Machine learning approach to transient-based leak detection of pressurized pipelines: Classification vs Regression

This study introduces a general framework for real-time leak detection in pipelines by coupling machine learning to transient hydraulics. In this framework, the performance of Support Vector Machines (SVM) as a superior pattern recognition algorithm is investigated in the experimental condition. First, a transient simulation model using the Method of Characteristics (MOC) is developed for the pipeline at hand. Then, the model is exploited to generate datasets containing the transient hydraulic responses at the measurement points. Afterward, the most efficient features and optimum SVM algorithm are selected through sensitivity analysis. The model is finally applied to an experimental reservoir-pipe-valve system, and the performance of the optimum SVM is evaluated taking both classification and regression approaches. In each case, the impact of applied kernels, the size of datasets, the length of the applied response signal, and the existence of various levels of uncertainty in the pipe system are studied. The results indicated that the model based on classification has higher performance and could detect leaks accurately and is stable and reliable against different types, and high levels of uncertainties.