Merging Fluid Transient Waves and Artificial Neural Networks for Burst Detection and Identification in Pipelines

Water transmission and distribution pipelines are critical infrastructure for modern cities. Owing to the sheer size of these pipelines and the fact that most of them are buried underground, the health monitoring and maintenance of this infrastructure is challenging. In addition, some water transmission pipelines cover long distances through remote areas that are not easily inspected on a regular basis. To monitor these systems, different noninvasive techniques have been developed to identify events that may put the functioning of a pipeline at risk. These techniques include visual observations (Thomson and Wang 2009), acoustic monitoring (Shimanskiy et al. 2003; Muggleton et al. 2006; Juliano et al. 2013), thermographic infrared inspection methods (Fahmy and Moselhi 2010), ground-penetrating radar methods (Hunaidi and Giamou 1998), and remote sensing (Agapiou et al. 2016; Martins et al. 2019). However, these techniques are time-consuming, do not provide permanent monitoring of the pipelines, or have a short inspection range along a given pipeline. Therefore, there is a need for a permanent monitoring method capable of identifying anomalous events and, potentially, their associated characteristics in near real time.

Hydraulic-based techniques have been developed based on the understanding of the movement of a fluid along a pipeline and are typically related to the measurement and analysis of two hydraulic variables: flow (or velocity) and pressure in the pipeline. These techniques can include volume-based methods coupled with alert systems (Mounce et al. 2003), pressure and flow analysis using statistical detection (Puust et al. 2008; Li et al. 2014; Lee et al. 2016; Wu et al. 2018b), system state estimation analysis (Andersen and Powell 2000), and transient-based methods (Misiunas et al. 2005). However, while each of these approaches has been moderately successful, they also have associated disadvantages. Some of these methods are only applicable for the detection of leaks, are not able to pinpoint the location of the abnormal event, or require an accurate numerical model of the pipeline, or the resolution of the data limits their ability to detect and locate anomalies in a timely manner.

Among the hydraulic-based techniques, transient-based methods have received attention because they provide for the inspection of a long section of pipe using only pressure measurements (Wang et al. 2002; Brunone et al. 2013; Gong et al. 2014). These methods are based on the interpretation of the effect that the occurrence of an anomalous event has on a measured transient pressure trace. The interpretation of the transient pressure trace can be conducted by visual analysis or by processing the transient pressure trace to identify reflections and detect the occurrence of abnormal events (Misiunas et al. 2007; Srirangarajan et al. 2011; Rashid et al. 2014; Rathnayaka et al. 2016). Nonetheless, a visual analysis cannot be conducted in real time and the processing techniques available require extensive numerical processing that makes a real-time monitoring system difficult.

This paper presents a new transient-based technique that employs artificial neural networks (ANNs) to interpret rapid changes in a transient head trace to identify, locate, and characterize bursts in water pipelines. Unlike current transient-based techniques, the proposed approach is data-driven and relies on no detailed information regarding the analyzed pipeline for the interpretation of the transient pressure trace in near real time. In addition, in contrast to many existing approaches, the proposed technique does not use an ANN as a metamodel of the transient phenomenon but instead uses an ANN as a tool to interpret the measured transient head traces to identify bursts. This technique analyzes short time segments of a potentially continuous transient head trace in previously trained ANNs through two different processes. A first ANN analysis detects whether a particular transient pressure head time window contains abnormal changes that could have been produced by the occurrence of a burst. This initial analysis also determines whether the information contained in the time window is enough to locate and characterize the potential burst. The second ANN analyzes the detected abnormal transient head time windows from the first ANN to predict the location and the characteristics of the occurring burst. Finally, the technique uses a transient flow numerical model to verify that the predictions of location and size of the burst are coherent and are similar to the measured transient head trace.

The approach proposed in this paper works with transient pressure head data to carry out the training of the ANNs. These pressure head data may be obtained from recorded data, from an available numerical model, or from a combination of these two sources. Once the ANNs have been trained, the use of this technique is data-driven in that it only requires measured transient head data to detect and identify a burst (by determining its location and size). Additionally, it can be applied in near real time considering that the computationally expensive processes are concentrated in the training of the ANNs and not in the processing of new transient pressure head traces, which can be tested almost immediately without retraining the ANNs.

