A neural network approach to burst detection

Water Science and Technology Vol 45 No 4–5 pp 237–246 © IWA Publishing 2002
S.R. Mounce*, A.J. Day**, A.S. Wood**, A. Khan**, P.D. Widdop** and J. Machell***
* Department of Computing, Phoenix Building, University of Bradford, Bradford BD7 1DP, UK
** Department of Mechanical & Medical Engineering, University of Bradford, BD7 1DP, UK
*** Yorkshire Water, ROCC1, Western Way, Halifax Road, Bradford BD6 2LZ, UK. (E-mail:
www.laps.ac.uk)

Abstract This paper describes how hydraulic and water quality data from a distribution network may be
used to provide a more efficient leakage management capability for the water industry. The research
presented concerns the application of artificial neural networks to the issue of detection and location of
leakage in treated water distribution systems. An architecture for an Artificial Neural Network (ANN) based
system is outlined. The neural network uses time series data produced by sensors to directly construct an
empirical model for predication and classification of leaks. Results are presented using data from an
experimental site in Yorkshire Water’s Keighley distribution system.
Keywords Bursts; detection; leakage; neural networks; online monitoring; water distribution networks

Introduction
Problem definition
The issue of leakage management is of serious consideration to water companies because
major water loss can have negative economic, environmental and political consequences.
In the past, a relatively high level of leakage has been typical as the loss of treated water was
considered by water utilities on a solely economic basis and the significant reduction of
leakage level was not recognised as being economically viable. It has been estimated that
the ten major UK water companies lose on average 30 per cent of their treated water
through leaks (Butler and West, 1987). However, in the UK the Office of Water Services
(OFWAT) has in recent years imposed company specific mandatory leakage targets which
are currently reviewed annually. The water companies additionally now have a statutory
duty to conserve water and publish details of leakage reduction performance.
Consequently, the requirement for reduction in leakage levels has led to an interest in tech-
nologies for pipeline-integrity monitoring and failure detection. Current leakage manage-
ment methodologies within the water industry tends to be predominantly a manual process
which is resource intensive. For example data logging, data collection and transfer to point
of use, data analysis, reporting and then mobilisation of leakage teams (one or more leakage
teams may require several days to locate the leak). Subsequently a contractor must be
employed to effect repair. This overall process may take weeks or in extreme cases months.
Assuming an automated system performed daily data analysis, a realistic target for signifi-
cant burst lifetime would be a reduction to 2 days or less. Ideally, a state of the art detection
system should be able to:
• detect and localise bursts within an entire distribution system
• differentiate between bursts and unusual demands or system events
• differentiate between bursts which occur instantly and those which develop gradually
(leaks).
District flow metering is the most common method of leakage control in England and
Wales. Martin and Farley (1994) commented that in recent times, due to the opportunities
afforded by computer technology, continual monitoring of leakage levels through district 237

Downloaded from https://iwaponline.com/wst/article-pdf/45/4-5/237/425218/237.pdf

by guest
on 20 May 2020
 metering had become commonplace. The principle of district metering is based on the sub-
division of the distribution system into discrete zones or district meter areas (DMA), by the
permanent closure of valves, and the measurement of the flows into each zone. A DMA will
generally comprise an area containing 1,000 properties. The flow will usually be measured
by a fixed flow meter at the inflow point although temporary insertion meters and data log-
gers can be used instead. In general, flow sensors log the flow every fifteen minutes and
assumptions are made about the average usage. The most widespread operational use of
S.R. Mounce et al.

