Detection and Prediction of Leakages in Water Distribution Networks

Mariaelena Berlotti

, Sarah Di Grande

, Salvatore Cavalieri

and Roberto Gueli

Department of Electrical Electronic and Computer Engineering, University of Catania, Viale A. Doria n.6, Catania, Italy

EHT, Viale Africa n.31, Catania, Italy

Keywords: Water Distribution Systems, Leakages, Anomaly Detection, Predictive Maintenance.

Abstract: Leakages are one of the main causes of water loss in a water distribution system (WDS). In recent years, the

increasing of streaming data coming from sensors installed in the water network, allows the monitoring the

health status of each asset of the WDS. In this paper, a preliminary data-driven approach for leakages detection

and prediction is proposed. Starting from the characteristics of a real water distribution network, a realistic

leakages dataset has been achieved. Using this dataset, unsupervised rule-based time series algorithms has

been trained for the detection and prediction of leakages.

1 INTRODUCTION

Water loss in the Water Distribution Systems (WDSs)

is a topic that has been attracted much attention in

recent years. The International Water Association

(IWA) defines leakages as an important form of loss

of water from a WDS due to leaks (Pearson, 2019).

Leak is considered as a failure causing an unplanned

loss of water from a network. The term is generic and

can be used to define leaks of any size and referred to

any type of asset from pipes and valves to reservoirs.

As concerns the sizes, leaks can be categorized as

abrupt leakages and incipient leakages (Vrachimis et

al., 2018). Abrupt leaks are leakages that occurs

suddenly in a water system and results in large

volumes of water coming out of the network in a short

period; generally, this type of leakages is associated

to pipe burst. Incipient leakages, instead, increase

gradually over time starting as background leakages

and developing into full-blown leakages

(Tornyeviadzi and Seidu, 2023).

Detection of incipient leakages is difficult as they

typically occur in pipes with smaller diameter,

casusing at the beginning low volumes of water loss;

this type of leaks growth slowly during the time

leading to huge losses if not discovered and repaired

in due time (Wan et al., 2022).

https://orcid.org/0009-0008-8895-2175

https://orcid.org/0009-0007-6564-704X

https://orcid.org/0000-0001-9077-3688

https://orcid.org/0000-0002-8014-0243

Leakage detection is defined as the process of

locating and pinpointing water leaks (NAIADES,

2022).

Among the existing approaches for leakages

detection there are those based on data-driven

models. These models rely on learning techniques

applied on a collection of data coming from the WDS

and for this reason, they do not require a domain

knowledge about the network. On the other hand, a

large amount of historical data is needed to perform

the analysis (Escofet et al., 2016).

Prediction of water leakages involves spotting

leaks before they happen (Cody, 2020). Leakages

prediction methods are used to identify areas and

pipes in the network with a high probability of

leakage, allowing water utilities to create an

appropriate active leakage control plan (Leu and Bui,

2016). Prediction of water leak is a challenging task:

tightness and invisibility of the hydraulic components

as the rarity and uncertainty of these events makes the

prediction of these faults events difficult (Wang et al.,

2022).

Data-driven approaches have emerged as a

powerful tool for predictive maintenance

applications; indeed, the increase of data availability

collected through the sensors and the smart meters

lead to the beginning of a digitalization process of the

436

Berlotti, M., Di Grande, S., Cavalieri, S. and Gueli, R.

Detection and Prediction of Leakages in Water Distribution Networks.

DOI: 10.5220/0012122000003541

In Proceedings of the 12th International Conference on Data Science, Technology and Applications (DATA 2023), pages 436-443

ISBN: 978-989-758-664-4; ISSN: 2184-285X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

water sector, known as Water 4.0 (Adedeji et al.

2022). Water 4.0 includes the service innovation of

water networks: maintenance becomes preventive

and predictive programmed on the basis of signals

data (Caldognetto et al., 2022).

In this paper, the authors are going to propose a

data-driven approach to realize both detection and

prediction of leakages, considering a real WDS of the

city of Milan (Italy). The machine learning models

aimed to realize this aim have been defined. Using

these models, a performance evaluation has been

carried out, comparing the leaks detected with the real

ones. Finally, the selected algorithm has been used for

leakages prediction and the relevant results have been

achieved.

