Available online at www.sciencedirect.com
ScienceDirect
IFAC PapersOnLine 51-24 (2018) 521–528
SysDetLok –
a leakage detection and localization system
for water distribution networks ⋆
W. Moczulski ∗ J. Karwot ∗∗ R. Wyczółkowski ∗∗∗ D. Wachla ∗
K. Ciupke ∗ P. Przystałka ∗ D. Pająk ∗
Institute of Fundamentals of Machinery Design, Silesian University
of Technology, 18a Konarskiego Str., Gliwice, Poland (e-mail:
wojciech.moczulski@polsl.pl)
∗∗
The Water and Sewage Limited Liability Company, 62 Pod Lasem
Str., Rybnik, Poland (e-mail: pwik@pwik-rybnik.pl)
∗∗∗
Institute of Production Engineering, Silesian University of
Technology, 26 Roosevelta Str., Zabrze, Poland (e-mail:
ryszard.wyczolkowski@polsl.pl)
∗
Abstract
The paper is focused on the leakage detection and localization system for water distribution
networks (WDNs). The architecture of the system is briefly presented and its most important
modules are discussed in detail. In the second part of the paper, a case study concerning
applications of the system to a real WDN located in one of the southern cities in Poland is
presented. A pilot version of the system has been running since 2013 in this city, which still
shows a high level of performance in leak detection, as well as reasonable efficiency for leak
location. The paper ends by emphasizing the merits and limitations of this solution from the
water and sewage company’s (WSC’s) perspective.
© 2018, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Keywords: water distribution systems, leakage detection and localization, fault diagnosis
systems, computational intelligence methods.
1. INTRODUCTION
One of the most vital municipal services is providing water.
Water supply systems must meet the demand resulting
from public, commercial and industrial activities. When
operational, such systems are exposed to uncontrolled
leakages of differing intensities and locations. Therefore,
leak detection is an increasingly important issue in practice. Leakages can be caused by natural processes of wear
and tear, the corrosion of inner and outer surfaces of
pipes, mechanical damage to pipes caused by excessive
loads, assembling errors, seasonal temperature changes,
movements of the subsoil and material defects in the pipes
(Puust et al. (2010)). Water loss depends on the technical
standards of the pipeline and can equate to 3-7% of total
water consumption in the case of developed countries and
up to 50% of consumption in developing countries (Puust
et al. (2010)).
Leakages not only waste water and energy, but can also
affect water quality. Therefore, there is a constant need to
develop methodologies for detecting and localizing leakages occurring in water supply networks.
⋆ The research described in this paper has been partially carried
out as part of the research project conducted within the framework
of the National Programme of Innovation Economy, financed by the
European Social Fund. This publication was financed from statutory
funds of the Institute of Fundamentals of Machinery Design at the
Silesian University of Technology.
Recently, several research studies on the detection and
localization of leakages on WDNs have been published.
Furthermore, many detection and location methods have
found practical implementation (Puust et al. (2010);
Geiger et al. (2003)): analysis of flow changes during
night-time; step testing; transient (Ferrante et al. (2007);
Vı́tkovskỳ et al. (2007)) and frequency (Mpesha et al.
(2002)) analysis of pressure waves; correlation analysis of
leakage noise; and inspection with a georadar. Unfortunately, these methods have numerous limitations: the need
to carry out experiments on the object; the requirement
for the diagnoser to have a significantly high level of
experience; the impossible automation of inspections; and
the need to apply specialized measuring equipment.
Many works have been published on leak detection and
isolation (localization) methods for WDNs (see e.g. Wu
and Farley (2011); Puust et al. (2010); Kim et al. (2015);
Meseguer et al. (2015)). Some of the methods were proposed and developed by the authors (see e.g. Moczulski
et al. (2016a); Wyczółkowski (2013); Ciupke (2016); Przystałka and Wyczółkowski (2011); Przystałka and Moczulski
(2012); Wachla et al. (2015)), as well as patented (Moczulski et al. (2016b)) and deployed by industry. The essence
of the idea is based on the discretization of the water
supply system to predefined areas, then identifying an
approximate location of where a leakage might occur.
2405-8963 © 2018, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Peer review under responsibility of International Federation of Automatic Control.
10.1016/j.ifacol.2018.09.626
522
W. Moczulski et al. / IFAC PapersOnLine 51-24 (2018) 521–528
The presented paper describes a real leakage detection and
localization system, developed by the authors. The system
is implemented in a WSC in Poland, which supplies water
to approximately 150,000 inhabitants, as well as companies and businesses, and is based on the proposed above
mentioned methodology concerning the discretization of
the water supply system to predeĄned areas.
