Sensor-fusion of hydraulic data for burst detection and location in a treated water distribution system

Abstract

This paper presents research into analysis and data fusion for sensors measuring hydraulic parameters (flow and pressure) of the pipeline water flow in treated water distribution systems. An artificial neural network (ANN) based system is used on time series data produced by sensors to construct an empirical model for the prediction and classification of leaks. A rules based system performs a fusion on the ANNs’ outputs to produce an overall state classification for a set of zones. Results are presented using data from an experimental site in a distribution system of a UK water company in which bursts were simulated by hydrant flushing. The ANN system successfully detected events and a study of the pressure gradient across the zone provided a more precise location within the zone.

Introduction

Treated water is distributed in the industrial world by means of pipeline networks. Leakage from a water distribution pipeline network can be defined as that water which, having been obtained from a source, treated and put into supply, leaks or escapes other than by a deliberate action. In the UK, between 20% and 30% of transported water was lost through leakage during the 1990s [1]. This figure is even higher for older pipes [2]. The loss of such large volumes of water is environmentally and economically damaging and this situation is likely to be a major issue during the 21st century as more extreme environmental fluctuations are experienced due to the climatic consequences of global warming. The UN’s Intergovernmental Panel on Climate Change has highlighted that water resources will be seriously affected in many areas [3].

The most common technique for identifying leakage is to conduct a water audit in which a detailed account of water flow into and out of the distribution system, or parts of it, is recorded. At the level of the whole system, this consists of a total water supply balance, i.e. the summation of all water consumed (metered and unmetered) and not consumed (leakage, theft, exports etc.) compared with the total distribution input. District flow metering extends this to monitoring individual zones [4], [5]. The distribution system (in urban areas or strategically important trunk mains) is subdivided into discrete zones or district meter areas (DMA), by the permanent closure of valves. A DMA will generally comprise 500–3000 properties. Flow (and sometimes pressure) sensors are placed on the DMA boundaries and collected data is subsequently analysed for leakage trends. The most popular operational use of flow data is the analysis of measured minimum night flows. Night flows (usually measured between midnight and 5:00 am) are used because water use is at a minimum and it is easier to identify and subtract legitimate flows. Any remaining unusual jumps in volumes will signify leakage in the absence of any other factors. If it is established that leakage has increased sufficiently to warrant further investigation, then a manual leakage detection is carried out on the entire DMA using methods such as step testing, sounding and leak noise correlation [6]. This detection takes approximately one day per 350 properties per leakage team.

In general, flow sensors log the flow every 15 minutes. Data from such flow metering has, until recently, always been gathered by manual download in the field with a downloading period of one or more weeks. In recent years, more sophisticated online data acquisition systems have become available to allow critical flows (for example reservoir outlets) to be monitored. This approach allows a minimisation of the time for leak detection by using online data from telemetry systems along the lines of a supervisory control and data acquisition (SCADA) system. In a leakage context, a SCADA system requires the use of a computer system, sensor data acquisition and an analysis algorithm to evaluate the sensor data and produce a leakage determination. A class of SCADA variant widely used in the oil industry is that of mass flow measurement methods which rely on various physical phenomena to obtain accurate measurements of the total mass flow between two points on a pipeline. However, the cost of sensors and data communication equipment necessary for implementing a full SCADA system on a large water pipeline network is currently prohibitive––where there is a lack of perceived environmental impact from a large scale leak (in contrast to fuel transportation for example).

A number of techniques are under investigation for automated leakage detection from real time telemetry systems. Recent work has seen attempts to link hydraulic models with online data from telemetry systems [7] and hence perform leak detection by solving an inverse steady state problem [8]. State estimation, widely used in the electrical power industry, has been investigated as a solution [9], [10], [11]. Research has been carried out on a method of using state estimation for leakage detection in a DMA structured system without measurements inside the DMA [12], [13]. The inverse transient technique has been proposed for leak location when a number of pressures within a DMA are present. This method is based upon the minimisation of the sum of the quadratic errors between the measured head values and those calculated by computer simulation in the hydraulic model [14]. All of these techniques rely on the accuracy of real-time hydraulic simulations, which are limited by modelling errors, a very limited number of sensors in most distribution systems, and accuracy and reliability of sensor data.

An alternative approach is to adopt a pattern recognition methodology in which a sensor’s output (discrete time series data) is treated as a signal and is analysed by techniques from the statistical and AI fields. AI-based systems can automate mundane tasks involved in the process, as well as presenting more ‘intelligent’ information to the operator. The concept is to analyse one or more sensor’s output with an ANN in order to identify meaningful patterns, and present a fusion of these patterns to the operator.

This paper outlines the development of a neural network knowledge-based system for automatically and continuously monitoring the signature (comprising the time series for one or more sensors) of a water pipeline distribution system for normal and abnormal behaviour. The output of this system will allow alarms to be raised for a DMA when failure or leaks are detected. The detection system adopts an empirical model based upon pattern recognition techniques applied to time series data. A rule-based module performs a fusion on the ANNs’ outputs to produce an overall state classification for a set of zones. The combination of these techniques comprises an innovative application in the field of leakage detection and location from water hydraulic time series data.

Section 2 provides theoretical background and a methodology for DMA level sensor monitoring and fusion. Section 3 describes an experimental trial conducted within a DMA in which bursts were simulated by flushing hydrants. Section 4 provides results of applying the DMA system to the experimental data. Section 5 presents a supplementary application of using fused data from the pressure sensors to provide a more precise localisation. Finally, Section 6 presents the conclusions.