STATISTICAL APPROACH TO DETECTION OF ANOMALIES IN WATER DISTRIBUTION NETWORKS

Abstract

This paper is devoted to solving the problem of developing an automated system for detecting anomalies in water distribution networks. The main causes of such anomalies are background leaks and pipe breaks. To address this problem, a statistical approach is proposed, which consists in testing the null hypothesis that the readings of pressure and/or water flow sensors received in real time correspond to the standard conditions of the network. The paper proposes a three-stage anomaly detection scheme, which includes: statistical profiling of network sensors; system calibration to achieve the desired ratio between the risks of false alarms and the omission of existing anomalies; determination of rules for drawing conclusions about the presence of anomalies. A methodology for statistical profiling and calibration of the system based on simulation modeling in the EPANET software environment using the WNTR software interface in Python was developed. In the process of such modeling, the distribution of pressure readings in the network sensors is obtained based on water demand fluctuations. As an example, the model of the L-Town water supply network was studied, which was developed for the BattLeDIM leak detection and isolation competition. The sensitivity of the anomaly detection results to the range of sensor values that are considered normal, as well as the number of sensors involved in the anomaly detection procedure, is investigated. The dependence of the number of sensors with anomalous readings on the size of the leak is analyzed. It is established that there is no combination of parameters of the proposed anomaly detection system that provides optimal results for all possible leakage sizes simultaneously, as a result of which it is proposed to calibrate the system using a man-machine procedure.