Development of the Methodology for Pipe Burst Detection in Multi-Regional Water Supply Networks Using Sensor Network Maps and Deep Neural Networks

Abstract

Multi-regional waterworks are large-scale facilities for supplying tap water to the public and industrial parks, and interruptions in the water supply due to leaks result in massive social and economic damages. Accordingly, real-time, around-the-clock accident monitoring is necessary to minimize secondary damage. In the present study, a section of a large-scale waterworks transmission mains system with frequent changes in its physical boundaries was defined for sensor network map-based deep learning input and output. A deep neural network (DNN)-based pressure prediction model, able to detect pipe burst accidents in real-time using short-term data collected over periods within 1 month, was developed. A sensor network map refers to a sensor-based hierarchy diagram, which is expressed using a hydraulically divided area. A hydraulically independent area can be determined using known value information (e.g., the known flow, pressure, and total head) in a complex water supply system. The input data used for the deep learning model training were: the water levels measured at 1 min intervals, flow rates, ambient pressure, pump operation state, and electric valve opening data. To verify the developed methodology, two sets of real-world data from past burst accidents in different multi-regional waterworks systems were used. The results showed that the difference between the pressure as measured by pressure meters and an estimated pressure was extremely small before an accident, and that the difference would reach a maximum at the time point when an accident occurs. It was confirmed that an approximate estimation of an accident occurrence and accident location could be estimated based on predicted pressure meter data. The developed methodology predicts a mutual influence between pressure meters and, therefore, has the advantage of not requiring past data covering long time periods. The proposed methodology can be applied immediately and used in currently operational large-scale water transmission main systems.

1. Introduction

Non-revenue water (NRW) is water that is lost before it reaches the customer. Losses can be divided into real losses such as through leaks or apparent losses, for example, through metering inaccuracies. The global volume of non-revenue water has been estimated to be 346 million cubic meters per day or 126 billion cubic meters per year [1]. Here, leakage is a key loss component that consists of NRW; therefore, a reduction in the leakage is a pillar in the cyclic economy regarding the water urban systems. In pipe-failure detection in waterworks systems today, real-time flow and pressure gauges are installed in the systems, triggering alarms in the case of a deviation from the upper and lower operational limits set for each gauge, based on the operational experience and statistical baselines. Furthermore, whereas multi-regional waterworks are branched pipe systems built for long-distance water supply, redundant pipelines have gradually been installed in recent times to minimize the nuisance from water cutoffs and to facilitate cleaning and management. Due to an increased demand from new town and industrial park developments, multi-regional waterworks are operated in coordination with other water supply systems, further increasing the complexity of the facilities and their operation and management. To date, pipe-accident detection is predominately reliant on customer complaints. The initial response to accidents that occur, including identifying the site of an accident, takes at least 1 h. The multi-regional waterworks of Korea are large-scale facilities for supplying tap water to national industrial parks and the country’s 113 municipalities, and interruptions in the water supply due to leaks result in massive social and economic damages. Accordingly, real-time, around-the-clock accident monitoring is necessary to minimize the secondary damage.

Methodologies for leak detection in waterworks pipe networks based on the utilization of flow rate and pressure data can be organized into three categories: transient flow analysis-based models, hydraulic models, and data analysis-based models. In the methods using a transient flow analysis, transient flows occurring along pipelines in the event of damage are analyzed to detect the occurrence and location of accidents. These require data collection using high-rate pressure sensors. Depending on the method of utilization, these can further be classified as directly detecting negative pressure waves [2,3], and burst-induced transient signals [4,5]. However, this method is operated in only a handful of actual pipe networks due to issues such as frequent data collection, high transmission costs and background noise, and this becomes particularly difficult as the network complexity increases. Hydraulic model-based methods [6,7,8,9] determine pipe failure using the differences between estimated values and real measurements. A prerequisite for accident monitoring using hydraulic models is the implementation of a hydraulic analysis model that is able to accurately simulate normal hydraulic behavior. Furthermore, such models need to be continually updated in keeping with the facilities and operation/management conditions. Water service providers are limited in their ability to continually update their hydraulic analysis models in house. This means that the prediction reliability gradually decreases after an initial implementation, and these systems are only continuously used in a few cases.

