Leak Detection Using Flow-Induced Vibrations in Pressurized Wall-Mounted Water Pipelines

Leak Detection Using Flow-Induced Vibrations in Pressurized Wall-Mounted Water Pipelines

Accelerometers

Clamps

leak detection

Leak detection

machine learning

Monitoring

Pipelines

sensor networks

vibration measurement

Vibration measurement

Vibrations

wall-mounted water pipelines

Wireless sensor networks (WSN) provide a powerful solution to the task of monitoring the operational conditions of buried and non-buried pipes of different lengths and materials. Due to the limited energy stored in the sensor nodes, the use of low-power vibration sensors becomes the preferred choice. However, the monitoring of vibrations for leak detection in wall-mounted pipelines, and the associated complexities are not adequately dealt with in the literature. This article offers to fill this gap by presenting a feasibility study of leak detection in wall-mounted water pipelines through vibrations measurements using low-power accelerometers. The work is divided into two steps: Firstly, a careful analysis is performed to understand the effect of various fittings such as clamps, bends, and leaks of various sizes, on the vibrations produced. Then this knowledge is used to find the best locations for placement of nodes in order to efficiently detect leaks of various sizes. This analysis revealed two important facts: (a) difficulty in detecting medium-size leaks as their vibrations and those from the no-leak condition are very indistinguishable, (b) vibrations measured away from the leak are of a small benefit to the leak detection process. Consequently, 3 different learning models are applied, all fed with information from multiple nodes, in order to reliably detect leaks and classify their size. Comparing the performances of these models shows that the Support Vector Machine (SVM)-based model gives the best results, in that for the worst case of medium-size leaks and with the use of one sensor, the worst accuracies for leak detection and leak size classification have remarkably been improved from being respectively 51% and 36% with one sensor, to being 88% and 93%, respectively, with only a moderate increase in the number of sensors to four.

188673-188687

M. -U. -R. A. Virk

M. F. Mysorewala

L. Cheded

I. M. Ali

IEEE Access

2020

journalArticle