Ensemble CNN Model for Effective Pipe Burst Detection in Water Distribution Systems

Various data-driven anomaly detection methods have been developed for identifying pipe burst events in water distribution systems (WDSs); however, their detection effectiveness varies based on network characteristics (e.g., size and topology) and the magnitude or location of bursts. This study proposes an ensemble convolutional neural network (CNN) model that employs several burst detection tools with different detection mechanisms. The model converts the detection results produced by six different statistical process control (SPC) methods into a single compromise indicator and derives reliable final detection decisions using a CNN. A total of thirty-six binary detection results (i.e., detected or not) for a single event were transformed into a six-by-six grayscale heatmap by considering multiple parameter combinations for each SPC method. Three different heatmap configuration layouts were considered for identifying the best layout that provides higher CNN classification accuracy. The proposed ensemble CNN pipe burst detection approach was applied to a network in Austin, TX and improved the detection probability approximately 2% higher than that of the best SPC method. Results presented in this paper indicate that the proposed ensemble model is more effective than traditional detection tools for WDS burst detection. These results suggest that the ensemble model can be effectively applied to many detection problems with primary binary results in WDSs and pipe burst events.