One example hydraulic system is considered in this paper to demonstrate the functioning and performance of the proposed technique. A single pipeline with the potential presence of a burst at any point along its length is analyzed. A numerical system with this configuration is used to present the operation of the technique and to train the ANNs. In addition, a laboratory validation is included to demonstrate the promising potential of the technique.

This paper includes a background in hydraulic and transient-based methods for identifying and locating bursts in pipelines. The example hydraulic system is described next, and the methodology to apply the proposed technique is then presented by explaining the two stages involved in the process. The results for the training, testing, and application of the proposed technique are presented for the numerical system. Finally, the experimental validation is described. The proposed technique is shown to be successful in analyzing a potentially continuous transient head trace to identify, locate, and characterize the occurrence of a burst in a pipeline.

The detection and location of bursts in water distribution systems is a complex task for water utilities. Various authors have proposed techniques for locating bursts in water pipelines using hydraulic-based methods. A first group of techniques includes the use of flow and pressure measurements with statistical methods to detect the occurrence of abnormal behavior in a system (Wu and Liu 2017). Ye and Fenner (2011) proposed the use of an adaptive Kalman filtering process to predict flow (or pressure) in a system at a district meter area (DMA) level. This statistical characterization of a dynamic system is able to model the normal hydraulic parameters that are compared to measured data to detect the occurrence of bursts. Similarly, Ahn and Jung (2019) proposed a hybrid statistical model that combines statistical process control with two univariate methods to enhance the performance of the burst detection technique in terms of the detection probability, the rate of false alarms, and the average detection time. Other authors have proposed the use of statistical risk functions (Cheng et al. 2018) and principal component analysis (Palau et al. 2012) to detect bursts in transmission mains and in DMAs. A different approach was described in Wu et al. (2018a), where a clustering-based method was used to identify bursts using only 1 day of historic measured data without using statistical methods to model the expected normal conditions of the system. Although these techniques are successful in detecting the occurrence of bursts, they are unable to accurately pinpoint the location of bursts and their range of effectiveness is often limited to a DMA level.

A second group of techniques uses supervised learning techniques coupled with hydraulic measurements to send an alert regarding the occurrence of a burst. Mounce and Machell (2006) used two ANN architectures (static ANN and time delay ANN) to detect the occurrence of bursts using flow data at a DMA level. The use of ANNs showed potential for identifying changes in the flow that corresponded to unusual fluctuations of this hydraulic variable. Mounce et al. (2010) proposed the use of support vector regression models to predict time series data in a moving time window and compare these series with measured data for the detection of anomalies. The use of this supervised learning technique was applied to historical data proving that 78% of the alerts corresponded to actual abnormal events in the system. Similarly, Romano et al. (2014) proposed a fully automated data-driven methodology at a DMA level using all the pressure and flow measurements available. This approach combined the use of an ANN for the short-term forecasting of hydraulic values and statistical processes to determine whether an abnormal event had occurred. The results obtained showed the potential of data-driven technologies for near-real-time incident management in water distribution systems.

Other authors have proposed the use of artificial immune systems not only for detection but also for an approximate localization of a burst. Tao et al. (2014) proposed the use of pressure data every 10 min and was able to detect and localize a burst in 48% of the cases in a real network. More recently, Huang et al. (2018) proposed the use of a random forest classifier to detect bursts in real time by analyzing successive time windows (every 15 min) of flow data at a DMA level. These contributions demonstrate the use of a supervised learning technique for the detection of bursts in pipelines; however, all the applications to date have been focused at a DMA level and using SCADA flow (or pressure) data, which are often available in intervals between 1 and 15 min.

