Professional Documents
Culture Documents
ANN-Fuzzy For DMA Flow Meter Data An PDF
ANN-Fuzzy For DMA Flow Meter Data An PDF
net/publication/221936179
An Artificial Neural Network/Fuzzy Logic system for DMA flow meter data
analysis providing burst identification and size estimation
CITATIONS READS
12 249
3 authors:
John Machell
The University of Sheffield
63 PUBLICATIONS 790 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
A feasibility study for the creation of DMAs within a Chinese water supply system View project
All content following this page was uploaded by Steve Mounce on 15 March 2019.
ABSTRACT: This research demonstrates the application of Artificial Neural Networks (ANNs) and Fuzzy
Logic for the detection and classification (leakage rate) of pipe bursts from DMA flow data. In contrast to ex-
isting methodologies that are reliant on human data analysis and interpretation, this automated approach is
able to provide efficient and consistent analysis of large data volumes. This Artificial Intelligence (AI) analy-
sis system was applied to a validation data set (generated by simulated flushing) to verify the burst estimation
capability and then to data from an eight month period of operational flow data. The technique successfully
identified and estimated the magnitude of burst events; some of which could not have been detected by man-
ual analysis. The results of the study show that the ANN system is a powerful tool for input-output mapping
and, when integrated with near real time communications, could be effectively used for rapid determination of
abnormal flow levels.
5.00
size confidence mate mate
l/s
0.6 4.00
0.4
3.00 (l/s) (l/s) (l/s)
2.00 20:58pm- 6 21:00pm-03:15am 4.89 3.71
0.2 02:00am (14/15th May)
1.00
0 0.00 (14/15th May) 98%
15/05
16/05
17/05
18/05
19/05
20/05
21/05
22/05
23/05
21:01pm- 6 20:45pm-02:30am 6.77 5.36
02:00am (15/16th May)
NFM ratio Flush rate l/s (15/16th May) 99.7%
22:00pm- 7 21:30pm-03:00am 7.26 5.37
01:59am (16/17th May)
Figure 1: NFM ratio with flushing rates
(16/17th May) 99.7%
16:23 pm- 6 02:00pm- 5.64 N/A
4.3 Results and discussion 16:56pm 05:00pm (16th
(16th May) May) 91%
The MDN system was applied to data measured by 22:20pm- 5 22:00pm-02:00am 5.44 5.37
the DMA meter. Previous work (Mounce, 2003 & 01:20am (17/18th May)
(17/18th May) 99.7%
2005) resulted in an expectation of a confidence of
95% or more for classification of 5-7 l/s events in It should be noted that there is inherent uncer-
this size DMA. Figure 2 show the FIS output for one tainty of actual flow rate due to inaccuracy with the
of the events for the DMA, for 90%, 95% and 99% hydrant flow mater. However, a brief perusal of Ta-
confidence intervals. The graph plots the flush rate ble 2 reveals that the AI system estimate more
(dotted line) and the FIS output for a 2.5 hr time closely matched the recorded flush size than the
window beginning at that output point on the time NFM – the average error for the four main flushes
axis (with the normalised flow rate shown for refer- being approximately 10% for the AI system, and ap-
ence). The analysis successfully detected all the proximately double at 20% for the NFM based esti-
events. mate. The short duration day time event was also de-
99% 95% 90% Normalised flow Burst
tected with a good sizing, albeit with reduced
1 7.00
confidence.
0.9
6.00
0.8
4.00
0.5
3.00
0.4
0.3 2.00
5.1 Data acquisition
0.2
1.00
Historical data was collected from 10 DMAs in a
0.1
water supply system for an eight-month period. The
0 0.00
data consisted of time stamped files of 15 minute
12:00 AM
1:15 AM
2:30 AM
3:45 AM
5:00 AM
6:15 AM
7:30 AM
8:45 AM
10:00 AM
11:15 AM
12:30 PM
1:45 PM
3:00 PM
4:15 PM
5:30 PM
6:45 PM
8:00 PM
9:15 PM
10:30 PM
11:45 PM
1:00 AM
2:15 AM
3:30 AM
4:45 AM
6:00 AM
readings.
