posted on 2017-09-01, 15:07authored byNicolas Cheifetz, Selma Kraiem, Pierre Mandel, Cédric Féliers, Véronique Heim
Monitoring water quality in a drinking Water Distribution Network (WDN) is typically based on several sensors that are deployed on different locations of the WDN. Each sensor measures one or multiple signals, and the real challenge resides in detecting significant contaminations by analyzing these water quality signals. In practice, some detection methods may fail due to a specific design to certain forms of contaminants as well as a generic formulation with model-free algorithms. Both cases might trigger false alarms or produce no detection when a contamination occurs. Moreover the problem is hardly addressed when the underlying hydraulic regime changes over time causing fluctuations in the detection statistics. Such variation can dramatically decrease the detection performance or result to misleading analysis. Events like source shifting or tank filling are normal situations and can occur frequently depending on the operating conditions. This paper aims to deal with the variability of the contamination events under various operational conditions in water networks. The procedure is fully data-driven and leads to the extraction of meaningful temporal patterns using various data analysis techniques. This methodology can be used as a preprocessing stage to improve the performance of any algorithm to detect contaminations. The proposed approach is illustrated on a large real-world network in France and a qualitative interpretation is given to highlight a better understanding of the hydraulic regimes over time.