jjlalli/Anzar

On-device acoustic leak detection for water pipes — ESP32 → STM32, no cloud. Named for the Amazigh rain deity. # Anzar Acoustic leak detection for water pipes. A low-cost device that listens to water pipes and finds hidden leaks. *Anzar (ⴰⵏⵣⴰⵔ) is the Amazigh word for rain, and the name of the rain deity invoked in North African drought rituals.* Tunisia loses close to a third of its drinking water to leaks in old underground pipes, and a leak stays invisible until a pipe bursts or a bill jumps. Anzar is a small sensor that clips onto a pipe and listens. A leak makes a faint, steady sound inside the pipe that people can't hear. Anzar picks it up, tells a leak apart from normal water use, and sends an alert so it gets fixed before the water is gone. 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 Three parts: a contact microphone clamped to the pipe, a microcontroller that runs the detection on-device, and a local alert. No cloud, no connectivity required — the classification happens on the MCU itself, which is what makes the unit cheap enough to put on a pipe and forget about. The hard part is the software that decides "leak" or "normal" from the sound. It looks at three things: how continuous the sound is, how much its energy varies over time, and where its energy sits in frequency. A leak is steady and broadband; normal use is short and bursty. ## Hardware The signal path was prototyped on an ESP32 with a piezo contact microphone. For the bench version I'm specifying an STM32 + MEMS microphone stack so inference runs on the microcontroller with industrial-grade sensing, and evaluating ST's edge-AI tooling for the on-device model. Sensor selection and the mechanical coupling to the pipe are the open questions. ## The detection demo `anzar_demo.py` builds simulated pipe signals (a steady leak hiss, normal taps, and a steady appliance hum as a hard case), pulls out those features, and trains a small classifier. On this simulated set it reaches an F1 around 0.88. Most of the errors …