Data-driven risk assessment on urban pipeline network based on a cluster model

The existing infrastructure system has collected multi-source attribute data for the Urban Pipeline Network (UPN) from sensors, while the failure records are much rare in a relatively short period. Supervised machine learning model confronts challenges for UPN risk assessment because of dramatic class imbalance between the major normal samples and the minor abnormal failures. The previous works apply the clustering as a pre-processing tool, by classifying samples to clusters for simplicity of the following manual risk level rating or failure prediction. In contrast, this work proposes a method based on clustering and statistical test for evaluating risk state via small volume of historical failure records. This framework achieves totally data-driven risk assessment in an unsupervised way, avoiding expensive manual labeling for pipelines’ risk states and probabilities estimation of elementary events in the probabilistic Bayesian approach. A case study is finally conducted on an urban gas pipeline network from a city, consisting of more than 13,000 pipelines (over 1700 km). The results demonstrate that the cluster of pipelines with highest risk have seven times higher accident rate than the lowest one. This method is hopeful to support decision-making regarding to routine inspection and restoration plan for the pipeline networks.