Assistant Professor Technische Universität Berlin, United States
Abstract Submission: Leakages in water distribution networks lead to substantial water and economic losses, and infrastructural damages. Over the last decades, both data-driven and model-based algorithms have been proposed in the literature to detect and localise leakages using sensor-based data. Change point detection algorithms are commonly employed to detect changes in the in-control target center of a monitored process. Under the assumption of independently and normally distributed data, the cumulative sum (CUSUM) control chart is a popular change point detection algorithm. CUSUM detects a specified magnitude of mean change from a known in-control mean. In the context of leakage detection, however, the standard CUSUM method might not be suitable in two cases: (i) the data analysed might not be normally distributed, if not all demands are included into the model forecast, and (ii) the magnitude of change might differ depending on the leak size, which makes the fast detection of incipient leaks challenging. In this work, advanced change point methods are comparatively tested on a variety of leakages from the L-Town benchmark data set. The methods include weighted CUSUM, transformation of the data, and combinations with the exponentially weighted moving average method. The methods are assessed with respect to their time to detection, false alarm rate, and implementation complexity. Our preliminary numerical results indicate that both the false positive rate as well as the time to detection heavily depend on the method type, and thus the method choice has to be made depending on which leaks are prioritized.
