Urban Water Journal

ISSN: 1573-062X (Print) 1744-9006 (Online) Journal homepage: https://www.tandfonline.com/loi/nurw20

Pipe network leak detection: comparison between

statistical and machine learning techniques

J. C. van der Walt, P. S. Heyns & D. N. Wilke

To cite this article: J. C. van der Walt, P. S. Heyns & D. N. Wilke (2018) Pipe network leak
detection: comparison between statistical and machine learning techniques, Urban Water Journal,
15:10, 953-960, DOI: 10.1080/1573062X.2019.1597375

To link to this article: https://doi.org/10.1080/1573062X.2019.1597375

Published online: 09 Apr 2019.

Submit your article to this journal

Article views: 490

View related articles

View Crossmark data

Citing articles: 9 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=nurw20
URBAN WATER JOURNAL
2018, VOL. 15, NO. 10, 953–960
https://doi.org/10.1080/1573062X.2019.1597375

RESEARCH ARTICLE

Pipe network leak detection: comparison between statistical and machine learning
techniques
J. C. van der Walt, P. S. Heyns and D. N. Wilke
Center of Asset Integrity Management, Department of Mechanical and Aeronautical Engineering, University of Pretoria, Pretoria, South Africa

ABSTRACT ARTICLE HISTORY

This paper investigates an inverse analysis technique to find leaks in water networks and compares Received 16 February 2018
different solution strategies. Although a number of strategies have been proposed by different authors Accepted 15 March 2019
to identify leaks on a vast selection of pipe networks, limited research has been done to compare KEYWORDS
strategies and point out their weakness. Three strategies, a Bayesian probabilistic analysis, a support Leakage; artificial
vector machine and, an artificial neural network were combined with the inverse analysis technique on intelligence; integrated
different numerical and experimental networks to point out each strategy’s weakness. Two numerical water management
networks are investigated and one experimental network. It is shown that the Bayesian probabilistic
analysis struggles to find unique solutions when a few observations are available, while the support
vector machine and the artificial neural network struggle when only flow measurements are available.
Additionally it is shown that the artificial neural network struggles to estimate unique solutions for leak
size and location.

1. Introduction
different sensor placement algorithms on a real-world net-
In the 2011
2012 National Non-Revenue Water assessment work. In their work they opened fire hydrants to simulate
(Mckenzie, Siqalaba, and Wegelin 2012) in South Africa the leakages within the network ranging between 0:25
1 l/s.
average national non-revenue water was found to be 37%. The six algorithms compared are shortest path 1, shortest
While the world average in that period was estimated to be path 2, Shannon entropy (de Schaentzen, Walters, and Savic
36:6% (Mckenzie, Siqalaba, and Wegelin 2012). Currently, to 2000), a binarized sensitivity matrix (Perez et al. 2011), a non
reduce non-revenue water in South Africa, pressure manage- binarized sensitivity matrix (Casillas et al. 2013), and SPuDU
ment systems are installed, minimum night flows are logged, (Steffelbauer et al. 2014).
and water balances are completed. The methods only reduce The following sub-sections show work done by different
the amount of leaked water, or raise the awareness. authors to compare several techniques, as well as work done
Contractors use listening sticks, geophones, ground penetrat- in this field by using artificial neural networks, support vector
ing radar, and noise loggers to help assist in finding the leaks machines and the Bayesian probabilistic analysis. These three
(Mckenzie, Siqalaba, and Wegelin 2012). techniques were chosen as a focus due to their and popularity.
In recent years research to find leaks using on-line
machine learning techniques with real time data has intensi-
1.1. Comparisons between different strategies
fied. Pèrez et al. (2014) applied this model-based methodol-
ogy to a real network in Barcelona. Their case study focused Nowicki, Grochowiski, and Dusinkiewichz (2026) investigated
on the Nova Icària DMA (district metered area), where a real data-driven models for fault detection in water distribution
leak occurred. Soldevila et al. (2016) investigated Bayesian networks. They investigated Kernel PCA (principle compo-
reasoning with the model-based methodology on the Hanoi nent analysis) and applied the method for fault detection to
DMA. They investigated four cases, leaks varying between the water distribution network of Chojnice, located in north-
25
75 l/s, 5% noise on pressure, 5% uncertainty in the ern Poland. They compared the results of the Kernel PCA with
demands and a case where all three of these cases were a normal PCA and a simple Control Chart. Romano, Kapelan,
combined. They compared their results with previously and Savić (2013) investigated geostatistical techniques for
found results using a k-Nearest Neighbors approach. burst detection in water distribution networks. They com-
Additionally they investigated using Bayesian classifiers pared four different techniques, namely inverse distance
(Soldevila et al. 2017) with the model-based methodology. weighted interpolation, local polynomial interpolation, ordin-
In this work they investigated two case studies, the Hanoi ary Kriging and ordinary CoKriging. They compared these
and Nova Icària DMA. techniques on a case study with a rural water network con-
Additional to the techniques used to find the leaks, sensor sisting of 17.8 km of pipes. They measured 13 pressure
placement for this method of leak detection is highly impor- measurements throughout the network and simulated bursts
tant. Fuchs-Hanusch and Steffelbauer (2017) compared six by opening fire hydrants.