In order to evaluate the effectiveness of the
Figure 2: DMA flow meter FIS output for 14th – 15th May analysis, operational and minimum night flow data
(burst 1) versus flushing period was collected for comparison. A record of repairs
was obtained from the Repair and Maintenance
The classification module in MDNGUI takes the (R&M) database for the period of interest. The mag-
FIS output and applies a user-defined threshold in nitudes of the events recorded in this register were
order to make the burst classification. Table 2 pro- unknown; ranging from a major burst to a minor re-
vides the time details of the classification for the pair of a valve. Only operations specifically tagged
DMA flow sensor, which are the limits of all 2.5 hr as mains repairs were considered. The R&M record
windows classed as containing the simulated burst logs the repair completion time only, hence it was
(the first confidence level to contain the full extent unknown when a burst commenced from this infor-
of the burst). In addition, the actual leak size is pro- mation, both start and end information is produced
vided from the hydrant flow meter. Finally, an esti- by the AI analysis. The NFM ratio was calculated as
mate for leak level for the evening is obtained by us- for the validation set, though from mid-night to 5am.
ing the NFM ratio.
5.2 Methods the basis of either an R&M record or significant
In order to make full use of the data sets, the time NFM ratio.
series was bisected and used for both training and The results show that the positive classifications
testing. Each half of the data, comprising around 4 by the AI system in some cases correspond to actual
months of readings, was used once as the training set bursts. Positive classifications then show a good cor-
and once as the test set. In this way the whole data relation between estimated burst size and the size
set could be analysed for bursts in an unseen man- indicated by a simplistic NFM ratio. There were two
ner. However, it should be noted that the non- exceptions to this when very large magnitude bursts
stationarity of some sensor data was an issue (a time were underestimated in size due to the technique
series is described as stationary if the behaviour of performing a truncation in the normalisation routine.
the series does not change with time i.e. the mean It will be possible to solve this problem by using raw
and variance do not change over time). The MDN data direct from the database in the next version of
was trained in accordance with the details in the software for the estimation (online operation).
Mounce et al. (2006). It is difficult to assess actual system performance
The burst estimate from NFM ratio was calcu- with these types of data sets as the work record is
lated as for the validation data. not comprehensive in the sense that not all events
will be recorded there. It is questionable how accu-
rately start and end times will be recorded. Further,
5.3 Results and discussion the size of burst that was repaired is unknown. It
Table 3 provides a summary of the results for the should be noted that the validation data set only con-
complete 8 month historic case study. A row is in- tained flushing events during the night which it
cluded in the table for any event in the R&M record might be argued make the events easier to detect,
(corresponding to actual repairs) and any event de- whereas the case study data sets are not similarly re-
tected by the AI system with an 85% confidence and stricted and thus demonstrate that the technique is
above (using a threshold of 0.8 on the defuzzified effective irrespective of time of day that the burst
output for the 48 window FIS). The table also in- occurs.
cludes the NFM ratio for these events/detections, the The NFM ratio is a useful indication of abnormal
table does not necessarily include all NFM devia- flow, however the ratio does not have a universal in-
tions from unity. Two columns report the correlation terpretation since its meaning varies for different
between the FIS and both the R&M record and NFM meter areas, taking no account of the variation for a
ratio. The next two columns report the estimated particular meter. This means that a value of 1.1 may
burst level for detected events for the AI system and indicate a very unusual event for one meter but not
the corresponding calculation (as from the validation for another (zone analysts do become familiar with
set) from the NFM ratio. It should be noted that the interpreting these values for particular DMAs). In
AI analysis is conducted over a window for historic contrast, the MDN can predict the conditional prob-
ability density function of the future flow from the
data, hence the burst estimate varies as abnormally
past data (the input vector) and the density function
high flow enters this window. Hence the AI results
can then be used to assign a confidence to the ob-
in table 3 provide the maximum estimated burst
served flow values. Hence, the MDN learns the dis-
level for the period with positive burst detection at
tribution of a particular meter potentially leading to
that confidence level.
improved analysis results.
From Table 3 it can be seen that the source data
For the case study, the training file was assem-
for comparison is limited, with discrepancies be-
bled using all the data so that the whole data set
tween the logged repairs and the NFM ratio leading
could be analysed for bursts in an unseen manner.
to uncertainty over the occurrence or otherwise of
Although this provided satisfactory results, these can
actual burst events. However, when the work record
be improved by retraining the ANN regularly with a
and NFM agree that an event did occur, the AI
slightly updated data set for the DMA. In this way,
analysis, apart from two exceptions which were an
current conditions could be continually built into the
artifact of poor training sets, returned an event and
model – particularly with access to regular online
provided close agreement between the recorded
data. If conditions for the supply system being moni-
event end and repair times. The correlation between
tored change drastically (e.g. valve position modifi-
the NFM and the AI analysis is predominately good
cations), then the training set will need to be started
(i.e. indicating abnormal flow for the DMA but not
anew or generated using mathematical simulation
necessarily indicated in the repair record). It should
techniques.
