A new alarm generation concept for water distribution networks based on machine learning algorithms

Water Distribution Networks are complex systems containing a large amount of sensors placed in the network. An important task of water quality sensors is to give information to the operators if a contamination has occurred in the network. Generally, the overall analysis of the sensor data is performed manually by the operators since data-driven alarm generation systems for protecting a network in real time are not established at water utilities. This has several reasons: At first, the parameterization of these modules is often very complex and time consuming. Secondly, in most cases these systems generate too many false positive alarms due to special operational actions like sensor calibration or flushing of pipes. In this paper an approach is presented which addresses both problems. First, a self-configuring alarm generation module is proposed which only needs a few parameters to be set. Next, using the module, it is shown that the amount of false positive alarms can be reduced, if known events are used for training the module. This approach is proved in experiments performed on a laboratory plant.