this flow data is in analysis of measured minimum night flows (NFM). The minimum night
flow is the lowest flow supplied to a hydraulically isolated supply zone. It is usually meas-
ured during the night hours between midnight and 5:00 am and includes leakage as well as a
certain minimum night consumption. Night flows are used because water use is at a mini-
mum and it is easier to identify and subtract the legitimate flows (e.g. 24 hr process loads).
If the night flow minus the legitimate flow is close to zero, the leakage must also be close to
zero. In contrast, unusual jumps in volumes will signify leakage in the absence of any other
factors. It is common to identify a nominal target nightline for each DMA. This target
implies a certain level of leakage activity with regard to the find and fix of hidden leaks. If it
is established that leakage has increased sufficiently to warrant further investigation, the
next stage involves a manual leakage detection which is carried out using methods such as
sounding, leak noise correlation and step testing.
This paper describes research into an automated computing system, based on artificial
neural network and sensor technology, for the detection and location of leakage in a water
pipeline distribution network. The methodology involves application of neural network(s)
to time series data generated by flow, pressure and other sensors (such as failure sensors
developed for the purpose). The target of the research is to automate district meter analysis
as well as allowing further localisation of the burst within the DMA prior to manual detec-
tion by leakage teams.
The complexity of a model for a water pipeline distribution network for a large urban
area is such that no single neural network can effectively monitor the behaviour of the
entire pipeline distribution network. Typically the distribution networks comprise several
DMAs with each zone containing one or more main pipelines, each of which feeds into one
or more sub-zones and finally into individual industrial and domestic consumers. Other
features such as a reservoir or pumping station may also be present. To manage such com-
plexity, it is necessary to build a number of neural networks arranged in a parallel and hier-
archical fashion to allow different types of information processing on the same data or
different subsets of data and at different levels of abstractions. This approach enables the
decomposition of a complex problem into a number of sub-problems. Each sub-problem is
then assigned to a functionally suitable network. At the top global level, level 1 (DMA
level), a neural network monitors zones and can detect leaks to within a zone (e.g. DMA)
but is unable to specify the exact position. Previous work has demonstrated the feasibility
of detection at this level using hydraulic data (Torsun et al. 1999). In the next level of
abstraction, level 2 (sub-zone level), a neural network will monitor sub-zones and can
locate leaks to within a sub-zone. Finally, at level 3 (pipe level), a neural network will mon-
itor one or more pipes and is able to accurately locate the position of the leak. Recent work
has seen attempts to link hydraulic models with on-line data from telemetry systems
(Skipworth et al., 1999; Orr et al., 1999). This type of online model could provide the pipe
level input. A very precise identification of leak location would rely on input to a neural
network of signals from mobile acoustic sensor instrumentation.
Initial work has focused on hydraulic data (from existing flow and pressure sensors) as
the principal data from which to provide features for leakage detection. Consultation with
238 domain experts, as well as analysis of historical data, had revealed hydraulic variables to be

Downloaded from https://iwaponline.com/wst/article-pdf/45/4-5/237/425218/237.pdf

by guest
on 20 May 2020
 good indicators of abnormal flow in the pipeline distribution system (including bursts and
leaks) at the DMA level. The accuracy of leak detection and location can be improved by
analysing signals from a greater number of sensors in a DMA, but the installation, opera-
tion and analysis resource costs are generally seen as prohibitive. Pressure sensors may be
deployed temporarily within a zone during field tests in order to provide data for hydraulic
model calibration. It has been proposed that if a burst occurs the variation in pressure signal
in association with the flow increase will enable more accurate location of leak (Machell,

S.R. Mounce et al.

1997). Research conducted at Bradford University for an EPSRC research project (as part
of the WITE programme) in association with Yorkshire Water is investigating the feasibil-
ity of using a number of failure sensors within a zone to improve leak localisation.
Consequently, we are now in a position to consider opacity (referring to the degree to which
a substance is opaque) of the water flow as an additional data source because low cost fail-
ure sensors have been designed and installed as described in an associated paper (“Low-
cost failure sensor design and development for water pipeline distribution systems”).
These sensors are providing data which can be correlated with actual events, with the ulti-
mate goal of leak localisation.

Experimental site
Yorkshire Water’s Keighley distribution system provides water services for around 26,000
customers. It contains approximately 110 km of pipeline and seven service reservoirs. The
distribution network has more recently been partitioned into 15 DMAs. Six of these cover
the city centre and immediate environs and one (K709) has been selected as the test bed for
a leakage detection case study. As well as state of the art hydraulic sensors (flow and pres-
sure are both logged at all DMA level sites in Keighley) the facilities were available to
install failure sensors designed and manufactured at Bradford University. The zone is
effectively disconnected from neighbouring zones resulting in an isolated environment for
tests – the area is supplied solely by Highfield SR. Figure 1 illustrates Zone K709 (from the
hydraulic model) showing the pipe geometry and sensor placement. Table 1 summarises
the sensors employed for the zone.

