Narrow AI in manufacturing: what actually works in 2020
An honest review of the tasks where AI in manufacturing already delivers measurable results - and where it does not yet.
When the conversation turns to AI in manufacturing, I hear two extreme positions. The first: "this is a revolution, everything will change". The second: "this is hype, we do not need it". Both are imprecise.
The reality is this: narrow AI - systems trained to solve a specific, well-defined task - already works in production environments where there is enough data and the task is clearly formulated. And it does not yet work where those conditions are absent.
What already works
Visual quality inspection. This is the most mature area. Machine vision systems based on deep learning detect surface defects, geometric deviations, incorrect assembly. They work faster than humans and without fatigue. The barrier to entry has dropped significantly over the past five years.
The success conditions here: enough labelled data, controlled shooting conditions, clearly defined defect classes.
Equipment failure prediction. If sensor data is being collected from equipment - vibration, temperature, current - ML models can find patterns that precede failures. This is not fortune telling in the general sense. It is statistical anomaly detection: this pattern in the past always preceded a failure.
The condition: enough history of sensor data with links to maintenance incidents.
Scheduling and load optimisation. Planning production jobs is a combinatorial optimisation problem where ML approaches compete with classical algorithms. For large problems with many constraints, modern methods show real gains.
Where it does not yet work
Where there is no data. If equipment is not connected, failure history is not systematically recorded, and defects are written in a notebook - there is nothing to train a model on. This is the first and most common barrier.
Where the task is not defined. "Have AI monitor quality" is not a task. A task is: "detect scratches on the surface of part A longer than 0.5 mm in zone B". Without specifics there is no model and no success metric.
Where there is no infrastructure to act on the result. The model predicted a failure - what happens next? Who gets the notification, in what form, through which system, what decision is made? If that is absent, the model runs in a vacuum.
How to assess readiness for a pilot
When evaluating a specific task:
- Is there historical data for this task - and covering what period?
- Is the output that needs to be predicted or detected defined?
- Who will use the model's output and in which process?
- Is there a team that will maintain the system after launch?
- What happens if the model makes an error - what is the acceptable error level?
If there are answers to all five questions, a pilot is worth starting. If not - get the answers first. An AI project without them turns into a demo that will not survive to production.