Detecting methane is easy. There are six or more off-the-shelf technology types available right now that will tell you methane is present. Some are sniffers that report a concentration in the air — put several out and triangulate roughly where the plume came from. Others see, smell, or otherwise detect the gas directly. A number of them will even quantify how much methane is coming off a location.
None of that is the hard part.
The problem is discrimination, not detection
What those systems could not do, in trial after trial, was tell the difference between process emissions — the venting and operation we already know about and expect — and fugitive emissions, the ones nobody knew about.
The consequence shows up the first week a system goes live. An alarm arrives, and it is for anything that is happening. It might be for something entirely routine. But if the system alarms on everything, an operator has no way to know when an alarm is the one that matters.
It alerted me about a needle in stacks of needles. Not just a needle in a haystack — everything looked like a needle.
That is the failure mode. Not missed detections; undifferentiated ones.
What an unactionable alert looks like
Triangulation puts the source “somewhere over there.” You do not know where it is. You do not know what it looks like. You do not know what parts, tools, or crew you would need to bring. You know only that some amount of methane was present, and that the system is telling you so continuously.
The result is alarm overload. The queue fills with events that are technically true and operationally useless. Field teams learn — correctly, given the information they are handed — that most alerts do not warrant a truck roll. Trust in the system erodes, and with it, the response times that justified buying it.
The requirement this creates
A monitoring system earns its place only if the alert it raises is one a person can act on. In practice that means three things the detection layer alone cannot supply:
- Discrimination. The system has to know the site’s normal process behavior well enough to stay quiet about it, and to escalate only what departs from that baseline.
- Context. An alert needs to arrive with the evidence attached — what the site looked like and sounded like at the moment of the event — so the decision to dispatch can be made before anyone drives out.
- Attribution. “Somewhere over there” is not a work order. Source location and equipment attribution are what turn a detection into a repair.
This is the gap Zensory.ai™ was built to close. Fusing video, audio, and optical gas imaging on a single edge platform lets machine learning establish what a site normally does — then discern true anomalies from the ordinary noise of production, and document each one well enough to defend it later.
Detection was never the bottleneck. Knowing which detections deserve a response always was.