The paper is structured as follows: Section 2 will

give an overview about the similar approaches found

in the literature, in order to point out the originality of

the work proposed. In Section 3 the authors will give

a description of the approach proposed for detecting

and predicting leakages. Section 4 presents the results

obtained from the application of the approach on the

leakages dataset. A final section will summarize the

conclusions.

2 RELATED WORK

The aim of this section is to give an overview of the

main approaches present in the current literature

about the use of data-driven approaches for leakages

detection and prediction in WDSs. This overview will

allow to point out the originality of the proposal.

Leakage detection methods can be broadly

classified into hardware methods and software-based

methods.

Hardware methods can be further categorized into

passive and active systems; if the former requires

vision and sensor utilization, the latter involves the

analysis of acoustic, vibration, flow or pressure

signals (Chan et al, 2018). In (Hunaidi et al. 1999),

leak identification through the acoustical signals

given by plastic pipes is presented. More recently,

(Cody et al. 2020) proposed a mixed approach where

deep learning is involved for the monitoring of

hydroacoustic spectrograms to pinpoint leaks on

pipelines. Finally, in (Wang et al. 2021) the authors

investigated the characteristics of acoustic signals

obtained by simulating leaks through an experimental

platform; then these signals have been passed to an

artificial neural network model for leak detection.

As concerns vibration signals, (Bentoumi et al.

2017) proposed a leak-detection model based on the

‘Haar’ continuous wavelet; the algorithm takes as

input vibration signals issued from a water pipeline

and decides if there is or not a leak in the network. In

(Yu et al. 2023), the authors presented a machine

learning models for leak detection on vibration

signals collected by wireless piezoelectric

accelerometers placed in real complex water

distribution systems.

Active systems comprise transient-based

approaches, hydraulic model-based approaches and

data-driven approaches. The core idea of transient-

based approaches is that any change in the physical

structure of the pipe can alter flow and pressure

measurement of a system (Wan et al., 2022). To

capture transient behavior, this type of analysis

requires big amount of data with high sampling

frequency that results in costly and too complex

process. For this reason, transient approaches are not

recommended for real-time monitoring of large

WDSs (Colombo, 2009).

Hydraulic model-based approach instead uses

mathematical functions and formulas to replicate the

operation of a network. Apart from requiring domain

knowledge to be built, these models need the

availability of large amount of historical data for

calibration. Another major drawback is that model-

based methods assume WDS conditions stability over

time; this is not true in real life scenarios since factors

as pipe ages and roughness coefficient as time goes

on increase and become increasingly influential on

leak occurrence (Perez et al., 2014).

Active systems include data-driven models; in

particular, three types of approaches can be actually

used for the detection of leakages: Supervised, Semi-

supervised and Unsupervised learning.

In Supervised learning methods binary or multi-

class classifier are trained using normal and abnormal

labeled data. This type of methods is rarely used for

leakages detection in reality due to the lack of labeled

hydraulic data. Moreover, if they let us reach high

accuracy for the identification of the leaks in small

and simple WDS, this is not the case for larger and

more complex network (Kammoun et al., 2022).

Semi-supervised learning requires only the

availability of normal labelled data; they have been

adopted for water quality applications (Barros, 2023).

Finally, unsupervised learning algorithms do not

rely on either normal or abnormal labeled data

availability. They are widely used in the field of

leakages detection since they are more flexible and

realistically feasible (Kammoun et al., 2022). In

particular, in this paper, the authors applied an

unsupervised RNN model for leaks detection and

localization on flow and pressure data coming from

different realist water demands scenarios of the Leak

Detection and Prediction of Leakages in Water Distribution Networks

437

DB dataset. In the NAIADES project an unsupervised

temporal and spatial anomaly detection approach is

applied to detect leakages on pressure and water flow

data of the Braila districts (NAIADES, 2022).

Although many studies have investigated over the

problem of the detection of the leakages in the water

networks, a very limited amount of researches

focused on prediction of leakages (Leu et al, 2016).

In the paper (Lijuan et al., 2012), the authors

presented a pipe leakages prediction approach based

on a radial basis function (RBF) neural network;

specifically, the authors analyze all the possible

factors influencing leaks and the possible relationship

existing between them that could facilitate in

predicting leakages more effectively. In their work,

(Leu and Bui, 2016) used a Bayesian network

learning (BNL) model with an updated failure

probability of each asset for leakages prediction.

