Parallel multi-layer sensor fusion for pipe leak detection using multi-sensors and machine learning

Abstract

Effective pipe leak detection is critical for maintaining the structural integrity and efficiency of water distribution systems and preventing damage such as sinkholes. Traditional leak detection methods often rely on single sensors, overlooking the advantages of multi-sensor configurations that capture diverse leak-related phenomena. To address this limitation, the study proposes an innovative machine learning-based sensor fusion approach called Parallel Multi-Layer Sensor Fusion (PMLSF), which leverages Convolutional Neural Networks (CNN) and Few-Shot Learning (FSL) to enhance leak detection. PMLSF integrates data from multiple sensors, including hydrophone, acoustic emission, and vibration sensors. The comparative analysis demonstrates that the PMLSF with multi-sensor systems substantially outperforms the CNN-based FSL (CNN-FSL) approach with single-sensor systems, achieving a leak detection accuracy of 97.1 % and leak location classification accuracy between 95.5 % and 97.4 %. Additionally, the study investigates the use of the acoustic emission sensor combined with CNN-FSL for early detection of material failure in pipes, demonstrated by a Pencil Test that achieved 92.3 % accuracy in detecting pencil breakage on the pipe. These results indicate that combination of CNN-FSL for the acoustic emission sensor and PMLSF offers a comprehensive solution for detecting and localizing existing leaks while predicting potential failures, thus laying a robust foundation for the development of reliable and efficient water distribution monitoring systems.