Methods
Data acquisition
Hydraulic data for the case study network was made available by Yorkshire Water. During
the initial research period this had to be via manual download from company databases but
later the opportunity to download directly using remote telemetry access was provided. The
data consists of 15 minute values for flow and pressure. Data collection for the opacity fail-
ure sensors was via manual download on site (with data loggers providing capacity for 25
days data at reading intervals of five minutes). In either case, time stamped files of readings
form the basis for further analysis.

Data pre-processing
Once data is acquired it is then passed through a number of filters (written in Matlab or
JAVA) as follows.
Table 1 Sensors in experimental site

Sensor type Number Logging interval Location

Flow 3 15 minutes Highfield SR, Albert St.

PRV, Gresley Rd. PRV.
Pressure 3 15 minutes As above.
Opacity failure sensors 10 5 minutes See Figure 1.
239

Downloaded from https://iwaponline.com/wst/article-pdf/45/4-5/237/425218/237.pdf

by guest
on 20 May 2020
 Highfield
Reservoir
500m
S.R. Mounce et al.

Albert St.
PRV Fixed Sensor Location
(flow and pressure)

Parkwood Failure Sensor Location

Gresley Rd. St.
PRV Low Mill Alternate installation for
Lane failure sensor

Figure 1 Zone K709 model with sensor locations

Data correction. First a filter removes any rogue values. Then, to ensure a continuous
data stream an ARIMA based filter fills in any periods of missing data. These missing
entries can be due to sensor failure or temporary power outages for example.
Normalisation. The input is normalised by means of linear re-scaling with mean and
standard deviation (Z-score) to a range of 0 to 1.
Reformatting. Finally, we reformat the input stream into a tapped delay line format in
order to prepare for neural network presentation.

ANN architecture and algorithms

For DMA level detection, a particular sensor (typically at the zone inlet) is modelled by one
neural network which learns to forecast the sensor output (time series prediction). A num-
ber of architectures were investigated for forecasting, including static and recurrent net-
works (Elman and Jordan networks) in a conventional manner. However, review of the
literature concerning Takens’ embedding theorem (Takens 1981) applied to forecasting
resulted in an alternative approach. Of course, the main result of Takens is that we can
predict future values of a time series xt+t solely on the basis of the observed past data
{xt, x t–t, . . .} without explicitly knowing the actual dynamics in the true system. However,
when we have a stochastic time series the best we can do is predict the conditional probabil-
ity P(xt+1Ωxt) which is the probability that we observe xt+1 at the next time step t+1 given
that we have observed the recent time series xt = (xt, xt–1,. . ., xt–m+1). Consequently neural
networks have been applied to the task of the prediction of the entire conditional probabili-
ty distribution of an unknown data generating process. These techniques have been applied
to time series prediction problems (Husmeier, 1999). A neural network model called the
mixture density network (MDN) as presented by Bishop (1994) has been selected for devel-
opment work. Essentially, the MDN is a mixture density model (semi-parametric model )
combined with a neural network. It consists of a two-hidden layer network – the first layer
with sigmoidal units and the second with Gaussian units. The target vector is clamped
directly to the Gaussian nodes. The network is trained using maximum likelihood (min-
imising the negative log of the likelihood). The mechanism for adapting the network
parameters is still backpropagation of error: the only difference being the error function.
240 The various parameters of the mixture models (mixing coefficients P(j), means µj and

Downloaded from https://iwaponline.com/wst/article-pdf/45/4-5/237/425218/237.pdf

by guest
on 20 May 2020
 variances sj) are governed by the outputs of the neural network which takes x as its input.
Once the network has been trained it can predict the conditional density function of the tar-
get data for any given value of the input vector i.e. we have a complete description of the
generator of the data, as far as predicting the value of the target vector. A GUI has been
developed for the MDN code developed using the NETLAB library (MATLAB toolbox).
The prediction generated by the MDN network is then used with the actual observed data to
form a comparative which is analysed by a classification module which indicates abnor-

S.R. Mounce et al.

malities in its output (based on a modifiable error sensitivity level). The values from the
classification stage from the various zones are analysed by a rules based system which indi-
cates overall system state. Figure 2 illustrates the sensor modelling sub-system.