Finally, in (Wang et al., 2022) the authors proposed a

five-dimensions digital twin model for both fault

diagnosis and predictive maintenance on hydraulic

system; to illustrate the effectiveness of their method,

they applied it to an hydraulic cylinder.

In this paper, the authors use the unsupervised-

based time series anomaly detection algorithms for

leakages detection and prediction. Differently from

the existing literature works, only one variable is used

to perform these tasks. Indeed, the core idea proposed

in the paper is to let the algorithms learn leakages

changes in the pressure nodes during the anomaly

detection step and to use then this information also for

the prediction. The advantage of this proposal is that

a reduced set of information is needed for the

detection and prediction of leakages, simplifying the

approach. The paper aims to provide a contribution to

the current literature concerning water leakages

prediction of water, considering the scarcity of

research available on this topic. Furthermore, the

proposed method for identifying and predicting

leakages is applicable to both incipient and abrupt

leakages. The authors believe that this additional

aspect should be considered when evaluating the

paper's contribution to the knowledge of predicting

and detecting water leaks, as current literature

primarily focuses on the detection of abrupt leaks.

3 APPROACH

In this paper we propose a two phases approach:

leakages detection and the leakages prediction. First

phase consists of the application of unsupervised

algorithms for the identification of leaks events.

Then, among the algorithms applied for detection we

choose the best performing one for prediction.

The analysis is divided in several steps: Data

Acquisition (Step 1), Data Pre-processing and

Transformations (Step 2), Leak Detection (Step 3)

and Leakages Prediction (Step 4). The Leakages

Prediction includes a Feature Engineering step, as

explained later.

In the next sections, each phase of the analysis

will be described.

3.1 Data Generation

Data plays a strategic role in a machine learning

approach, as known. For the problem of leakages

detection and prediction in WDSs data about real

losses is needed. Availability of data relevant to

losses is very difficult as many time data is missing;

this happens for different reasons, among which there

is the lack of digital support systems to store the

maintenance activities inside the water distribution

system assets. To solve this problem, in the present

paper the data needed to run the machine learning-

based solution was synthetically generated.

In details, data was created using the Water

Network Tool for Resilience (WNTR), a Python

package designed to simulate and analyse resilience

of water distribution networks (Klise, 2018).

Simulations data is related to the actual water

distribution network of Milan, Italy. For the analysis,

we consider a reduced version of the original urban

water network, obtained through a skeletonization

process that allows us to remove those pipes and

nodes that have a minimum impact on the system

behaviour.

Figure 1 and Figure 2 show a planimetry of the

Milan network and of its reduced version.

The analysed WDS is made by: 12,354 nodes,

17,548 pipes, 26 pumping stations, 95 booster pumps.

To simulate the behaviour of the system in different

days of the week, the coefficients of variation of

water demand should be taken into account. To get

these coefficients, the authors adopted the following

approach. First of all, the real water demand

coefficients recorded by the Supervised Control and

Data Acquisition system (SCADA) each minute on a

particular day, have been considered. SCADA are

industrial applications for the control and monitoring

of assets (either machines or single components of an

equipment).

Then, the original coefficients have been

aggregated to half an hour using the mean as

aggregating function. Next, the results of this 30-

minutes aggregation have been stored, summed to a

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

438

random value between + 1.5 and – 1.5 and multiplied

for a percentage (0.05). This percentage corresponds

to a noise factor that lets us reproduce possible

fluctuations of the water demand curve without

altering the daily patterns of water consumption.

Figure 1: Milan WDS.

Figure 2: Skeletonized Milan WDS.

This method is repeated to generate the coefficient

of variation of the half an hour water demand for the

other days. These coefficients are given as input to the

simulation tool to obtain hydraulics data.

Figure 3 shows the obtained water curves demand.

Figure 3: Water demand curves.

In Figure 3, the black-line curve represents the

original water demand curve, obtained starting from

the real-world coefficients. The other curves in the

plot (Demand-Pat2, Demand Pat3, Demand Pat4 and

Demand Pat5) are the simulated water demand

curves.

WNTR package includes the possibility to add

leaks in the water system.

Leakages are simulated at network nodes

randomly selected. The leakage magnitude varies due

to the assignment of a random leakage hole diameter.

