Utilities / Acoustic Classification

Water Pipe Leak Detection

Working as embedded members of a client's data-science team to classify water-leak acoustic signals from hydrophones — distinguishing leak size, flow rate, pipe size and material from noisy, pressure-dependent sound.

The challenge

Leak noise is often treated in the industry as effectively random: it varies day-to-day with water pressure, and a small leak rushing through a small hole can be just as loud as a large leak flowing slowly through a larger one. The goal was to classify acoustic signals from hydrophones to determine flow rate, pipe size and material — and so tell large leaks from small.

Our approach

Embedded within the client's data-science department, we ran a pre-project workshop to refine the approach, then carried out a literature review and acoustic-data preparation. We guided the team through several modelling routes — direct classification, flow-rate inference, time-based anomaly detection, and general versus region-specific models — alongside training and consultancy, building toward classifiers in Azure ML Studio.

Results

We delivered reports, documentation, results and recommendations, equipping the client's team to take the most promising classification approaches forward.

Technologies

• Neural networks
• Machine learning
• Classification
• Azure ML Studio
• Signal processing

← Back to work