Experimental investigation into techniques to predict leak shapes in water distribution systems using vibration measurements

Water loss from leaking pipes represents a substantial loss of revenue as well as environmental and public health concerns. Leak location is normally identified by placing sensors either side of the leak and recording and analysing the leak noise. The leak noise contains information about the leak’s characteristics, including its shape. Whilst a tool which non-invasively provides information about a leak’s shape from the leak noise would be useful for water industry practitioners, no tool currently exists. This study evaluates the effect of various leak shapes on the vibration signal and presents a unique methodology for predicting the leak shape from the vibration signal. An innovative signal processing technique which utilises the machine learning method Random Forest classifiers is used in combination with a number of signal features in order to develop a leak shape prediction algorithm. The results demonstrate a robust methodology for predicting leak shape at several leak flow rates and backfill types, providing a useful tool for water companies to assess leak repair based on leak shape.