Considering the data obtained through simulations,

the authors used the pressure data recorded each 30

minutes for the nodes of the skeletonized network,

and the leak history used to evaluate the performances

of the algorithms.

3.2 Data Pre-Processing

Data pre-processing was needed before proceeding

towards the detection and prediction of the leakages.

To realize simulated leakages the WNTR

simulator adopts the following method: first it divides

the randomly selected pipe in two parts. Then, it adds

a new node where we want to locate the leak. For this

reason, duplicate columns will be present on the data

achieved by the simulation: one containing the

normal conditions pressure values, and the other

showing a decrease in pressure data when the leak

occurs (e.g., “N04755” column and

“N04755_leak_node” column). We retain columns

whose values reflect the occurrence of a leak. For the

pressure and the leak history datasets, the time given

in seconds was converted in a date format.

Table 1 and Table 2 show the final datasets after

transformations.

Table 1: Pressure dataset.

Timestamp Abbiategrasso Anfossi ...

2009-11-18 00:00:00 66.4090 64.3000 ...

2009-11-18 00:30:00 66.6450 64.3740 ...

2009-11-18 01:00:00 62.2259 63.9548 ...

... ... ... ...

2009-11-22 23:00:00 67.6521 64.6836 ...

Table 2: Leak history dataset.

End Node Start Time Diamete

N03185 2009-11-18 01:30:00 0.3717

N22998 2009-11-18 16:00:00 0.1328

... ... ...

N01352 2009-11-21 02:30:00 1.2112

N10174 2009-11-22 04:00:00 0.1174

Table 1 contains a total of 12,355 columns: first

column represents the Timestamp while the

remaining ones, named as the nodes of the Milan

network, contain 30 minutes-pressure values for each

of the nodes. Table 2 is made by three columns: the

‘End Node’ column containing the names of the leak-

nodes, the ‘Start Time’ column containing the time at

which the leak happens, and the ‘Diameter’ column

giving us information about the size of the leak in the

pipe (in meters). Pressure data was standardized using

the Standard Scaler function in order to normalize

Detection and Prediction of Leakages in Water Distribution Networks

439

features by removing the mean and scaling to unit

variance (Scikit-learn). The proposed choice is based

on the fact that the hydraulic features look like

standard normally distributed data.

3.3 Leak Detection

Detection and prediction of leakages are performed

through the use of the Anomaly Detection Toolkit

(ADTK), a Python package for unsupervised/rule-

based time series anomaly detection (ADTK, 2023).

The algorithms implemented have been widely used

in recent years for anomaly detection applications in

time series data (Gopali and Namin, 2022), (Otte et

al., 2022), (Ameli et al., 2022), (Beliichovski et al.,

2022).

For all these models, there is only one

hyperparameter that must be fixed that is the c-factor.

The c-factor establishes when an observation has to

be considered normal or anomalous basing on the

historical interquartile range. We leave it to the

default value c = 3.0; this means that when an

observation is 3-times greater than the interquartile

range of the n-previous observations, it is classified

as anomalous. As Table 3 shows, the output of the

trained algorithms is a data table containing for each

end node, the timestamp when the leak has been

identified.

Table 3: Example Output Anomaly Detection Algorithms.

End Node Timestamp

Anfossi 2009-11-18 12:00:00

Abbiategrasso 2009-11-18 16:00:00

N05636 2009-11-19 02:00:00

... ...

N00780 2009-11-22 19:30:00

Table 4 summarizes the performances of the

algorithms described in the previous section.

Table 4: Performances Anomaly Detection Algorithms on

Unbalanced Dataset.

Algorithms Accuracy Precision Recall F1

PersitAD 0.8452 0.0007 0.1728 0.0015

LevelShiftAD 0.9007 0.0007 0.1043 0.0014

GeneralizedESD TestAD 0.9985 0.0004 0.0005 0.0005

InterQuartile RangeAD 0.9981 0.0029 0.0055 0.0038

Auto RegressionAD 0.8498 0.0007 0.1637 0.0018

Local Outlier Facto

0.9164 0.0005 0.068 0.0011

Isolation Fores

0.9041 0.0009 0.1325 0.0004

-Means 0.9413 0.0001 0.0166 0.0003

Affinity Propagation 0.9911 0.0002 0.0023 0.0014

When dealing with anomaly detection problems,

we would like to know how good the anomaly

detection algorithm was in identifying the anomalous

events. This measure is expressed by the precision.