CONTACT J. C. van der Walt [email protected]

© 2019 Informa UK Limited, trading as Taylor & Francis Group
954 J. C. VAN DER WALT ET AL.

1.2. Artificial neural networks A model error can now be written as:
Caputo and Pelagagge (2003) proposed a method of using eij ¼ xij
xij ðθÞ; (1)
artificial neural networks to estimate the leak location in pip-
ing networks. They performed tests on a network where they where xij are the actual measurements from the system.
generated input data for leaking and non leaking states. Two Another parameter was added to be optimized, namely σ.
neural networks were used in their proposal, the first identify- This parameter represents the uncertainty within the error.
ing the leaking branch and the second estimating the leakage Using Bayes’ theorem and applying the uncertainty and
amount and location. Applying the neural networks, they parameter set to be quantified by a probability density func-
found that the leaking branch could be correctly identified tion, πðθ; σÞ, it follows that:
with the leak size estimated to between 2
10% of the actual Pðθ; σjxÞ ¼ c1 Pðxjθ; σÞπðθ; σÞ; (2)
value. The location of the leak could be estimated within 50

100m of the actual leak location. where c1 is a normalization constant, so that Pðθ; σjxÞ
Mounce and Machell (2007) performed tests on an actual integrates to one. Assuming the model error eij is indepen-
water supply network in the UK. Bursts in the network were dent, normally distributed with a zero mean, and standard
simulated by opening fire hydrants. Two sensor locations were deviation of σ, the likelihood of Pðxjθ; σÞ can be written as:
used: one at the input of the network and one at the output Y
L Y
N ðxij
xij Þ2
1
going to the neighboring DMA. The sensors measured both Pðxjθ; σÞ ¼ pffiffiffiffiffiffiffiffi e
2σ2 ; (3)
pressure and flow. Five different burst locations were simu- i¼1 j¼1 2πσ
lated. They found that they could locate the bursts with an
where L is the total number of monitoring locations and N is
accuracy of 98:33%.
the total number of flow tests. This equation can be simplified
Salam et al. (2014) investigated an on-line monitoring sys-
by assuming the initial probability density function πðθ; σÞ is
tem to detect leakages in pipe networks. They used a network
constant and substituting Equation (3) into Equation (2). The
from Makassar in Indonesia. They used pressure measure-
simplification can be written as:
ments at each junction as input data. The input data were 0 P 1
N
generated by simulating leaks in the network. They used jjxij
xij jj2
a radial basis function neural network which could detect B j¼1 C
@
2σ2
A
the leak location and sizes with an accuracy of 98%. 1
Pðθ; σjxÞ ¼ c2 pffiffiffiffiffiffiffiffi LN e : (4)
ð 2πσÞ
1.3. Support vector machines By maximizing this function the most probable leak location
De Silva, Mashford, and Burn (2011) investigated support vec- can be found within a network. Poulakis considered a network
tor machines to act as pattern recognizers to detect leaks in that consists of 50 pipes, 31 nodes, and 20 loops. The network
pipe networks. They started with a SVM (support vector forms a grid network supplied by one reservoir with one leak.
machine) as a regressor to try and predict emitter coefficients. They went on by introducing variation in pipe roughness
Six monitoring nodes were used to act as sensor locations. coefficients, variation in the assumed demands, and
They selected 10 candidate leaking nodes and generated a variation in the model measurements. They found that
a data set with varying emitter coefficients. The SVM could, when the model measurements had an uncertainty of 2%,
after training, achieve a testing accuracy of 76:8%. the location could be calculated. If the uncertainty in the
They then used 40 candidate leaking nodes and created model measurements was increased to 5% the model was