be noted that the AI column reports all events de-
The techniques described in this paper are only
tected by the system in the period of analysis (>85%
applicable to ‘standard’ DMA, residential, and in-
confidence) and the correlation with NFM suggests
dustrial monitoring points. The system has been ap-
minimal false positives. No false positives were re-
plied as a first step to flow data only. Analysis
ported for the DMAs with complete test data sets, on
Table 3. ANN/FIS analysis burst detection for historical data set with estimated burst size
DMA R&M record repair NFM AI positive classification period % AI / R&M AI / NFM Burst esti- Burst esti-
date ratio Conf Agreement? Agreement? mate mate
NFM (l/s) AI (l/s)
DMA A MR35 24/11/05 0.94 None >85% - NO YES -
DMA B MR35 21/11/05 0.79 None >85% - NO YES -
MR35 26/11/05 1.20 None >85% - NO ? -
MR35 4/01/06 1.09 None >85% - NO YES -
MR35 9/03/06 1.12 None >85% - NO YES -
MR35 28/04/06 0.99 None >85% - NO YES -
MR35 4/05/06 1.00 None >85% - NO YES -
DMA C None 1.5-2 12/3-26/3 None >85% - YES NO -
DMA D None 19.81 06/03/06 09:15 – 07/03/06 00:45 99 NO YES 2.01 2.254
None 17.56 20/03/06 09:15 – 21/03/06 01:15 99 NO YES 1.77 2.203
DMA E MR35 19/04/06 1.08 None >85% - NO YES -
MR35 22/04/06 1.92 21/04/06 17:00 – 22/04/ 07:30 99.7 YES YES 0.73 2.124
DMA F None 1.39 30/12/05 18:45 – 31/12/05 06:45 85 NO YES 0.44 0.684
MR35 10/01/06 1.63, 1.55, 1.50 07/01/06 16:00 – 08/01/06 04:45 97 YES YES 0.71 0.999
MR35 26/01/06 1.31 None >85% - NO ? -
MR35 19/02/06 9.34 16/02/06 19:30 – 17/02/ 06 21:15 99.7 YES YES 9.40 2.592
(17/02 but cancelled)
DMA G MR35 24/12/05 1.62 23/12/05 14:00 – 24/12/05 06:15 96 YES YES 0.95 0.74
MR35 09/02/06 0.93 None >85% - NO YES -
MR30 12/05/06 19.38 11/05/06 21:15 – 12/05/06 11:45 99 YES YES 28.24 3.126
None 1.21 03/06/06 14:30 – 04/06/06 05:30 99 NO ? 0.33 0.602
DMA H MR35 22/11/05 0.87 None >85% - NO YES -
MR35 13/01/06 1.24, 1.19 None >85% - NO ? -
None 1.37 03/06/06 09:30 – 03/06/06 22:00 92 NO YES 0.60 1.301
DMA I None 1.65 27/12/05 20:45 – 28/12/05 11:15 97 NO YES 0.52 1.120
None 1.26 15/01/06 21:15 – 16/01/06 10:30 94 NO ? 0.21 0.940
None 1.75 19/02/06 18:30 – 20/02/06 07:15 99 NO YES 0.60 1.131
None 1.47 01/03/06 19:15 – 02/03/06 09:15 98 NO YES 0.37 1.055
None 1.29 27/04/06 08:30 – 27/04/06 21:15 97 NO ? 0.13 1.155
DMA J None 1.54 06/06/06 13:45 – 07/06/06 06:30 90 NO YES 0.34 0.708
within the context of the overall distribution system based on NFM ratio although the actual burst rate
is confused by lack of system knowledge and de- was unknown.
tailed operational data. To obtain added value and a
clearer picture of events, future work will assemble a
system knowledge base able to interpret both flow
and pressure analysis (via a rules-based system). REFERENCES
These techniques will offer improved interpreta-
tion leading to better identification, for example in- Bishop, C.M. (1994). “Mixture density networks”. Technical
terpretation of results from cascading DMA struc- report NCRG/94/004, Department of Computer Science and
Applied Mathematics, Aston University, Birmingham, UK.
tures reducing the potential of false alarms. The sys- Blueprint for Water (2006).
tems could also be readily expanded for other avail- http://www.blueprintforwater.org.uk/
able data such as pressures and operational data i.e. House of Lords (2006). Water Management, Volume I: Report,
pump status and reservoir levels. It is envisioned that House of Lords Science and Technology Committee, HL
this will be achieved through integration of ANN, Paper 191-I, 6th June 2006.
