You are on page 1of 7

International Conference on Engineering, Science and Technology

October 7-10, 2019 www.iconest.net Denver, CO, USA

A Contamination Risk Model for Water Distribution Networks


Thikra Dawood
Purdue University, U.S.A.

Emad Elwakil
Purdue University, U.S.A.

Hector Mayol Novoa


National University of St Augustin of Arequipa, Peru

José Fernando Gárate Delgado


National University of St Augustin of Arequipa, Peru, josegarate@gmail.com

Abstract: The intrusion of contaminants into water supply networks can compromise the water quality and may
cause fatal hazards to urban communities. The risk of water quality failure has been de facto an intricate and
challenging field of research. This complexity is due to the fact that these networks are subjected to various
physical, environmental, and operational mechanisms. In addition to the limited and uncertain data, pertaining to
the deterioration and performance of these structures. The objective of this paper is to develop a risk-based
assessment model grounded in Artificial Neural Networks (ANNs) and pattern recognition techniques to estimate
the risk of failure in water distribution networks. The model is developed using seven years of historical data
collected from the city of El Pedregal in Peru. The ANN-based model is validated against different statistical error
terms, in which the results proved to be satisfactory with an average validity performance of 91%. The proposed
model is a decision support tool, expected to assist infrastructure managers and civil engineers in their future plans
and decision making.

Keywords: Pipe Failure, Risk Analysis, Prediction Models, Water Networks

Introduction
The safety and quality of drinking water is the main concern of water management authorities. A standard water
supply system encompasses; the surface water or groundwater, transmission pipelines, treatment plants, and a
water distribution network (WDN) (Sadiq et al. 2008). The WDN is the most costly part of the water supply
system (Giustolisi et al. 2006), which comprises pipelines and distribution tanks. Although contamination can
occur at any point in the water supply system, the risk of failure in the distribution phase is considered the most
critical because it is the phase of delivery to the customer (Sadiq et al. 2008).

Generally, the risk is defined as a measure of the probability of an occurring event and the amount of adverse
consequences (Kirchhoff and Doberstein 2006). The process of estimating the rate of occurrence and physical
consequences of undesirable incidents is referred to as the risk analysis (Ricci et al. 1981). The risk of water
contamination and condition assessment have been very complex tasks (Ana & Bauwens, 2010), because WDNs
are subjected to numerous physical, environmental, and operational mechanisms. These buried networks consist
of extended pipes of various materials and ages, with limited and ambiguous records related to their structural
performance. Moreover, the failure rates and water contamination are very difficult to be diagnosed due to the
huge time interval between the failure incidence and the consequence observation (Sadiq et al. 2007). This
accentuates the need for a decision support tool capable of predicting pipes’ contamination and water quality
(Dridi et al. 2009). This process may include various techniques ranging from simple statistical methods to highly
sophisticated systems such as machine learning and artificial intelligence.

This paper proposed a novel model grounded in machine learning techniques and ANNs to estimate and predict
the risk of contamination in water distribution networks. First, the sources of failure in water supply systems are
investigated and categorized. Second, the historical data pertaining to water pipelines and contamination
frequency are collected and analyzed. Third, Artificial Neural Networks (ANNs) are leveraged and trained to
estimate the risk of quality failure in WDNs. The developed model provides a framework to prioritize and optimize
maintenance, rehabilitation, or replacement action of water infrastructure assets.

31
Dawood, T., Elwakil, E., Novoa, H.M. & Delgado, J.F.G. (2019). A Contamination Risk Model for Water Distribution Networks. In
M. Shelley & V. Akerson (Eds.), Proceedings of IConEST 2019--International Conference on Engineering, Science and Technology
(pp. 31-37). Monument, CO, USA: ISTES Organization.
International Conference on Engineering, Science and Technology
October 7-10, 2019 www.iconest.net Denver, CO, USA

Background
Factors that Contribute to Water Pipes Failure

The InfraGuide (2003) report classified the causes of water systems deterioration and failure into three groups;
physical, environmental, and operational, each of which included several factors, as presented in Figure 1.

1) Physical: pipe material, pipe wall thickness, pipe age, pipe vintage, pipe diameter, type of joints, thrust restraint,
pipelining and coating, dissimilar metals, pipe installation, and pipe manufacture.

