The terms machine learning and AI are scattered about liberally at the moment, simultaneously the answer to all our prayers as well as a potential future threat to mankind.
Think SkyNet (forgive me, like many data scientists, I’m a sci-fi fan).
The truth is a bit more prosaic. But first, there’s a difference between machine learning and AI. With machine learning a machine, typically software, learns a task which it can perform repetitively and mechanically without error. These tasks tend to fall into one of the three Ds – dirty, dull or dangerous.
True AI on the other hand is more clever. Here the machine transfers its knowledge of one task to learn another, and another – generating more and more value from the data it’s analysing.
The application of AI to the huge data sets being created by the water industry is one of the central tenets of systems thinking. Like many sectors, water companies find themselves data rich thanks to massive investment in real-time monitors and alarm systems across vast networks of assets. The problem is that sifting through and extracting value from that data is incredibly hard for human beings to do.
In the case of leakage, companies are being challenged to make the next great leap in performance but, despite huge investment in monitoring, the process remains frustratingly manual and error prone.
This is where AI comes in. The manual task of listening to noise files is subjective and easy to get wrong. We realised there was an opportunity, with ample access to a data set, to do a cluster analysis and evolve new modelling technology to identify the real leaks.
Typically, noise graphs are shown in two dimensions as frequency over time. But in reality, the frequency profile of a noise is a lot more complex and interesting. Working with United Utilities, as part of its Innovation Lab, we took lots of files, which had been physically verified as either a leak or not, and pre-processed them to create a multi-dimensional graph with each as a point plotted in virtual space.
After much experimentation with cutting edge mathematical techniques, true leak signatures eventually clustered in the same part of the virtual space. This was a recognisable pattern, and Fido AI was born. As more and more verified data was entered into Fido’s neural network (or algorithms – it’s all right to interchange those terms), the AI got more and more accurate. Until, finally it was able to analyse even unverified sounds and, by recognising the features, get its decision right more than 90 per cent of the time. It really doesn’t matter where the noise comes from. Any sensor type will do just as well.
In research, when we asked experienced leakage analysts to listen to the verified files, they only agreed 60 per cent of the time. That means that about 40 per cent of files are difficult for humans to identify. And that’s very experienced analysts with 30+ years of proven expertise. With new members of staff the error rating is even higher.
But that was just the start. Now Fido could learn, we could train it to recognise other patterns. First was leak size. We did this by matching specific Fido analysed leak files with the corresponding reduction in water flow in the network when the leaks were repaired. The AI began to recognise new clusters of noise and kinetic data points linked to the size of the leak.
With enough verified data, Fido was soon correctly predicting leak size more than 90 per cent of the time. It no longer needs to flow data to confirm it.
It’s hard to say where the noise of this breakthrough has been loudest – in the pipes, or in the water industry itself. Knowing your largest leaks transforms resource prioritisation towards the ones that matter – the largest – cutting leak run time and water loss.
The next goal is to turn individual leaks into data intelligence assets and use them to help predict and even prevent future leaks. It’s not far away. Fido’s complex multi-relationship model stores the knowledge about leak behaviour, pipe material, ground conditions, temperature and pressure from the information it gleans from things like sensor files, flow data, even the dig team, to spot trends. This important data is usually too fragmented to be useful and overlooked once a leak is repaired. By capturing and persisting that data – things like what behaviour did the leak present, what caused it – Fido can analyse patterns to work out where it came from and therefore predict it.
We’ll soon also be able to identify accurate leak location between sensor points and even tell material type and pipe contents – all by learning to identify cluster patterns in leak sound files. This gives Fido applications beyond the water industry too. That’s the power of true AI. Dirty, yes. Dangerous, yes. But never dull.