What two sensors hear along a leaking pipe
In a Sheffield lab, one accelerometer stayed beside an open leak while a second moved along the pipe. We analysed all twenty recordings to see how far the two sensors could still pick up the same vibration.
Explore the measurements
Move the second sensor
Both sensors recorded the pipe at the same time. At each frequency, a coherence near 1 means the two recordings vary together closely. A value near 0 means they share little consistent vibration at that frequency.
At 1 m, mean coherence across the displayed frequencies was 0.521 in the first recording and 0.516 in the repeat.
The dotted line pairs the fixed sensor from one recording with the downstream sensor from the other. Both runs include the open leak. Pairing them this way removes their shared timing.
What the distance changed
Mean coherence across the displayed frequencies fell from 0.519 at one metre to 0.207 at ten metres, averaging the two recordings at each distance. At ten metres, both runs still had a peak near 420 Hz where coherence exceeded 0.78. Much of the shared vibration had fallen away while this narrow band remained.
Our water-main model asks where a robot should listen when two pipes fit the first reading. The robot also needs a place where enough shared vibration reaches both sensors to make a timing comparison meaningful. Whether choosing that place reduces wasted visits on a buried network still needs a field trial.
Try the water-main modelMeasurements and method
Sheffield’s team drilled a 3.6 mm hole in a 63 mm plastic pipe and held the water at 2.8 bar. They recorded both accelerometers at 8,192 samples per second for 30 seconds, repeating the experiment at each metre from one to ten. The fixed sensor sat directly over the known leak in this above-ground loop. Every recording includes the leak, so this series cannot measure how often a detector would give a false alarm or send a crew to the wrong pipe.
We computed magnitude-squared coherence with 2,048-sample Hann windows and 50% overlap, following the window length and overlap in the researchers’ paper. The chart covers 4 to 1,000 Hz in 4 Hz steps. Each summary is the arithmetic mean across those 250 frequencies. The two cross-run pairings are averaged to draw the dotted line.
Results for all ten distances
| Distance | First run | Repeat | Different runs |
|---|---|---|---|
| 1 m | 0.521 | 0.516 | 0.006 |
| 2 m | 0.447 | 0.460 | 0.005 |
| 3 m | 0.333 | 0.332 | 0.005 |
| 4 m | 0.313 | 0.315 | 0.005 |
| 5 m | 0.319 | 0.320 | 0.004 |
| 6 m | 0.282 | 0.290 | 0.004 |
| 7 m | 0.231 | 0.219 | 0.005 |
| 8 m | 0.190 | 0.190 | 0.004 |
| 9 m | 0.199 | 0.194 | 0.005 |
| 10 m | 0.201 | 0.212 | 0.005 |
Recordings by Mohammad Reza Shekofteh. The laboratory photograph is Figure 1a from the paper by Mohammad Reza Shekofteh, Kirill V. Horoshenkov, Edward John, Claire Gowdy, Andrew Blenkharn and Joby B. Boxall (2026), cropped and converted to grayscale. Minsky Robotics calculated the spectra shown here. Original recordings · Published experiment · CC BY 4.0