Inspiration
Underground water pipes leak quietly for months before anyone notices — often only after a sinkhole or a burst pipe appears. Manual acoustic surveys are slow and expensive, and streaming raw audio from every sensor to the cloud is too costly in bandwidth and battery life. We wanted to build a system that could listen to pipes locally, only "wake up" and think hard when something sounds like a leak, and pinpoint the leak's location without needing GPS or constant data streaming.
What it does
Korteweg AI is a network of small sensors clamped onto pipes (at hydrants, valves, or joints) that:
- Listens to the pipe using a piezo sensor.
- Filters for candidate leak sounds using a lightweight AI model (a temporal convolutional network) running directly on the chip.
- Verifies real events with a second, more powerful AI model (a compact spectrogram transformer) — only when something looks suspicious, to save power.
- Talks to neighboring nodes over radio to compare timing of the sound arriving at each sensor.
- Calculates the leak's location using the speed of sound in the pipe and the time difference between nodes.
Instead of sending raw audio anywhere, each node only sends small bits of information: "I think there's a leak, here's the timing, here's how confident I am."
How we built it
- Sensing & hardware: A contact piezo sensor picks up vibrations from water escaping the pipe, feeding into an analog front end (amplifier + filters) and a fast analog-to-digital converter.
- Two-stage AI pipeline: A small always-on model watches for candidate leak sounds cheaply, and only triggers a bigger, more accurate model when it thinks it's found something — this saves battery.
- Physics-based speed calibration: We used a formula (the Korteweg equation) that accounts for pipe material, wall thickness, and water properties to figure out how fast sound travels through that specific pipe segment, since it's not the same everywhere.
- Timing without GPS: Since these sensors are often buried or indoors, we can't use GPS. Instead, nodes send timed radio "ping-pong" messages to each other to work out the tiny clock differences between them, accurate to millionths of a second.
- Locating the leak: Once we know (a) how fast sound travels and (b) the tiny time difference between when each sensor heard the leak, we can calculate roughly where between the two sensors the leak is — and how confident we are in that estimate. If we're not confident enough, the system honestly reports "leak is somewhere in this stretch of pipe" instead of guessing a precise spot.
- Handling noisy/messy signals: Pipes echo sound in confusing ways, so we used a signal-processing technique (GCC-PHAT) that's good at finding the true time delay even with echoes and noise.
Challenges we faced
- Battery vs. accuracy tradeoff: Constantly analyzing audio would drain the battery fast, so we had to design a "lazy" system that only does heavy computation when needed.
- Different pipes behave differently: Metal pipes, plastic pipes, different wall thicknesses — sound travels differently through all of them, so a single fixed "speed of sound" doesn't work. We had to calibrate per pipe segment.
- No GPS, no synced clocks: Getting multiple independent sensors to agree on exact timing without a shared clock source was one of the trickiest parts — solved using a four-way timestamp exchange between nodes.
- Radio network congestion: With many sensors on one pipe network, we had to carefully budget "airtime" on the radio so all the nodes could talk without messages colliding or timing out.
- Being honest about uncertainty: Rather than always giving a confident-sounding number, we built the system to admit when it isn't sure and report a range instead of a fake precise point — because a wrong "precise" answer is worse than an honest "somewhere around here."
What we learned
We learned how to combine physics, embedded hardware constraints, and machine learning into one working pipeline — and that real engineering means being upfront about what your system doesn't know yet, not just what it can do.
Log in or sign up for Devpost to join the conversation.