Looking at Table 4, the InterQuartileRangeAD

stands out as the top-performing algorithm in terms

of precision score. However, this algorithm suffers

from a Quadratic Time Complexity O(n

), requiring a

greater computation effort. On the contrary, the

Isolation Forest algorithm, despite its lower

computational complexity (Linear Time Complexity

- O(n)), exhibits lower precision compared to the

InterQuartileRangeAD.

Difference in ranges between accuracy and the

other evaluation metrics is motivated by the

unbalancing of the dataset, where observations

referring to normal conditions of the system are

present in large quantities with respect to leakages

events. The analysis of the data generated in the first

tests showed that leakage duration was limited to 24

hours. It was, therefore, decided to extend this

duration to obtain leakages with a minimum duration

of 8 hours and a maximum of 72 hours thus increasing

the size of the data set in the presence of losses.

Table 5 shows the performances of the anomaly

detection algorithms tested on the balanced dataset.

Table 5: Performances Anomaly Detection Algorithms on

Balanced Dataset.

Algorithms Accuracy Precision Recall F1

PersitAD 0.8450 0.4370 0.0300 0.0563

LevelShiftAD 0.8451 0.4074 0.0147 0.0284

GeneralizedESD TestAD 0.8462 0.5455 0.0040 0.0080

InterQuartile RangeAD 0.8470 0.7222 0.0087 0.0172

Auto RegressionAD 0.8403 0.3399 0.0405 0.0723

Local Outlier Facto

0.8354 0.1790 0.0194 0.0350

Isolation Fores

0.7834 0.2310 0.1752 0.1993

-Means 0.8441 0.4383 0.0475 0.0857

Affinity Propagation 0.8443 0.3889 0.0211 0.0400

Looking at Table 5, it can be seen that the

performances of all the algorithms (in terms of

precision) trained on the balanced dataset improved.

Also in this case, the InterQuartileRangeAD

algorithm exhibits the highest precision score

compared to the others. For this reason, the authors

selected the InterQuartileRangeAD algorithm for the

leak prediction task.

3.4 Leak Prediction

Before passing to prediction, a Feature Engineering

step was needed.

Feature Engineering is a machine learning

technique that leverages data to produce new

information by combining features. To reach this

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

440

goal, a mathematical function f is applied to data

(Gutschi, 2018).

In this step, the authors performed a rolling

window aggregation that consist in aggregating data

into equally sized windows for all the timestamps.

Rolling aggregation allows us to build a dataset ready

for prediction so that, starting from the pressure data

which represents the system behaviour over historical

time widows, we can anticipate leakages occurrence.

In the analysis, the rolling aggregation process has

been applied on the test dataset containing the

simulated pressure data of the 24

November 2011.

Table 5 reports the leak history test set.

Table 6: Leak History Test Set.

End Node Start Time

N02197 2009-11-24 23:00:00

N15337 2009-11-24 04:30:00

N10848 2009-11-24 10:00:00

N11577 2009-11-24 16:30:00

N26758 2009-11-24 16:30:00

N06633 2009-11-24 04:00:00

N19500 2009-11-24 03:30:00

N00971 2009-11-24 03:00:00

N04628 2009-11-24 20:30:00

As Table 5 shows, the leak history dataset

contains two columns: the ‘End Node’ that is nodes

where the leak occurred, and the start time that is

when the leak occurred.

For leakages prediction, two different lag

windows were considered: a shorter prediction

window of 1 hour and a longer prediction window

with duration of 3 hours. We used the median as

aggregating function since compared to the mean, it

is more robust to the outliers.

4 RESULTS

In the present section we will present the results given

by the application of InterQuartileRangeAD

algorithm for leakages prediction.

In order to evaluate the performances of the

InterQuartileRangeAD algorithm in predicting

leakages, we use the start time information shown in

Table 5. In other words, we will verify if the

algorithm raises a warning before the true leak time.