Data-based models can be classified according to the methods of data analysis and the utilization into classification models, prediction-classification models, and statistical models. The classification method relies on the classification of normal data and abnormal data occurring in the event of a pipe burst. Mounce and Machell [10] studied pipe burst detection through static and time-delayed artificial neural network models with varied neural network structures, using real-time flow rate data. Aksela et al. [11] introduced an artificial SOM (self-organizing map) neural network-based leak detection technique, improved using leak detection functions. These classification models are supervised learning models, requiring both normal and abnormal (that is, from bursts or leaks) data for training; however, the volume of abnormal data available is insufficient for training. Even if data have been gathered over prolonged time periods, the physical boundaries of pipe systems continually change, and using data from far in the past to predict the near future may be inappropriate.

The prediction-classification model involves unsupervised training-based methods which make normal-state predictions to classify the deviant, abnormal values. Ye and Fenner [12] used weight least squares-based polynomial functions for fitting to past data, allowing for the prediction of a normal flow rate or pressure values. The method was based on the fact that when a pipe burst occurs in district metered areas (DMAs), sensor measurements may be extremely divergent from the predictions which use normal data. Bakker et al. [13] used an adaptive learning model for normal DMA demand patterns to generate a water demand prediction model, then detected pipe burst accidents through the analysis of the deviance from the measured flow rates. In the training process, the results of the deviance analysis accumulated over 1 year were used as the threshold values for classifying the normal and abnormal values. These prediction-classification models must be trained selectively using only data from normal operational states, and accordingly, the data pre-processing step is critical.

Methods under the statistical model are based purely on statistical theory, centered on statistical process control (SPC). Jung et al. [14] applied a univariate and three multivariate SPC methods to pipe burst and leak detection. The univariate methods based on source data were reported as being better performing than the multivariate methods. Of the three univariate methods—Western Electric Company (WEC), Exponentially Weighted Moving Average (EWMA), and Cumulative Sum Method (CUSUM)—applied, the EWMA was reported to have the best detection efficiency. Meanwhile, applying SPC methods requires a local calibration to the variables measured in each pipe system, the system size (total demand), and the operational status. Additionally, Palau et al. [15] implemented a principal component analysis (PCA) method in which the dimensionality of the flow data is reduced.

Meanwhile, the machine learning-based methods for data analysis are reported to require large amounts of training data. However, pipe rehabilitation and the installation of parallel pipe work is continually ongoing in multi-regional waterworks systems to address the aging of facilities and improve the ability to provide a stable water supply. Accordingly, the physical boundaries of pipe systems may change multiple times in the space of a year, limiting the feasibility of applying machine learning-based models, which require training data over long periods.

In recent times, various studies that combine and merge the various data analysis-based techniques introduced above, or that adopt deep learning-machine learning improved with the Fourth Industrial Revolution technology techniques, have been introduced. Zhou et al. [16] proposed a pipe burst accident detection method involving a pressure prediction model trained using the pressure patterns from a simulated accident generated by the EPANET 3 pipe network analysis model. FL-DenseNet (BLIFF), which is an adaptation of DensNET, a type of deep learning-based convolutional neural network (CNN) technique, was used. Wang et al. [17] developed a pipe burst monitoring technology that uses the difference between actual measurements and DMA inflow rate predictions made with long short-term memory (LSTM) techniques, and verified the technique using simulated data including opened fire hydrants and arbitrarily generated and synthesized accident data. Lee and Yoo [18] used a LSTM to generate a flow rate (demand) prediction model, established the threshold conditions using the Shewhart Control Chart, a univariate statistical process management technique, and proposed a methodology where the cases of predicted error-exceeding threshold conditions were determined to indicate pipe burst. Some recent studies [19,20,21,22] proposed leak detection methodologies combining data-based and model-based methods. Quiñones-Grueiro et al. [19] proposed a combined model using a deep neural network model to detect the occurrence of leaks, and locate leaks based on inverse problem solutions, while Ares-Milián et al. [20] used a support vector machine and inverse solution-based procedures to accurately determine the leak location. Wu et al. [21] proposed a leak detection methodology using well-calibrated pipe network analysis data and deep machine learning techniques, and applied the same to a small-scale network in Singapore. Daniel et al. [22] proposed a methodology where the linear regression analysis of data from two pressure gauges is used to identify potential leaks, and leaking pipes are identified and located using the mixed- integer programming method. While these deep learning and combined model methodologies are advantageous in situations where measurement data is readily available and the amount of data is large, most of the latest studies are limited in that the model congruence is tested using a simulated leak accident, and in that the real-world networks adopting these methods are rather limited in size [21,22]. Recently, Wan et al. [23] performed an in-depth review of various data-based methods to detect leaks in water supply network systems. They suggested that there still remains a challenging area for model improvement by considering the uncertainty of the change in hydraulic behavior according to the system size and the variety of leak types and operating conditions.

