Your cart is currently empty!
The Science of
“Digital Water Signatures”
Most water monitors simply measure volume – how much water you used.
Sonic goes further. It measures behaviour.
Our proprietary AI technology doesn’t just see “10 litres of water”.
It sees the unique pattern of how that water was drawn, using a process we call Digital Water Fingerprinting.
1.The Digital Fingerprint
Every appliance has a unique hydraulic personality. Sonic analyses the flow of water 8 times every second, looking for 102 different parameters that then result in five distinct variables that create a “signature” for every event:
-
Attack: How fast does the water start? (A toilet flush hits hard and fast; a tap opens gradually).
-
Sustainability: How long does it run? (A bath is long and steady; a dishwasher pulses).
-
Falling: How does it stop? (A toilet float valve closes slowly and proportionally; a solenoid valve snaps shut).
-
Approach Speed: Is the flow changing over time? (A slow, creeping increase often signals a pipe freezing/bursting).
-
Flow Variability: Is the pressure erratic or stable? (Washing machines have high variability; leaks are dangerously constant).
2. De-Multiplexing: Untangling Simultaneous Events
The biggest challenge in water monitoring is the “Cocktail Party Problem” – what happens when a toilet flushes while the shower is running?
Sonic uses an advanced Signal Decomposition Algorithm (De-multiplexing). It views the water flow as a complex waveform and uses gradient methods to search for “critical points” (extremes and inflections) in the signal.
This allows the AI to mathematically “separate” the signals. It can identify that a complex wave is actually two distinct “child events” – a steady shower overlaid with a short toilet flush – and classify them individually. In our development tests, we achieved a 92% success rate in correctly separating these simultaneous events
3. The Engine: Hybrid Neural Networks
Behind the scenes, Sonic uses a sophisticated hybrid AI architecture to make these decisions:
Dual Convolutional Neural Networks (CNNs)
We use two deep learning networks that analyse raw water flow data and pressure readings, extracting over 1,800 distinct data features from every event.
Expert Feature Engine
These deep learning features are combined with logical “Expert Features” (human-defined rules about volume, peak density, and duration).
Final Classification
All this data feeds into a Fully Connected Neural Network (FCNN) which outputs the final probability.
The Result: Continuous Learning
During our intensive Beta testing programme, where we trained the engine on over 175,000 verified events, Sonic achieved an average recognition efficiency of 98.36%.
In the real world, we face a wider variety of plumbing configurations and new appliance models that the system has not seen before. While this naturally means initial accuracy may fluctuate as the system encounters these “unknowns”, our platform is built for continuous learning.
As we ingest more data and retrain our algorithms on these new appliance signatures, we project a sustained accuracy rate in excess of 90% for broad market deployment.