Figure 4 and Figure 5 show the performances of

InterQuantileRangeAD algorithm with 1-hour

prediction window. In this case the algorithm was

able to predict the leak events for two nodes of the

network: N00971 and N02197. As shown in Figure 4,

for the node N00971 the algorithm generated two

warnings, at 02:00 pm and 02:30 pm, before the true

leak time that is at 03:00 pm.

The same for the node N02197 shown in Figure 5,

where the leakage event occurred at 11:00 pm and the

first warning was generated by the algorithm at 10:00

pm.

Figure 4: 1 Hour Prediction Node N00971.

Figure 5: 1 Hour Prediction Node N02197.

In Figure 6, we report the performances of the

algorithm with 3-hours prediction window. In this

case, the InterQuantileRangeAD algorithm generated

the first warning at midnight that is 3 hours before the

true leak time (2009-11-24 03:00:00).

Figure 6: 3 Hours Prediction Node N00971.

5 CONCLUSIONS

In this paper an approach for leaks detection and

prediction has been presented. The authors applied an

Detection and Prediction of Leakages in Water Distribution Networks

441

unsupervised approach for detecting leakages in the

Milan WDS. In our data leaks represent a minority

class: as in real-world cases, values representing

normal conditions of a water system are present in

large quantities with respect to the ones referring to

leakages, which make them an unrepresented class in

data.

Unbalancing of data given as input to the trained

algorithms for detection and prediction let to obtain a

high gap between values of accuracy with respect to

precision. To solve this problem in the present case

study the leak duration has been extended, analysing

leak events with a minimum duration of 8 hours. This

let us have a significative improvement in the

precision score of the trained algorithms. One future

step could be that of considering data in the night

period, normally defined as being between midnight

and 6 am. During night, flowrate is low while

pressure assumes maximum values. For this reason,

the minimum night flow is the most meaningful piece

of data as far as estimating night leakage is

concerned. Giving to the algorithm flowrate data in

addition to pressure data could be another possible

improvement.

Finally, for future works we plan to simulate more

data in order to include in the analysis weekly and

yearly seasonality in water consumption.

ACKNOWLEDGEMENTS

The research results presented in this paper have been

achieved inside the Water 4.0 project, named

“Technologies for the convergence between industry

4.0 and the integrated water cycle”. This research

project is currently running and is funded by the

Ministry of Enterprises and Made in Italy

(https://www.mimit.gov.it/en/).

REFERENCES

Adedeji K.B., Ponnle A. A., Abu-Mahfouz A.M., Kurien A.

M. (2022). Towards Digitalization of Water Supply

Systems for Sustainable Smart City Development—

Water 4.0. In Applied Sciences, vol. 12, issue 18.

MDPI, pp. 1-25.

Ameli M., Becker P.A., Lannkers K., Ackeren M., Bahring

H., Maab W. (2022). Explainable Unsupervised Multi-

Sensor Industrial Anomaly Detection and

Categorization. In 2022 21st IEEE International

Conference on Machine Learning and Applications

(ICMLA). IEEE Xplore, pp. 1468-1475.

ADTK, (2023) Anomaly Detection Toolkit 0.6.2

documentation. Available online: Anomaly Detection

Toolkit (ADTK) — ADTK 0.6.2 documentation

(Accessed on 28 March 2023).

Barros D., Almeida I., Zanfei A., Meirelles G., Luvizotto E.

Jr., Brentan B. (2023). An Investigation on the Effect of

Leakages on the Water Quality Parameters in

Distribution Networks. In Water, vol. 15, issue 324.

MDPI.

Belichovski M., Stavrov D., Donchevski S., Nadzinski G.

(2022). Unsupervised Machine Learning Approach for

Anomaly Detection in E-coating Plant. In 2022 IEEE

17th International Conference on Control &

Automation (ICCA), on June 27-30 2022 (Hybrid)

Naples, Italy. IEEE Xplore, pp. 992-997.

Bentoumi M., Chikouche D., Mezache A., Bakhti H.

(2017). Wavelet DT method for water leak-detection

using a vibration sensor: an experimental analysis. In

IET Signal Process. vol.11, issue 4. IET Journals - The

Institutions of Engineering and Technology, pp. 396-

405.

Caldognetto N., Evangelisti L.P., Poltronieri F., Russo M.,

Stefanelli C., Tenani S., Toboli S., Tortonesi M. (2022).