A different group of techniques involves the use of fluid transients to detect the occurrence of bursts in water systems. These techniques are able to detect and locate anomalies in pipelines such as leaks (Wang et al. 2002; Lee et al. 2007a, b; Capponi et al. 2017), blockages (Rubio Scola et al. 2017), or wall deterioration (Gong et al. 2013) and have obtained accurate results. Liggett and Chen (1994) proposed an inverse transient algorithm to detect leaks using the method of characteristics (MOC) and the Levenberg-Marquardt method to find friction and leak parameters in a system. These authors proposed, for the first time, the use of transient pressure measurements to locate pipe ruptures using an event algorithm with the capacity to approximately pinpoint burst locations to locate additional nodes in the numerical model of systems and conduct inverse transient analysis. Misiunas et al. (2005) used the principles of time domain reflectometry for detecting and locating abrupt pipeline breaks using a single pressure measurement point in a pipeline. A system of continuous monitoring of the pressure at a high sampling frequency, coupled with a cumulative sum test and prefiltering techniques, was used to detect changes in the data. In addition, an offline analysis of a short time window was conducted to interpret the pressure changes to determine the burst location. The results of this research have demonstrated that transient pressure signals can be used to detect bursts in water systems.

More recently, Srirangarajan et al. (2011) described the use of a wavelet-based multiscale analysis combined with a focusing algorithm and a graph-based search algorithm to detect and locate a burst event. This technique showed the potential for application of transient pressure techniques if the characteristics of the system are known and presented the first application in a network layout. Other authors have proposed the use of embedded and distributed event processing algorithms to detect transient events in a system and stroke-based transient recognition algorithms to classify transient events as bursts (Hoskins and Stoianov 2014). Although several authors have proposed techniques that use transient pressure signals to detect and localize bursts in pipelines, the existing techniques require offline processing that can delay the detection of bursts or require specific prior knowledge about the systems.

The analysis of transient pressure signals using supervised learning algorithms has not been widely explored. Bohorquez et al. (2020) presented a technique that uses ANNs to predict the presence of different features (leaks and junctions) in a pipeline after the generation of a controlled transient event. The obtained results demonstrated the potential of combining the principles of transient-based techniques with ANNs to interpret the pressure traces. A similar application was proposed by Perera and Rajapakse (2011) for the identification of transient faults in power transmission networks using hidden Markov models and probabilistic neural networks (PNNs) to classify transients as faults of normal switching events. Considering the current literature, this paper presents a methodology that for the first time merges ANNs and fluid transient waves to detect, locate, and characterize bursts in water pipelines.

The proposed methodology has been applied to a single pipeline under the assumption that a burst can occur at any point along its length. The pipeline, of diameter $D$, is connected at the upstream end to a reservoir, and at the downstream end there is a closed inline valve (Fig. 1). The length of the pipeline is $L_T$, and the location of the burst is characterized by a distance $x$ measured from the upstream end of the pipeline. The bursts are modeled as circular orifices of diameter $D_B$.

Two pipelines with the configuration presented in Fig. 1 were analyzed. A numerical pipeline was considered with the characteristics presented in Table 1. In this case, the head at the reservoir is defined as $H_0=55\mathrm{m}$ at the beginning of the simulation, and a sinusoidal fluctuation of this head is considered to model gradual changes in pressure head that can be observed in pipeline systems. Steady-state friction was considered using a Darcy-Weisbach friction factor $f$ with a pipeline roughness of $\epsilon =0.01\mathrm{mm}$. The different burst locations and sizes that have been considered are described in what follows, as part of the methodology (Steps A.2 and A.3 in Fig. 2) of the proposed technique.

A second pipeline was analyzed as part of the experimental verification of the proposed technique. The characteristics of this pipeline are shown in Table 2 and explained in detail in the experimental results section of this paper.

The effect that a burst has on a transient head trace is characterized by a sharp drop in the head that propagates in both directions away from the burst location (Misiunas et al. 2005; Bohorquez et al. 2018). The proposed methodology for detecting and identifying bursts is presented in Fig. 2. Two stages have been included in this diagram. The first stage comprises the ANN model development (Stage A in Fig. 2), which is carried out first and can be updated regularly depending on the availability of new pressure head data. The application stage (Stage B in Fig. 2) includes the required steps to process and interpret a continuous transient pressure head trace using the ANNs for detecting, identifying, and verifying the occurrence of a burst.

