It extends it into areas where the volume or the complexity is beyond what people can see. The plants that get the most from AI are usually those with the strongest basics, because their data describes a process that behaves the same way tomorrow as it did yesterday.