Machine-Learning-Based Leakage-Event Identification for Smart Water Supply Systems

In this article, we are interested in leak identification (LI) for water supply pipelines using transient-wave (pressure) measurement data. This is challenging since water pipeline system conditions are usually uncertain in practice. For instance, the pipeline diameter, the friction factor, and the...

Full description

Saved in:
Bibliographic Details
Published in:IEEE internet of things journal 2020-03, Vol.7 (3), p.2277-2292
Main Authors: Zhou, Bingpeng, Lau, Vincent, Wang, Xun
Format: Article
Language:English
Subjects:
Algorithms
Artificial neural networks
Diameters
Failure rates
Feature extraction
Friction factor
Fusion
Kernel
leak identification
Machine learning
machine learning (ML)
neural network
Neural networks
Noise levels
Optimization
Pipelines
Signal to noise ratio
System dynamics
Systems design
Time-domain analysis
Training
Transient analysis
transient-wave model
water pipeline system
Water pipelines
Water supply
Water supply systems
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In this article, we are interested in leak identification (LI) for water supply pipelines using transient-wave (pressure) measurement data. This is challenging since water pipeline system conditions are usually uncertain in practice. For instance, the pipeline diameter, the friction factor, and the pipeline shape will vary. The conventional signal propagation model-based LI methods rely on a deterministic system model with perfectly known and fixed-value parameters, which limits their application in general cases. To address this challenge, we design a novel deep neural network (DNN)-based machine learning approach to solve the LI problem. First, we propose a novel fusion-enhanced stochastic optimization algorithm for the DNN training, which can greatly improve the DNN training performance and hence the LI accuracy, without increasing the computational cost. Second, we design a novel convolutional-based pooling network to extract the stable texture feature of transient-wave samples, thus achieving a reliable LI solution against the pipeline system dynamics. It is shown in experiments that, thanks to the above system design, the proposed DNN-based LI method can achieve a failure rate lower than 6\times 10^{-4} when the signal-to-noise ratio is 0 dB, which outperforms the conventional LI methods.
ISSN:2327-4662
2327-4662
DOI:10.1109/JIOT.2019.2958920