The methodology presented in Fig. 2 was applied to the system described in Fig. 1 and Table 1. A burst detection ANN and a burst identification ANN were designed, trained, and tested to detect and identify bursts. This section presents the results for both the model development stage and the application stage.

A series of experiments in the Robin Hydraulics Laboratory of the University of Adelaide were conducted to validate the proposed methodology for detecting and identifying bursts in pipelines using ANNs. The layout of the experimental pipeline system replicates the system configuration presented in Fig. 1. The upstream end of the pipeline is connected to a pressurized tank, and the downstream end has a closed inline valve. The characteristics of the pipeline are shown in Table 2.

To validate the methodology described in Fig. 2, all the steps of the model development stage were replicated for the pipeline in the laboratory. The architecture of the ANNs and the number of samples to create the input data set for the ANNs were not modified; however, new transient head traces were numerically generated taking into consideration the pipeline characteristics presented in Table 2.

The fast opening of a side-discharge solenoid valve was used to model the occurrence of a burst in the pipeline. Therefore, the numerical modeling of the transient head traces included a typical opening curve for a solenoid valve with an opening time of 10 ms. The downsampled frequency selected was 5 kHz considering that the pressure head in the experiment was measured at the downstream end of the pipeline using a PDCR 810 (Druck, Groby, Leicester) pressure transducer with a 10-kHz sampling rate.

Once the numerical transient head traces were generated, the time windows were obtained, classified, and used to train and test the burst detection and burst identification ANNs. Because the same procedure was explained for the numerical results, the results of the training and testing of these ANNs are not included here. However, considering that stochastic gradient descent algorithms were used for the ANN training, each time that this process is conducted, the resulting weights are different. Therefore, both ANNs were trained multiple times (without modifying its architecture) to assess the robustness of the obtained results. In the case of the burst detection ANN, the results were not sensitive to changes in the ANN weights, so only the results from one training attempt are presented. For the burst identification ANN, results will be presented in what follows in this section for three of these training processes. Each training process for the burst detection ANN took approximately 3 h, while the training of the burst identification ANN took 15 h on average.

The transient head trace obtained from the experimental test is shown in Fig. 11. This trace corresponds to a burst (modeled as the opening of a side-discharge solenoid valve) located 11.06 m downstream of the pressurized tank and has a size of approximately 2.0 mm. The initial pressure head before the occurrence of the burst was 29.48 m. Fig. 11 also presents the length of the time windows for this pipeline. The length selected was $2.5L/a$ seconds, which corresponds to 0.072 s.

The application stage was validated on the transient head presented in Fig. 11, following an individual analysis of each time window in sequence. Results from the burst detection analysis (Step B.4 in Fig. 2) are presented in Fig. 12. This figure presents the category predicted by the burst detection ANN for the first 577 time windows of the complete transient head trace. In addition, the correct category for each of these time windows is included with the objective of illustrating where the burst detection ANN misclassified a time window.

As can be seen in the figure, most of the time windows are correctly classified, except for one time window that belongs to Category N and was classified as Category Ab-I. In addition, some other time windows were classified in Category N, but belong to Category Ab-I (showing a time lag in the ability of the burst detection ANN to identify the occurrence of a burst), and two time windows that belong to Category Ab-I were classified into Category Ab-C (not visible in the plot).

This figure also shows the time windows that were considered in the consistency test. A total of 144 time windows were classified as Category Ab-C and so are enclosed in the rectangle in the figure. These 144 time windows were processed by the three different burst identification ANNs, and the distribution of the predictions for the burst location and size are presented in Fig. 13.

Fig. 13(a) shows that the three burst identification ANNs provide different distributions of results, but for the three cases, the median predicted location is within 1.50% of error in terms of the total length of the pipeline (37.24 m), as shown in Table 3. The closest prediction to the real location of the burst was obtained for the third burst identification ANN in which the burst location was predicted to be only 0.06 m away from the real location. In a similar way, Fig. 13(b) presents the distribution in the predictions of the burst size. The three burst identification ANNs predicted different burst sizes, resulting in a median error between 6% and 16% of the true burst size. Although these results are less accurate than the burst location results, it is important to note that the largest median error corresponds to an absolute error of only 0.32 mm. In addition, the computational time required to obtain these results was only a few seconds.