In the present study, a section of a large-scale waterworks transmission mains system with frequent changes in its physical boundaries was defined for sensor network map-based deep learning input and output. A deep neural network (DNN)-based pressure prediction model, able to detect pipe burst accidents in real-time using short-term data collected over periods within 1 month, was developed. For the implementation of the real-time pipe burst monitoring system, a new sensor network map, based on the hydraulic connections among gauges, was proposed and implemented. This map was used as the zoning and input model for pipe burst monitoring. The impacts of transient hydraulic fluctuations due to pipe burst become dispersed along the transmission main, and the hydraulic connection can be disrupted at service reservoirs and booster station sumps. Therefore, a sensor network map needs to be capable of a water balance analysis by isolating the downstream inflow and outflow rates for managing the deep learning model computing load and improving the model management efficiency. Pressure prediction does not predict the near-future pressure based on time-series predictions, but rather estimates the pressure determined according to a normal-state pipe system energy balance; therefore, it is classified as a prediction-classification model. That is, the pressure being predicted is determined by the ambient flow rate, pressure, and operational state of the hydraulic facilities. A pressure prediction model trained using ambient measurement data during normal operation is able to estimate the pressure at a given pressure gauge using other nearby data in the pipe burst monitoring stage. In the event of a pipe burst accident, a sudden change in the flow rate and pressure at an unknown point in which the system is not trained occurs, and a pressure different from normal is estimated. It is through the residual here that the occurrence of pipe burst accidents can be detected.

2. Methodologies

There are three necessary considerations for creating a burst detection system in a multi-regional water transmission system using a pressure-estimating data-based model. First, there is insufficient data on past burst accident events for training the data-based model; therefore, hydraulic anomalies at the time point of the burst accident events cannot be extracted from the monitoring data. Second, long-term training is impossible because the physical boundaries of the data-based model frequently change. Third, and most importantly, is the requirement of a high reliability. Systems with high false error rates are avoided by operators. For real-time burst event detection with the three considerations above in mind, a three-stage methodology was adopted, implemented and tested. The methodology included: setting unit zones for real-time monitoring through a sensor network map; data collection and pre-processing; and pressure prediction for deep learning-based burst event detection.

2.1. Drawing a Sensor Network Map for Real-TimeBased Pipe Burst Detection

Multi-regional waterworks, which supply water to two or more municipalities, are systems with extremely long pipe lengths and with hydraulic retention times of up to 72 h. Currently, the Geographic Information System (GIS), Supervisory Control and Data Acquisition (SCADA), and online pipe network analysis systems have been implemented and are being used for managing operations. However, delays in updating the GIS and hydraulic analysis models make it difficult for the state of operations to be known to persons other than workers in the field; therefore, those who are not field operators have a limited ability to determine the current state of a pipe system’s operation. Additionally, the time requirements for performance measurement and GIS input following pipe work in the field make keeping physical systems up to date a challenge. The SCADA system collects data at 1 min intervals for the real-time monitoring of hydraulic behavior at key points and hydraulic facilities such as treatment plants, booster stations and service reservoirs, and the control of pumps and electric valves. Moreover, the customers of multi-regional waterworks systems are mostly the service reservoirs of municipalities and industrial parks. The supply flow rate and pressure are measured in real-time at some larger and more important facilities for billing and water supply service quality control.

For the implementation of the real-time pipe burst monitoring system, a new sensor network map, based on the hydraulic connections among gauges, was proposed and implemented in this study. This map was used as the zoning and input model for the pipe burst monitoring. The impact of transient hydraulic fluctuations due to pipe burst become dispersed along the transmission main, and the hydraulic connection can be disrupted at service reservoirs and booster station sumps. Therefore, the sensor network map needs to be capable of a water balance analysis by isolating the downstream inflow and outflow rates for managing the deep learning model computing load and improving the model management efficiency.

