Hear the bearing go before the line does.
Most plants have assets with no digital output at all: an analog gauge, a motor that only tells you it failed by stopping. This layer gets data off them and learns what normal sounds and looks like, so the exception is what reaches a person.
What it catches
- Vibration, thermal and acoustic drift away from an asset’s own baseline
- Analog gauges, dials and seven-segment displays read on a schedule
- Process values wandering outside the envelope for that product run
- Leaks, drips and surface changes visible before an alarm exists
- Assets running outside their duty cycle or rated load
Where it reports
- CMMS work orders with the reading and the evidence attached
- Historian and time-series stores through governed HTTPS APIs
- Existing condition-monitoring dashboards via outbound webhooks
- Bounded telemetry to the PLC through a Zephyr companion
- A retained record of every reading, with the model that produced it
What actually runs on the device
- Task
- Compact CV for gauge and display reading, plus signal models on sensor streams
- Typical target
- Raspberry Pi 5, x86_64 gateway, or QCS6490 reference design
- Mode
- Scheduled reads, with continuous sampling where a sensor exists
- Reach
- Whatever you grant on that asset. Read-only by default
Every one of these passes the Oddy test against your named device before it is admitted. A candidate that does not fit the device envelope never receives an edge directive.
Common questions
We have no sensors on these assets. Does this still work?
That is the common case, and it is why gauge reading is part of this. A camera pointed at an existing analog gauge turns an asset with no digital output into a time series without touching the asset or its controls. Where sensors do exist, the agent reads those instead — the point is to work with what the asset already has.
How is this different from the anomaly detection our historian already does?
Threshold alarms in a historian fire on values you already knew to watch. The useful cases are usually the ones nobody wrote a rule for. This layer establishes a baseline per asset and reports departures from it, with the evidence attached — and it can read assets your historian has no connection to at all.
Can it schedule the maintenance itself?
It can open a work order if you grant it that reach, and many deployments do exactly that — a CMMS write path scoped to creating records is a low-risk grant. Approving, scheduling and closing the work stays with your maintenance team unless you decide otherwise.
How long before it knows what normal looks like?
Long enough to have seen the asset across its real operating range, which usually means at least one full production cycle including changeovers. We would rather establish that on your asset than quote a number that will not hold.
Start with one line.
Send your SOPs and we will audit them free — the agent that fits your procedures, and the hardware it runs on.