You are on page 1of 12

Water Research 166 (2019) 115058

Contents lists available at ScienceDirect

Water Research
journal homepage: www.elsevier.com/locate/watres

Deep learning identifies accurate burst locations in water distribution


networks
Xiao Zhou a, c, Zhenheng Tang b, Weirong Xu a, Fanlin Meng c, *, Xiaowen Chu b,
Kunlun Xin a, d, **, Guangtao Fu c, e
a
College of Environmental Science and Engineering, Tongji University, 200092, Shanghai, China
b
Department of Computer Science, Hong Kong Baptist University, Hong Kong, China
c
Centre for Water Systems, College of Engineering, Mathematics and Physical Sciences, University of Exeter, Exeter, EX4 4QF, UK
d
Shanghai Institute of Pollution Control and Ecological Security, 200092, Shanghai, China
e
The Alan Turing Institute, 96 Euston Road, London, NW1 2DB, UK

a r t i c l e i n f o a b s t r a c t

Article history: Pipe bursts in water distribution networks lead to considerable water loss and pose risks of bacteria and
Received 1 April 2019 pollutant contamination. Pipe burst localisation methods help water service providers repair the burst
Received in revised form pipes and restore water supply timely and efficiently. Although methods have been reported on burst
31 August 2019
detection and localisation, there is a lack of studies on accurate localisation of a burst within a potential
Accepted 5 September 2019
district by accessible meters. To address this, a novel Burst Location Identification Framework by Fully-
Available online 6 September 2019
linear DenseNet (BLIFF) is proposed. In this framework, additional pressure meters are placed at limited,
optimised places for a short period (minutes to hours) to monitor system behaviour after the burst. The
Keywords:
Burst localisation
fully-linear DenseNet (FL-DenseNet) newly developed in this study modifies the state-of-the-art deep
Deep learning learning algorithm to effectively extract features in the limited pressure signals for accurate burst
DenseNet localisation. BLIFF was tested on a benchmark network with different parameter settings, which showed
Pipe burst that accurate burst localisation results can be achieved even with high model uncertainties. The
Water distribution network framework was also applied to a real-life network, in which 57 of the total 58 synthetic bursts in the
potential burst district were correctly located when the top five most possible pipes are considered and
among them, 37 were successfully located when considering only the top one. Only one failed because of
the very small pipe diameter and remote location. Comparisons with DenseNet and the traditional fully
linear neural network demonstrate that the framework can effectively narrow the potential burst district
to one or several pipes with good robustness and applicability. Codes are available at https://github.com/
wizard1203/waternn.
© 2019 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).

1. Introduction billion m3 of drinking water (a value of over $9 billion) is lost per


year in Asia (Frauendorfer and Liemberger, 2010) and about 7.85
Water loss is a key issue in the management of water distribu- billion m3 of water was lost in 2017 in China (CUWA, 2017). Pipe
tion systems as in addition to the depletion of water resources, it bursts in water distribution networks (WDNs) are a main source of
also causes additional energy and chemical usage for water treat- water loss (AWWA, 2009); moreover, they often lead to social im-
ment and supply and poses risks of bacteria and pollutant pacts like water service interruptions and traffic delays (Berardi
contamination (Fontanazza et al., 2015). For example, about 29 et al., 2008). As such, it is critical that the location of a pipe burst
can be determined effectively for efficient restoration of water
supply.
* Corresponding author. Centre for Water Systems, College of Engineering,
With the development of supervisory control and data acqui-
Mathematics and Physical Sciences, University of Exeter, Exeter, EX4 4QF, UK sition (SCADA) systems, real-time monitoring of pressure and/or
** Corresponding author. College of Environmental Science and Engineering, flow data are being widely used for the detection and localisation of
Tongji University, 200092, Shanghai, China. pipe bursts in WDNs. The deviation of real-time data from the
E-mail addresses: m.fanlin@exeter.ac.uk (F. Meng), xkl@mail.tongji.edu.cn
normal range of system behaviour is a key principle adopted in
(K. Xin).

https://doi.org/10.1016/j.watres.2019.115058
0043-1354/© 2019 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
2 X. Zhou et al. / Water Research 166 (2019) 115058