The sensor network map refers to a sensor-based hierarchy diagram, which is expressed using hydraulically divided areas. Hydraulically independent areas can be determined using fixed grade or boundary node information (e.g., the known flow, pressure, and total head) in a complex water supply system. A “fixed grade node” is generally defined as a point at which the head generally does not change when simulating a water supply system. The sensor network map is used to identify the status of the input data for the model design. In the developed model, DNN models were created for all the pressure gauges in the pipeline system. In the case of the example in Figure 1, if the sector is not divided, 22 input nodes are required for the prediction of the P2. Conversely, if the sector is divided, only 10 input nodes are required, reducing the computing load by more than half.

2.2. Data Collection and Preprocessing

The data including the past pipe burst events that occurred in the three water systems were collected, and validated, cleansed, and normalized. Data cleaning was conducted for predefined extreme minimum and maximum thresholds that could not physically occur, such as negative values and exceeding the facility capacity, and the data with different scales were normalized to apply the model. In the case of the data normalization, numerical data such as the pressure and flow values were changed to a value between 0 and 1 using the MaxMinScaler, and nominal variables such as the pump operation were digitized as values of 0 or 1. If there was a false reading or missing data in even one sensor in the collected dataset, all the data measured at that time was deleted. As reported in numerous studies on waterworks’ operating pattern characteristics, waterworks’ operating patterns vary among the weekdays, Saturdays and Sundays [13]. In Korea, with four pronounced seasons, differences are also apparent from season to season and between the vacation and school seasons. In the present study, a minimum time interval of 1 week was established for the dataset.

2.3. Pressure Prediction Model Using Deep Learning Algorithms

2.3.1. Pipe Burst Detection Concept from Hydraulic Balancing

The hydraulic analysis of waterworks pipe systems is carried out in accordance with the two basic governing principles of the conservation of energy (i.e., the sum of energy loss in each loop is 0) and the conservation of mass (i.e., an equal inflow and outflow at each node). Accordingly, based on these two basic principles, a steady flow waterwork pipe system analysis model calculates the head and pipe flow rate for a specific time point at a given point under fixed head (reservoir level, and tank level) and fixed water demand conditions. The process of finding the flow rate and pressure is called hydraulic balancing, and this can be calculated using iterative methods for linear functions.

Not all outflow from the system is measured, and there are uncertainties such as background and unreported leakages commensurate to the percentage of the outflow measured. However, under conditions where no changes are made to the physical system, a hydraulically balanced pattern is formed among the flow rate and pressure measurements at the respective measuring points. In other words, when these basic principles of pipe network analysis are applied to the measurement data, there exist relationships that allow an estimation of the values at a given point using nearby measurement data. Meanwhile, fluctuations are caused in the flow rate by artificial intervention on the part of the operator, and pressure fluctuations are caused as a result of changes in the flow rate and pumps being switched on or off. This means that once the control patterns are known, data can be relatively easily calculated from normal, hydraulically balanced patterns.

In this study, a pressure estimation model able to calculate the pressure at a given pressure gauge using same-time data from nearby, hydraulically-connected measuring instruments was generated for the immediate detection of pipe burst accidents upon their occurrence. The pressure estimation model was created using a deep learning known for being useful on imperfect data and which is good at extracting patterns in input data [24]. That is, the pressure prediction models generated for each pressure gauge lose their normal prediction accuracy due to events that disrupt the hydraulic balance such as pipe bursts. Such temporary increases in the prediction residuals can be used to detect the occurrence of pipe burst accidents.

2.3.2. Design of Deep Learning Structure and Hyperparameter Setting

A deep neural network (DNN) is a type of artificial neural network (ANN) in which numerous hidden layers exist between the input layer and output layer. Researchers have proposed thousands of types of specific neural networks that revise or adjust the existing models. There is a class of theoretical frameworks for deep learning, such as a multi-layered feed-forward neural network (FFNN), recurrent neural network (RNN) and convolutional deep neural network (CNN).

