A Leak Localisation System In Water Distributed Network Using Machine Learning

Png, Wen Hao (2022) A Leak Localisation System In Water Distributed Network Using Machine Learning. PhD thesis, UTAR.

Preview

PDF Download (4Mb) | Preview

Abstract

Water crisis has preponderated worldwide due to non-effective resource management and uncontrolled water losses. Leakage appears to be one of the major factors of the water losses. In the past decades, various detection and localisation approaches have been established to solve the problem, acoustic sensing is an effective localisation technique which has been extensively implemented in single pipeline system. However, the conventional acoustic sensing technique faces multiple challenges such as analytics complexity and time-consuming issue in complex piping network system. In this thesis, two leak localisation systems based on machine learning and remote-acoustic sensor network were developed as the solution. In the first part of thesis, a multi-level analytics framework (MLAF) was formulated for adaptive leak localisation in piping networks. The MLAF analyses the complex spatial acoustic signals from the sensor network and predicts the leak location through multi-level hierarchical analyses and sequential reasoning processes. The system aggregated path analysis, timecorrelation location analysis, and machine learning methods to predict the leak location. The processes were handled by an automated flow control to ensure time-effective prediction without needs of human supervision. The performance of MLAF in varying shaped piping networks has been validated by a set of characterisation tests. The result of a field prediction in a local district metered area (DMA) with excellent root mean square error (RMSE) has further confirmed the feasibility of the MLAF. In the second part of thesis, a mixed-model deep neural network (MDNN) was designed for alternative leak localisation. The system serves for non-adaptive application in a targeted piping network. The MDNN model comprises multi-layer classification and regression neural networks which was constructed based on Keras module. The MDNN was trained with a set of simulated leak data of the targeted piping network to identify the leak segment and location. Various neural network’s hyperparameters such as tensor shape, batch size and sample size were tuned during the training processes to identify the optimal model of prediction. The optimal MDNN model was validated with 2.3% mean absolute percentage error (MAPE) and 0.99 training accuracies. Finally, the optimal MDNN model was implemented to predict the leak location in the local DMA and achieved an excellent result of an average 3.2% location error.

Actions (login required)