Differential analysis and machine-learning in the detection of water leaks

By <p>Founder and Chief Technical Officer</p>

Home > News > Differential analysis and machine-learning in the detection of water leaks.

FIDO is an implementation of Neil Edward’s white paper Applying Differential Analysis and Machine Learning Techniques to Temporal Data Series in Order to Determine Fluid Leakage from Pipelines.

This abstract explains how FIDO became the practical embodiment of Neil’s work.

Water loss: A global problem

As the leakage of fresh consumer water in the water main distribution network steadily increases it is becoming a pressure point for the water companies to resolve the loss of in excess of 13 billion litres a day. Identifying leaks is only the first part of the solution. After identification the leaks then require a physical resolution and a manual excavation is the only answer.

At present large numbers of the ‘digs’ are dry and find nothing, resulting in large amounts of wasted effort. FIDO seeks to resolve this by accurately identifying the size and location of leaks in a normally pressurised system, using multiple recorded data streams and differential analysis to give unprecedented sensitivity.

Neil Edward’s solution

When multiple similar data sources are compared to each other they give much more accurate reporting of variances, like leaks, than a single data source performing absolute measurements would be able to.

Using several free-flowing small intelligent data collection tools inside water mains could therefore identify the various data signatures created by leaks and pinpoint their location.

The first free intelligent domain observer (FIDO) devices combined micro-electromechanical systems (MEMS) technology and differential analysis to accurately determine leak size and location to less than a metre.

Proof of concept

Three FIDOs were deployed a few metres apart, each simultaneously collecting movement (rotation, acceleration, turbulence), audio, visual and pressure information for the duration of their journey inside a test pipe section.

The data was then uploaded to a cloud analysis platform for processing and the output was a marked up .fid file for importing in to such applications as Unity3D, processing.org and Google Maps to display the specific geographic locations for the ground works needed to fix the leak as well as a virtual in-pipe fly through.

A number of post collection techniques were used to analyse the data collected by the FIDOs on their different paths through the pipe. And because the multi path data analysis uses variances, not absolutes, it was much more accurate on leak location and size.

As FIDO acquires more and more real fix data the machine-learning algorithm is improving and re-applying old data to look for other data patterns. For example; running successive tests at different pressures can identify optimum operational pressure.

FIDO today

FIDO is designed to act as both a stand-alone product or a decision support and verification tool to supplement top level network analysis tools and identify with the best accuracy possible the physical locations of leaks and their size.

FIDO does not require highly skilled operators in the field. FIDO is part of the engineer’s diagnostic tool kit.