2. ARCHITECTURE OF A LEAKAGE
DIAGNOSIS SYSTEM
The SysDetLok system is the result of work carried out by
a group of researchers interested in the various methods
involved in the operation of this system. The creators of
the individual methods were also responsible for building
and testing the respective part of the system that implements each method. In addition, because the system
was created as a prototype, while the short time spent on
its implementation was partly offset by research work on
methods and software development work, it was decided to
adopt the modular and open architecture of the SysDetLok
system. The system consists of independently functioning
modules communicating with each other via messages that
are placed on and read from the system’s blackboard,
which is implemented in PostgreSQL (Fig. 1).
The synchronization module is responsible for the whole
operation of the program and cyclical contact with the
measurement database, as well as generating a message
about the appearance of the new measurement in the
database and the registration of individual modules and
mediating the transfer of data from/to the blackboard.
Thanks to this assumption, it was possible to easily
exchange individual modules for new versions, along with
making changes in the message system and the structure
of the blackboard without any changes to the individual
modules. Descriptions of the other modules are provided
in the next part of this paper.
Figure 1. Modules of the SysDetLok system
not). In addition, if a leak is detected, information about
its approximated size is also recorded. This information
is taken over by the leakage location module, which, from
among the predefined zones by which the water supply
network has previously been divided, indicates the likely
leakage zones. The way of dividing the network into predefined zones is described by Wyczółkowski (2013).
Leakage location is the basis for the operation of the
M WIZ module, which prepares data for visualization
through external SCADA and GIS systems. As there was a
need to erase the system-generated alarms by the network
operator, and SysDetLok had no communication interface,
it was necessary to develop an additional interface to
perform this operation. This was prepared as a webpage
by which the operator cleared the alarm, while the corresponding module (M DEL WIZ) cleared the data for
visualization in the GIS system and generated a signal,
which the alarm ceased in the SCADA system.
In order not to burden the network operator with the use
of additional tools, the built-in system did not have its
own interface with which to indicate leakage locations; the
alarms and visualization of the results were used by the
company’s SCADA and GIS systems. Upon detection by
the leak system, the corresponding information (warning
or alarm) was generated and added to the SCADA alarm
list. This resulted in the signalling of an event on the screen
of this system. An additional benefit of this architecture is
the solution to the data security problem. Authentication
of system users, providing different levels of access to data
is solved by the mechanisms implemented in the systems
already in use in the company and preserves the already
implemented level of rights for groups of users already
defined.
The operation algorithm for the SysDetLok system is
presented in Fig. 2. The program works in a closed loop.
At any given time, the synchronization module checks for
new data to be processed and generates a corresponding
message. The appearance of new data triggers two leak
detection modules, working independently of each other.
Their results are collected by the M KLAS DET module,
which fuses the information and generates one final message about the degree of certainty regarding leak detection
and the diagnostic signal (i.e., whether a leak is detected or
Figure 2. Diagram of the operation algorithm of the
SysDetLok system
2.1 M DET IOM - leakage detection based on in/out
models
An information redundancy technique, corresponding to
a model-based fault detection scheme, is proposed for
W. Moczulski et al. / IFAC PapersOnLine 51-24 (2018) 521–528
the first leakage detection module. In this module, neural
network-based non-linear autoregressive exogenous models
are employed to map the relation between water flow rates,
whereas classic parametric models are utilized to estimate
uncertainties of neural models of water flows. Both types
of models are then used for generating residuals and computing adaptive thresholds required for early and reliable
leakage detection. Input-output neural models are created
to represent dynamic behaviours of water flows at different
locations in the WDN. It is assumed that flowmeters are
installed at particular nodes of the network by means of
the optimal placement method (Przystałka and Moczulski
(2012)). A non-linear part of the models is obtained using
a multilayer neural network with one hidden layer of neurons, for which the hyperbolic tangent activation functions
are used. Neural models are trained using the LevenbergMarquardt algorithm and the simulation data related to
a faultless state of the system. On the other hand, linear
autoregressive models are applied for the approximation
of adaptive thresholds. These models are developed using
the least-squares technique. In this module, the value of
the residuum is compared with adaptive threshold values
representing the acceptable levels of leakage in the WDN.
If the residuum is greater than the threshold, a diagnostic
signal equal to 1 is returned by the module (in the other
case, the diagnostic signal is equal to 0). It should be
stressed here that a belief factor is computed for the diagnostic signal. In terms of the outcome, the diagnostic signal
and corresponding belief factor are sent to the blackboard
of the SysDetLok system.
