Methods for the localization of a leak in open water channels

Nadia Bedjaoui, Erik Weyer, Georges BastinNetworks Heterog. Media 2009 · DOI · OpenReview

Abstract

Leak localization in open water irrigation channels is a critical task for water resource management. This paper explores model-based approaches, specifically a bank of open-loop models and a bank of nonlinear observers, to identify the position of lateral outflows. While these methods achieve high accuracy in ideal simulations, their real-world application is hampered by the mathematical ill-conditioning of the problem and high sensitivity to parameters like leak start time and channel friction.

Introduction

In global agriculture, water is an increasingly scarce resource. Much of this water is transported through vast networks of open channels, which are prone to leaks caused by aging infrastructure, failing gates, or unauthorized "offtakes." While detecting that a leak has occurred is a well-studied problem, localizing that leak within a reach that may be several kilometers long remains a major challenge. This research aims to provide a mathematical framework to estimate a leak's position using only existing sensors at the upstream and downstream ends of a channel reach (water levels and gate positions).

Physical Modeling and the Saint-Venant Equations

The authors model water flow using the Saint-Venant equations, a pair of coupled, non-linear hyperbolic partial differential equations (PDEs) representing mass and momentum balance. A leak is treated as a lateral outflow $w$ at a specific point along the channel.

Schematic of an open water channel reach with a leak $w$ at an unknown location.

The physical setup involves measuring water levels ($Y$) and flow rates ($Q$) or gate positions ($p$) at the boundaries ($x=0$ and $x=L$). To analyze the stability and behavior of these systems, the authors utilize Riemann coordinates, which transform the PDEs into a form that describes how information waves propagate upstream and downstream.

Proposed Localization Schemes

Two main schemes for localization are proposed:

  1. Bank of Models (Method 1): The channel is divided into $N$ sections. For each section, a model is run assuming the leak is located there. The "winner" is the model that minimizes a quadratic cost function—essentially the difference between the model's predicted water levels and the actual measured levels at the boundaries.
  2. Bank of Observers (Method 2): This is similar to Method 1, but instead of pure open-loop models, it uses nonlinear observers. These observers use feedback from boundary measurements to "reconstruct" the internal state of the channel.
Block diagram of the Bank of Models approach for leak localization.

The Inherent Difficulty of Localization

A key insight of this research is that leak localization is an "inherently difficult" problem. In subcritical flow (typical for irrigation), waves travel upstream and downstream at nearly the same speed. Because the effect of a leak on boundary water levels is very similar regardless of its origin point, the problem is mathematically ill-conditioned.

Furthermore, the system is extremely sensitive to the leak's start time; an error of just two minutes can shift the predicted location by hundreds of meters. This is because the algorithm relies on the subtle timing of when the "drawdown" wave hits the upstream versus downstream sensors. Additionally, variations in the Manning coefficient (representing channel friction) significantly degrade localization accuracy.

Results and Real-World Performance

The methods were evaluated using both simulated data and experimental data from Pool 4 of the Coleambally Channel 6 in Australia.

  • Simulation Success: Under ideal conditions with no noise or model uncertainty, both the model-bank and observer-bank methods correctly identified the leak location (Section 35 out of 50).
  • Real-World Struggle: When tested on the Coleambally Channel data, the methods were less successful. The actual leak was at Section 50, but the algorithms placed it at Section 33. This discrepancy was attributed to unavoidable uncertainties in real-world friction parameters and sensor noise.

Conclusion and Future Work

While model-based banks provide a robust framework for state estimation, the localization task remains highly sensitive to model mismatch and timing errors. The authors suggest that data observed long after a leak starts carries very little information; the most useful data occurs in the narrow window immediately following the leak's onset.

Future work may involve expanding model banks to include variations in both position and start time, or utilizing a simplified time-difference approach based on the lag between when the leak's effect is seen at the upstream end versus the downstream end.

Summary updated

Abstract

In this paper, we present two methods for determining the positionof a leak in an open water channel. The available measurementsare the water level and the gate position at the upstream anddownstream end of a channel reach. We assume that the size of theleak and the time it started are already estimated by aleak-detection method. Both of the proposed methods make use of anonlinear Saint-Venant equation model of the channel where theleak is modelled as a lateral outflow. The first method makes useof a bank of $N$ models corresponding to $N$ possible positions ofthe leak along the channel. The estimated position of the leak isdetermined by the model which minimizes a quadratic cost function.The second method is based on the same principle except that ituses observers instead of pure models. The methods are tested onboth real and simulated data from the Coleambally Channel 6 inAustralia. It is further shown that the determination of theposition of a leak is an inherently difficult problem.
BibTeX
@article{bedjaoui2009methods,
  author = {Nadia Bedjaoui and Erik Weyer and Georges Bastin},
  title = {{Methods for the localization of a leak in open water channels}},
  journal = {Networks Heterog. Media},
  year = {2009},
  doi = {10.3934/nhm.2009.4.189},
}
Figure