In the present study, a FFNN was used for generating the pressure estimation model, with the pressure estimation problems calculated by nearby pressure sensors in a normal equilibrium state. That is to say, the data set order was not given consideration, and supervised deep learning was used as the multiple regression model. As for the training data for the generation of the pressure prediction models for each pressure gauge, the data inputs used at 1 min intervals were: the water levels measured, flow rates, ambient pressure, pump operation state, and electric valve opening data. The pressure prediction formula of the resultant deep learning model is represented in Equation (1) below:

$$\begin{gathered} P_{p_x}=w_{f_1}{F}_{f_1}+\cdots +w_{f_n}{F}_{nf}+ \\ {w}_{p_1}{P}_{p_1}+\cdots +w_{p_{\left(x-1\right)}}{P}_{p_{\left(x-1\right)}}+w_{p_{\left(x+1\right)}}{P}_{p_{\left(x+1\right)}}+\cdots +w_{p_n}{P}_{p_{np}} \\ {w}_{b_1}{B}_{b_1}+\cdots +w_{b_n}{B}_{b_{nb}}+ \\ {w}_{v_1}{V}_{v_1}+\cdots +w_{v_n}{V}_{v_{nv}}+b \end{gathered}$$

(1)

Here, $F_i$is the flow rate, $P_i$ is the pressure ith, $B_i$ is the pump on/off, $V_i$ is the electric valve opening rate (%) at the ith sensor, nf is the number of flowrate gauges, np is the number of pressure sensor, nb is the number of pumps, nv is the number of electric valves in the system, $w$ represents the impact (weighted) of each gauge, and $b$ is a bias.

Figure 2 shows the concept diagram of the model, which means that the flow rate (F), pressure (P), pump (B), and electric valve (V) are used in the pressure prediction model for the pipe breakage accident detection. In the deleted equation, t means the “target” to be predicted, and the equation has been deleted because there is room for confusion. Therefore, when there are 1 to n pressure gauges, the symbols of the remaining pressure gauges, except for the pressure gauge t to be predicted, are expressed as P(t − 1) and P(t + 1).

Therefore, in the case study presented in Figure 3, a total of 18 input nodes, 6 flowmeters, 5 pressure gauges (except for one point for prediction), and 7 pump states, were created to create one pressure prediction model.

Data from the previous three weeks were used as the validation data for training/validation and the sensitivity analysis from the time point for an accident judgment, and also one week after the accident was used as the test data. That is, data for the 3 weeks before the accident were randomly sampled as 8 (training): 2 (validation) and performed.

The pipe burst detection model of the present study must be able to provide reliable anomaly detection while also being capable of regular updates to reflect the latest pipe system operating states. This means that large amounts of training data cannot be used, and the model must be regenerated using the latest data. In our study, the weekly re-generation of a new pressure prediction model was assumed. As for the network structure-related hyperparameters of the number of hidden layers, the number of neurons in each layer and the type of activation function were determined, while a review of previous literatures [18,25,26,27,28] and a simple sensitivity analysis were carried out. The sensitivity analysis was performed using combinations of 5 numbers of hidden layers, 3 activation functions, and 10 cases of hidden nodes in a hidden layer. In the “Rule of thumbs” of the above references [25,26,27], the number of neurons in the hidden layer was 2/3 of the number of input nodes or a value plus the number of output nodes, or less than twice the number of input nodes, or between the number of input nodes and output nodes that were recommended. Based on the literature, this research team conducted a sensitivity analysis of the forward approach method, which is performed when the number of neurons is increased from a small number. The number of hidden layers used in the sensitivity analysis was 8 (with the basic input nodes ×0.5, ×1, ×2, ×4, ×6, ×8, ×10, ×12, ×14, and ×16) cases performed.

From the sensitivity analysis results, it was decided that better results could be expected if using a rectifier function as the activation function and with 144 nodes in the hidden layer, which is 8 times the number of nodes in the input layer, and 4 hidden layers.

Looking at the previous research related to burst recognition, Lee and Yoo [18] proposed a structure with 3 layers and 128, 64, and 48 layer-by-layer neurons, and the duration of the learning data used in that case was 9 days (12,960 data). Additionally, Capelo et al. [28] proposed a structure with three layers and 55, 50, and 45 neurons, and suggested that the result with the densest layer and neurons had the best result. They concluded that a denser layer was needed, and that caution should be exercised against overfitting. The authors believe that the optimal structure of deep learning can vary depending on the nature and situation of a problem through the development and application of such prior research and this research model.