2) Environmental: pipe bedding, trench backfill materials, soil type, groundwater existence, climate, pipe location,
disturbances in the vicinity, stray electrical currents, and seismic activity.

3) Operational: water pressure, leakage, water quality, flow velocity, backflow potential, and operational and
maintenance practices.

Physical

Pipe Pipe Wall Pipe Pipe Pipe Type of Thrust Pipe Lining Dissimilar Pipe Pipe
Material Thickness Age Vintage Diameter Joints Restraint & Coating Metals Installation Manufacture

Risk
Environmental
of Failure

Stray
Pipe Trench Soil Ground Pipe Seismic
Climate Disturbances Electrical
Bedding Backfill Type Water Location Activity
Currents
Operational

Water Water Flow Backflow O&M


Leakage
pressure Quality Velocity Potential Practices

Figure 1. Factors of Water Quality Failure

Artificial Neural Networks (ANNs)

The artificial neural network is a soft computing technique for predicting the risk of failure through analyzing
complex nonlinear relationships between the input and output datasets. ANNs are inspired by the biological neural
system of the human brain in processing the information (Ampazis et al. 1999), and they have a significant
tolerance for noise. One of the most common ANN configurations are the Back Propagation Neural Networks
(BPNNs). The BPNN comprises input, hidden, and output layers, each of which includes a number of nodes. The
nodes are connected to the nodes of the adjacent layer through connections (Elwakil et al. 2009). The ANN trains
itself through an iterative learning process of adjusting weights and biases. Thus, it can recognize the patterns
between pipes attributes (inputs) and pipes breaks (output) (Jafar et al. 2010).

Numerous researchers have been inspired to explore the ANN computational capabilities, precisely to predict
pipes breaks and failure rates. Winkler et al. (2018) applied the boosted decision trees to model the pipe failure in
water distribution networks. Sattar et al. (2016) proposed a model to predict time to failure of three water mains
types (i.e., Ductile Iron, Cast Iron, and Asbestos Cement) via Gene expression algorithms. Kabir et al. (2015)
considered structural integrity, hydraulic capacity, water quality, and consequence factors to evaluate the risk of
water mains failure. The soft computing techniques were leveraged by Deng et al. (2011) to generate primary
probability assignment, and evaluate the risk of contamination in water supply systems. Fares and Zayed (2010)
incorporated 16 risk factors to develop a hierarchical fuzzy expert system for urban water networks. Christodoulou
et al. (2010) integrated artificial neural networks (ANNs) and survival analysis to propose a proactive risk
assessment model. Sadiq et al. (2007) analyzed the risk of water quality failures and system contamination by
combining the fuzzy set theory and the evidential reasoning approach.

Data Collection
In order to construct the contamination risk model, data were specifically collected from the water supply system
of the city of El Pedregal, Peru. The geographical location of El Pedregal is shown in Figure 2. The water systems
of El Pedregal was built in August of 2012. The main distribution network has 21266.4 ft. of PVC pipes with

32
International Conference on Engineering, Science and Technology
October 7-10, 2019 www.iconest.net Denver, CO, USA

diameters ranging from 3.9 inches to 5.9 inches, while the secondary network consists of 36512.5 ft. of pipes with
diameters ranging from 3.0 to 3.9 inches. Data on both networks, i.e., main and secondary over the period of 7
years, were collected, tabulated, and analyzed.

It was reported that in El Pedregal, there was a disease outbreak due to water consumption that caused infectious
intestinal disease. The incidents reported in the years 2014, 2015, 2016, 2017, and 2018 were 3308, 3120, 2550,
2358, and 2830 cases, respectively. According to Kleiner (1998), water quality may be compromised because of
one or a combination of the following mechanisms: a) contaminants intrusion into the water system; b) pipes
corrosion or leaching; c) disinfection byproducts formation; d) microbial and/or chemical permeation; e) organic
components penetration; f) biofilm regrowth; and g) intentional contamination.

Accordingly, the registered incidents database, along with the factors that caused the quality failure of the network,
were taken into account to develop the contamination risk model. These factors encompassed; pipe age, pipe
diameter, wall thickness, water pressure, water quality, operational and maintenance (O & M) practices, and the
number of quality failures.