a data set, for which a testing accuracy of 57:2% was achieved. unsure about the actual leak location.
They found that the predicted leak location was within 500 m
of the actual leak location in all cases for a network that could
fit into a 1000 by 1100 m square box. 1.5. This research
They went on to investigate whether the SVM could detect There was found that various theoretical and practical net-
small leaks in the network. The smallest leak registered by works have been tested with machine learning techniques
EPANET to generate a pressure difference was a leakage of such as neural networks, support vector machines, and other
90l=hour. A new data set was created to which the SVM was statistical approaches. This paper investigates the pressure-
trained. A testing accuracy of 35% was found. flow deviation method with different solution strategies to
find the application of each strategy for specific networks
and an overall comparison between the strategies.
1.4. Bayesian probabilistic framework
Three strategies are investigated: the Bayesian probabilistic
Poulakis, Valougeorgis, and Papadimitriou (2003) investigated analysis, a support vector machine, and an artificial neural
a Bayesian probabilistic framework to detect leaks in a water network. The three investigated strategies are then applied
pipe network. The derivation starts by assigning θ as the to three water networks, of which two are numerical and the
parameter to be optimized. This parameter includes the leak- other an experimental network, to gain a deeper understand-
ing pipe, location and size of the leak. It can also be written as ing of the networks and strategies.
xðθÞ to indicate the measured values such as pressure and The first two tested networks simulate a numerical trans-
flow for a given set of leak parameter. portation and distribution network. The solutions for these
 URBAN WATER JOURNAL 955

networks are idealized, with no model or measurement error. 2.3. Strategy 3: artificial neural network
This is to find limitations within the solution strategies. The
The artificial neural networks considered the same input and
experiential network contains calibration of the model and
output parameters as the SVMs. The sizes of the ANNs (artifi-
therefore contains model and measurement errors. The experi-
cial neural networks) vary between problems with the sizes
mental model is based on a simple distribution network to
being chosen for each problem to result in the optimum
offer some complexity to the problem.
results. This was completed by increasing the size of the
ANN until the accuracy of the prediction stopped increasing
or before over fitting occurred. The output for the classifica-
2. The algorithms
tion ANN returns a string of numbers which suggest pipes
Three strategies investigated in this research are discussed in with possible leaks.
this section. Each pipe in the network has its own regression ANN which
calculates the leak location and size. The number of hidden
layers and nodes in a layer is identical to that of the classifica-
2.1. Strategy 1: Bayesian probabilistic analysis tion ANN. The output for the regression ANN is the leak size
and location of the leak. In Python the ANNs are created using
This method was proposed by Poulakis, Valougeorgis, and
the scikit-learn (Pedregosa et al. 2011) package similar to the
Papadimitriou (2003). It uses the mean squared error (MSE)
SVMs.
calculation to which a probability is calculated. An extra opti-
For both the SVMs and ANNs the training and testing data
mization parameter is introduced, namely σ, which is the
are split with a 70:30 ratio. Within the training algorithm 10%
uncertainty in the MSE calculation. The formulation is writ-
of the training data is used as validation to ensure over fitting
ten as:
does not occur, which is then validated using the testing
  data set.
1

