Playbooks: how to apply it

Playbook: edge AI next to your existing PLC

For companies with machines that have a PLC and often SCADA, and that want more out of that data. In five steps you put a first model next to one machine, without changing the control system.

Sound familiar?

The same failure comes back every few weeks and nobody sees it coming.
There is a historian full of data, but nobody looks at it until something goes wrong.
A quality problem only shows up at final inspection or at the customer.

Step 1. Pick one machine and one question

Start with a problem that costs money and that you can count. A bearing that seizes unexpectedly. A line that makes rejects without anyone knowing why. A compressor that draws more power than last year. Write down how often it happens and what it costs each time. You need that number in step 5.

Step 2. List what is already measured

Ask your PLC programmer or the supplier for the list of tags on that machine, with unit and meaning. Check in the historian how often each value is stored. If a value is stored only once a minute, or only when it changes enough, you will not see fast signals. Note for each tag whether it says anything about your question from step 1.

Step 3. Add what is missing

Often the measurement that matters is exactly the one missing. A vibration sensor on a bearing housing, a current clamp on a motor, a camera above the belt. ISO 20816 sets out how to measure and evaluate machine vibration, by its size and by how it changes. You connect the sensors to an edge computer in the cabinet, which reads the PLC through OPC UA or Modbus. Nothing changes in the PLC program.

Step 4. Let the model only watch first

In this phase the model controls nothing. It raises alerts that you compare with what really happened. How many failures did it see coming? How often did it raise a false alarm? Track that for a few weeks, together with the people on the floor. They know which alert is right.

Step 5. Decide what happens with the result

Only when the alerts are right do you agree where they go. An alert to the maintenance team. A value the PLC uses within its own limits. An output that ejects a product. Emergency stop and safeguarding stay in the safety chain certified for that. Put the result next to the number from step 1.

What it costs

An industrial edge computer and a few sensors are a one-off expense. Most of the time goes into steps 2 and 4: working out what the data means and testing the model against practice. Weigh it against what the failure from step 1 costs per year.

How we do it ourselves

In SevenTwin, our digital twin platform, every measurement passes physical checks before it may enter the model. If an installation reports more output than its capacity allows, the measurement is rejected and you see why. A model that learns from bad data gives bad alerts. That is why the check sits as close to the source as possible.

Frequently asked questions

What if our PLC has no OPC UA?

Older controllers often speak Modbus, PROFIBUS or their own protocol. A gateway translates that without changing anything on the machine. Otherwise you read in plain 24V signals.

How much data does a model need?

Enough examples of the problem itself. A failure that happens twice a year gives little to learn from. Then you start with spotting deviations: the model learns what normal looks like and reports what differs from it.