A model that spots a defect three seconds after the part has left the station is a report, not a control. On the factory floor, timing, connectivity and safety decide where AI can run.
That is why most useful industrial AI ends up at the edge: on gateways, industrial PCs or embedded devices next to the line, with the cloud used for training, fleet management and analysis.
Why the edge, not only the cloud
Plants have constraints that office systems do not:
- Latency — vision inspection and process control need answers in milliseconds
- Connectivity — links to the cloud can be slow, shared or intermittent
- Bandwidth — streaming raw high-resolution video from every camera is expensive
- Data sensitivity — process parameters and recipes are often trade secrets
- Uptime — the line cannot stop because a remote service is unavailable
Edge inference handles the time-critical decision locally. The cloud receives summaries, samples and exceptions.
Common use cases that work
- Visual inspection for surface defects, assembly errors and missing components
- Predictive maintenance from vibration, temperature and current signatures
- Energy monitoring and anomaly detection on utilities
- Process parameter recommendations reviewed by an operator
- Safety zone monitoring, alongside, never instead of, certified safety systems
Each has a clear signal, a measurable outcome and a person who acts on the result.
Respect the OT network
Operational technology networks were built for stability, not for frequent change. Many devices speak industrial protocols and run for decades.
Good edge designs:
- Read data through established industrial protocols and historians rather than poking PLCs directly
- Follow the plant’s zone and conduit model, with edge devices in a defined segment
- Push data outward through controlled gateways; avoid inbound connections from the internet
- Leave control logic in the PLC; AI recommends or flags, the control system acts
- Align with industrial security standards such as IEC 62443 as the plant interprets them
Bring the controls and maintenance engineers in early. They know which signals are trustworthy and which sensors have been miscalibrated for years.
Data is the hard part
Industrial data is messy. Sensors drift. Timestamps disagree between systems. Labels for defects depend on which inspector was on shift.
Before model work:
- Align clocks and timestamps across sources
- Agree on a defect taxonomy with quality teams and label against it
- Capture the context: product variant, batch, shift, maintenance events
- Keep raw samples for the cases operators dispute
Lighting changes, new suppliers and tooling wear all shift what the camera sees. Plan for retraining as routine, not as an emergency.
Operating models across many sites
One pilot line is a project. Twenty lines across several plants is a fleet.
- Package models and runtime together, versioned and signed
- Roll out in stages with the ability to roll back per device
- Monitor device health, model latency and prediction distribution centrally
- Keep the device useful when offline, queueing data until the link returns
- Document who on site can restart, update or bypass the system
Keep operators in the loop
The person at the station knows when a flagged part is fine. Make it easy for them to confirm or reject a prediction, and feed those decisions back into training data. A system operators trust gets used. One that stops the line on false alarms gets switched off.
Where to start
- Pick one station with a clear, costly problem and an engaged supervisor
- Confirm the data you need is available and time-aligned
- Run the model in shadow mode alongside existing inspection
- Agree success criteria with quality and production before go-live
- Design the fleet update and rollback path before the second line
Incubics builds edge AI and industrial IoT systems that fit existing OT networks and keep running when the cloud does not.