Figure 2. Study Area Location on Google Earth

Research Objective and Methodology


The main objective of this paper is to create a model for predicting the risk of contamination in urban water
systems. To achieve this objective, the following sub-objectives are defined:

1) Identify and study the types and causes of water quality failure.
2) Build and design the ANN model architecture.
3) Develop a contamination risk of failure model.
The proposed methodology encompasses several steps starting from the data collection to the modeling of water
system contamination. The overall framework of the methodology is represented in Figure 3. First, the data
obtained from the city of El Pedregal are classified and analyzed. The factors that have a significant impact on the
deterioration of water pipelines are identified as 1) pipe age, 2) pipe diameter, 3) pipe thickness, 4) water pressure,
5) water quality, 6) O & M, and 7) the number of water quality failures. These seven factors are categorized as
the inputs to the ANNs, whereas the output is the risk of quality failure. It is worth noting that some of the model’s
attributes are qualitative in nature (water quality, O & M, risk of quality failure). However, the rest of the attributes
have quantitative values but with different units. Therefore, it is imperative to convert the qualitative into
quantitative values, as well as to normalize the whole model factors, as illustrated in Table 1. The second step
involved supplying the ANN machine with the input and output data sets, and each data set is divided into three
sets; namely training, validation, and testing.

In general, the training dataset is used to train the ANN so that it can recognize the pattern between data entries
and output variables. The validation dataset is utilized to validate the results produced in the training stage,

33
International Conference on Engineering, Science and Technology
October 7-10, 2019 www.iconest.net Denver, CO, USA

whereas, the testing dataset is used to measure the network’s performance during and after the training stage
(Dawood et al. 2018). It followed by designing the ANN architecture and setting the network for training in order
to fit the inputs to outputs. The ANN will train itself through different scenario patterns; hence, it will be used to
predict the risk of water quality failure.

Pipeline Failure Factors ANN Architecture

Artificial Neural
Networks

Data Base
Cross Entropy

Performance NO
Optimized?
Testing Validation Training
Data Data Data
YES
Select Best ANN
Based on
Gradient Value

Testing Error NO
Accepted?

YES

Risk of Failure

Figure 3. Flowchart of the Proposed Method

34
International Conference on Engineering, Science and Technology
October 7-10, 2019 www.iconest.net Denver, CO, USA

Table 1. Model Factors’ Values


Factor Category Value
Pipe Age (Years) 0-1 1
2–3 2
4-5 3
6-7 4
Pipe Diameter (Inches) 0-4 1
5-8 2
9 - 12 3
13 - 16 4
Pipe Thickness (Inches) 0 – 0.5 1
0.6 – 1.0 2
1.1 – 1.5 3
Water Pressure (Bars) 0-1 1
2–3 2
4-5 3
Water Quality Good 1
Medium 2
Poor 3
O & M Practices Good 1
Medium 2
Poor 3
No. of Quality Failures 0-1 1
2-3 2
4-5 3
Risk of Quality Failure Low 1
Moderate 2
High 3

Model Implementation and Validation


The previously explained methodology and shown in Figure 3 was implemented in MATLAB utilizing the
embedded statistical package. The ANN model was fed by the input and output datasets, which were divided
randomly into three sets; training (70%), validation (15%), and testing (15%). A Back Propagation Neural
Network (BPNN) that comprised of three layers (i.e., input, hidden, and output) was used to evaluate the level of
contamination in the water system. After numerous iterations using the cross-entropy algorithm, the best
validation performance was generated based on the gradient value. The training stopped with the best validation
performance registered as 0.012787 at Epoch 12, as shown in Figure 4(a). Additionally, the trained network was
tested against the magnitude of gradient and the number of validation checks to evaluate different interactions
pertaining to building the ANN model. Fig. 4(b) shows a training gradient of 0.0015659 at Epoch 18 with six
validation checks, which reveals a satisfactory degree of accuracy.

(a) (b)
Figure 4. ANN Validation: (a) Training Performance; (b) Training State

35
International Conference on Engineering, Science and Technology
October 7-10, 2019 www.iconest.net Denver, CO, USA

The efficiency of the model could be investigated with different performance indicators. In this perspective, the
developed model for risk assessment was tested via four different statistical indicators, namely, mean absolute
error (MAE), root-mean-squared error (RMSE), average invalidity percent (AIP), and average validity percent
(AVP). Table 2 demonstrates the validation results of the model, which states that the MAE is 0.09, the RMSE is
0.17, the AIP is 9, and the AVP is 91. All the investigated performance indicators delivered acceptable results,
which indicate the consistency of the model.