Training
Algorithm. In the NETLAB library the mixture model used is Gaussian with a single
covariance parameter for each component. The mixture coefficients are computed from a
group of softmax outputs, the centres are equal to a group of linear outputs, and the vari-
ances are obtained by applying the exponential function to a third group of outputs. The net-
work is trained with a scaled conjugate gradient optimiser with the error function being the
negative log likelihood of the training data:

N N ÏÔ M ¸Ô
E = - ln L = -  ln pÊË x n ˆ¯ = -  ln Ì Â pÊË x n | jˆ¯ P j ˝
Ô j =1 Ô˛
()
n =1 n =1 Ó
Ï || x - m ||2 ¸
1 Ô Ô
( )
where p x | j = d /2
exp Ì -
2
2s j
j
˝
Ê 2ps 2 ˆ Ô
Ó ˛Ô
Ë ¯
The input is a lag vector and the target is a one day ahead prediction. The prediction evalu-
ated can either be the most “likely value” (centre of the highest component) or we can take
an average across distributions.

Training data. The MDN network is trained with a period of historical data for the sensor in
order to learn the distribution. This data set should be significant (e.g. at least several
weeks, preferably months) and should show reasonable well behaved behaviour. Every so
often (either automatically or at the behest of a user) the network is re-trained with a slight-
ly updated data set. In this way, current conditions are continually built into the model. If

Figure 2 Sensor modelling sub-system for DMA monitoring 241

Downloaded from https://iwaponline.com/wst/article-pdf/45/4-5/237/425218/237.pdf

by guest
on 20 May 2020
 conditions for the area being monitored change drastically (e.g. system alterations such as
valve modifications) then the training set will need to be started anew. In the example we
will consider the Albert St. PRV flow meter. This is the critical flow meter for monitoring
Zone K709. Figure 3 shows the training set used in the example covering a period of 10
months. Missing data has been filled (missing data segments appear dark). As can be seen,
the minimum night flow is fairly constantly just below 10 l/s for this period. First, the file
was normalised and reformatted as required for a tapped delay input. This data set was then
S.R. Mounce et al.

used to train a dedicated MDN. Experimentation revealed that a suitable number of hidden
units was 8 and a set of three Gaussians was sufficient. The main effect of variation of
hidden units was convergence time and increasing the number of Gaussians above 3 had
negligible effect on performance. There was little appreciable reduction of the error term
after around 100 complete cycles of the training set (containing 26,208 presentations). The
one day ahead prediction for the seen data was satisfactory at this stage. To test the per-
formance a test set of data was used for a period of approximately one month (26th
December 2000 to 30th January 2001). Results for this test file and supplemental sensor
data will be described in the next section.

Results and discussion

The test file was presented to the MDN network. Essentially, the output consists of a file of
one day ahead prediction for a moving one day window. The comparative formed by use of
the actual observed reading one day later should indicate the significant in-front of a burst.
Figures 4–9 illustrate the test data. Figure 4 illustrates the MDN one day ahead unseen
prediction output for a number of days in the test set for a well behaved period with no
abnormal events (hence the prediction is very close to the actual observed data). In Figures
5 and 6 the output is illustrated for burst events which occurred in the zone. As can be seen,

Training file for K709 after filtering

60

50

40
flow l/s

30

20

10

0
04-May
15-May
25-May
01-Feb
12-Feb
22-Feb
03-Mar
14-Mar
24-Mar

06-Sep
16-Sep
26-Sep
03-Apr
14-Apr
24-Apr
12-Jan
22-Jan

26-Aug
04-Jun
15-Jun
25-Jun
06-Jul
16-Jul
26-Jul
06-Aug
16-Aug

07-Oct

date

Figure 3 Training data for K709 Albert St. flow meter

Predicted v actual
0.7
0.65
Normalised flow

0.6
0.55 Actual
0.5 predicted
0.45
0.4
0.35
0.3
04/01/01
05/01/01
05/01/01
06/01/01
06/01/01
06/01/01
07/01/01
07/01/01
07/01/01
08/01/01
08/01/01
08/01/01
09/01/01
09/01/01
10/01/01
10/01/01
10/01/01
11/01/01
11/01/01
11/01/01