Lastly, a burst verification analysis was conducted. Three different transient head traces were generated taking into consideration the prediction of the three burst identification ANNs and they were compared to the experimental head trace presented in Fig. 11. The comparison is shown in Fig. 14(a), and the obtained values for the NRMSE are presented in Table 3. There is significant agreement between the transient head trace obtained in the laboratory and the numerical transient head traces generated from the predictions of the burst identification ANNs in terms of the occurrence of the head drop and recovery due to the arrival of the burst wave to the measurement point. This is due to the accurate prediction of the burst location. However, differences are visible in the magnitude of the head drop for the three predictions. These differences are due to the error in the prediction of the burst size given that the magnitude of this head drop is directly related to the burst size. Following the proposed methodology, a burst prediction adjustment was applied taking into consideration the discrepancy between the transient head traces in Fig. 14(a) and the fact that the values for NRMSE are larger than the selected threshold.

Taking into consideration the prediction location obtained from the burst identification ANNs, a series of transient head traces with different burst sizes were generated covering the complete range of possible burst sizes (ranging from 0.4 to 3.5 mm in increments of 0.1 mm). For all of these transient head traces, the NRMSE was calculated when compared to the experimental transient head trace. Fig. 14(b) shows the transient head trace with the lowest NRMSE, which corresponds to a burst size of 2.0 mm (obtaining an error of 1.47%). It is important to highlight that only 31 transient head traces were generated for this adjustment, which does not represent an impractical computational time. Thus, the proposed technique was able to locate and characterize the occurrence of a burst in the laboratory with enough accuracy based only on the interpretation of the measured transient head trace.

This paper presents a novel technique to detect and characterize the occurrence of bursts in pipelines by merging the use of fluid transient waves and artificial neural networks. A complete methodology was described and divided into two main stages: model development and model application. The model development stage includes the generation of a transient head sample data set and the design, training, and testing processes to obtain two different ANNs for (1) the detection and (2) identification of a burst. The application stage includes the processing of a potentially continuous transient head trace into individual time windows to be processed through the two ANNs and a final stage of verification of the obtained results.

The technique was applied to a numerical example and validated with data obtained from a laboratory test at the University of Adelaide. In both cases, the technique was successful at detecting and identifying the occurrence of a burst in a pipeline. In the numerical application, a sharp burst was considered. The burst detection ANN accurately classified the analyzed time windows in one of three possible categories with only a short time lag of 0.016 s to identify transient head traces in the third category. The burst identification ANN predicted the occurrence of a burst 0.85 m from the real burst location (an error of 0.085% on the 1,000-m pipeline) with an error in the prediction of the burst size of 0.17 mm.

For the experimental laboratory application, the burst was modeled as a circular orifice that did not open instantaneously in order to test the technique with transient head traces that do not contain perfectly sharp transient waves. In this case, the burst detection ANN accurately classified most of the time windows. However, there was a slightly longer time lag in the identification of the first head deviation from normal conditions. For the burst identification ANN, three different trainings were conducted to test the robustness of the predictions. Each burst identification ANN predicted different burst locations and sizes; however, all the errors in the location prediction were within 1.5% of the total length of the pipeline. The burst size predictions were less accurate, oscillating between 6% and 16% of the real burst size.

Because this is the first application using fluid transient waves and ANNs, the training samples were obtained from a numerical model. However, future applications could use historical data from real systems and consider more general situations in terms of the burst characteristics and hydraulic systems analyzed. An application of this technique in a more complex hydraulic system might involve the development of more robust ANNs, the generation of more training samples, and the placement of more sensors in the system to collect enough information.

The numerical samples used for the ANN training and the laboratory data that support the findings of this study are available from the corresponding author upon reasonable request.