MSEðθÞ
Pðθ; σjx Þ ¼ c2 pffiffiffiffiffiffiffiffi LN e 2σ2 ; (5)
ð 2πσÞ
3. Numerically simulated networks
where θ is the optimized leak parameter. The optimization of
this algorithm requires the equation to be maximized for the The strategies were tested on two different numerical water
largest probability. Therefore the log-likelihood of this func- networks. These networks include a simple single pipe net-
tion can be calculated as: work and a distribution network found in the literature. The
simple problem is chosen to offer an in depth understanding
MSEðθÞ LN of the performance of the solution strategies for the most
gðθ; σÞ ¼
lnðPðθ; σjxÞ ¼ þ lnðσ2 Þ; (6)
2σ2 2 simple problem. The second numerically investigated problem
is a complex network that is reproducible from the literature.
where L is the total number of monitoring locations, N is the Both these networks are idealized with no model or measure-
total number of flow tests and σ is the uncertainty within the ment errors. This is to ensure this method of leak detection
error. The log-likelihood given by gðθ; σÞ can now be mini- has the ability to solve the leak size and location from the
mized. For this strategy no data set is generated. Within the measured data.
optimization algorithm the EPANET model is simulated with
the estimated parameters which results in the pressure and
flow measurements to calculate the error. This is repeated 3.1. Problem 1: simple single pipe system
until the error is minimized.
This network consists of two reservoirs connected together.
Figure 1 shows a diagram of the network layout. R1 and R2
indicate the two reservoirs, N1 and N2 indicate the two nodes
2.2. Strategy 2: support vector machine
at which pressure will be measured and P1 indicates the pipe
For the support vector machine a data set is generated. SVMs with a leak on it.
aim to solve the problem as a classification or inverse regres- The height of R1 and R2 is chosen to be 50 m and
sion problem. The input parameters for the SVM are the
pressure and flow measurements from a simulated model,
and the output parameters are the leak location and size.
Two SVM types are used, the first estimating on which pipe
a possible leak occurred, and the second to find the possible
leak size and location on all the pipes in the network.
For the classification SVM the outputs are an integer value
depending on the number of pipes in the network. For the
regression SVM, the outputs are the leak size and leak location
on the length of the pipe. All the SVMs use a RBF kernel
function and they are generated using scikit-learn
(Pedregosa et al. 2011) which is a Python package used for
machine learning. Figure 1. Diagram of problem 1.
956 J. C. VAN DER WALT ET AL.

