Ruggedised, industrially-rated compute platforms running AI inference physically close to the process — in the control room or on the plant floor — rather than in a distant cloud region.
Embedded edge-AI inference in transmitters
Field transmitters with an onboard inference chip that classifies faults, drift and process anomalies locally, before the signal ever reaches the control system.
Short entry. This page summarises our current placement. Full, sourced analyses are published first for the technologies closest to practical use; see the list of full analyses.
How it works
A conventional transmitter converts a sensor signal into a process value and a handful of status bits. An edge-AI transmitter adds a small inference engine, usually running on a microcontroller-class or low-power neural processor, that looks at the raw high-frequency signal before it is filtered and averaged away.
The model is trained offline on labelled examples, compressed and quantised to fit the device's memory and power budget, then loaded as part of the firmware. In service it outputs a classification or a health index alongside the primary measurement: bearing wear from a vibration spectrum, cavitation from an acoustic signature, or entrained gas from the drive signal of a flowmeter.
Why it matters
Moves anomaly detection to the millisecond, zero-latency edge — useful for vibration, acoustic and flow signatures where the interesting event is a shape, not a threshold.
Where it fits
The natural first applications are measurements where the useful information is in the shape of a waveform and where sending raw data upstream is impractical: vibration, acoustic emission, and flow or level devices with rich internal signals. It suits remote or bandwidth-limited points more than a well-connected unit where the same analysis can run on a server.
Why we place it at the Emerging Idea stage
First commercial pilots are appearing in vibration and acoustic monitoring; instrumentation vendors are still standardising on model formats and update mechanisms. Estimated time to the Proven in Use stage: 5–10 yrs. What the stages mean.
Key risk
Model drift, firmware-update governance and functional-safety certification of an AI-containing transmitter are all unresolved for SIL-rated loops.
Questions to ask before committing
- Does the AI output ever influence the measured process value, or is it strictly a separate diagnostic channel?
- How are models updated in the field, who signs them, and can a device be rolled back to a previous model?
- What evidence exists of detection and false-alarm rates on equipment similar to yours, not just on a demo rig?
- Is the device intended for any loop with a safety integrity requirement, and if so what does the certificate actually cover?
Standards and terms to know
- Inference
- Running an already-trained model on new data, as opposed to training it.
- Quantisation
- Reducing the numerical precision of a model so it fits on small, low-power hardware.
- Model drift
- Loss of accuracy over time as the process or equipment moves away from the training conditions.
Related technologies
See this entry on the full maturity map · All Instrumentation entries · How placements are made. Stage and time-horizon placements are editorial judgements based on public information as of October 2026, not engineering, safety or investment advice; see the Terms of Use.