Date
242 Figure 4 Predicted v actual unseen data for standard conditions

Downloaded from https://iwaponline.com/wst/article-pdf/45/4-5/237/425218/237.pdf

by guest
on 20 May 2020
 a discrepancy immediately results (due to the one day ahead nature of the prediction) since
the abnormally high flow does not tally with the predicted level. Figure 7 shows the
calculated comparative (QONLINE/QOFFLINE) for the whole test period. These signals form
the input to a general classification module which analyses the signal and outputs a class,
for example burst/ no burst (Figure 8). This module has been trained on a selection of
comparatives from various meters. Both the two major events were classified as significant
bursts commencing 31st December and 22nd January respectively. Table 2 lists the

S.R. Mounce et al.

recorded repairs of mains conducted in the zone in January.
During the test period analysed, data from opacity failure sensors was also logged at a
number of sites around the zone (see Figure 1). Due to technical problems in several of the
installations, a full data set was not available for the whole of January. However, significant

Predicted v actual (burst 1)

0.7
0.65
0.6
Normalised flow

0.55 Actual
0.5 predicted

0.45
0.4
0.35
0.3
30/12/00 31/12/00 01/01/01 02/01/01
Date

Figure 5 Burst occurs 31st December

Predicted v actual (burst 2)

0.8
0.75
0.7
Normalised flow

0.65
0.6 Actual
0.55 predicted
0.5
0.45
0.4
0.35
0.3
21/01/01 22/01/01 23/01/01 24/01/01
Date

Figure 6 Burst occurs 22nd January

Figure 7 Comparative used for test period 243

Downloaded from https://iwaponline.com/wst/article-pdf/45/4-5/237/425218/237.pdf

by guest
on 20 May 2020
 NN output Post
processed
S.R. Mounce et al.

Burst
Alert

Figure 8 Classification module decision

1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
29/12/00
29/12/00
29/12/00
29/12/00
30/12/00
30/12/00
30/12/00
30/12/00
31/12/00
31/12/00
31/12/00
01/01/01
01/01/01
01/01/01
01/01/01

Figure 9 Correlation for diurnal cycle between opacity and flow in Gresley Rd. sub-zone (normalised data
from Figure 10 29/12–01/01)

Table 2 Repaired mains January 2001 in K709

Burst 1 Burst 2 Burst 3

Type Repaired main Repaired main (dowel piece) Repaired main

Emergency allocated 4th January 16th January 25th January
Task finished 7th January 16th January 29th January
Priority 3 1 1
Description Leakage (visible) Leakage (visible) Leakage (not
visible)

correlation between abnormal flows and the opacity was witnessed on several sensors dur-
ing the period. Figures 9 and 10 illustrate the output for two of the failure sensors, along
with the Gresley Road flow meter which monitors a sub-area in K709. The data has been
sampled at a 15 min interval rate. The graph shows two interesting events. Firstly, the 31st
December to 3rd January abnormal flow previously discovered is noticeable on the Gresley
Road flow meter. Note that the opacity failure meters have a visible diurnal cycle which ini-
tially increases significantly in response to the velocity increase through the pipe. Even
with 15 minute reading interval there is a noticeable phase difference for these two geo-
graphically close sensors. Secondly, observe the sudden flow surge on the 3rd of January
and the corresponding spikes in the opacity data. The source of this event is unknown,
though it is possibly industrial usage as this section of the zone has several industrial cus-
tomers. Figure 9 shows the first 4 days data of Figure 10 which has been normalised to illus-
244 trate the high correlation for the diurnal cycle. Note that the opacity cycle is lagged behind

Downloaded from https://iwaponline.com/wst/article-pdf/45/4-5/237/425218/237.pdf

by guest
on 20 May 2020
 Sensor output 29/12/00 to 5/01/01
0.7 8

7
0.6

6
0.5

S.R. Mounce et al.