literature for burst detection/localisation. To establish the normal powerful tool for feature identification and pattern recognition.
performance expected from a WDN, models have been developed Comparing with the traditional machine learning algorithms, very
based on data driven techniques, such as machine learning models complex functions can be learned from raw data with the key
(Mounce et al., 2002, 2010a; Romano et al., 2014) and clustering features automatically extracted. Convolutional neural networks
analysis (Wu et al., 2016). The deviation can also be analysed using (CNNs) are the most widely used deep learning methods. More
hydraulic models (Bicik et al., 2010; Meseguer et al., 2014; Wu et al., recently, researchers found that CNNs can be deeper, thus more
2010; Zhou et al., 2018). The data-driven techniques were found to accurate and efficient to train if they have shorter connections
be effective in burst detection but incompetent in burst localisation between former layers and latter layers (Srivastava et al., 2015).
(Wu and Liu, 2017). The model-based approaches rely heavily on Based on this, Huang et al. (2017) proposed a novel densely con-
the accuracy of the hydraulic model (Hutton et al., 2014) and do not nected convolutional network (DenseNet) architecture, which im-
perform as well in larger WDNs as in small ones. Other burst proves the flow of information and gradients among layers,
localisation methods using pressure/flow measurements have also intensifies the propagation of features, and encourages feature
been proposed based on spatial analysis of the change in pressure/ reuse. Thus, DenseNet provides better parameter efficiency and is
flow data across a WDN. For example, Mounce et al. (2003) used a easier to be trained with better accuracy.
triangle-based cubic interpolation to create a pressure drop surface CNNs (including DenseNet), which employ convolutional layers
during a burst to estimate its location. In Farley et al. (2013), the to detect local conjunctions of features from the previous layers and
location of a burst within a district metering area (DMA) was pooling layers to merge semantically similar features (LeCun et al.,
inferred by the sensitivities of different pressure/flow measure- 2015), show higher accuracy in identifying objects with abundant
ments with respect to a burst outflow. Romano et al. (2013) local features (e.g. language, images or videos) than global features.
employed multivariate Gaussian mixtures-based graphical model However, the global features of a WDN pressure signal such as its
and geostatistical techniques to exploit data from a larger number amplitude and time variation could be more important, and the
of pressure meters to determine the potential positions of a burst convolution and pooling procedures may disorganise the global
within a DMA. Nevertheless, due to the insensitivity of pressures to features. In this study, we proposed a new fully-linear DenseNet
bursts (Bakker et al., 2014), and as normally random demand (FL-DenseNet) by replacing the convolutional layers in DenseNet
fluctuations could mask the hydraulic variations caused by bursts, with linear connections. Benefiting from the dense connectivity
these methods can only narrow down the potential locations of a pattern, FL-DenseNet enables the key features in pressure signals to
burst to a rough district as large as hundreds of meters including at be automatically extracted and more efficiently learned.
least several tens of pipes, which is insufficient for timely repair of This study aims to propose a novel Burst Location Identification
pipe burst and service restoration. Framework by FL-DenseNet (BLIFF). It can effectively identify the
For more accurate localisation of pipe bursts, new, targeted accurate location (a single or a few pipes) of a burst from a larger
monitoring is usually required to collect detailed information (e.g. potential district (e.g. a DMA). BLIFF uses the hydraulic model of a
acoustic or transient signals) of system behaviour at the potential WDN to simulate bursts to train FL-DenseNet. The newly collected
locations to detect the abnormality. Acoustic equipment such as pressure data at the optimised meter locations are fed to the
listening sticks and noise correlators are widely used for manual trained FL-DenseNet for prediction of burst location. The frame-
inspection. Despite the effectiveness and easy implementation, work is applied on two case studies and its robustness against the
they are very labour intensive. By temporarily placing accelerom- accuracy of system hydraulic model is examined.
eter sensors and analysing the acoustic signals, the method pro-
posed by Kang et al. (2018) can detect the location of a burst/ 2. Methodology
leakage automatically. However, its reliability is sensitive to the
characteristics of the leak conditions (e.g. water pressure, leak flow Built on the hydraulic model of a WDN, FL-DenseNet extracts
rate), and the detection range is limited by the clarity and corre- features in pressure patterns when a burst occurs at each pipe. This
lation of acoustic signals. Transient-based methods (Colombo et al., is based on the fact that 1) for any burst, the responses in pressure
2009) locate bursts by analysing transient wave features such as drop caused by a burst outflow vary at different nodes across a
propagation, reflection and damping. However, the background WDN, and 2) for any node, the pressure variations when the burst
noise or other activities in a system may shadow the burst-induced occur at different pipes also vary. By matching the learnt features
transient signals, especially if the number of pipes to be analysed with new monitoring data, the likelihood of burst occurring at
increases (Wu and Liu, 2017). Moreover, some methods in this every pipe can be predicted. The structure of the proposed BLIFF is
category need computationally demanding mathematical simula- introduced in Section 2.1 with a few key steps described in Sections
tions and the high frequency sensors employed are costly. As such, 2.2 to 2.5.
the current acoustic or transient-based methods are not suitable for
accurate burst locating from a reasonably large area (e.g. a DMA or 2.1. Burst Location Identification Framework by FL-DenseNet
tens of pipes) to a single pipe. Pressure meters can be a good
alternative as they are easy to install, e.g. at any hydrant or other Fig. 1 shows the flow chart for the proposed BLIFF. As high-
tapping points, and are less costly than flow meters. A possible lighted in the top box (in dashed lines), a key assumption is that a
solution to overcome the insensitivity of pressures is to mutually burst has already been detected and the potential district of the
complement key features in pressures at different moments, and burst (a DMA, or tens to hundreds of potential pipes) is given. BLIFF
fuse and identify the features in a continuous period of measure- consists of eight steps, the first four of which can be done either
ments by pattern recognition algorithms. offline or simultaneously with the latter four after a burst is
Although machine learning algorithms such as traditional arti- detected.
ficial neural networks (ANNs), clustering, and support vector ma- The first step of BLIFF is to choose the optimal locations for the
chines have been used to detect and/or locate bursts (Mounce et al., added pressure meters in the potential burst district so that effec-
2002, 2010a; Romano et al., 2014; Wu et al., 2016), they need to tive and efficient collection of burst information can be made using
hand design suitable feature extractors and are difficult to learn the least number of meters. The meters should be sensitive to flow
complex features. In recent years, as a newly-developed subfield of changes occurring in the potential burst district. The optimisation
ANNs, deep learning technology (LeCun et al., 2015) has become a and sampling design are described in detail in Section 2.3. The
X. Zhou et al. / Water Research 166 (2019) 115058 3

blocks and transition layers) to transfer, analyse and weight the


input information, and an output layer. An illustrative flowchart is
shown in Fig. 2 and the detailed description is provided as fol-
lows.where ½l0 ; l1 ; …; li1  is a concatenation of the previous layers
0; 1; …; i  1; Ci is a composite function, which includes three
consecutive operations: batch normalisation (BN) (Ioffe and
Szegedy, 2015), rectified linear unit (ReLU) (Schmidhuber, 2015),
and a linear connection.

1) Input layer: it receives input information (e.g. preprocessed


pressure measurements) in the form of matrices or vectors.
2) Dense block: it contains a number of layers which are connected
by a dense connectivity pattern. Namely, there are direct con-
nections from any layer to all subsequent layers (Huang et al.,
2017), and a layer li can receive the outputs of all preceding
layers at its input:

li ¼ Ci ð½l0 ; l1 ; …; li1 Þ (1)

3) Transition layer: It is also a combination of BN, ReLU and linear


Fig. 1. A flowchart of BLIFF.
connection and is used to link the blocks to fuse different
characters.
4) Prediction: A softmax classifier is attached at the end of the last
hydraulic model of a WDN built on EPANET3 (Rossman, 2010; Yan dense block to give the prediction results. In this study, a
et al., 2019), which allows for pressure-driven analysis (PDA), is probability value of burst occurring is generated for every pipe
employed to simulate pipe bursts. Further details are provided in in the potential burst district.
Section 2.4. The simulation results are processed and fed to train FL-
DenseNet (described in Section 2.2). The data preprocessing, as
described in Section 2.5, can improve the learning performance by 2.3. Optimisation of pressure meter locations
amplifying key features and improving comparability in pressure
measurements. After a pipe burst is detected, pressure meters are In a WDN, the sensitivity of nodal pressures to the changing of
placed at the optimised nodes to collect dynamic pressure changes nodal outflows can be indicated by a Jacobian matrix, which con-
within a prescribed duration (in minutes to hours, further tains the gradient information of nodal pressure with respect to
described in Section 2.4.3). The collected monitoring data are pre- nodal outflow as presented in Eq. (2):
processed according to the procedure described in Section 2.5 and 
then fed to the trained FL-DenseNet model for prediction. Vq H ¼ vH vq (2)

where Vq H is the Jacobian matrix, of which the elements represent


2.2. Fully-linear DenseNet the sensitivities of nodal heads with respect to nodal outflows; H is
the vector of nodal heads; q is the vector of nodal outflows. An
FL-DenseNet differs from the original DenseNet by replacing explicit method proposed by Liu et al. (2017) is employed in our
convolutional layers with linear connections and omitting the study to solve Vq H.
pooling layers. Similar to other ANNs, FL-DenseNet is based on a Three parameters are proposed to measure the effectiveness of
collection of connected processors (i.e. artificial neurons), which meters, which are coverage limit (CL ), coverage number (CN ) and
form an input layer, a number of hidden layers (i.e. a series of dense coverage rate (CR ). For node i and pressure meter at node j, if

Fig. 2. A flowchart of FL-DenseNet.


