Pipeline leakage alarm via bootstrap-based hypothesis testing

Abstract

To date, the research of transient-based pipeline leakage detection has been focused on the estimation of leak location and size, assuming at least one leak exists in the pipe system. However, due to the presence of aleatoric and epistemic uncertainties, fictitious leaks always appear in the estimation results even if the pipe is intact, such that the existing methods cannot quantitatively decide whether the estimated leaks really exist. This paper proposes a hypothesis testing framework so that the probability of the existence of leakage can be evaluated. To address the aleatoric uncertainties, the bootstrap method is employed to extract the information of the probabilistic distribution of leakage feature (i.e., the test statistics) from a very limited number of real-time transient measurements. The complex epistemic uncertainties are estimated from a baseline signal obtained from the same pipe system before leaks emerge. The distribution of the test statistics under the no-leak condition is then constructed so that the existence of leakage under the current condition can be determined with a given false alarm probability. Experimental results show that the proposed method accurately pinpoints the appearance of small leak and distinguishes it from the no-leak case, where the leak location and size are estimated simultaneously.