20 m respectively. The elevation of N1 and N2 is the same and solution was found for this problem. The minimum error of
at 0 m. The length of P1 is 100 m with a diameter of 32 mm. the ANN solution is found to be 0:02. This results in a leak
The location to the leak is measured from N1 . The leak is length of 42.9 m and an emitter coefficient of 0:244. The
modeled with an emitter coefficient of 0:3. The flow though difference between the Bayesian probabilistic analysis and
this network was calculated by EPANET as 2:68m/s, giving the SVM or ANN is that the SVM and ANN use the absolute
a Reynolds number of 85,760 resulting in turbulent flow in values of the pressures while the Bayesian probabilistic analy-
the pipe. The Hazen-Williams model was used to solve the sis uses the pressure difference over the pipe.
EPANET model.
To solve this problem with the SVM and ANN, a data set of
3.2. Problem 2: benchmark network
1000 samples was generated with random leaks at locations
between 0--100 m and leak sizes with emitter coefficients The network consists of 20 loops, 30 nodes with demands, one
between 0
2. Pressure and flow measurements could be reservoir supplying the network and 50 pipes. This network
generated by simulating the EPANET model with the randomly was introduced by Poulakis, Valougeorgis, and Papadimitriou
generated leak parameters. For this problem the input data for (2003) and they applied the Bayesian probabilistic framework
the strategies is the pressures, flows and the leak location. The to it. The same network was used by Nasirian, Maghrebi, and
leak location was added to the input to help with the pre- Yazdani (2013) to benchmark their new heuristic genetic algo-
sentation of the results. The output data for the SVM and ANN rithm methodology to find leaks. In 2016 the same network
is simply the diameter of the leak. was used by Asgari and Maghrebi (2016) where they investi-
gated a new method of locating a leak by calculating a leak
index.
3.1.1. Solutions
The network is depicted in Figure 3. The supply to the
The solution found by the three solution strategies can be
network comes from reservoir R1, which has a static head of
seen in Figure 2. The dot in the figures shows the actual
52 m. Each junction has a demand of 50 l/s, which are num-
modeled leak. From the figure the Bayesian probabilistic ana-
bered from N1
N30. The pipe lengths are 1000 m and
lysis error can be seen as 0 throughout the length of the pipe.
2000 m respectively, as depicted in the figure. The pipe dia-
This indicates there are an infinite number of solutions for this
meters vary from 600 mm, to 450 mm, and finally 300 mm as
strategy. This is due to the Bayesian probabilistic analysis
the flow decreases through the network.
optimizing to find the minimum error for the problem, which
Two cases are considered: the first only pressure is mea-
is the difference between the two measured pressures. For this
sured and the other only flow rate is measured. For case A the
problem a leak can be added anywhere on the pipe which will
pressure observation nodes are marked with circles around
result in the error being zero, giving an infinite amount of
the nodes. For case B the flow rate observation pipes are
solutions. Adding the flow within the pipe to the input of the
marked with rounded rectangles. A leak of 22.8 l/s is added
algorithms could solve this problem since the problem is no
to the network. The leak is modeled at the center of P25, at
longer ill-posed. The dot in the figure indicates the actual leak
node N55, which is marked with a square. This leak is modeled
location.
as a demand. The same sensor placement was used as intro-
From the solution for the SVM it can be seen that the SVM
duced by Poulakis, Valougeorgis, and Papadimitriou (2003).
gives a unique solution for this problem which is where the
The flow regime within this network was calculated and
error is calculated to be 0:0005, at its minimum. The estimated
found to be turbulent, therefore the Hazen-Williams formula
solution was found at a length of 58:4m and an emitter coeffi-
is used within EPANET.
cient of 0:413. For the ANN it can be seen that a unique
Similar to the previous problem, a data set was generated
for different leak cases. The data set was used for both the SVM
and the ANN, which contained 10,000 data samples. The data
samples were randomly generated with leak sizes between 10 l/
s and 50 l/s. This data set contained two outputs, the leaking
pipe and size. The inputs for this data set were simulated with
the EPANET model depending on the tested case.

3.2.1. Solutions
The solution found by the solution strategies can be seen in
Figure 4. This figure shows the solution for the Bayesian
Probabilistic analysis for case A, where the most probable
leak location was found at P25, with a probability of 43:1%.
The leak size was estimated to be 22.8 l/s. For case B the most
probable leak location was calculated at pipe P25, with
a probability of 39:9%. The leak size was estimated as 22.8 l/s.
The solution found after training the SVM to the generated
data set can also be seen in this figure. For case A it can be
seen that a probability of a leak at P25 was calculated as
Figure 2. Problem 1: solutions found by the three strategies. 54:3%, with the adjacent pipes with leak probabilities of
 URBAN WATER JOURNAL 957

Figure 3. Diagram of problem 2.

Figure 4. Problem 2: solutions found by the three strategies.

17% and 8%. For case B it was found that the actual leak
3.3. Discussion of results
location was given a probability of 1:8%, with other pipes
having higher probabilities of leaks. The leak size estimation A comparison between these strategies showed that the
of the SVM was calculated between 29 l/s and 31 l/s, while the Bayesian probabilistic analysis could not find leaks in simple
actual leak size was 22.8 l/s. pipe networks due to the ill-posed nature of the problem,
The solution found by the ANN after training shows while the SVM and ANN could solve the simple pipe net-
a perfect classification solution as well as a leak estimation of work accurately by algorithmic regularization. This indicates
22.8 l/s for case A. For case B it can be seen that an incorrect that the Bayesian probabilistic analysis has difficulties sol-
classification solution for the location was found. The leak size ving the problem when too little information is known
estimation for this case was found to be 22.7 l/s. about the network.
958 J. C. VAN DER WALT ET AL.

