Optimal Sensor Placement for Pipe Blockage Detection in Urban Drainage Networks Using Graph Signal Processing

During the long-term operation of urban drainage networks (UDNs), the accumulation of sediments and debris can cause partial or complete blockages. Blockages within the pipe network may lead to flooding at manholes, and such overflows can disrupt traffic and increase pollution emissions in urban areas. Identifying the locations of partial blockages and implementing a regular pipe monitoring and cleaning plan support proactive maintenance and operation of UDNs. This approach helps increase network resilience, prevents complete pipe choking, and reduces flood inundation during high-intensity rainfall over long-term operation. With Internet of Things IoT-based smart urban water systems, sensor data can be used for the UDN anomalies detection, such as blockages, through water-level monitoring and time-series analysis. To improve the accuracy of blockage detection, determining the optimal sensor locations while considering implementation costs is a fundamental part of IoT-based UDN anomaly detection. Therefore, this work focuses on the optimal placement of sensors in UDNs for anomaly detection, supported by graph signal analysis. Water elevation variation data are modeled as signals on a graph representation of the UDN, providing a robust framework for effective anomaly detection.

In this research, at first, graph clustering is applied to divide the UDN into monitoring zones corresponding to the number of sensors. Subsequently, the optimal sensor locations within the monitoring zones are determined. For that, a genetic algorithm (GA) is used to determine the optimal location of each sensor within its corresponding cluster (monitoring zone). Therefore, the sensor network is modeled as a graph in which vertices correspond to sensor locations at manholes, and edges represent the minimum shortest paths connecting these locations. The objective function for optimal sensor placement is based on the graph Fourier analysis of that sensor network subgraph. Finally, water elevation variation data are assigned to the graph nodes as node signals. Using the graph Fourier transform (GFT), the graph Fourier coefficients of these signals are computed. The proportion of high-frequency components, defined as the energy contained in the largest 50% of Laplacian eigenvalues relative to the total signal energy, is used as a metric for anomaly detection efficiency. Nodes exhibiting high-energy components at these large eigenvalues are more suitable for blockage detection, as such high-frequency variations indicate localized disturbances. These variations have greater potential for accurately identifying pipe blockages and reducing misinterpretation in the sensor network under multiple blockage scenarios. The proposed method is implemented in a real-world UDN in an alpine region, and the performance of the sensor placement strategy is validated through sensitivity analysis under modelling multiple pipe blockage scenarios and varying numbers of sensors.

Funding: The project “RESTORE” is funded by the Austrian Science Fund (FWF) P 36737-N.