Lodestar

Thousands of machines, no wiring budget, and nobody to change batteries.

The client services rotating equipment across a few hundred plants. Their technicians were walking routes with handheld analysers, which meant a bearing could start failing the day after a visit and run for a month before anyone knew. They wanted a node they could stick on anything, in numbers, without an electrician, and then forget about.

The brief in one line: five years on a primary cell, sampling often enough to catch a fault a week out, in a housing that survives wash-down, oil and a dropped wrench.

Started with the power budget and let it argue with everyone.

Battery life is a systems problem. We built the energy model on day one and kept it live through the program: every sensor choice, every radio decision, every firmware wake-up had to justify itself against the five-year line. The model, not opinion, settled the debates.

On the electrical side that meant a MEMS accelerometer with a hardware wake-on-threshold, a sub-GHz radio with a duty cycle measured in seconds per day, and a boost stage that stays efficient as the lithium thionyl chloride cell ages. The firmware does spectral analysis on the node so the radio carries features, not waveforms.

Mechanically, the client's technicians told us the same thing three different ways: they mount these on ladders, with one hand. So the base is a rare-earth magnet with a stud-mount option, the housing is a single glass-filled nylon moulding with a laser-welded lid, and there is nothing to align. Put it on the machine, press the one button, walk away.

We ran the vibration, thermal shock and ingress testing in-house, then took the design through radio certification and to a regional contract manufacturer with an end-of-line fixture that programs, tests and pairs each unit in forty seconds.

• The magnet base is bonded and mechanically captured, because bonded alone failed our thermal cycling at the third round. That test paid for itself.
• The antenna is a trace on the main board, tuned with the housing and the steel it sits on, not in free space. Range doubled once we tested it on an actual pump.
• Spectral features are computed on the node and sent as fourteen bytes. The waveform is kept locally and can be requested when something looks wrong.
• The single button confirms mounting with a short flash, and nothing else. Every other interaction happens on a phone.
• Firmware updates over the air were designed in from the start, with a fallback image, because five years is a long time to be wrong.

The first production run went into the client's own service contracts. The second, larger run followed nine months later once the field data showed the energy model was holding.