The solutions for Problem 2 showed that the Bayesian a location of 2 m and 1 m. The valves between the pipes to
probabilistic analysis can solve the problem accurately using change the network layout are closed.
either pressure or flow measurements. This is not the case for Figure 6 shows the 3 mm leak applied to the network. The
the SVM and ANN as both of these strategies could not find leak is applied by clamping a saddle over the pipe, and drilling
the leaking pipe or its size when only flow measurements are the correct diameter hole through the pipe. The demands
used. Indicating that the error optimization methods have an applied to each pipe were created in the same way, resulting
advantage when using flow measurements. in the network being pressure driven. The pressure sensors
had a full scale error of 1:5% and could measure pressure from
0--5 bar. The flow sensors used could measure between 1--25
4. Experimental network
l/min and had an error rating of 3%.
The solution strategies were tested on an experimentally built Figure 7 shows the EPANET model used to calibrate the
network. The experimentally measured values are calibrated experimental measurements. In this figure P1
P12 indicate
and tested to find the actual leak for two different cases. the locations of the pressure sensors while F1
F7 indicate
the locations of flow sensors. L1
L6 indicate the leaks on the
pipe. The calibration process consists of calculating an error
4.1. Setup
between the experimental measurements and the model mea-
The experimental network built can be seen in Figure 5. The surements when there were no leaks. These measurements
experimental networks used twelve pressure sensors, one at include all 12 pressure and seven flow measurements. The
the start and end of each pipe. Additionally, seven flow meters parameters optimized include the roughness coefficient,
are used: one measuring the input of the network while six a loss coefficient on each pipe, an emitter coefficient simulat-
measure the output flows of each pipe. In this figure it can be ing the demand in the network and the pump efficiency.
seen that six possible leaking pipes were used. These pipes are
fed by a pump from a reservoir. The demands in the network
was modeled with 3 mm holes, which fed back to the
reservoir.
The reservoir used is a simple container holding 50 l of
water and cannot be seen in this photo. A Pentax CM 210
pump was used which is capable of supplying the network
with the necessary pressure and flow. The lengths of each pipe
is 3 m with a diameter of 10 mm. The pressure in this network
with the specified supply achieved a pressure in the network
ranging between 2
3 bar, with an average flow of about
0:1
0:2 l/s through each pipe, resulting in a Reynolds
Number ranging between 100
200. This results in the flow
to be laminar in this network, therefore the Darcy-Weisbach
formula was used within EPANET.
Three leak locations were added to the network of which
only the first two leak locations were considered. The leaks
were located on the first and third pipe with diameters of
3 mm and 2 mm respectively. The leaks were located at Figure 6. Actual leak of 3 mm applied to the experimental network.

Figure 5. Photo of the built experimental network.

 URBAN WATER JOURNAL 959

The ANN classified the two leaking pipes with probabilities

of 32:4% and 23:2%. For the second leak case the sixth pipe
was estimated to be leaking with a probability of 30:2%. The
leak size for the two leak cases was estimated as 0.18 l/s and
0.31 l/s. The location of the leaks was estimated to be at
1.39 m and 1.53 m.