2.2 M DET AR SVM - leakage detection based on SVM
models
The M DET AR SVM module was developed as the second tool in the SysDetLok system for supporting leakage
detection tasks. Its operation is based on analytical redundancy, where the observed process variables and the analytical models of these variables are used for fault detection
by means of analysis of the residuum signal. In the proposed solution, the support vector machine (SVM) method
(see Vapnik (1995) and Schölkopf and Smola (2001)) was
used to define adequate analytical models. Furthermore, it
was assumed that: (a) the monitored WDN has only one
inflow (in this case, any leakage in the WDN causes an
increase in water flow at the WDN inflow where the level
of water flow is equal to the level of the leakage); (b) the
models are trained to represent the water flow at the main
inflow to the WDN under normal conditions; and (c) the
models have an autoregressive structure.
The leakage detection procedure realized by the M DET AR SVM module consists of a number of steps. Firstly,
an adequate data set from the database of the SysDetLok
system is gathered. The features of this data set (e.g., a
process variable specification, size of data) are defined by
the structure of the models and by the size of the window
for calculating the simple moving average of the residuum
signal. At the second stage, the residuum signal is computed and the simple moving average is evaluated. Next,
the estimated value of the residuum moving average is
compared with a threshold value representing the acceptable level of leakage in the WDN. If the residuum moving
average is greater than the threshold, the probability of
523
difference between the average of the residuum signal is
equal to 0 (assumed for normal conditions of the WDN)
and the actual average is computed by means of the tStudent statistic (Rice (2006)). Finally, the obtained diagnosis is sent to the blackboard of the SysDetLok system.
2.3 M KLAS DET - fusion of leakage detection results
Having more than one classifier means there is a problem
in combining their results. The combining process is also
called classifier fusion. Fusing classifiers’ decisions can
improve the performance of a leakage detection system.
Many combination schemes have been proposed in the
literature according to the type of information provided
by each classifier, as well as their training and adaptation
abilities. Some fusion methods operate on the classifiers,
while others operate on their outputs (Ruta and Gabrys
(2000)). We were able to distinguish three types of classifier (Bezdek et al. (2006)):
• crisp classifier,
• fuzzy classifier,
• probabilistic classifier.
According to these types, different fusion methods are
applied. In our case, two classifiers are used for the detection problem. The outputs are described with the degree
of certainty. After an investigation process, it transpired
that, in some cases, classification based on in/out models
works better, while, in others, classification based on SVM
models is more sensitive. Finally, then, the fusion of results
is implemented in a simple way, which can be described
as follows: if any of the classifiers detects a leakage, the
alarm signal is indicated. The certainty degree of the signal
is calculated as the maximum value of both calculated
certainty degrees.
2.4 M LOK - leakage location based on hydraulic models
and soft computing
The following assumptions were made for the construction
of the leak location module:
• the leakage will be localized to the predefined area
• the division of the entire network into areas will
be made by analysing the sensitivity of individual
sensors to the appearance of leakage at a particular
location
These assumptions were verified experimentally. During
the experiment, at chosen locations, leakage at the specified intensity of 5m2 /h was simulated using a fire hydrant.
The expected changes in flow, calculated with the district
metering area (DMA) hydraulics model, were compared
with changes detected in the measured flow signal.
Fig. 3 shows DMA 5 with a partitioned view and leakage
simulation locations. Table 1 shows the values of expected
flow changes, while Fig. 4 shows the measured flow rate
obtained from the st1 flowmeter during the experiment
(red line), with flow prediction (blue line) and moments
marked when changes were expected.
As the experiment confirmed the assumptions, a leakage
localization module was prepared.
524
W. Moczulski et al. / IFAC PapersOnLine 51-24 (2018) 521–528
produces a similarity between the real condition and the
virtual leakage in the following way:
n
1
△pj = [(P rij − P mij )2 ],
n i=1
pj =
1
,
1 + △pj
(1)
(2)
where: △p is theaverage distance of the data model and
the actual data, p is the value of the similarity between
the model and actual flows, P r is the flow rate measured
by the actual flowmeter, P m is the flow rate predicted by
the virtual flowmeter, i is an index of the flowmeter, j is an
index of the area, and n is the total number of flowmeters.
