Sensor placement in water distribution networks using centrality-guided multi-objective optimisation

This paper introduces a multi-objective optimisation approach for the challenging problem of efficient sensor placement in water distribution networks for contamination detection. An important question is, how to identify the minimal number of required sensors without losing the capacity to monitor the system as a whole. In this study, we adapted the NSGA-II multi-objective optimisation method by applying centrality mutation. The approach, with two objectives, namely the minimisation of Expected Time of Detection and maximisation of Detection Network Coverage (which computes the number of detected water contamination events), is tested on a moderate-sized benchmark problem (129 nodes). The resulting Pareto front shows that detection network coverage can improve dramatically by deploying only a few sensors (e.g. increase from one sensor to three sensors). However, after reaching a certain number of sensors (e.g. 20 sensors), the effectiveness of further increasing the number of sensors is not apparent. Further, the results confirm that 40-45 sensors (i.e. 31 - 35% of the total number of nodes) will be sufficient for fully monitoring the benchmark network, i.e. for detection of any contaminant intrusion event no matter where it appears in the network.HIGHLIGHTS center dot It is possible to significantly reduce the number of undetected events by deploying only a few more sensors.center dot Placing sensors on 31 - 35% of nodes is sufficient for full monitoring of the case study network.center dot Maximising the opportunity to detect events prioritises the selection of nodes that neither have the highest centrality nor the lowest.center dot Minimising the detection time of events prioritises nodes with centrality at/close to the extremes.

Automated summary

This paper introduces a multi-objective optimization approach for efficient sensor placement in water distribution networks. By minimizing the expected detection time and maximizing the detection network coverage, the minimal number of required sensors can be identified. The approach is tested on a benchmark problem with 129 nodes, and the results demonstrate the significant improvement in detection network coverage by deploying only a few sensors. Additionally, the study shows that 40-45 sensors are sufficient for fully monitoring the benchmark network.