4 X. Zhou et al. / Water Research 166 (2019) 115058

  pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Vq H ij
 CL (3) qburst ¼ Cd Aori 2gHori (7)

then node i is considered to be covered by meter at node j. As such, where qburst is the burst flow; Cd is the outflow coefficient,
CL determines the minimum sensitivity that the meters should Cd ¼ 0:5  0:7 depending on the Reynolds number and the shape
meet. If node i is covered by more than CN meters, node i is of the orifice (Lambert, 2001), and is set to be 0.6 in this study; Aori
considered to be sufficiently covered. The higher the value of CN , is area of the orifice and Hori is the pressure. It is assumed that Aori ¼
the richer the information can be used to train FL-DenseNet. CR is gAij , where Aij is the cross-sectional area of corresponding pipe ij; g
the ratio of sufficiently covered nodes with respect to all nodes in is an intensity coefficient, which is determined by the intensity of a
the potential burst district. For example, if CL , CN , and CR are set as burst. Therefore, we can simulate different scales of burst flow on a
0.02, 2, and 85% respectively, then there should be at least 85% of pipe by assigning different values of g. For example, for a pipe with
nodes are covered (½Vq Hij > 0:02) by no less than two meters. a diameter of 200 mm and a pressure of 30m, g ¼ 10% corresponds
The optimal pressure meter locations are those which can to a burst flow of 45.7L/s and g ¼ 30% corresponds to 137.1L/s.
satisfy the threshold limits defined by CL, CN , and CR with the least Different values of g are used and examined in this work.
number of meters. The optimisation model can be expressed as The embedded leakage simulation model in EPANET3, which is
below. based on Eq. (8), is used to calculate the orifice outflow equation
and simulate burst flows,
min jSj (4)
S∢f1;…;np g qleak
ij ¼ aij Hbij ,Lij (8)

Subject to
where qleak
ij , Hij and Lij are the leakage flow, pressure and length of
pipe ij, respectively; aij and b are parameters to be determined. If aij
1 X
np X    
c c Vq H ij  CL  CN  CR (5) and b are defined according to Eq. (9),
np i¼1 jεS pffiffiffiffiffiffi
Cd gAij 2g
aij ¼ ; b ¼ 0:5 (9)
where np is the number of nodes in the potential burst district; set S Lij
contains the selected meter locations, element j in S denotes that a
meter is placed at node j; jSj denotes the cardinality of the set S; c: then the right side of Eq. (8) will be the same as that of the orifice
R1f0; 1g is defined as cðaÞ ¼ 1 if a 0, and 0 otherwise. The outflow formula in Eq. (7) Therefore, using the parameters defined
optimisation problem is solved by the genetic algorithm (GA) in Eq. (9), we can employ EPANET3 to simulate burst (orifice
(Holland, 1992). GA is a metaheuristic algorithm which is widely outflow) on pipe ij, and g is the only parameter to be determined.
used to solve optimisation problems in urban water management As PDA is employed in this study, the burst flow rate is dynamic
(Meng et al., 2016; Fu et al., 2013). Inspired by natural selection, GA with the change of water pressure on a pipe.
creates a set of solutions in the feasible region as individuals, which
evolve into better individuals in the next generation by means of
bio-inspired operators such as crossover, mutation, and selection. 2.4.1. Uncertainties of model parameters
This process is repeated until satisfactory results are obtained. The There are no perfect/accurate hydraulic models of any WDNs
GA Toolbox in MATLAB is used in our study. due to the various sources of uncertainty (Hutton et al., 2014; Savic
et al., 2009). For example, there are inherent random patterns in the
behaviour of water usage and the scale/characteristics of a pipe
burst; moreover, some parameters (e.g. pipe roughness co-
2.4. Training data generation
efficients) cannot be precisely measured. To make realistic as-
sumptions for BLIFF and test its robustness against the accuracy of
EPANET3 is employed to build the hydraulic model of a WDN
the hydraulic model, the following parameters are simulated in a
and perform extended period simulations in this study, as it enables
stochastic manner.
PDA which can produce a more realistic representation of the
pressure deficient situations under pipe bursts (Yan et al., 2019). A
 Pipe roughness coefficients and nodal demands. White Gaussian
nodal outflow is dependent on its nodal pressure in a PDA model,
noises Nð0; s2 Þ are added to the existing roughness coefficients
which is calculated according to the Wagner's formulas (Wagner
and diurnal nodal demand curves, with the standard deviation
et al., 1988):
values for the two parameters being sC and sq , respectively.
8 Smaller standard deviation should be assigned if a hydraulic
>
> 0 H*i < Hmin model is well calibrated. Otherwise, a larger value should be
>
>
i
>
> !1=h
>
< used.
H*i  Hmin  Intensity coefficient g. Pipe burst flow rate can vary in a wide
q*i ¼ req
qi i
Hmin < H*i < Hi
req
(6)
>
> req
Hi  Hmin
i
range in real life and it is hard (if possible) to know the exact
>
> i
>
> outflow rate of a burst even after its detection. The burst flow is
>
:
H *i > Hi
req req
qi assumed to be affected by the diameter and the intensity coef-
ficient g of a burst pipe in this work; g is simulated as a random
req
where q*i and qi is the available outflow and required demand at number which follows the uniform distribution Uðgmin ; gmax Þ.
node i, respectively; H *i is the pressure at node i; h is the head
req
exponent, generally h ¼ 1:5  2; H min i and H i is the minimum
pressure and required pressure, respectively. The values of h, H min
i
req
and H i are set as 2, 0m and 30m in this study (He et al., 2016). 2.4.2. Simulation of pipe burst
The flow rate of a burst on a pipe is assumed to follow the orifice The procedures for the simulation of a burst are summarised in
outflow formula in Eq. (7). six steps described below.
X. Zhou et al. / Water Research 166 (2019) 115058 5

(1) Set pipe ij as the pipe where a burst occurs. 2.5. Data preprocessing
(2) For each pipe, draw a noise value from Nð0; s2C Þ and add it to
the original roughness coefficient. The pressure data at a node (meter) m can be expressed as:
(3) For each node, draw noise values from Nð0; s2q Þ and add them
to elements of the nodal demand curve respectively. H m ¼ ½Hm1 ; Hm2 ; …; Hmn 0 (10)
(4) Randomly select intensity coefficient g from Uðgmin ; gmax Þ
and calculate the parameter aij, and assign leakage parame- where H m denotes the monitored pressures at node m with fixed
ters aij and b to pipe ij. time intervals. The length of H m is decided by the Data collection
(5) Employ EPANET3 to run the hydraulic simulation with the duration (Td ) and interval. For example, if Td ¼ 24h with a time
assigned parameter values. interval of 15 min, then the length of H m is 96 (n ¼ 96). The pres-
(6) Record pressure values at selected monitoring nodes. sure measurements at all available meters can be expressed as a
matrix as shown below
2 3
2.4.3. Generation of training data
H11 / Hm1
H ¼ ½H 1 ; H 2 ; …; H m  ¼ 4 « 1 « 5 (11)
Besides sC , sq and ðgmin ; gmax Þ mentioned in Section 2.4.1,
H1n / Hmn
values of the following parameters also need to be defined for the
generation of the training data set.