Figure 3. A DMA layout with measured points
Table 1. Flow rate changes [m3 /h] expected
during the experiment
No.
predef. area no.
st1
st2
st6
st7
st11
st12
1
2
3
4
5
44
21
37
50
9
0
1.5
1.5
0.5
0
0
-1
-1
-1
0
2.5
0
0
0
0
0
0
0
2.5
0
0
-1.5
-1.5
1
0
0
0
5
0
0
.
Figure 4. A sample flow rate measured during the experiment
The first method is based on the flow data obtained from
simulating the hydraulic model for virtual leaks localized
in particular areas of the district (Tomasik (2012)). For
each leakage, virtual flow values are calculated corresponding to the nodes where flowmeters are physically located
in the network. The locations of virtual measurement
nodes are also determined, based on the method proposed
by Przystałka and Moczulski (2012). The size of a leakage
is determined with the use of the hydraulic model. This is
done by comparing the virtual flow values calculated for
the input of the zone (it is calculated using the model for
normal conditions), with the nominal values measured by
the real flowmeter, which is installed on the supply pipe of
the zone. In the next step, a comparison of the values of the
virtual flow with the values of the real flow measurements
Areas with the higher value of similarity (2) are potentially
considered to be the ones where the leakage could occur.
The proposed method assumes that the leak is simulated
at one node in the area. Thus, the number of emitters is
equal to the number of control areas. If the areas are well
isolated (they are highly distinguishable, even for small
leaks), this assumption is correct because all places in the
area are equally representative, which, in the same way,
reveals the appearance of leaks anywhere in the area where
the given flowmeter is installed. This method allows for
very fast computations. This approach makes possible the
location of individual leaks in the water supply network.
The second solution developed for leakage location and implemented in the M LOK module was described by Wachla
et al. (2015). In this solution, the process of leakage location is divided into a set of simple tasks of leakage
detection, where every single leakage detection task is
conducted for a specified subarea of the considered WDN
by classifying different signals between actual water flows
at selected nodes of the WDN and water flows estimated
using adequate analytical models. For this purpose, the
procedure of leakage detection is composed of two steps.
The first step involves the generation of residual signals
using a SVM for the approximation of water flow, while
the second involves the leakage location using neuro-fuzzy
classifiers with residual signals as inputs. In particular,
each classifier for leakage detection is associated with only
one subarea of the WDN where the partition of the WDN
into subareas is defined a priori by domain experts.
The structure of the classifiers for leakage location is
hierarchical and consists of two elements: the adaptive
network-based fuzzy inference system ANFIS (MATLAB
(2009)) and a simple binary classifier for comparing the
output of ANFIS with a threshold value. The neurofuzzy system application, whose purpose is to evaluate
the residuum signals according to the level of disturbances
occurring in the WDN, is used to generate signals that
are associated with two states: an occurrence of leakage
in the associated subarea of the WDN (value 1) and a
faultless state in this subarea (value 0). The output signal
of the neuro-fuzzy system is a signal whose values vary
continuously; therefore, in the structure of the proposed
classifiers for leakage location, an additional binary classifier is applied in order to obtain an output with the
adequate decision value.
W. Moczulski et al. / IFAC PapersOnLine 51-24 (2018) 521–528
525
2.5 M WIZ - data preprocessing for visualization purposes
When the system generates information about the detection and identification of the area where the leakage
may be located, the system passes the identifiers of the
relevant network segments to the three GIS tables, which
were created previously and correspond to the successive
fault locations (with the qualitative belief factor: very
likely, likely, unlikely location). This was possible because
each predefined differentiated spill area in the system
database was assigned identifiers representing the network
segments. The system operator, who is notified about the
leak, must go to the GIS system and trigger a view.
The screens of both systems showing example leakages
are shown in Fig. 5. After resolving the problem, the
operator must clear the alarm, which is correlated with
the automatic clearing of the corresponding tables in the
GIS system.
Figure 6. Layout of DMAs with measured points
a school. For these customers, their specific water consumption patterns were determined. In special cases, when
the consumption was particularly large and irregular, the
measurement was carried out in the online mode.
3.2 Examples of leakage detection
Figure 5. The screens of the SCADA and GIS systems with
leakage signalization
3. CASE STUDIES
The SysDetLok system has been developed to work with
big DMAs, that is, parts of an entire water distribution
system, separated by measuring devices, for which the
amount of water can be measured.
3.1 Description of diagnosed object
The system has been implemented and tested for different
DMAs, which supply water to various districts of the city.
Basic information about these DMAs is summarized in
Table 2. The layout of the DMAs is shown in Fig. 6. In
this figure, the measurement points are marked as well.
