Production and manufacturing
Visual inspection, maintenance planning, fault analysis: where AI holds up on the shop floor, what data it takes, and where the Machinery Regulation applies.
What this is about
Four tasks cover most applications:
| Task | What happens |
|---|---|
| Visual inspection | A camera detects scratches, cracks, missing parts |
| Predictive maintenance | Sensor values indicate an approaching failure |
| Fault analysis | Reports and logs are grouped by cause |
| Work instructions | Manuals become searchable and are turned into simple steps |
Introduction
- 01
Capture situation first
Fix camera, distance, angle, lighting and background, mechanically. Only then does data collection begin.
- 02
Collect defect images, including the rare ones
Good parts are plentiful, bad ones rare. The rare ones are the expensive ones. A few hundred per defect type is the target.
- 03
Pre-sort rather than decide
The system splits into "certainly good", "certainly bad" and "inspect". The value comes from the third group being small.
- 04
Collect the corrections
Every human correction is a new label. Collected systematically, after a year you hold a dataset nobody can buy.
The trade-off
- A defect that gets through costs a complaint, rework and trust.
- A false alarm costs inspection time and scrap.
- Both costs belong quantified before the threshold is set.
- The chosen threshold belongs documented and must not be adjusted in operation without that being logged.
Regulatory frame
Three circles interlock:
| Circle | When engaged |
|---|---|
| AI Act | Always as deployer; high risk when a safety component under Annex I |
| Machinery Regulation | When the AI performs a safety function |
| Product liability | For the manufactured product, regardless of the inspection method |
The decisive point is Annex I of the AI Act: an AI system used as a safety component of a product that is itself subject to conformity assessment counts as a high-risk system. A visual inspection judging only appearance routinely does not fall under it. An inspection deciding release of a safety-relevant component very much does.
Evidence
- Model version, checksum and date per inspection lot, not per day.
- The threshold used, with reasoning and sign-off.
- Sampling results from human re-inspection, continuously.
- Error rates broken out by defect type, not as an overall figure.
- Calibration state of camera and lighting, because a shift there changes quality.
Predictive maintenance, realistically
The most common reason projects fail is too thin a data basis on the failures themselves. A defensible model needs failures, and failures are rare and expensive.
- Count first how many documented failures of the same kind exist. Below 30, a model is not sensible.
- Rule-based thresholds on vibration, temperature and current draw solve a large share of cases without a model.
- The value comes from the warning time gained. That time belongs measured, not the hit rate.
Choosing a task in production
FREE ACCOUNT
Selection sheet and templates: production
A sheet that picks the first task out of the ones production actually has, with two tested templates and the review steps for them.
Worksheet3 items