Industrial AI and physical AI, explained

Hardly a trade fair or keynote gets by without two terms at the moment: industrial AI and physical AI. Both sound like the future. Both are often mixed up. And neither tells a company with a machine park much about what it can do differently on Monday.

Here is a sober classification.

Industrial AI: AI that works with machine data

Industrial AI means AI for industry and manufacturing. The difference from AI in the office lies in the data. A language model in the office reads emails, contracts and minutes. Industrial AI also works with measurements, time series and the knowledge from maintenance and production: vibration, temperature, power draw, output counts, fault messages.

Typical tasks are:

• Condition monitoring: how is the machine doing right now?
• Predictive maintenance: what is building up before the line stops?
• Energy and evidence: where is energy lost, and how do you prove it to customers or auditors?
• Knowledge at the machine: which maintenance instruction applies, and what happened last time?

The benefit rarely comes from a single measurement. It comes when measurements, documentation and experience meet. A bearing that runs warmer is a number. A bearing that runs warmer although it was replaced three weeks ago is information.

Physical AI: AI that perceives the physical world

Physical AI is the younger term, coined mainly by chip and robotics makers. It means AI that does not only process text and images but perceives and understands the physical world: through sensors, cameras and measurements.

The term has two levels, and the distinction is worth making:

1. Perceiving and understanding. The AI captures what happens at the machine and explains it.
2. Acting. The AI controls things itself, for example as a robot or an autonomous vehicle.

Headlines are mostly about the second level. For most companies, physical AI starts with the first. Before an AI may change anything on a machine, it has to understand the machine reliably. And that understanding alone already saves downtime.

Edge AI: computing where the data is created

Industrial AI and physical AI create a lot of data. A single vibration sensor easily delivers millions of data points a day. Sending this data to a remote data centre costs bandwidth, time and control.

Edge AI turns this around: the computing happens where the data is created, at the machine and in the plant. Nothing leaves that does not have to leave. For machine data this is more than a technical question. Measurements reveal cycle times, recipes and utilisation. Those are trade secrets.

How to recognise a good offer

Anyone who wants to introduce industrial AI should ask a few questions before talking technology:

• Where does the data run? In the plant or at the vendor? And who can access the raw data when the collaboration ends?
• Does it fit the machines you have? Can it be retrofitted to existing equipment, or does it require a new machine park?
• Does the AI explain where its statement comes from? A warning without a reason is quickly ignored on the shop floor. A warning with a measurement and a source is taken seriously.
• Does it connect measurements with knowledge? Sensor data alone is a dashboard. Only with the maintenance instruction and the history does it become an answer.
• Can you get out again? Open interfaces and models you can swap are the insurance against dependency.

How we do it with MINERVA

MINERVA is our answer to these questions: industrial AI in the sense of the first level of physical AI. We develop the hardware ourselves. A probe measures as a sensor at the machine, a node computes at the line, a hub in the plant. On top sits an AI that brings measurements together with maintenance instructions and history and answers questions about the machine with sources. All in the plant, without the cloud.

MINERVA is in the design partner program. We develop it with companies that bring their requirements from the shop floor.

Industrial AI, physical AI, edge AI: the terms are new. The question behind them is old. How do you get more out of your machines without giving up control over your data?