Table 2. Basic information on diagnosed DMAs
DMA ID
No. of residents
No. of consumers
No. of flow meters
2
∼20,000
∼3,000
18+3
5
∼5,000
∼1,300
11+1
14
∼5,000
∼900
11+1
In addition, in the middle of Zone 2, there is a pump
station equipped with water tanks, which raises the level of
water in order to supply higher-level buildings comprising
multi-family housing. The estate also has an additional
power supply, which is powered by water on two sides.
The filling of tanks at the pump station is carried out at
night, which greatly increases the flow of water throughout
the zone.
The remaining zones, which have a one-sided power supply, mainly comprise single-family houses. In each of the
zones, there are industrial and non-typical customers, who
use more water, such as a concrete plant, a bakery and
The model-based fault detection procedure of the leakage
detection module (M DET IOM) was created as follows.
For DMAs described above, input-output neural models
were created to represent dynamic behaviours of water
flows at different locations. As mentioned in the previous section, the flowmeters were installed at particular
nodes of the network by means of the optimal placement
method. Neural network-based, non-linear autoregressive
exogenous models with one hidden layer (including neurons with the hyperbolic tangent activation function) were
created. The mean absolute percentage error (MAPE) and
the normalized root-mean-square error (NRMSE) were
used to estimate the accuracy of each neural model. Moreover, the Bayesian information criterion (BIC) was used
for the selection of neural model structures. It should be
noted that the MAPE was less than 5% for each neural
model. On the other hand, linear autoregressive models
were successfully applied for the approximation of adaptive
thresholds. These models were developed using the leastsquares technique (the mean-square error was acceptable
for each model). The final results of this stage are presented in Table 3. Diagnostic signals were generated using
residual signals, while adaptive thresholds were obtained
by means of neural and linear autoregressive models. The
parameters of the decision block (which is needed for the
evaluation of residuals) were calculated using the threesigma rule.
Next, the performance of the leak detection scheme was
evaluated by applying false and true detection rates.
Validation tests were carried out for 25 trials with a
random leakage occurring at one point of the DMA. The
average false detection rate was less than 1%, while the
average true detection rate was greater than 97%.
The implementation of the module was realized with the
use of the prepared models and by taking into account the
outcomes of the verification tests. The performance results
of the M DET IOM module are presented in Figs. 7-9. The
figure legend is as follows: the blue dashed line is used to
visualize the flow measurements, the red line denotes the
diagnostic signals and the solid yellow line is used to show
the belief factor. As one can observe, a validation stage was
526
W. Moczulski et al. / IFAC PapersOnLine 51-24 (2018) 521–528
Table 3. The accuracy evaluation of in/out
models and error models
Figure 9. Leak detection results for DMA 14
carried out for almost three months. During this period
of time, several leakages occurred. These leakages were
detected by the module and hence the diagnostic signals,
as well as values of the belief factor, which were sent to
the blackboard of the SysDetLok system.
Figure 7. Leak detection results for DMA 2
software (Rossman (2000)). This model was used to simulate operation of the DMA under faultless conditions and
with instances of leakage uniformly distributed over the
whole DMA. In particular, it was assumed that the number
of leakage emissions at every single subarea would be equal
to four. The first four emissions were the source of the
training data, while the fourth emission was the source
of the testing data. Due to the dynamic properties of the
DMA and the requirements of the approximators of the
water flow, one assumed that a simulation of the DMA
operation would be four weeks with a time interval ∆t =
15 minutes. The response of the network was observed at
nodes related to the locations of flowmeters installed at the
real DMA. Furthermore, the DMA was partitioned into 23
subareas based on expert knowledge.
The next step was to identify the approximators of the
water flow at selected nodes of the DMA. In particular, a
regression formulation of the SVM was applied (Schölkopf
and Smola (2001)). The SVM models were trained for
faultless conditions of the DMA using adequate simulation
data. The linear structure of the identified SVM approximators with the following output and inputs was assumed:
Qid (k) = f (Qid (k − ∆t), Qid (k − 664∆t)) ,
(3)
where: Qid represents the water flow rate at a measurement point identified by node id in the hydraulic model
of the DMA, ∆t is a sampling interval (15 min), n∆t is
the time delay and k is the time point id. Twelve SVM
approximators of the water flow at 12 specified nodes of
the DMA were identified.