 Number of samples per pipe (Ntr ). To support feature identifi-


2.5.1. Nonlinear mapping
cation of a pipe burst against various sources of uncertainties,
The nonlinear mapping procedure is applied to facilitate feature
the simulation of pipe burst on any single pipe is performed for
extraction as it can amplify the time-variation features of an orig-
Ntr times with stochastic parameter settings. As such, the total
inal data set and increase the dimension of the original feature
number of samples in the training data set is Ntr  mp if there are
space. For each H m , defines
mp pipes in the potential burst district.
 Duration of data collection (Td ). After a burst occurs, pressure S m ¼ diagðH m Þ,H m (12)
needs to be monitored at the optimised locations (described in
Section 2.3) for a prescribed duration Td . As nodal pressures are where the operator diag converts the vector to a diagonal matrix, of
likely to show distinctive features at different times of a day which the diagonal is the original vector and other elements are 0.
driven by water usage behaviours, the time parameters of the For all the measurements, S ¼ ½S 1 ; S 2 ; …; S m .
training data should correspond to those of actual bursts for
effective machine learning. For example, assuming that a burst 2.5.2. Data normalisation
occurs at 2pm and field pressure data is measured from 3pm for As pressures of different meters may have diverse ranges
24 h (Td ¼ 24h) with a 15-min interval, when generating affected by the location and elevation of the meters, we need to
training data, the burst should also starts at 2pm and the report normalise the values to make the pressures at different meters
start time, time step and duration (corresponding to the time more comparable. Besides, normalisation can help to improve the
parameters in EPANET3) should be 3pm, 15 min and 24 h, accuracy and efficiency of FL-DenseNet (Al Shalabi and Shaaban,
respectively. A longer Td is likely to yield a higher prediction 2006). The z-score normalisation is employed to normalise the
accuracy as more data are captured. However, it also means data. Taken H m as an example,
longer waiting time (for the burst localisation) before a burst can
be repaired, thus a trade-off needs to be made. Different values g H m  mm
H m ¼ (13)
of Td are tested in this work. sm
 Number of pressure meters (Nm ). Pressures are monitored at the
Nm selected nodes both for the generation of training data and where Hgm is the normalised vector; mm and sm are the mean and
for the field monitoring. The number and location of meters are standard deviation of all the measurements collected at node m.
decided by the values of CL , CN , and CR , as described in Section Both H and S are normalised and fed to FL-DenseNet.
2.3. The accuracy of model prediction is likely to be higher with
a larger Nm, but this increases the expenditures on the sampling 3. Case studies and results
equipment. Although a specific value of Nm needs to be deter-
mined for the generation of a training data set, similar to Ntr and Two cases (a benchmark network and a real-life WDN) are
Td , different values are tested for comprehensive analysis and studied to demonstrate the reliability and applicability of BLIFF. The
discussion. following default settings are used: sC ¼ 10, sq ¼ 10%, ðgmin ; gmax Þ
¼ ð10%; 30%Þ, Ntr ¼ 1800, Td ¼ 24h, Nm ¼ 4. The default sampling
The four steps summarised below are used in our study to interval is 15 min. The neural network structure and hyper-
generate training data. parameters (whose values are set before the learning process be-
gins) used in the case studies are summarised in Table 1. For
(1) Choose parameter values for sC, sq , ðgmin ; gmax Þ, Ntr , Td and example, the hidden layers of the FL-DenseNet used in case 1
Nm . consist of four dense blocks, which contain 8, 16, 16, and 12 dense
(2) Set pipe ij as the pipe where a burst occurs, simulate the layers respectively; the learning rate we used for this network is
burst and record pressure values at monitoring nodes. 0.6. To demonstrate the effectiveness of the proposed FL-DenseNet,
(3) Repeat step (2) for Ntr times, each of which uses newly four different kinds of ANNs are used for case 2, which are FL-
randomly generated noise values and intensity coefficient. DenseNet, FL-DenseNet with the activation function (ReLU)
Ntr burst records are generated for pipe ij. replaced by sigmoid function (Schmidhuber, 2015), DenseNet, and
(4) For all the mp pipes in the potential burst district, repeat the traditional fully linear neural network (FLNN). As the FLNN does
steps (2)e(3), and Ntr  mp burst records are generated as not contain dense blocks, its hidden layers are constituted by three
the training data. fully connected layers. The hyper-parameters of the ANNs are the
6 X. Zhou et al. / Water Research 166 (2019) 115058

Table 1
Network structure and hyper-parameters used in the case studies.

Dense blocks Layers in dense blocks Hidden layer type Activation function Learning rate

case1: FL-DenseNet 4 8, 16,16, 12 Linear ReLU 0.6


case2: FL-DenseNet 4 8, 16, 24, 16 Linear ReLU 0.1
case2: FL-DenseNet (sigmoid) 4 8, 16, 24, 16 Linear Sigmoid 0.15
case2: DenseNet 4 6, 12, 24, 16 Convolutional ReLU 0.0005
case2: FLNN - (3) Linear ReLU 0.00002