4.3. Discussion of results

For the experiment the model calibration could be completed
by optimizing 39 coefficients. These coefficients included the
Figure 7. EPANET model used to calibrate experimental measurements.
roughness coefficient, minor losses, emitter coefficients and
the pump efficiency which resulted in the model error
Calibrating these parameters resulted in a total error of 2.56% of 2:56%.
when there was no leak. The three strategies could all identify the first leak case
To solve the experimental problem limits to the optimiza- correctly, with only the SVM finding the correct leaking pipe
tion of the Bayesian probabilistic analysis and the data sets for for the second leak case. Both the Bayesian probabilistic ana-
the SVM and ANN had to be set. For the three strategies, the lysis and the SVM estimated the leak to be nearly double its
leak size was limited to an emitter coefficient between 0
0:1, actual size, with the ANN training to constant values for the
while the leak location was limited between 0--3 m. For the leak size and location in both cases.
SVM and ANN the data set contained 1500 data samples per
pipe, with a total of 9000 samples.
5. Conclusion
The pressure-flow deviation method was used to find leaks in
4.2. Results of experimental measurements
various water distribution and supply networks. The three
The results found by the three strategies for the two leak cases solution strategies used could find leaks in all the networks
can be seen in Figure 8. For the two leak cases, the actual leak tested. The algorithms used were the Bayesian probabilistic
size and location was marked with a black dot. The leak size analysis, a support vector machine, and an artificial neural
for the two leaks were calculated from the flow measurements network.
as 0.23 l/s and 0.0935 l/s. The leak locations were at In the first numerical network it was seen that the ANN and
2 m and 1 m. SVM outperform the Bayesian probabilistic analysis since the
In these figures it can be seen that the Bayesian probabil- Bayesian probabilistic analysis makes use of error calculations,
istic analysis could find the leaking pipes with probabilities of removing the absolute pressure and flow values from its
27:9% and 23:4%. For the second leak case the leak was estimation. In the second numerical problem it was seen
identified on the sixth pipe with a probability of 25:4%. The that the ANN and SVM struggled to find the leaking pipe as
Bayesian probabilistic analysis calculated the leak size to be well as its size when only flow measurements were used,
0.48 l/s and 0.21 l/s for the two leak cases, with the location although the Bayesian probabilistic analysis could find the
calculated at 0.74 m and 2.49 m. leak in both cases.
The SVM could identify the leaking pipe accurately for both In the experimental network the SVM could identify the
cases with a probability of 51:0% and 45:1%. The leak size for leaking pipes for both cases, while the Bayesian probabilistic
the two cases was estimated to be 0.40 l/s and 0.24 l/s, with analysis and the ANN only identified the leaking pipe for the
the location of the leaks estimated at 1.44 m and 1.35 m. first case. Furthermore it was seen that the Bayesian

Figure 8. Experimental network: solutions of the two leak cases.