Figure 8. Leak detection results for DMA 5
3.3 Examples of leakage location
The application of the soft-computing approach for leakage
location was presented using DMA 14. Both the identification stage of the SVM approximators of the water
flow through the selected nodes of the DMA and the
process of training and testing the neuro-fuzzy classifiers
for leakage location were described in detail. The data for
training and testing water flow approximators and leakage
classifiers were obtained using a hydraulic model of the
considered DMA, which was developed using EPANET
Based on the identified SVM models and simulated data,
including instances of leakage in each subarea of the DMA,
the residual signals were calculated. They were used to
form learning examples for training and testing neurofuzzy classifiers. In particular, the learning examples were
created as follows. From each residuum, signal frames with
96 samples (one day) were extracted. Next, the extracted
frames were connected together to form input signals for
the neuro-fuzzy classifiers. Given the imbalance between
the frames with regard to the considered leakages and
the other frames, the first were repeated several times.
A detailed description of the preparation of such learning
examples was provided by Wachla et al. (2015).
In the last stage, the process of training and testing classifiers for leakage location was initiated by establishing the
initial structure and parameter values of the neuro-fuzzy
models. This was realized using the subtractive clustering method (Chiu (1994)). In particular, a squash factor
W. Moczulski et al. / IFAC PapersOnLine 51-24 (2018) 521–528
Table 4. The confusion matrix of classifiers for
leakage location in subareas of the WDN no.14
(part 1)
Cl.
ID
A1
A2
A3
A4
C1
1.0 0.0 0.99 0.01
C2
0.0 1.0 0.0 0.0
C3
1.0 0.0 1.0 0.22
C4 0.13 0.0 0.38 0.99
C5
0.0 0.0 0.0 0.0
C6 0.67 0.69 0.83 0.83
C7
0.0 0.0 0.0 0.0
C8
0.0 0.0 0.0 0.0
C9
0.0 0.0 0.0 0.0
C10 0.0 0.0 0.0 0.0
C11 0.0 0.0 0.0 0.0
Subarea ID
A5 A6 A7
0.0
0.0
0.0
0.0
1.0
0.0
0.0
0.0
0.0
0.0
0.0
0.88
0.0
0.93
0.15
0.0
0.97
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
1.0
0.0
0.0
0.0
0.0
A8
A9 A10
0.0
0.0
0.0
0.0
0.0
0.0
0.0
1.0
0.0
1.0
1.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
1.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
1.0
0.0
1.0
1.0
Table 5. The confusion matrix of classifiers for
leakage location in subareas of the WDN no.14
(part 2)
Cl.
ID
Subarea ID
A11 A12 A13 A14 A15 A16 A17 A18 A19 A20 A21 A22 A23
C11
C12
C13
C14
C15
C16
C17
C18
C19
C20
C21
C22
C23
0.81
0.93
0.16
0.0
0.0
0.3
0.0
0.13
0.26
0.15
0.0
0.96
0.0
0.91
1.0
0.2
0.0
0.0
0.4
0.0
0.11
0.21
0.02
0.0
1.0
0.0
0.94 0.0
0.0
0.0
0.0
0.0 0.25 0.8
0.96 0.0
0.0
0.0
0.0
0.0 0.16 0.09
1.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0 0,99 0.0
0.0
0.0
0.0
0.0
0.0
0.0
1.0
1.0
0.0
0.0
0.0
0.0
0.0
0.02 0.0
0.0 0.94 0.98 1.0 0.67 0.0
0.0
0.0
0.0 0.64 1.0 0.59 0.0
0.0
0.01 0.04 0.03 1.0
1.0
1.0 0.83 0.0
0.0 0.03 0.03 0.0
0.0
0.0 0.99 0.88
0.0
0.0
0.0
0.0
0.0
0.0 0.13 1.0
0.0
0.0
0.0
0.0 0.02 0.0
0.0
0.0
0.08 0.0
0.0
0.0
0.0
0.0 0.95 0.0
0.0
0.0
0.0 0.17 0.96 0.13 0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
1.0
0.0
1.0
0.89 0.0
0.99 0.0
0.13 0.0
0.0
0.0
0.0
0.0
0.48 0.97
0.0
1.0
0.15 1.0
0.39 0.0
0.01 0.0
0.0 0.03
1.0
0.0
0.0 0.99
equal to 1.25, an accept ratio equal to 0.5, a reject ratio
equal to 0.15 and Gaussian membership functions on each
input were used. Finally, the structure of the neuro-fuzzy
inference system was generated. It consisted of three rules,
three membership functions on every single input and
three linear membership functions on the outputs. Subsequently, the ANFIS (Jang (1993)) algorithm was used to
train the neuro-fuzzy models. This algorithm (MATLAB
(2009)) transforms a fuzzy model into a neural network
and calculates the network’s weights using the backpropagation method and the gradient descent (Jang (1993))
method.