optimal parameters identified from a series of trials. The diameters of pipes range from 200 mm to 400 mm, and the
To test the performance, for each pipe in the potential burst base demands of nodes range from 12.5 L/s to 63.8L/s. Different
district, we simulate burst on it for Nte times and generate Nte  mp diurnal demand patterns are allocated to different nodes.
burst data as the test samples, with the parameters be the same as The potential burst district is assumed to be the whole network,
the training data. Nte is set to 200. After being trained, BLIFF pre- that is, every pipe in the network is a potential location of the burst.
dicts the location of the burst in the test samples. The prediction The least required number of pressure meters varies when different
accuracy of a test sample i can be assessed by pre1i and pre5i . BLIFF threshold values are assigned for CL, CN , and CR . For example, if CL ,
will evaluate a burst happening probability for each pipe. If the pipe CN , and CR were set as 0.01, 1, and 0.85 respectively, one meter is
with the highest probability is the true burst pipe, pre1i ¼ 1; sufficient if it is placed at node 9. If CL , CN , and CR are set as 0.02, 2,
otherwise pre1i ¼ 0. If the top five pipes include the true burst pipe, and 0.85 respectively, at least four meters need to be placed and the
pre5i ¼ 1, otherwise pre5i ¼ 0, in which case the prediction is optimal locations are found to be nodes 3, 12, 15 and 17. The optimal
considered as failed. The overall performance of BLIFF can be locations corresponding to different Nm are presented in Table 2
assessed by Eqs. (14) and (15). and marked in Fig. 3.
To provide insights on the practical application of BLIFF, its
X
Nte mp
1 performance under different network physical or hydraulic condi-
acc1 ¼ pre1i (14) tions are analysed by using one-at-a-time sensitivity analysis. In the
Nte  mp i¼1
analysis, one parameter setting is changed to a different value at a
time, while other parameters remain at default settings. As Nte ¼
X
Nte mp
1 200, a total of 200  34 test samples are generated with each
acc5 ¼ pre5i (15) parameter setting. As the uncertainties in pipe roughness co-
Nte  mp i¼1
efficients and nodal demands highly affect model accuracy in a joint
manner, the performance of BLIFF under different combinations of
where acc1 and acc5 are the overall prediction accuracy of all test
samples corresponding to pre1i and pre5i .
sC and sq are analysed for a comprehensive representation of
different levels of hydraulic model accuracy. The results are shown
in Fig. 4.
3.1. Case 1 As shown in Fig. 4a and b, the performance of BLIFF improves if
the uncertainties in pipe roughness and nodal demands decrease,
The Anytown (Walski et al., 1987) network, which is a highly- i.e. acc1 (in Fig. 4a) and acc5 (in Fig. 4b) increases with lower values
looped small-sized WDN (Fig. 3), is employed to illustrate the of sC and sq . If sC ¼ 5 and sq ¼ 5%, acc1 and acc5 are 99% and 100%
application of BLIFF and its performance under different settings. respectively (i.e. the top left corner of Fig. 4a and b). If sC and sq
The network consists of 16 nodes, 34 pipes, 2 tanks, and 1 reservoir. increases to 20 and 25% respectively, the accuracy reduces with
acc1 ¼ 81% and acc5 ¼ 98% (i.e. the bottom right corner of Fig. 4a
and b). Nevertheless, satisfactory prediction results are still ob-
tained which indicates the high robustness of BLIFF against the
accuracy of the hydraulic models.
Results under different burst flow rates are presented in Fig. 4c.
It can be seen that the accuracy increases with the growth in ðgmin ;
gmax Þ, which corresponds to larger burst flows. If
ðgmin ; gmax Þ ¼ ð1%; 5%Þ (corresponding to a 4.6e22.9L/s burst flow
for a 200 mm pipe with 30 m pressure), acc1 is 35% and acc5 is 70%.
However, acc1 and acc5 become higher than 80% and 95% if ðgmin ;
gmax Þ increases to ð5%; 10%Þ (corresponding to a 22.9e45.7L/s burst
flow for a 200 mm pipe with 30 m pressure). This is expected as the
disturbance in pressure is more distinct with a higher burst flow,
which facilitates pattern recognition and burst localisation.
BLIFF can complement the features of pressures at different

Table 2
Optimal meter locations under different parameter sets.

CL CN CR Nm Optimal Meter Locations

0.02 2 0.85 4 Node 3,12,15,17


0.015 2 0.85 3 Node 4,11,17
0.01 2 0.85 2 Node 8,11
Fig. 3. Layout of the Anytown network and the optimal monitoring locations if
0.01 1 0.85 1 Node 9
different numbers (Nm ) of pressure meters are placed.
X. Zhou et al. / Water Research 166 (2019) 115058 7

Fig. 4. The performance of BLIFF with different values of factors. (aeb) Joint effect of pipe roughness and nodal demands uncertainty; (cef) the effects of burst flow range, data
collection duration, number of meters, and time difference respectively.

moments to achieve higher accuracy, therefore, longer Td will in- chosen meters should meet. Larger CL, CN and CR results in more
crease the prediction accuracy because pressures at more moments meters, which will increase the performance of BLIFF but also will
are collected. However, shorter Td allows faster field data collection lead to more costs. As shown in Fig. 4e, when there is only one
hence the burst can be more rapidly located. The relationship be- meter in the network, acc1 and acc5 are 62% and 96%, respectively;
tween Td and prediction accuracy are shown in Fig. 4d. When Td ¼ if four meters are installed, the acc1 and acc5 increased to 97% and
0:25h (i.e. 15 min, which equals to the sampling interval, so only 2 99%, respectively. acc5 is very high even with only one meter, which
pressure measurements are collected per meter), the acc1 and acc5 indicates that it is also feasible to use fewer meters but employs
are 70% and 98%, respectively; when Td increases to more than 10 h, more manpower to locate burst (e.g. with listening sticks) within
acc1 reaches 96% and higher. The accuracy rates are higher than the five identified pipes.
80% if Td is 1 h, which is satisfactory and fits with regulations in the In generating the training data, the simulated start time of the
water industry, such as the 3 h limit of response time in the UK bursts should ideally be the same as the actual situation. However,
(Diao et al., 2016). as FL-DenseNet could be trained in advance and in some cases we
Meters should be placed at nodes where a burst can have suf- cannot clearly confirm when the burst started, the start time of
ficient influence on the pressures. In our study, CL , CN and CR are burst simulation (ts ) may lag behind the real burst time (tb ). To test
used to define the minimum sensitivity requirement that the its effect on the performance of BLIFF, the burst time of training
8 X. Zhou et al. / Water Research 166 (2019) 115058

Fig. 5. The performance of BLIFF with (a) different number of samples and (b) training epoch.

data (ts ) is set to 24:00 and the real burst time (tb ) of the test ranging from 100 mm to 400 mm CL , CN and CR are set as 0.03, 2 and
samples is set to 12:00 to 24:00. Results are presented in Fig. 4f. It 0.9 respectively, and four optimal pressure meters are required.
can be seen that the accuracy decreased with larger lag times. The Their locations are optimised according to Section 2.3.
main reason could be that the continuous burst outflow causes the The training data and test samples are generated using the
water level of the tanks to drop, hence the overall pressure distri- default parameter settings described at the beginning of Section 3.
bution during simulation is inconsistent with the actual situation. As the potential burst district is connected with and affected by the
However, even with a lag of 2 h, the accuracy of BLIFF is still higher rest of the WDN, the uncertainties in the roughness coefficient and
than 80%. nodal demand are also simulated for pipes/nodes outside the po-
It is expected that the learning is more effective with larger Ntr tential burst district. To compare the performance of FL-DenseNet
and training epoch. Training epoch indicates the number of times with other ANNs, four different ANNs (as shown in Table 1) are
that FL-DenseNet repeatedly learns the same training data set. The trained by the same training data and are evaluated with the same
default training epoch used in our study is 120. The effects of Ntr test samples. Table 3 reports the results on acc1 and acc5 of the four
and training epoch are presented in Fig. 5. As shown in Fig. 5a, acc1 ANNs. Changing the activation function of FL-DenseNet from ReLU
and acc5 are 85% and 99% respectively if Ntr is 100 and increase to sigmoid has minimal effect on the results. Nevertheless, using
with larger Ntr. However, the increase rate becomes very small if Ntr sigmoid is less computationally efficient than ReLU because
is over 1000. Fig. 5b shows that BLIFF can achieve good results frequent exponential operations in sigmoid, which are computa-
(acc1 > 90%) after 25 training epochs, with minor improvement tionally demanding, need to be performed. For DenseNet,
(acc1 > 97%) if over 85 training epochs is used. In this case, with acc1 ¼ 31% and acc5 ¼ 67%, these accuracies are about half of those
Ntr ¼ 1800, processing each epoch takes about 3.5 min on an i7- of FL-DenseNet. The results demonstrate the validity of replacing
4700MQ CPU, GTX850M GPU computer. the convolutional layers in DenseNet with linear connections,
which can avoid the global features in pressure signals being dis-
3.2. Case 2 organised. Besides, using linear connections requires less compu-
tational time than convolutional layers. Although the accuracy of
A real network (though the burst data are synthetic) is used to FLNN in this case seems acceptable, in trial simulations it is found
test the reliability of BLIFF under realistic settings. The network that the performance of FLNN is very sensitive to the settings of the
(layout presented in Fig. 6) lies in eastern China, Jiangsu province. It hyper-parameters, such as the learning rate and the number of
consists of 567 pipes (total length is147km), 480 junctions and 4 hidden layers. It only provides satisfactory results in a quite limited
reservoirs, and supplies 57,000 tons of water per day. A burst is range of hyper-parameters. This implies that the algorithm is less
assumed to happen at the lower left district of the network. This robust and thus limited in practical application.
potential burst district contains 58 pipes with different diameters To further explore the performance of BLIFF on different pipes,

