1. Introduction
Pipe burst events are widespread in water distribution systems (WDS). Detection of pipe burst events has usually been by direct visual observation or customer reports. These methods are time-consuming and are not reliable. Water companies must locate burst events quickly and accurately to reduce water loss and damage from pipe bursts, such as water waste, secondary pollution, and real estate damage. Various techniques have been devised to detect burst events. These techniques can be broadly divided into equipment-based methods and software-based methods [
1].
Equipment-based methods use hardware (e.g., listening sticks, leak noise correlators, vibro-acoustic techniques, gas injection techniques, ground penetrating radar) to detect and locate bursts. These techniques are effective and fairly accurate. However, their use is limited by financial costs and detection time, and their accuracy is usually biased by worker experience.
Software methods have been important in burst detection since they were introduced in the 1990s. Transient detection methods use pressure wave signals for burst detection. These methods include transient analysis, inverse transient analysis, and time-frequency domain analysis. To find leaks, Pudar and Liggett [
2] developed a method of inverse analysis based on water pressure, and they suggested it be used to supplement traditional leak detection methods. VÃtkovský et al. [
3] used a Levenberge–Marquardt algorithm to estimate the magnitude and location of a single leak in a pipeline. This method is also suitable for situations where there are two or more leaks. Misiunas [
4] continuously monitored a laboratory pipeline for pressure transients to detect and locate burst events; the method can be used in an operational pipeline system. Kapelan et al. [
5] created an inverse transient model to detect leaks. They calibrated pipe roughness by incorporating parameterized prior information and found that the use of prior information improved the results given by the model. Lee [
6] devised a frequency domain method to locate a burst event. This method determines the exact location and magnitude of one or more leaks using frequency response curve. Methods using fluid transients can analyze the hydraulic behavior of the WDSs for burst detection. However, they require a lot of data, which are very expensive to collect, and these methods lack enough field trials for adequate validation.
Other software-based methods, such as statistical and artificial intelligence methods, have been widely used as hydraulic models, and data analysis techniques have been developed. These methods are promising and have recently become more popular as they can extract burst-related information from large amounts of data to identify bursts. Mounce and Machell [
7] found a relationship between pipe burst events and fluctuations in pressure and flow by using artificial neural networks (ANNs). The results show that both static and time-delayed ANNs can identify bursts, and the effectiveness of the ANN is determined by the quantity and quality of data. Bicik [
8] used Dempster–Shafer (D-S) evidence theory, which combines the outputs of different models, to provide increased confidence in the results of individual models and rapidly localize bursts. Cheng et al. [
9] also used D-S theory combined with pressure and flow risk functions to identify bursts on a water main trunk. A burst detection system was built using this method and correctly identified all bursts from one-year historical records. Ye and Fenner [
10] developed an adaptive Kalman filter for flow and pressure observations at a district meter area to detect burst events. The burst events that were detected matched observed events, and the authors used the pressure-based detection method to confirm the flow-based method. Jung and Lancey [
11] developed a nonlinear Kalman filter which was robust in a variety of operations to detect burst events. Palau et al. [
12] identified anomalous behaviors in water use, burst events, and illegal connections using principal component analysis. This method was very sensitive in detecting bursts and other abnormal events. Romano et al. [
13] developed an event recognition system (ERS) for automated burst detection using statistical and artificial intelligence techniques. Several cases showed that the ERS can detect burst events quickly and reliably. Jung et al. [
14] compared six statistical process control methods in terms of burst detection capability and found that an exponentially weighted moving average method was the most reliable. Besides, the method had the shortest detection time. However, these methods that use data mining have a high false alarm rate because they could be affected by uncertainty in the monitored data, which makes their widespread use impractical.
Burst detection methods and algorithms have been extensively and thoroughly studied, and progress has been made in recent years. In this paper, our focus is not on a new burst identification method. Instead, we introduce a burst detection process that explores the impact of a simulated burst event on the WDS and the reduction of the false alarm rate of the current monitoring system. This paper details two aspects of our work. First, the method of determining the background noise, using observed data from supervisory control and data acquisition (SCADA), is described. A burst event can be accurately detected only when the impact of the burst can be differentiated from the background noise. Based on this assumption, the threshold for burst detection and the minimum pipe diameter for which a burst event is detectable can be determined for any given WDS. Second, the method of determining the minimum number of sensors is described. This determination will improve the acceptability rate of correctly detected burst events.
The rest of this paper is organized as follows. The main requirements for building a comprehensive burst detection system are described in
Section 2. In
Section 3, we estimate the level of background noise in the data that have been collected and determine the thresholds of pressure that identify abnormal fluctuation from the cumulative probabilities of pressure fluctuations and head loss fluctuations. In
Section 4, we define the minimum pipe diameter for which a burst event is detectable; that is, when burst event data are discernible from background noise. Only burst events in sufficiently large diameter pipes would cause a noticeable pressure drop. Finally, we identify the number of sensors required to determine a burst location by analyzing burst records collected over one year and suggest a simple burst detection method by comparing pressure drop and threshold values in
Section 5. The main conclusions are given in
Section 6.
6. Conclusions
This paper introduces some basic principles and requirements for the design of a burst event detection system. The thresholds for burst detection and the minimum pipe diameter for which a burst event is detectable were determined. A simple burst detection method was suggested as a result of the statistical analysis of monitoring data and evaluation of the impact of burst events on the WDS. The method was validated using historical records, showing that this method can quickly identify the burst event and give a reasonably accurate burst location. The minimum number of sensors necessary for accurate burst detection was determined for the Guangzhou WDS. The main conclusions and recommendations are as follows:
- (1)
Monitoring data contain background noise from WDS consumption use and from monitoring equipment. The effect of a detectable burst event should be greater than the level of background noise for a burst event to be correctly identified. The background noise of a pipe system can be determined by statistical analysis of the monitored data using the 3-sigma rule.
- (2)
The minimum pipe diameter for which a burst event is detectable can be determined by model simulation. Monitoring burst events in pipes of greater than the minimum diameter will improve the burst detection and reduce monitoring costs.
- (3)
In a detection zone (i.e., the part of the network close to and containing the burst event), a minimum number of sensors is necessary for an acceptable rate of correctly detected burst events. We found that data from two or more sensors close to the burst event location are required to reduce errors due to undetected events and false alarms.
The methods described in this paper for determination of burst thresholds, minimum pipe diameters, burst detection and localization, and minimum sensor numbers were validated using the data for one year from Guangzhou, but they can be used in other applications. The discussion of burst detection algorithms is not within the scope of this paper, so we did not compare the algorithm we used with other methods. However, we think that incorporating evidence theory and Kalman filters would make the prediction of burst events more accurate and thus improve the effectiveness of the SCADA infrastructure.