Twenty-three classifiers were obtained, with their leakage
detection ability in every single subarea of the WDN
presented in the form of a confusion matrix (Tab. 4 and
Tab. 5). The values in the confusion matrix vary from
0 to 1 and represent the leakage detection probability.
The confusion matrix clearly shows that every classifier
properly detects leakage occurring in the related subarea
(the main diagonal of the confusion matrix). On the other
hand, some are also more sensitive to leak occurrence in
other subareas than their nominal one. For example, it is
not possible to identify the right subareas A1, A3 and A6
with regard to leakage by means of classifier C1. For this
reason, other classifiers must be considered.
4. DISCUSSION AND CONCLUSION
In this paper, the complex methodology of designing and
tuning a system for leakage detection and location in
water supply systems has been presented. This methodology has been implemented in a WSC in Poland, which
527
supplies water to approximately 150,000 inhabitants, as
well as companies and businesses. One of the prominent
parts of this work has been the SysDetLok system, which
takes advantage of data acquired from GIS databases,
the measurement database and the maintenance and servicing database. The hardware part of the system has
been implemented in three separated zones of the water
delivery system. Each zone has a different character, with
one covered by individual houses and another dominated
by multi-family buildings with many floors, which require
additional technical systems to guarantee the necessary
water pressure on the highest floors. Research and development work carried out within the last decade has
allowed the WSC to draw conclusions from the models and
simulations developed in the framework of related research
and implementation projects.
The investments carried out in last five years have also
allowed the WSC to modernize the water supply network
over a range of about 100 km and develop systems to manage the water supply process for residents of the city. In addition to the standard replacement investments, the WSC
has invested in a variety of measuring devices, sensors,
servers and a teleinformatic infrastructure, which facilitate
data acquisition and collection, as well as their further
interpretation and subsequent decision-making with the
highest probability of success. This activity has reduced
the number of failures from 490 in 2012 to 190 in 2016 and
reduced water losses from 8.52% in 2012 to 4.10% in 2016.
The water supply system is constantly being rebuilt and
the currently deployed intelligent integrated water management system, in a configuration elaborated at the end
of the projects, needs to be reconfigured while awaiting a
new calibration formula that includes all network rebuilds
and additionally installed measurement and data transmission equipment. The analysis of collected historical and
topical data will allow the system to perform self-learning
and self-adjustment in order to take constantly changing
flow directions into account, along with the ongoing process of pressure optimization.
The undisputed advantage of implementing this system,
together with the earlier development, and testing the
water supply network models in selected locations is the
change in the organizational culture of the company, which
continues to run continuously, allowing the relevant data
gathering services to develop alongside it.
To sum up, the amount of data collected through the exploitation of this system has allowed the company to draw
optimal conclusions and make decisions that anticipate the
emergence of any critical situation, which it is strongly
advised to avoid in order not to cause serious disruption to
the water supply process in the area of operation of the enterprise. The experience we have gained has enabled us to
continually develop the system and keep it operational at a
technical level in order to meet the needs of the enterprise,
for the benefit of the local community, in a reliable way by
implementing the sustainability formula and, in particular,
the smooth operation of the water supply system, which
is currently expected from companies of this kind.
Further work will be focused on reducing implementation
costs and developing algorithms for self-learning and selftuning models that are used for the detection of leakages.
528
W. Moczulski et al. / IFAC PapersOnLine 51-24 (2018) 521–528
Furthermore, more accurate methods for locating leaks
will be sought. To improve data acquisition, new ideas
related to the Internet of Things will be implemented,
especially in the hardware part of the system. Finally,
respective methods of big data analysis and processing will
be developed and implemented.
REFERENCES
Bezdek, J.C., Keller, J., Krisnapuram, R., and Pal, N.
(2006). Fuzzy models and algorithms for pattern recognition and image processing, volume 4. Springer Science
& Business Media.
Chiu, S. (1994). Fuzzy model identification based on cluster estimation. Journal of Intelligent & Fuzzy Systems,
2(3).
Ciupke, K. (2016). Leak detection using regression trees.
In International Congress on Technical Diagnostic, 311–
321. Springer.
Ferrante, M., Brunone, B., and Meniconi, S. (2007).
Wavelets for the analysis of transient pressure signals
for leak detection. Journal of hydraulic engineering,
133(11), 1274–1282.
