Anzar

Anzar is a small sensor that sits on a water pipe, feels it vibrate, and tells a leak apart from normal water use. It also catches taps left running. Everything runs on the microcontroller, on a battery pack, with no internet connection.

Anzar (ⴰⵏⵣⴰⵔ) is the Amazigh word for rain, and the name of the rain deity invoked in North African drought rituals.

Why

Roughly a third of Tunisia's drinking water never reaches a tap: by the national utility's own 2023 figures, the transmission network delivers 86.8% of the water produced and the distribution network 77.7% of what it receives (SONEDE 2023 report, as summarised by WMC). The same thing happens at small scale in a house or on a farm: a leak behind a wall or a tap left open in a field keeps running until someone happens to notice.

I'm from Tunisia, where water is scarce and a lot of it is lost this way, so this is the problem I wanted to work on.

How it works

1. A MEMS accelerometer on the pipe samples vibration at 1.67 kHz, in windows of 512 samples on 3 axes (about 0.3 s each).
2. A NanoEdge AI classifier running on the STM32 labels every window: no flow, use, or leak.
3. The firmware smooths those labels with a vote over the last five windows and keeps time:
   • leak-level flow that lasts 30 minutes raises a leak alert;
   • use-level flow that lasts 20 minutes raises a tap left running alert.
4. Alerts show on the board's LED (steady for a leak, blinking for a tap) and can sound a buzzer. With a computer connected, they are also printed on the serial port.

Over any short stretch, a tap left running feels exactly like a tap in normal use. The only difference is how long it lasts, so the classifier names the flow and a timer makes the call.

Hardware

• STM32 NUCLEO-F401RE board
• X-NUCLEO-IKS01A3 MEMS shield (LSM6DSO 3-axis accelerometer)
• a 7-12 V battery pack on the board's VIN pin
• an enclosure, clamped or strapped to the pipe

Parts, assembly and mounting: docs/hardware.md.

Repository

firmware/app/        application code, drops into an STM32CubeIDE project
firmware/nanoedge/   where the trained NanoEdge library goes (not committed)
tools/capture.py     records labelled training data from the board
tests/               host tests for the decision logic
docs/                hardware, training guide, diagrams

Getting started

1. Set up the firmware project and flash the datalog build: firmware/README.md
2. Record training data and build the model: docs/training.md
3. Flash the detect build.

The decision logic is plain C and has tests that run on any computer with a C compiler:

sh tests/run_tests.sh

Status

Prototype. Units are running on household plumbing and on our family's olive grove in Regueb, Sidi Bouzid. The thresholds above suit a home; long legitimate flows such as irrigation need a longer tap threshold (see docs/training.md).

License

MIT, see LICENSE. ST's LSM6DSO driver and the NanoEdge AI libraries are not part of this repository and keep their own licenses.