The first water pipelines were made of wood, clay or stone, and date back to ancient civilizations. The Roman aqueducts are perhaps the most well-known example of early water pipelines. Over the years the material used to make pipelines evolved and in the 1800s, water pipelines were mostly made of metals like cast iron. The development of water pipelines allowed for the growth of cities and the rise of industry. Today, water pipelines are made of a variety of materials, including PVC and HDPE.
There are over 3.5 million km of pipeline around the world today, with some stretching thousands of kilometers to deliver water to hard reach places or areas with water scarcity. Pipelines have made water more accessible, cities have designed complex WDN to ensure the most efficient water deliver system. However, regardless of the material, regardless of the pipe network, the pipes are still prone to leaks.
Leaks are detrimental; environmentally, economically and socially. Pipelines periodically lose 20-30% of water transported, and can reach up to 50% in some areas. Leaks are estimated to account for up to 70% of these water loses. Considering the large volumes of water transported daily over the world, the loss due to leaks is substantial for one of our most important resources. So I set out to design a leak detection and localization system using AI.
Now I was not inventing the wheel (nor re-inventing it). There already exists a number of solutions to this issue using AI and IoT devices. These solutions however, are either very complex or very expensive but usually both. My initial goal was to take this same technology and with minimal loss to its accuracy and precision, make it less complex and more accessible. However, Covid-19 hit and made it rather difficult to implement and test.
With the pandemic preventing me from accomplish my objective, I had to switch things up. Instead of building a smart leak detector I focused on a simulation to find out how accurately a classification model could detect leaks and a deep learning algorithm to localize the leak. I used vibration as the parameter to track (although not practical to use for good reasons from an engineering point of view, but if the advantages outweighed the drawbacks, it could be worth investigating further). My model was able to detect leaks with an accuracy of 82% and could localize 85% of them within a 1m radius for a pipe of 10m long. I had tried to use a classification model to determine the pipe material however this negatively impacted the leak detection and localization which needed this information to proceed with its task.