Blockages have been reported to account for a significant proportion of reported failures in sewer networks. The malfunctioning of the sewer network from blockages and the subsequent disruption to other public services and flooding may constitute a risk to the environment and human health. Due to the complex nature of underground sewer networks, a reactive approach to blockage maintenance is typically employed. However, although proactive maintenance strategies have been developed, both approaches could be expensive and highlight the need to address the problem with analytics-based methods.
The overall aim of this thesis was to contribute with new knowledge, approaches and methods that can support improved proactive maintenance planning of blockages in sewer networks. Methods and approaches were investigated in relation to asset management planning levels. Some of the methods and approaches investigated include Poisson and partial least squares regression, network k-function, geographically weighted regression (GWPR) and random forest ensembles. A conceptual framework was also proposed to assess the state of data-driven integrated asset management (IAM), based on data quality assessments, interoperability evaluations between IAM tools, and data collection and informational benefits analysis.
Results showed that the average blockage rate in medium sized networks was 2-3 times the rate in other sewer networks in Sweden. Furthermore, sewer maintenance strategies were suspected to be ineffective, and increased proactive strategies may improve maintenance efficiency. Results also indicated that the clustering of recurrent blockages may be linked to an increased need for flushing-related maintenance in sewer pipe networks. The degree of influence between investigated factors and increased blockage propensity indicated that these relationships were not global (not the same in all locations) within and between the sewer networks. The networks with relatively more substantial spatial clusters of blockages, higher data quality and availability were observed to have a higher mean prediction accuracy.