Geiger, G., Werner, T., Matko, D., et al. (2003). Leak detection and locating-a survey. In PSIG Annual Meeting.
Pipeline Simulation Interest Group.
Jang, J. (1993). Anfis: Adaptive-network-based fuzzy
inference systems. IEEE Transactions on Systems, Man,
and Cybernetics, 23(3), 665–685.
Kim, Y., Lee, S.J., Park, T., Lee, G., Suh, J.C., and
Lee, J.M. (2015). Robust leakage detection and interval estimation of location in water distribution network. IFAC-PapersOnLine, 48(8), 1264 – 1269. doi:
https://doi.org/10.1016/j.ifacol.2015.09.142. 9th IFAC
Symposium on Advanced Control of Chemical Processes
ADCHEM 2015.
MATLAB (2009). Fuzzy Logic Toolbox User’s Guide. The
MathWorks, Inc.
Meseguer, J., Mirats-Tur, J.M., Cembrano, G., and Puig,
V. (2015). Model-based monitoring techniques for leakage localization in distribution water networks. Procedia
Engineering, 119(Supplement C), 1399 – 1408. doi:
https://doi.org/10.1016/j.proeng.2015.08.1000. Computing and Control for the Water Industry (CCWI2015)
Sharing the best practice in water management.
Moczulski, W., Wyczółkowski, R., Ciupke, K., Przystałka,
P., Tomasik, P., and Wachla, D. (2016a). A methodology
of leakage detection and location in water distribution
networks - the case study. In 2016 3rd Conference on
Control and Fault-Tolerant Systems (SysTol), 331–336.
doi:10.1109/SYSTOL.2016.7739772.
Moczulski, W., Karwot, J., Wyczółkowski, R., Ciupke, K.,
Przystałka, P., Tomasik, P., and Wachla, D. (2016b).
Detection and location of faults in closed water supply systems using devices measuring water flow in the
network. Silesian University of Technology, Water and
Sewage Limited Liability Company, patent pl no. 224
049 edition.
Mpesha, W., Hanif Chaudhry, M., and Gassman, S.L.
(2002). Leak detection in pipes by frequency response
method using a step excitation. Journal of hydraulic
research, 40(1), 55–62.
Przystałka, P. and Moczulski, W. (2012). Optimal placement of sensors and actuators for leakage detection and
localization. IFAC Proceedings Volumes, 45(20), 666–
671.
Przystałka, P. and Wyczółkowski, R. (2011). Detection
of small leaks in water supply systems using the uncertainty modelling method (in polish). Measurements
Automation Control, 57, 1307–1310.
Puust, R., Kapelan, Z., Savic, D., and Koppel, T. (2010).
A review of methods for leakage management in pipe
networks. Urban Water Journal, 7(1), 25–45.
Rice, J.A. (2006). Mathematical Statistics and Data
Analysis. Duxbury Press.
Rossman, L.A. (2000). EPANET 2 Users Manual.
Ruta, D. and Gabrys, B. (2000). An overview of classifier
fusion methods. Computing and Information systems,
7(1), 1–10.
Schölkopf, B. and Smola, A.J. (2001). Learning with
Kernels: Support Vector Machines, Regularization, Optimization, and Beyond. MIT Press, Cambridge, MA,
USA.
Tomasik, P. (2012). Hydraulic model-based method for
leak location in urban water distribution networks (in
polish). Diagnostyka, (1), 45–50.
Vapnik, V. (1995). The Nature of Statistical Learning
Theory. Springer-Verlag, New York.
Vı́tkovskỳ, J.P., Lambert, M.F., Simpson, A.R., and
Liggett, J.A. (2007). Experimental observation and
analysis of inverse transients for pipeline leak detection.
Journal of Water Resources Planning and Management,
133(6), 519–530.
Wachla, D., Przystałka, P., and Moczulski, W. (2015).
A method of leakage location in water distribution networks using artificial neuro-fuzzy system.
IFAC-PapersOnLine, 48(21), 1216 – 1223.
doi:
https://doi.org/10.1016/j.ifacol.2015.09.692. 9th IFAC
Symposium on Fault Detection, Supervision andSafety
for Technical Processes SAFEPROCESS 2015.
Wu, Z.Y. and Farley, M. (2011). Water Loss Reduction.
Bentley Systems.
Wyczółkowski, R. (2013). A methodology of a detection
and localization of the water supply networks failures
with using approximate models (in Polish). 493. Publishing House of Silesian University of Technology, Gliwice.