Two different claims travel under one name
"Edge AI" is used for two things that deserve separating.
The first is running inference locally: a model small enough to execute on a microcontroller or a small embedded processor, classifying what the sensors are doing without a round trip. That is an architecture decision, and for safety work it is usually the right one.
The second is the marketing claim that a product is intelligent, which most often means a threshold with a good dashboard. It is worth asking any vendor which of the two they mean, and what happens to the intelligence when the site loses its uplink.
What running on the device actually buys
Three things, and they are worth stating separately because only one of them is about speed.
- Independence. The alarm does not have a network dependency, so an outage costs visibility rather than protection.
- Latency. No round trip, and no queue behind whatever else is using the link.
- Data residency. Readings from a production floor describe how a plant runs - throughput, shift patterns, when things go wrong. Keeping the analysis local means that description does not have to leave the site to be useful.
Where a model helps, and where it must not decide
There is a real job for learned models in this domain, and it is recognising shape. A hazard has a characteristic movement across several channels at once - which ones rise, in what order, how fast, and what happens to the ones that do not. Distinguishing that shape from ordinary process variation is exactly the sort of problem a model is good at and a threshold is bad at.
What a model should not do is hold the safety decision. Not because models are unreliable in general, but because a safety function has to be auditable: after an incident somebody has to be able to say precisely why the system did what it did, and demonstrate that it will do the same thing again under the same inputs. A learned function that cannot be inspected line by line cannot carry that.
So the architecture here separates authority from analysis explicitly.
The deterministic layer is authoritative: the absolute-limit floor, the behavioural signature layer and the local alarm decide the safety response. The generative-AI layer is advisory only, and structurally so - it has no software interface, permission or code path by which it can invoke any actuator, and the system runs undegraded when it is switched off.
Why generative models are the advisory layer here
Generative models are used in this system for one thing: turning a decision the deterministic layer has already made into an explanation a person can act on at three in the morning.
That is a genuinely useful job. "Combustible gas at 2.3 times baseline, rising, air quality following, temperature flat" is correct and nearly useless to a shift supervisor. "A solvent vapour release, not a fire - the gas channels are moving and the thermal channels are not - probably from the paint line, spreading toward the packing bay" is the same information in a form somebody can act on.
But the model receives the decision; it does not make it. It has no software interface to any actuator, no permission that would let it acquire one, and the system runs undegraded with it switched off. That constraint is what makes it safe to use a probabilistic component in a safety product at all.