Fig. 6. A real-life network in China and detailed layout of its potential burst district].
X. Zhou et al. / Water Research 166 (2019) 115058 9

Table 3 second category (i.e. pre1 ¼ 0, pre5 ¼ 1). This suggests that the
Performance of different ANNs in case 2. feature of bursts at small pipes may be harder to distinguish from
FL-DenseNet FL-DenseNet (sigmoid) DenseNet FLNN similar, nearby pipes. Fig. 7d shows an example of failed prediction.
acc1 (%) 61.23 62.35 30.87 42.43
A few pipes near the true burst location are predicted with high
acc5 (%) 98.38 98.58 66.75 82.22 probabilities of burst. This indicates that the pressure responses at
the monitoring nodes are similar when the burst occurs at
branched pipes or nearby pipes. Nevertheless, BLIFF provides a
we create a new set of test samples with Nte ¼ 1 (i.e. each pipe in good direction though not the accurate location in this case.
the potential burst district is assumed as burst pipe for once There could be more than one burst concurrently happening in
respectively, i.e. a total of 58 burst events are generated). The real life. To test the performance of BLIFF with multiple bursts, two
prediction results are shown in Fig. 7a. 37 pipes (i.e. 64% of the 58 bursts are simultaneously simulated in the potential burst district
pipes, presented in green in Fig. 7a) are accurately targeted with with default parameters and at two randomly selected locations.
pre1 ¼ 1. 20 pipes (35%, presented in yellow in Fig. 7a) are included 11600 (equals to Nte  mp ) simulations are conducted as test sam-
in the top five likely locations (i.e. pre1 ¼ 0, pre5 ¼ 1). Only one ples. It should be noted that FL-DenseNet in this test is trained by
pipe fails to be correctly located (i.e. pre1 ¼ 0, pre5 ¼ 0), as high- data with only one burst (and default parameters) rather than two.
lighted in red in Fig. 7a. Fig. 7b to d show three examples of bursts For each test sample, if at least one of the bursts is correctly
in the three categories mentioned above. The likelihood of burst located, the sample is deemed as successfully predicted. An overall
happening is shown by the line colour and the true location of the result of all the test samples is acc1 ¼ 29% and acc5 ¼ 77%. It can be
burst is marked with a star. seen that the accuracy of burst localisation decreases when more
As shown in Fig. 7a, pipes with larger diameters (i.e. thicker than one burst occurs. This is expected as WDNs are nonlinear
lines) or closer to the monitoring meters are accurately predicted in systems thus the impact of two bursts is not the addition of two
general. The two large pipes in the top left corner of the district single bursts, hence the accuracy is low with the model trained
have large diameters but are not as well predicted. This may be using only single bursts. Three representative cases are illustrated
because the pipes are linked to each other with very few branches, in Fig. 8. If the bursts happen at pipes with large diameters and are
hence they may yield similar patterns of pressure change when a far from each other (so there are less mutual hydraulic effects), both
burst occurs. Fig. 7b gives an example of a correctly located burst, of of them can be targeted by BLIFF. In the example shown in Fig. 8a,
which the pipe with the highest predicted probability is the burst one burst is successfully located at the pipe with the highest
pipe. Fig. 7c shows another example of pipe burst that falls in the probability, and another included in the top five likely pipes. Fig. 8b

Fig. 7. (a) Burst localisation results of different pipes; (b) the prediction with the highest probability is the burst pipe; (c) the top five predictions contain the burst pipe while the
first prediction misses; (d) the top five predictions miss the real burst pipe.
10 X. Zhou et al. / Water Research 166 (2019) 115058

Fig. 8. Burst localisation results when two bursts concurrently happen (a) the prediction with the highest probability is one of the burst pipes; (b) the top five predictions contain
one of the burst pipes; (c) the top five predictions miss both bursts.

