Drinking water distribution networks are exposed to malicious or accidental contamination. One level of responses consists of installing a sensor network to monitor the system in real time. Once a contamination has been detected, it is also important to take appropriate counter-measures. The SMaRT-OnlineWDN project relies on modelling to predict both hydraulics and water quality. An online model makes it possible to identify the contaminant source location and perform a simulation of the contaminated area. The sensor system is intended for detection by an early warning system. The sensor placement is described in the deliverable reports D3.1 and D3.2. The objective of this deliverable is to demonstrate the early-warning sensor placement optimisation method for the three end-users of the SMaRT-OnlineWDN project. For the CUS network (Strasbourg Eurométropole), a network graph is obtained after simplification that includes 16,000 links and 14,000 junction nodes. It is the basis of water quality multi-probe sensor for early-warning detection system. 5,000 uniformly distributed contamination events were generated by random sampling. Pareto optimal fronts were proposed to the water utility for average time to detection versus the sensor cost, and for detection likelihood versus the sensor cost. This makes possible the placement of additional 1 to 200 water quality multi-probe sensors. It was found that for 94 additional sensors there is no significant improvement for the detection likelihood, 95% of the generated contamination events were detected in less than 5 hours. BWB supplied the Hochstadt Ost part of the Berlin WDN to the Partners of the SMaRT- OnlineWDN project. Several nested WDN graphs were worked out from the more detailed one to the simplest with important crossroad nodes and head resource nodes. There was no information about number of connections, pipe accessibility, sensitive consumers and capacity to connect to the SCADA system; also the optimisations were only about minimization of average time to detection and likelihood of detection. The designs for 200 sensors were studied for the detailed model of Hochstadt Ost with 59,158 nodes and 63,828 links and for the supergraph with 9,557 nodes and 14,228 links. It takes several days to generate 50,000 contamination events for the detailed network and only few hours for the supergraph and 10,000 events. So the intention was to validate the calculation on the simplest network with contamination events and sensor placements located at important crossroad nodes (or path nodes). It was found that calculation on the simplest network, which makes focus on path nodes, gives similar results both in term of scores and sensor spatial distribution. Residence times in general are shorter on the supergraph because the forest where are the lowest velocities has been removed; also the average time to detection was slightly underestimated for consideration of contamination events in the forest. The SEDIF network was decomposed in 11 hydraulic models. Around 200 sensors are deployed on the whole SEDIF network: about 100 probes based on expert knowledge and the remaining 100 using the proposed methodology. In this work, VEDIF has used the greedy algorithm designed by Irstea in order to minimize the expected fraction of the exposed population to contaminations, but also the average time to detection, and to maximise the detection likelihood. Moreover, expert knowledge was used to limit sensor location site of interest for both goals, following the technical recommendations of the sensor supplier and speeding up the optimization by selecting only important crossroad nodes. With the deployment of about 200 sensors, the normalized criterion of unexposed population is from 75% to 90% for each of the eleven models. Finally, a GUI has been developed by Veolia Eau d'Ile de France to display the solutions of the sensor placement.