Water 4.0: enabling Smart Water and Environmental

Data Metering. In NOMS 2022-2022 IEEE/IFIP

Network Operations and Management Symposium.

IEEE.

Chan T. K., Chin C.S., Zhong X. (2018). Review of Current

Technologies and Proposed Intelligent Methodologies

for Water Distributed Network Leakage Detection, In

IEEE Transactions on Knowledge and Data

Engineering, vol.6. IEEEAccess, pp. 78846-78867.

Cheng Z., Zou C., Dong J. (2019). Outlier Detection using

Isolation Forest and Local Outlier Factor. In

Proceedings of International Conference on Research

in Adaptive and Convergent Systems, China, September

24–27 2019 (RACS ’19). ACM Digital Library, pp.

161-168

Cody A. R., Tolson B. A., Orchard J., Detecting Leaks in

Water Distribution Pipes Using a Deep Autoencoder

and Hydroacoustic Spectrograms. In Journal of

Computing in Civil Engineering, vol. 34. No. 5. ASCE.

Cody A. R., Dey P., Narasimhan S. (2020). Linear

Prediction for Leak Detection in Water Distribution

Networks. In Journal of Pipeline Systems Engineering

and Practice, vol. 1, issue 1. ASCE, pp. 1-16.

Colombo A. F., Lee P., Karney B.W. (2009). A selective

literature review of transient-based leak detection

methods. In Journal of Hydro-environment Research,

vol.2. ELSEVIER, pp. 212-227.

Escofet, M.A.C., Quevedo, J., Alippi, C., Roveri, M., Puig,

V., García, D., Trovò, F. (2016). Model- vs. data-based

approaches applied to fault diagnosis in potable water

supply networks. In Journal of Hydroinformatics,

vol.18, No.5. IWA PUBLISHING.

Fitrianto A., Wan Muhamad W.Z.A., Kriswan S., Susetyo

B. (2022). Comparing Outlier Detection Methods using

Boxplot Generalized Extreme. In Aceh International

Journal of Science and Technology, vol.11, issue 1.

Graduate School of Syiah Kuala University, pp-38-45.

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

442

Frey B.J., Dueck D. (2007). Clustering by Passing

Messages between Data Points. In Science, vol. 315,

issue 5814. JSTOR, pp. 972-976.

Gansukh C., Yoo K.H, Bazarbaev M., Nasridinov A.

(2021). Feasibility Study of Outlier Detection in Smart

Manufacturing Applications. In Advances in Intelligent

Information Hiding and Multimedia Signal Processing:

Proceeding of the 16th International Conference on

IIHMSP in conjunction with the 13th international

conference on FITAT, vol.2, November 5-7, 2020,

Vietnam. Springer, pp.283-290.

Gopali Saroj, Namin A.S. (2022). Deep Learning-Based

Time-Series Analysis for Detecting Anomalies in

Internet of Things. In Electronics, vol. 11, MDPI, pp.

1-16 .

Gutschi C., Furian N., Suschnigg J., Neubacher D,

Voessner S. (2018). Log-based predictive maintenance

in discrete parts manufacturing. In Procedia CIRP,

vol.79. ELSEVIER, pp. 528-533.

Hunaidi O., Chu W.T., (1999), Acoustical characteristics

of leak signals in plastic water distribution pipes. In

Applied Acoustics, vol. 58. ELSEVIER, pp. 235-254

Kammoun M., Kammoun A., Abid M. (2022). Experiments

based comparative evaluations of machine learning

techniques for leak detection in water distribution

systems. In Water Supply, vol.22, Issue 1. IWA

PUBLISHING, pp. 628–642.

Kammoun M., Kammoun A., Abid M.

(2022). LSTM-AE-WLDL: Unsupervised LSTM

Auto-Encoders for Leak Detection and Location

in Water Distribution Networks. In Water Resources

Management, vol.37, Issue 2. Springer, pp.731-746.

Klise K., A., Murray, R., Haxton, T. (2018). An Overview

of the Water Network Tool for Resilience (WNTR).

Leu S.S., Bui Q.N. (2016). Leak Prediction Model for

Water Distribution Networks Created Using a Bayesian

Network Learning Approach. In Water Resources

Management, vol. 30. Springer, pp.2719-2733.