shows an example that a large burst and a small burst happen offline BLIFF is more suitable for small networks or those with DMA
simultaneously. In this case, the hydraulic response of the smaller configurations, as the potential burst district can be easily targeted
burst is masked by the larger one, hence BLIFF failed to target it. On (e.g. within a DMA). As the burst time in training data should aim to
the other hand, the localisation accuracy of the larger burst is be the same as actual bursts, a series of FL-DenseNets need to be
affected due to the disturbance of the smaller one. However, the trained with different burst time in the offline BLIFF. As shown in
larger burst pipe is still correctly identified within the top five Fig. 4f, the accuracy is still high if there is a 1 h lag in the actual and
predictions. In the example shown in Fig. 8c, BLIFF incorrectly assumed burst time, hence 24 FL-DenseNets (i.e. burst occurring at
locate the burst to the middle district of the two burst pipes. This is 00:00, 01:00, … and 23:00 respectively) should be sufficient for a
because similar patterns are produced when a burst occurs in the given WDN.
middle of two bursts. However, BLIFF still identifies the correct area The online BLIFF is more suitable for large WDNs without DMAs
of bursts. or where sufficient pressure meters cannot be installed. In these
cases, mobile meters can be employed when a burst is detected.
Concurrently, training of FL-DenseNet is performed on the provided
4. Discussions
potential burst district. Longer time is required for burst local-
isation in the online BLIFF than the offline because extra time is
Although previous studies suggest that pressure values are less
needed to place the mobile meters. However, the online BLIFF is
sensitive to a burst or leak event (Bakker et al., 2014; Mounce et al.,
more economically favourable as no pre-installed meters are
2010b), the proposed BLIFF which is built on the latest deep
required.
learning technology can identify features in a continuous period of
Different parameter settings should be used on different bursts
pressure measurements to accurately locate a burst to one or
to save costs and achieve higher localisation accuracy. Large bursts
several pipes. Based on the two case studies, comprehensive dis-
can be located more easily as the localisation accuracy of bursts
cussions about BLIFF are presented as follows.
with g > 10% is very high as shown in Fig. 4c. Based on the two case
A WDN model needs appropriate parameters such as pipe
studies, a large burst can be deemed as one which causes a pressure
roughness coefficients and nodal demands to represent the real
drop of more than 3 m at any pressure meter or with an outflow of
network accurately. The uncertainty in the parameter values are
over 100L/s. Due to the economic and social consequences of large
especially high in less calibrated models and poorly maintained
bursts, the efficiency of the localisation is more important. If the
networks, which undermines the performance of BLIFF. However,
offline BLIFF is applied, a shorter Td such as 15 or 30 min can be
even assuming considerably large uncertainties in case 1, BLIFF can
chosen to achieve an almost real-time localisation of large bursts.
still provide satisfactory results. This may be attributed to the
For the online BLIFF, in addition to using shorter Td, employing a
following reasons. Firstly, for a network in poor conditions with an
smaller number of mobile meters is also critical as it saves time for
uncalibrated model, larger sC and sq are chosen to generate the
the manual placement of the meters; this is unlikely to undermine
training set, hence it enables FL-DenseNet to learn in a wider
the accuracy of the localisation of large bursts, as satisfactory re-
feature space and cover the features of the network despite the
sults are obtained even with only one meter in case 1 as shown in
large uncertainty. Secondly, as the pressure data are collected for a
Fig. 4e. Besides, the training process can be faster by using small Ntr
continuous period of time rather than a single moment, features at
and training epochs. In comparison, the localisation of smaller
different moments can complement each other, which provides
bursts is harder and needs greater efforts and careful design. Small
more thorough information and the effect of uncertainties can be
bursts have limited disturbance on the WDN hydraulic state and
reduced.
can be easily masked by normally random demand fluctuations.
In practical application, BLIFF can be set up as either ‘offline’ or
Reducing the parameter uncertainties by good maintenance of the
‘online’. In the offline BLIFF, FL-DenseNet is trained in advance.
network and model calibration could be the key to improve the
Moreover, pressure meters can be installed in the optimal locations
accuracy of locating small bursts. Besides, setting more meters and
for a given district even without any bursts (i.e. the first four steps
choosing larger Td also help in improving the accuracy. In addition
in Fig. 1 are done offline). If a burst occurs in this area, the pressure
to small bursts, bursts which happen at branched pipes at the end
records can be extracted and fed to the trained FL-DenseNet. In this
of a network could also have lower localisation accuracy because
case, the efficiency of burst localisation solely depends on the
they are prone to be confused with adjacent bursts. In practice, if
duration of data collection process (Td ). As discussed in Fig. 4d, even
the predictions include the branched end pipes, the nearby pipes
if Td is within 1 h good prediction results can be yielded, hence the
should also be checked.
offline BLIFF can achieve near real-time burst localisation. The
X. Zhou et al. / Water Research 166 (2019) 115058 11

