AI in the factory starts with the data you already have

Most factory AI projects that go nowhere start with "which AI should we use?" The first question should be "which decision do we want to make better?"

Start from a decision, not an algorithm

A production line already makes decisions every day: which lot goes on the machine first, which machine is likely to fail next week, whether there is enough material for the next shift. If you can write the decision as one sentence, you know what to predict and how to measure it.

Most of the data already exists

A plant that has logged production for years has more data than it thinks: which machine each lot passed through, how long it took, why it stopped. In my research the starting point was almost 18 million rows from systems already in use. Not one new sensor was needed.

The real work is cleaning: dropping incomplete rows and choosing features that mean something on the floor, such as process code, lot number and production time. About 4 million usable rows remained.

Measure everything, not just accuracy

High accuracy can mislead. If some groups are rare, a model that always answers the common one still scores well. Look at macro-averaged precision, recall and F1, and use cross-validation to check the result is stable. Comparing three models, Logistic Regression reached almost 80% accuracy but only about 24% average recall: it missed nearly all of the small groups.

Put it where people work

A model in a notebook helps no one. I exposed mine as an API for the scheduler and built an Android app that keeps working when the signal drops, because the shop floor does not have good coverage everywhere.

Leave the last word to people

A good system explains its suggestion and lets people override it. Every correction becomes a training example for the next round, so the AI grows with the knowledge of the people in the plant instead of replacing it.