Exploration of deep learning leak detection model across multiple smart water distribution systems: Detectable leak sizes with AMI meters

Abstract

Numerous deep learning (DL) models have been developed for leak detection in water distribution systems (WDSs). However, significant lack of knowledge still remains concerning their detectability and the smallest detectable leak sizes across various WDSs. To address these research gaps, this study explores the performance of a DL leak detection model across eleven smart WDSs. A convolutional neural network (CNN) is employed to identify leaks using the spatially distributed pressure response images derived from the difference between advanced metering infrastructure (AMI) measurements and predictions from a well-calibrated hydraulic model (i.e., digital twin). Ten leak magnitudes are evaluated for each WDS, and three performance metrics (recall, precision, and F1 score) are calculated to assess the detectability and the detectable leak sizes of the CNN. The analysis reveals that the DL model's detection ability is highly impacted by WDS type, whether transmission- or distribution-oriented. The former networks exhibit low accuracy in identifying leaks due to the indistinguishability of pressure response images between normal and leak conditions. On the other hand, the latter networks generally achieve higher precision and recall results and can detect smaller leaks. Moreover, the smallest detectable leak sizes are more sensitive to WDS structural parameters (pipe diameter and length) than system hydraulics (system demand). Examining pipe characteristics along the leakage flow path provides most useful information in determining the detectability of leaks.