FIS and mathematical simulation of the network (su- European Parliament (2000). EU Water Framework Directive,
Directive 2000/60/EC, 23/10/2000. Official Journal (IJ L
ing hydraulic modelling software) in order to pro- 327), 22 December 2000.
vide wider system performance knowledge. Hornick, K., Maxwell, S. & Halbert, W. (1989) Multilayer
The work described does not rely on hydraulic feedforward networks are universal approximators. Neural
modelling software, however it is seen as a com- Networks 2: 359-366.
ponent of an overall system that would include such Husmeier, D., and Taylor, J.G. (1998). “Neural Networks for
Predicting Conditional Probability Densities: Improved
models for derivation of greater system knowledge. Training Scheme Combining EM and RVFL”, Neural Net-
It is proposed that the 'intelligent' analysis system works, 11(1): 89-116.
developed can be readily applied to an online en- ICE (2006). State of the Nation 2006, ICE.
vironment providing the capability to proactively http://www.ice.org.uk/state_of_the_nation/water_wastewater_u
manage water supply systems. This is being ex- k_overview.asp
plored in conjunction with water company trials of a IPCC WG I (2007). Summary for policymakers Climate
change 2007: Impacts, Adaptation and Vulnerability.
GRPS communications system uniquely sup-plying Available from WWW: http://www.ipcc.ch/
the temporal resolution and updating required. Mamdani, E.H. and Assilian, S. (1975). “An experiment in lin-
guistic synthesis with a fuzzy logic controller.”. Interna-
tional Journal of Man-Machine Studies, 7(1): 1-13.
6 CONCLUSIONS Masters, T. (1993). Practical Neural Network Recipes in C++.
San Diego: Academic Press, CA.
Mounce S.R., Khan A, Wood A.S., Day A.J., Widdop P.D.,
− A system has been developed that can accept and and Machell J. (2003). “Sensor-fusion of hydraulic data for
store flow data from external sources and auto- burst detection and location in a treated water distribution
matically convert it into useful operational infor- system.” International Journal of Information Fusion, 4(3):
mation i.e. for the detection and size estimation of 217-229.
bursts. Benefits include: robustness to noisy or Mounce, S.R. (2005). A hybrid neural network fuzzy rule-
based system applied to leak detection in water pipeline dis-
incomplete data; large capacity for routine and tribution networks, PhD thesis, University of Bradford.
repeatable analysis; data driven self learning and Mounce, S.R., and Machell, J. (2006). “Burst detection using
updating; confidence interval for classifications hydraulic data from water distribution systems with artifi-
produced. cial neural networks.” Urban Water Journal, 3(1): 21-31.
− A validation data set (from a series of artificial Mounce, S. R., Machell, J. & Boxall, J. (2006) Development
Of Artificial Intelligence Systems For Analysis Of Water
flushing trials) demonstrated the accuracy of size Supply System Data. Proceedings of the 8th Water Distri-
estimation routines. The error for estimation of bution System Analysis Symposium, Cincinnati USA, Au-
the actual flushing rate was only approximately gust 27-30.
10% for the AI system compared to that of Nabney, I. (2001). NETLAB: Algorithms for pattern recogni-
around 20% from a calculation based on an Night tion, Springer-Verlag, UK.
Flow Minimum (NFM) ratio. Water Research Centre, Engineering and Operations Commit-
tee (1994a) UK Water Industry: managing leakage sum-
− The system was successfully applied to a case mary report (Report A). Swindon.
study. A significant number of events detected by
the AI system were exactly correlated with re-
pairs described in the repair and maintenance da- ACKNOWLEDGEMENT
tabase. A higher number were also correlated
with the NFM ratio: an indicator of abnormal The authors wish to acknowledge the support given for this re-
flow for the DMA but not necessarily captured in search by Yorkshire Water Services Ltd., UK.
the repair data base. From the combined correla-
tion with the repair and maintenance database and
NFM no false positives were reported. The leak
size estimates correlated closely with an estimate