960 J. C. VAN DER WALT ET AL.

probabilistic analysis and SVM estimated the leak to be nearly Mckenzie, R., Z. N. Siqalaba, and W. A. Wegelin. 2012. The State of Non-
double its size. The ANN was seen to train to constant estima- Revenue Water in South Africa. Technical Report. Water Research
tions for the leak size and location. Commission. August.
Mounce, S. R., and J. Machell. 2007. “Burst Detection Using Hydraulic Data
From this work it can be seen that no strategy consistently from Water Distribution Systems with Artificial Neural Networks.” Urban
outperforms any of the others. This begs the question whether Water Journal 3 (1, February): 21–31. doi:10.1080/15730620600578538.
these strategies should not be combined to work together to Nasirian, A., M. F. Maghrebi, and S. Yazdani. 2013. “Leakage Detection in
find better predictions. This leak detection technique can also Water Distribution Network Based on a New Heuristic Genetic
be combined with on-line monitoring systems allowing for Algorithm Model.” Journal of Water Resource and Protection 5 (March):
294–303. doi:10.4236/jwarp.2013.53030.
quick and accurate detection of leaks. Although this technique Nowicki, A., M. Grochowiski, and K. Dusinkiewichz. 2026. “Data-Driven Models
requires further investigation to find accurate leak sizes and for Fault Detection Using Kernal Pca: A Water Distribution System Case
exact location estimations, it can accurately find leaks in net- Study.” Applied Mathematics and Computer Science 22: 939–949.
works and identify the pipes they are on. In addition, this work Pedregosa, F., G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel,
shows more research needs to be completed on model cali- M. Blondel, et al. 2011. “Scikit-Learn: Machine Learning in Python.”
Journal of Machine Learning Research 12: 2825–2830.
bration techniques to help with the detection of the leaks, Perez, R., J. Quevedo, V. Puig, F. Nejjari, M. A. Cuguero, G. Sanz, and J. M. Mirats.
their sizes and locations. 2011. “Leakage Isolation in Water Distribuion Networks: A Comparative
Study of Thwo Methodologies on a Real Case Study.” In 19th
Mediteranean Conference on Control Automation, Corfu Greece, 138–143.
Disclosure statement Perez, R., G. Sanz, V. Puig, J. Quevendo, M. A. Cugeuro Escofet, F. Nejjari,
J. Meseguer, G. Cembrano, J. M. Mirats Tur, and R. Serrate. 2014. “A
No potential conflict of interest was reported by the authors. Model-Based Methodology Using Pressure Sensors Applied to A Real
Network in Barcelona.” In IEEE Control Systems Magazine, 24–36. August.
Poulakis, Z., D. Valougeorgis, and C. Papadimitriou. 2003. “Leakage
Detection in Water Pipe Networks Using a Bayesian Probabilistic
References Framework.” Probabilistic Engineering Mechanics 18 (4, July): 315–327.
doi:10.1016/S0266-8920(03)00045-6.
Asgari, H. R., and M. F. Maghrebi. 2016. “Application of Nodal Pressure
Romano, M., Z. Kapelan, and D. A. Savić. 2013. “Geostatistical Techniques
Measurements in Leak Detection.” Flow Measurement and
for Approximate Location of Pipe Burst Events in Water Distribution
Instrumentation 50 (June): 128–134. doi:10.1016/j.
Systems.” Journal of Hydroinformatics 15 (3): 634–651. doi:10.2166/
flowmeasinst.2016.06.009.
hydro.2013.094.
Caputo, A. C., and P. M. Pelagagge. 2003. “Using Neural Networks to
Salam, A. E. U., M. Tola, M. Selintung, and F. Maricar. 2014. “On-Line
Monitor Piping Systems.” Process Safety Progress 22 (2, June):
Monitoring System of Water Leakgage Detection in Pipe Networs
119–127. doi:10.1002/(ISSN)1547-5913.
with Artificial Intelligence.” ARPN Journal of Engineering and Applied
Casillas, M. V., V. Puig, L. E. Garza-Castanon, and A. Rosich. 2013. “Optimal
Science 9 (10, October): 1817–1822.
Sensor Placement for Leak Location in Water Distribuion Networks
Soldevila, A., R. M. Fernandez-Canti, J. Blesa, S. Tornil-Sin, and V. Puig.
Using Genetic Algorithms.” In Conference on Control and Fault-
2016. “Leak Localization in Water Distribution Networks Using
Tolerant Systems, Nice France, 61–66. June.
Model-Based Bayesian Reasoning.” In European Control Conference,
de Schaentzen, W. B. F., G. A. Walters, and D. A. Savic. 2000. “Optimal
Aalborg Denmark, 1758–1763. June.
Sampling Desing for Model Calibraion Using Shortest Path, Genetic and
Soldevila, A., R. M. Fernandez-Canti, J. Blesa, S. Tornil-Sin, and V. Puig.
Entropy Algorithms.” Urban Water. 2 (June): 114–152.
2017. “Leak Localization in Water Distribution Networks Using Bayesian
De Silva, D., J. Mashford, and S. Burn. 2011. Computer aided leak location
Classifiers.” Journal of Process Control 55 (April): 1–9. doi:10.1016/j.
and sizing in pipe networks. Technical Report 17. Urban Water Security
jprocont.2017.03.015.
Research Alliance. April.
Steffelbauer, D., M. Neumayer, M. Günther, and D. Fuchs-Hanusch. 2014.
Fuchs-Hanusch, D., and D. Steffelbauer. 2017. “Real-World Comparison
“Sensor Placement and Leakage Localization considering Demand
of Sensor Placement Algorithms for Leakage Localization.” In XVIII
Uncertainties.” Procedia Engineering 89: 1160–1167. doi:10.1016/j.
International Conference on Water Distribuion Systems Analysis,
proeng.2014.11.242.
Cartagena Colombia, 499–505.