PipePulse: AI Leak Detection Clamp Sensor

What PipePulse is

PipePulse is a small sensor that clamps onto a water pipe and listens to it. A tiny machine-learning model running on the Nordic nRF54LM20 DK decides whether the pipe sounds normal or like it has a leak, then sends the result to your phone over Bluetooth. No cloud, and no cutting into the pipe.

Why I'm building it

Leaks are boring until they aren't. They start small, they hide in walls and under floors, and you usually find out when the water bill jumps or something starts to rot. The EPA says a faucet dripping once a second can waste more than 3, 000 gallons a year, and that's one faucet.

When I read the Sustain the System theme, I went through a bunch of ideas: pump health, motor wear, 3D printer failures, even beehive monitoring. I picked pipes because the problem is easy to understand, easy to test on a desk, and cheap to fix if you catch it early and expensive if you don't. It also suits what the DK does well: read a sensor, run a small model on the device, and talk over Bluetooth without burning power.

How it will work

Listen. An accelerometer on the pipe wall, wired to the DK, records short bursts of vibration a few thousand times per second.

Understand. The DK turns each burst into frequency features, and the Axon NPU runs a small model trained in Edge Impulse. It sorts each burst into idle, normal flow, or leak-like.

Remember what normal is. The node also keeps a baseline for that specific pipe and watches for slow drift. That's the predictive part: it should flag a joint starting to weep, not just a pipe that's already leaking badly.

Tell you. The status goes out in Bluetooth advertisements, so a phone can read it without pairing and the node can sleep between bursts. Connecting gives you history and scores. A small Web Bluetooth page shows the alert in plain words, like "possible leak, check the coupling".

How I'll test it

I'll build a short PVC section with a coupling and a tap. I'll record labeled data for no flow, normal flow, and a controlled drip from slightly loosening the coupling. Then I'll train in Edge Impulse, deploy to the NPU, and see if it can tell them apart on recordings it hasn't seen before.

What I'm unsure about

Whether a cheap accelerometer is sensitive enough to hear small leaks on low-pressure pipes. That's the first thing I'll test. If vibration alone isn't enough, I'll add a small microphone to the same node. I'll post the results here, including if it doesn't work as well as I hoped.

Where it stands

Concept stage. The parts list and plan are done, and I'll update this page as the hardware arrives.

Notes

The 3, 000-gallon figure is from EPA WaterSense, which says a faucet dripping once per second wastes more than 3, 000 gallons per year. Link it in the story: https://www.epa.gov/watersense/fix-leak-week (epa)

Fill in or delete the bracketed line. I didn't invent a personal story for you, and a made-up one is worse than none.