The reason for such a result is that this study has two characteristics unlike other deep learning problems. First, it is difficult to predict the pressure exactly in a water supply system. The water supply system can calculate the flow rate in the pipeline and the pressure in the demand point according to the basic equations, the law of the conservation of mass and the law of the conservation of energy, but the relationship between the head loss and flow in the pipeline is nonlinear. Therefore, there is a significant barrier to predicting the pressure value of a very complex (more than 1000 nodes and pipelines) system with a simple system of equations.

Second, such a water supply system fluctuates every minute and every hour according to the users’ water use, and there is no periodicity depending on the operation of the ancillary facilities such as pumps and valves. For this reason, we developed this model on the premise of building a learning model using only data within 3 weeks instead of building a predictive model according to the continuation of long-term learning, which is a characteristic of a general deep learning model. The reason for using the learning model with a short data period was that more than 1000 models are needed to be individually developed in order to practically predict whether a leak is for the entire metropolitan water supply system.

The performance of the FFNN according to the change in the number of layers was performed during the hyperparameter selection process as shown in

Table 1. Error evaluation metric results depending on the number of hidden layers.

Number of Layers	Error Evaluation Metric
	Coefficient of Correlation (R2)	Mean Absolute Error (MAE)	Mean Squared Error (MSE)
1	0.807897	0.121182	0.034666
2	0.929306	0.076717	0.012905
3	0.938509	0.068533	0.011218
4	0.939177	0.065202	0.010971
5	0.653996	0.144021	0.063374

, and it was numerically confirmed that the multi-layer performance was superior to that of the shallow layer.

3. Application Results

3.1. Application Event Data Set and Drawing a Sensor Network Map

To verify the developed methodology, two sets of real-world data from past burst accidents in different multi-regional waterworks systems were used. The dataset included a week of test data including a burst accident event, and 3 weeks of pre-event training data. The datasets for Case #1 and Case #2 represent raw water supply networks where the entire system must be analyzed as a single sector. The sensor network map and pipe network information used for the pipe burst detection model implementation are shown in Figure 3 and Figure 4. Case #1 was a pipe breakage event with a clearly cut pipe connection and leakage due to a pipe breakage that occurred instantaneously up to about 7500 CMH. In Case #1, about 750 CMH of leakage occurred due to a pipe leakage event.

The 3 weeks of data (30,240 data points) gathered for the model training and validation were checked for outliers, then pre-processed. Anomalies owing to extreme error data were not found in the data pre-processing. Missing and holding data occurring in the data transmission process were deleted from the training data.

3.2. Application Results of Case #1

The accident in Case #1 was an accident occurring in a 45-year-old D1000 mm ductile cast iron pipe of a raw water supply pipe system. The monitoring zone is a pipe system that supplies raw water from a booster station to three water treatment plants and industrial facilities. The system is operated by seven pumps, with a total of six pressure gauges and flow meters installed in the system. The results of applying the developed model are as shown in Figure 5. The results showed that the difference between the pressure as measured by the six pressure meters and the estimated pressure was extremely small before the accident, and that the difference reached a maximum at the time point when the accident occurred.

The prediction errors over 4 days up to 13 September 2019, prior to the accident, are shown in

Table 2. Time series for measured and estimated pressure for Case #1.

Pressure Meter	Apparent Event Detection	R2	MAE	MSE	RMSE	MAPE (%)
P1	Yes	0.9914	0.0448	0.0024	0.0492	0.7672
P2	Yes	0.9611	0.1039	0.0136	0.1168	1.8428
P3	Data Missing *	0.9728	0.1563	0.0828	0.2878	4.4182
P4	No	0.9998	0.0040	0.0000	0.0054	0.3053
P5	Yes	0.9911	0.0647	0.0067	0.0818	4.2245
P6	No	0.9994	0.0057	0.0001	0.0083	0.4364

* Data was not properly collected at P3 in this event.

. The “Apparent Event Detection” presented in Table 2 indicates whether a reliable leak was possible, and “Yes” means that the difference between the predicted value and the measured value at the time of the burst was clearly distinguished, as shown in Figure 5. The variability of pressure values at each gauge is shown in Figure 6. The average residual (MAPE: mean absolute percentage error) was within 5% in all the pressure prediction models except for P4, confirming a good pressure prediction.