5. Conclusions Press, Beijing.


Diao, K., Sweetapple, C., Farmani, R., Fu, G., Ward, S., Butler, D., 2016. Global resil-
ience analysis of water distribution systems. Water Res. 106, 383e393. https://
A novel deep learning framework called BLIFF is proposed in this doi.org/10.1016/j.watres.2016.10.011.
study for accurate localisation of pipe bursts within a potential Farley, B., Mounce, S.R., Boxall, J.B., 2013. Development and field validation of a
burst district, based on additional real-time pressure monitoring at burst localization methodology. J. Water Resour. Plan. Manag. 139 (6), 604e613.
https://doi.org/10.1061/(ASCE)WR.1943-5452.0000290.
limited places for a short period. In the framework, FL-DenseNet, a Fontanazza, C.M., Notaro, V., Puleo, V., Nicolosi, P., Freni, G., 2015. Contaminant
modified version of the popular deep learning algorithm DenseNet, intrusion through leaks in water distribution system: experimental Analysis.
is employed to effectively extract the features in pressure signals in Procedia Eng. 119, 426e433. https://doi.org/10.1016/j.proeng.2015.08.904.
Frauendorfer, R., Liemberger, R., 2010. The Issues and Challenges of Reducing Non-
WDNs. Procedures are also incorporated in BLIFF for optimising the revenue Asian Development Bank: Philippines. http://hdl.handle.net/11540/
locations for the additional monitoring. BLIFF is tested on two case 1003.
studies and the key findings are summarised below. Fu, G., Kapelan, Z., Kasprzyk, J., Reed, P., 2013. Optimal design of water distribution
systems using many-objective visual analytics. J. Water Resour. Plan. Manag.
139 (6), 624e633. https://doi.org/10.1061/%28ASCE%29WR.1943-5452.0000311.
1) Accurate predictions can be made based on pressure signals He, P., Tao, T., Xin, K., Li, S., Yan, H., 2016. Modelling water distribution systems with
using the newly developed deep learning algorithm FL- deficient pressure: an Improved Iterative methodology. Water Resour. Manag.
30 (2), 593e606. https://doi.org/10.1007/s11269-015-1179-4.
DenseNet. The pressure meters are cheaper than flow meters Holland, J.H., 1992. Adaptation in Natural and Artificial Systems: an Introductory
and easier to install, indicating the practicality of BLIFF for real- Analysis with Applications to Biology, Control, and Artificial Intelligence. MIT
life applications. press.
Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q., 2017. Densely connected
2) BLIFF is trained using the data generated from hydraulic models,
convolutional networks. In: The IEEE Conference on Computer Vision and
but is found to be robust to high model uncertainties. Pattern Recognition, pp. 4700e4708.
3) The one-at-a-time sensitivity analysis shows that the prediction Hutton, C.J., Kapelan, Z., Vamvakeridou-Lyroudia, L., Savic, D.A., 2014. Dealing with
accuracy is high under a wide range of parameter settings of uncertainty in water distribution system models: a framework for real-time
modeling and data assimilation. J. Water Resour. Plan. Manag. 140 (2),
BLIFF, highlighting its robustness for effective burst localisation. 169e183. https://doi.org/10.1061/%28ASCE%29WR.1943-5452.0000325.
Ioffe, S., Szegedy, C., 2015. Batch normalization: accelerating deep network training
Although a few key uncertainty sources in modelling have been by reducing internal covariate shift. In: 32nd International Conference on Ma-
chine Learning.
considered, further study is recommended to examine the effects of Kang, J., Park, Y.J., Lee, J., Wang, S.H., Eom, D.S., 2018. Novel leakage detection by
other uncertainty sources such as inaccurate pump curves, un- ensemble cnn-svm and graph-based localization in water distribution systems.
known valve states, and pipe and node elevations. The Wagner's IEEE Trans. Ind. Electron. 65 (5), 4279e4289. https://doi.org/10.1109/Tie.2017.
2764861.
formulas and orifice outflow equation are used in this study to Lambert, A., 2001. What do we know about pressure-leakage relationships in dis-
simulate a burst with PDA, however, other burst simulation tribution systems. In: Proceedings IWA Conference “System Approach to
methods can be investigated for improving the performance of Leakage Control and Water Distribution Systems Management”.
LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521 (7553), 436e444.
BLIFF. https://doi.org/10.1038/nature14539.
Liu, N., Du, K., Tu, J., Dong, W., 2017. Analytical solution of Jacobian matrices of WDS
Declaration of competing interest models. Procedia Eng. 186, 388e396. https://doi.org/10.1016/j.proeng.2017.03.
236.
Meng, F., Fu, G., Butler, D., 2016. Water quality permitting: from end-of-pipe to
The authors declare that they have no known competing operational strategies. Water Res. 101, 114e126. https://doi.org/10.1016/j.watres.
financial interests or personal relationships that could have 2016.05.078.
Meseguer, J., Mirats-Tur, J.M., Cembrano, G., et al., 2014. A decision support system
appeared to influence the work reported in this paper. for on-line leakage localization. Environ. Model. Softw 60, 331e345. https://doi.
org/10.1016/j.envsoft.2014.06.025.
Mounce, S.R., Boxall, J.B., Machell, J., 2010a. Development and verification of an
Acknowledgements online artificial intelligence system for detection of bursts and other abnormal
flows. J. Water Resour. Plan. Manag. 136 (3), 309e318. https://doi.org/10.1061/
This work was financially supported by National Key Research (ASCE)WR.1943-5452.0000030.
Mounce, S.R., Day, A.J., Wood, A.S., Khan, A., Widdop, P.D., Machell, J., 2002. A neural
and Development Plan (Grant No. 2016YFC0400602), National
network approach to burst detection. Water Sci. Technol. 45 (4e5), 237e246.
Natural Science Foundation of China (Grant No. 51678425, https://doi.org/10.2166/wst.2002.0595.
51808396), and the UK Royal Society through an industry fellow- Mounce, S.R., Khan, A., Wood, A.S., Day, A.J., Widdop, P.D., Machell, J., 2003. Sensor-
ship to the last author (Ref: IF160108). Xiao Zhou's visiting at the fusion of hydraulic data for burst detection and location in a treated water
distribution system. Inf. Fusion 4 (3), 217e229. https://doi.org/10.1016/S1566-
University of Exeter is sponsored by China Scholarship Council 2535(03)00034-4.
(201806260130). Mounce, S.R., Mounce, R.B., Boxall, J.B., 2010b. Novelty detection for time series data
analysis in water distribution systems using support vector machines.
J. Hydroinf. 13 (4), 672e686. https://doi.org/10.2166/hydro.2010.144.
References Romano, M., Kapelan, Z., Savic, D.A., 2013. Geostatistical techniques for approximate
location of pipe burst events in water distribution systems. J. Hydroinf. 15 (3),
Al Shalabi, L., Shaaban, Z., 2006. Normalization as a preprocessing engine for data 634e651. https://doi.org/10.2166/hydro.2013.094.
mining and the approach of preference matrix. In: 2006 International Confer- Romano, M., Kapelan, Z., Savic, D.A., 2014. Evolutionary algorithm and expectation
ence on Dependability of Computer Systems, pp. 207e214. https://doi.org/10. maximization strategies for improved detection of pipe bursts and other events
1109/DEPCOS-RELCOMEX.2006.38. in water distribution systems. J. Water Resour. Plan. Manag. 140 (5), 572e584.
AWWA, 2009. Water Audits and Loss Control Programs. American Water Works https://doi.org/10.1061/(ASCE)WR.1943-5452.0000347.
Association, Denver. Rossman, L.A., 2010. An overview of EPANET version 3.0. In: Water Distribution
Bakker, M., Vreeburg, J.H.G., Van De Roer, M., Rietveld, L.C., 2014. Heuristic burst Systems Analysis 2010, Tucson, Arizona, USA. https://doi.org/10.1061/41203%
detection method using flow and pressure measurements. J. Hydroinf. 16 (5), 28425%293.
1194e1209. https://doi.org/10.2166/hydro.2014.120. Savic, Kapelan, Z.S., Jonkergouw, P.M.R., 2009. Quo vadis water distribution model
Berardi, L., Giustolisi, O., Kapelan, Z., Savic, D.A., 2008. Development of pipe dete- calibration? Urban Water J. 6 (1), 3e22. https://doi.org/10.1080/
rioration models for water distribution systems using EPR. J. Hydroinf. 10 (2), 15730620802613380.
113e126. https://doi.org/10.2166/hydro.2008.012. Schmidhuber, J., 2015. Deep learning in neural networks: an overview. Neural Netw.
Bicik, J., Kapelan, Z., Makropoulos, C., Savic, D.A., 2010. Pipe burst diagnostics using 61, 85e117. https://doi.org/10.1016/j.neunet.2014.09.003.
evidence theory. J. Hydroinf. 13 (4), 596e608. https://doi.org/10.2166/hydro. Srivastava, R.K., Greff, K., Schmidhuber, J., 2015. Training very deep networks. In:
2010.201. Advances in Neural Information Processing Systems, vol. 28, pp. 2377e2385.
Colombo, A.F., Lee, P., Karney, B.W., 2009. A selective literature review of transient- Wagner, J.M., Shamir, U., Marks, D.H., 1988. Water distribution reliability: simulation
based leak detection methods. J. Hydro Environ. Res. 2 (4), 212e227. https://doi. methods. J. Water Resour. Plan. Manag. 114 (3), 276e294. https://doi.org/10.
org/10.1016/j.jher.2009.02.003. 1061/(ASCE)0733-9496(1988)114:3(276).
CUWA, 2017. Urban Water Statistics Yearbook 2017 (In Chinese). China Statistics Walski, T.M., Brill, E.D., Gessler, J., et al., 1987. Battle of the network models:
12 X. Zhou et al. / Water Research 166 (2019) 115058

epilogue. J. Water Resour. Plan. Manag. 113 (2), 191e203. https://doi.org/10. application to a district water system. J. Water Resour. Plan. Manag. 136 (1),
1061/(ASCE)0733-9496(1987)113:2(191). 116e128. https://doi.org/10.1061/(ASCE)0733-9496(2010)136:1(116).
Wu, Y., Liu, S., 2017. A review of data-driven approaches for burst detection in water Yan, H., Wang, Q., Wang, J., Xin, K., Tao, T., Li, S., 2019. A simple but robust
distribution systems. Urban Water J. 14 (9), 972e983. https://doi.org/10.1080/ convergence trajectory controlled method for pressure driven analysis in water
1573062X.2017.1279191. distribution system. Sci. Total Environ. 659, 983e994. https://doi.org/10.1016/j.
Wu, Y., Liu, S., Wu, X., Liu, Y., Guan, Y., 2016. Burst detection in district metering scitotenv.2018.12.374.
areas using a data driven clustering algorithm. Water Res. 100, 28e37. https:// Zhou, X., Xu, W., Xin, K., Yan, H., Tao, T., 2018. Self-Adaptive calibration of real-time
doi.org/10.1016/j.watres.2016.05.016. demand and roughness of water distribution systems. Water Resour. Res. 54 (8),
Wu, Z.Y., Sage, P., Turtle, D., 2010. Pressure-dependent leak detection model and its 5536e5550. https://doi.org/10.1029/2017WR022147.

You might also like