Opacity ntu

0.4
Low Mill Lane

Flow l/s
4 Parkwood St.
Gresley Rd (flow)
0.3
3

0.2
2

0.1
1

0 0
29/12/00 30/12/00 31/12/00 01/01/01 02/01/01 03/01/01 04/01/01
date

Figure 10 Correlation between opacity and flow in Gresley Rd. sub-zone

the flow. Also, observe the high opacity response the morning following the commence-
ment of abnormal flow.

Conclusions
Many water utility companies are beginning to amass a great deal of data by means of
remote sensing of flow, pressure and other attributes. In the past, this information has
mainly been used for district level management including water audit and calibration of
hydraulic models. However, the opportunity exists to build on existing analysis of mini-
mum night flow for leakage detection using an automated neural network based pattern
recognition system. In this paper, by means of an ongoing experimental field study, we
have indicated how this may be accomplished at the district meter level. Further, we have
demonstrated that additional in-zone failure sensors can provide a set of associated signals
possessing magnitude and phase information which, with suitable quantities of example
training data, will enable further understanding and localisation of events within a DMA.
Future work will extend the system by constructing a neural network for analysis of failure
sensor output. This will require an expanded data set: to be accomplished both by continua-
tion of monitoring within the zone (along with discovery of system events) and by specific
controlled trials in the area in which leak events are simulated. A multi-input TDNN is the
expected architecture for this component of the system. Finally, analysis of an online
hydraulic model’s pipe level data will be attempted for a more refined location.

Acknowledgements
The authors wish to acknowledge the support given for this research by the Engineering and
Physical Sciences Research Council (UK) under its Water Infrastructure and Treatment
Engineering (WITE) Initiative. The authors also acknowledge the support of Yorkshire
Water and are grateful to the directors of the company for permission to use the data and
publish this paper. Finally all the authors wish to acknowledge the intellectual lead provid-
ed to this project by the late Professor I.S. Torsun.

References
Bishop, C.M. (1994). Mixture density networks. Technical report NCRG/94/004, Department of Computer
Science and Applied Mathematics, Aston University, Birmingham, UK. 245

Downloaded from https://iwaponline.com/wst/article-pdf/45/4-5/237/425218/237.pdf

by guest
on 20 May 2020
 Butler, M. and West, J. (1987). Leakage prevention and system renewal, Pipeline Management Seminar,
Pipeline Industries Guild, 1987.
Husmeier, D. (1999). Neural Networks for Probability Estimation, Springer-Verlag.
Khan, A., Widdop, P. D., Day, A. J., Wood, A. S., Mounce, S. R., and Machell, J. (2002). Low-cost failure
sensor design and development for water pipeline distribution systems. Wat. Sci. Tech. 45(4–5) 207–215
(this issue).
Machell, J. (1997). “Research and development proposal for an on-line leakage detection system”, YWS
S.R. Mounce et al.

internal report.
Martin, M. and Farley, L. (1994). Improving water management in leakage control. Proc. of Hydrotop 94,
Marseille, France.
Orr, C., Bouulos, P., Stern, C. and Liu, P. (1999). Developing real-time models of water distribution sys-
tems. Modeling and optimisation applications (Water engineering and management series, 3–4) by D.
Savic (Editor) and G. Walters (Editor). Vol. 1, Research Studies Pr.
Skipworth, P.J., Saul, A. and Machell, J. (1999). Practical applications of real time hydraulic modelling of
water distribution networks. Comadem, 1999.
Takens, F. (1981). Detecting strange attractors in fluid turbulence. Dynamical Systems and Turbulence,
Berlin, D. Rand and L. Young Eds. Springer.
Torsun, I.S., Day, A.J., Wood, A.S., Khan, A. and Mounce, S.R. (1999). Neural network for leak detection
in pipeline distribution networks. Presented at the 1st Annual WITE Workshop, part of International
Conference on Computing and Control for the water industry CCWI’99 held at University of Exeter, UK.

246

Downloaded from https://iwaponline.com/wst/article-pdf/45/4-5/237/425218/237.pdf

by guest
on 20 May 2020