A heatmap of the overall prediction error is shown in Figure 7. The predicted or estimated error value in the heatmap could be calculated from Equation (2). “Time” on the left side of the figure is the result in 1 min units and expresses the “Hour: Minute: Second”:

$$Estimated error(\%)=\frac{P_{e_i}-P_{m_i}}{P_{m_i}}\times 100$$

(2)

Here, $P_{e_i}$ is the predicted pressure in gauge i, and $P_{m_i}$ is the measured pressure in gauge i.

The accident occurred at 21:45 and 21:46. Data was not properly collected at P3, closest to the accident site, and the largest prediction error was observed at P1, which was the second closest. When an accident occurred, the predicted pressure learned from the normal data appeared higher than the measured pressure; therefore, only the (+) error was judged to be related to the accident. If the ID was +/−1, it meant that it was an adjacent instrument; however, when an actual accident occurred, a sudden increase in the prediction error occurred in two or more measuring instruments that were close to the accident point and were hydraulically close. If only one measuring instrument was regarded as an accident, there was a possibility that it may have been caused by a malfunction of the instrument; therefore, there was a high possibility of generating a false alarm. In the case of the threshold for judging anomalies, the prediction reliability was different depending on the installation location for each pressure prediction model. Currently, using the MAE for the validation data for each model, an abnormality judgment standard would be operated; therefore, a case in which the error rapidly increased in two or more adjacent measuring instruments would be classified as an accident. Through this process, the model would be able to approximately locate the site of an accident.

3.3. Application Results of Case #2

The monitoring zone of Case #2 is part of a water conveyance pipe system supplying raw water at a facility capacity of 2,600,000 m³/day. The system supplies raw water to two water treatment plants and nearby municipalities. The accident to be predicted was in a D800 mm pipe due to damage from other work. Six flow rate data and 10 pressure data were used for the accident prediction. The results of applying the developed model are as shown in Figure 5. The variability of the pressure values at each gauge is shown in Figure 8. The results showed that the difference between the pressure as measured by the 10 pressure meters and the estimated pressure was extremely small before the accident, and that the difference reached a maximum at the time point when the accident occurred.

The variability of the pressure values at each gauge is shown in Figure 9. At P11, there was a temporary increase in the prediction error due to a coordinated operation with another system not trained through V4. P12, a pressure gauge installed at the leading section of a booster station sump, was found to be affected by the booster station valve controls.

A heatmap of the overall prediction error is shown in Figure 7. The variability of the pressure values at each gauge is shown in Figure 10. Unlike in Case #1, the prediction error for the pipe system in Case #2 was relatively low; however, as shown in the figures, abnormal and sudden increases in error were found at points P6 and P7, which were closest to the point of the accident, confirming a good accident detection performance.

4. Conclusions

The present study proposes a pressure prediction methodology through real-time monitoring using a sensor network map, and a methodology comprised of establishing unit zones, data collection and pre-processing, and deep learning-based burst event detection. The developed methodology was applied to two large-scale real-world accidents that occurred in the past. The significance of the study is as follows. For the implementation of a real-time pipe burst monitoring system, a new sensor network map based on hydraulic connections among gauges was proposed and implemented. This map was used as the zoning and input model for pipe burst monitoring. The proposed sensor network map can be used effectively in waterworks systems prone to frequent operating condition changes and with changes in their boundary conditions such as the demand/supply balance. Additionally, unlike some previous studies, the performance of the model of the present study was tested using actual past accident data; therefore, the model can be applied immediately to currently operational systems. Previous studies have focused on the simulation and prediction of artificial accidents through pipe network models, and time-series predictions from single flow meters and pressure gauges. In this study, the model was trained to predict pressure values using data obtained from multiple measuring instruments positioned across the entire pipe network. The methodology proposed by our study uses a prediction model based on a sensor network map and it does not require the input of a long period of past data accumulated over long time periods or predefined demand patterns. Ultimately, the model has the advantage of being able to deduce reasonable results even with frequent changes to pipe system boundary conditions. Although the proposed method has been proven to work in two test cases, more test cases or synthetic control experiments are needed in the near future for a generalized efficiency evaluation. In particular, follow-up studies that can derive a rational model operation plan based on real-time model operation results are to be expected, since the currently-proposed model is installed and operated in real-time. In addition, the methodological development of the proposed method is necessary through a comparison between various deep learning models in the future. Additionally, the reason for why the multi-layer-based feed forward neural network used in this study is more effective for this problem than other deep learning techniques, should be derived through a detailed analysis of additional, practical application results.