Deep learning identifies leak in water pipeline system using transient frequency response

Abstract

Pipeline leak identification method using transient frequency response (TFR) has been researched in the past two decades. To extend this method to a more general water pipeline system with hydraulic uncertainties, this work (1) introduces deep learning (DL) into the TFR-based leak identification framework and (2) develops extended TFR equations in matrix form for DL learning set generation. In this framework, TFR equations are firstly solved in a pre-calibrated hydraulic model of the system to extract frequency response function (FRF) for the training set preparation. Then the simulated FRFs are fed to train fully linear DenseNet (FL-DenseNet) for feature recognition. Finally, the measured FRF of the system is fed to the trained FL-DenseNet to identify a leak to a pipe in the suspected leak area. A study on a hypothetical small system shows that the proposed framework has robustness against uncertainties of friction coefficient, wave speed, and leak flow. A significant advantage is also observed over the existing method with an inaccurate model. Then the framework is applied to a larger network. Over 90% of the synthetic leaks are identified in 5 of the 149 pipes. These results presented in the paper indicate the potential of applying this framework to a water pipeline system.