Lijuan W., Hongwei Z., Zhiguang N. (2012). Leakage

Prediction Model Based on RBF Neural Network. In

Software Engineering and Knowledge Engineering:

Theory and Practice, vol 114. Springer, pp 451-458.

F.T., Ting K.M., Zhou Z.H. (2008). Isolation Forest. In

2008 Eighth IEEE International Conference on Data

Mining. IEEE Computer Society, pp. 413-422

Liu F.T., Ting K.M., Zhou Z.H. (2012). Isolation-Based

Anomaly Detection. In ACM Transactions on

Knowledge Discovery from Data (TKDD), vol.6, Issue

1, No. 3. ACM, pp.1-39.

Naiades Project. A holistic water ecosystem for digitization

of urban water sector. Available online: https://

www.naiades-project.eu/ (Accessed on 20 January

2023).

Otte T., Posada-Moreno A.F., Hubenthal F., Habler M.,

Bartels H., Abdelrazeq A., Hees F. (2022). Condition

Monitoring of Rail Infrastructure and Rolling Stock

using Acceleration Sensor Data of on-Rail Freight

Wagons. In Proceedings of the 11th International

Conference on Pattern Recognition Applications and

Methods (ICPRAM 2022). SCITEPRESS, pp.432-439.

Pearson, D. (2019). Standard Definition for Water Losses,

IWA Publishing, London.

Perez R., Sanz G., Puig V., Quevedo J., Escofet M..C.,

Nejjari F., Meseguer J., Cembrano G., Tur, J.M.M.,

Sarrate R. (2014). Leak Localization in Water

Networks A Model-Based Methodology Using

Pressure Sensors Applied to a Real Network in

Barcelona. In IEEE Control Systems Magazine, vol.34,

issue 4. Applications of Control, pp.24-36.

Philips S.J. (2002). Acceleration of K-Means and Related

Clustering Algorithms. In Algorithm Engineering and

Experiments, 4th International Workshop, ALENEX

2002, San Francisco, CA, USA, January 4-5, 2002,

Springer, pp.166-177.

Scikit-learn, Scikit-learn:machine learning in Python, 1.2.2.

Available online: scikit-learn: machine learning in

Python — scikit-learn 1.2.2 documentation (Accessed

on 3 April 2023).

Rosner B., (1983). Percentage Points for a Generalized

ESD Many Outlier Procedure. In Technometrics,

vol.25, No. 2. ASQ, pp.165-172.

Shukla S., Naganna S. (2014). A Review on K-Means Data

Clustering Approach. In International Journal of

Information & Computation Technology, vol.4, No.17,

Springer, pp.1847-1860

Tornyeviadzi H.S., Seidu R. (2023). Leakage detection in

water distribution networks via 1D CNN deep

autoencoder for multivariate SCADA data. In

Engineering Applications of Artificial Intelligence,

vol.122. ELSEVIER.

Vrachimis G.S., Kyriakou M.S., Eliades D.G.,Polycarpou

M. M., (2018). LeakDB: A benchmark dataset for

leakage diagnosis in water distribution networks. In 1

International WDSA / CCWI 2018 Joint Conference.

Wan X., Kuhanestani P.K., Farmani R., Keedwell E.

(2022). Literature Review of Data Analytics for Leak

Detection in Water Distribution Networks: A Focus on

Pressure and Flow Smart Sensors. In Journal of Water

Resources Planning and Management, vol.148, issue

10, ASCE.

Wang L. Liu Y., Yin H., Sun W. (2022). Fault diagnosis

and predictive maintenance for hydraulic system based

on digital twin model. In AIP Advances, vol. 12. AIP

Publishing.

Wang W., Sun H., Guo J., Lao L., Wu S., Zhang J. (2021).

Experimental study on water pipeline leak using In-

Pipe acoustic signal analysis and artificial neural

network prediction. In Measurement, vol.186.

ELSEVIER.

Xu D., Tian Y. (2015). A Comprehensive Survey of

Clustering Algorithms. In Annals of Data Science,

vol.2, No. 2. Springer, pp.165-193.

Yu T., Chen X., Yan W., Xu Z., Ye M. (2023). Leak

detection in water distribution systems by classifying

vibration signals. In Mechanical Systems and Signal

Processing, vol. 185. ELSEVIER.

Detection and Prediction of Leakages in Water Distribution Networks

443