Table 2. Statistical Performance Tests


Indicator MAE RMSE AIP AVP
Results 0.09 0.17 9 91

Conclusion and Recommendations


Proactive risk of failure and deterioration models for water distribution networks are substantial because they can
assist water utilities in pinpointing critical deficiencies, thus optimizing the management of their funding policies.
This paper proposed an ANN-based model to estimate the risk of quality failure in water distribution networks.
Different contamination and failure mechanisms were identified and fed to the neural networks. In this regard, the
information collected from the city of El Pedregal in Peru was channeled to the ANN engine to be trained.
Subsequent to generating the validation performance, the ANN model was able to satisfactorily predict the failure
of water contamination with an average validity performance of 91%. The accuracy of the network was also
checked using numerous validation performance schemes.

Several promising research topics could be conducted in the future. Some of the future research may improve the
current directions, while other challenges may address the fundamental issues of water networks, using
technology-based applications and procedures. Future work may incorporate additional factors and more data sets
in order to augment the coherency of the model. Other interesting topics may use computer vision-based
techniques along with soft computing to assess the condition of WDNs. Through the fusion of multiple algorithms
and advanced technologies, a consistent inspection of underground utilities could be achieved by saving time and
money, in addition to offering clean potable water to urban communities.

Acknowledgements
This work is supported by the collaboration of the Universidad Nacional de San Agustín (UNSA) in Arequipa,
Peru, and Purdue University in Indiana, USA, through Discovery Park's Center for the Environment (C4E).

References
Ampazis, N., Perantonis, S.J., and Taylor, J.G. (1999). “Dynamics of multi-layer networks in the vicinity of
temporary minima.” Neural Networks, 12, 43–58.
Ana, E.V., and Bauwens, W. (2010). “Modeling the structural deterioration of urban drainage pipes: The state-of-
the-art in statistical methods.” Urban Water Journal, 7 (1), 47–59. doi:10.1080/15730620903447597.
Christodoulou, S., Agathokleous, A., Charalambous, B., and Adamou, A. (2010). “Proactive risk-based integrity
assessment of water distribution networks.” Water Resour Manage, 24, 3715–3730. DOI 10.1007/s11269-
010-9629-5.
Dawood, T., Zhu, Z., and Zayed, T. (2018). “Computer vision-based model for moisture marks detection and
recognition in subway networks.” Journal of Computing in Civil Engineering, 32 (2). DOI:
10.1061/(ASCE)CP.1943-5487.0000728.
Deng, Y., Jiang, W., and Sadiq, R. (2011). “Modeling contaminant intrusion in water distribution networks: A
new similarity-based DST method.” Expert Systems with Applications, 38 (1), 571-578.
https://doi.org/10.1016/j.eswa.2010.07.004.
Dridi, L., Mailhot, A., Parizeau, M., and Villeneuve, J.P. (2009). “Multiobjective approach for pipe replacement
based on Bayesian inference of break model parameters.” J. Water Resour. Plann. Manage., 135 (5), 344–
354. https://doi.org/10.1061/(ASCE)07339496(2009)135:5(344).

36
International Conference on Engineering, Science and Technology
October 7-10, 2019 www.iconest.net Denver, CO, USA

Elwakil, E., Ammar, M., Zayed, T., Mahmoud, M., Eweda, A., and Mashhour, I. (2009). “Investigation and
modeling of critical success factors in construction organizations.” Construction Research Congress,
Seattle, Washington, US. https://doi.org/10.1061/41020(339)36.
Fares, H., and Zayed, T. (2010). “Hierarchical fuzzy expert system for risk of failure of water mains.” Journal of
Pipeline Systems Engineering and Practice, 1 (1). https://doi.org/10.1061/(ASCE)PS.1949-1204.0000037.
Giustolisi, O., Laucelli, D., and Savic, D. A. (2006). “Development of rehabilitation plans for water mains
replacement considering risk and cost-benefit assessment.” Civil Engineering and Environmental Systems,
23(3): 175-190. https://doi.org/10.1080/10286600600789375.
InfraGuide (2003). “Deterioration and inspection of water distribution systems.” A Best Practice by the National
Guide to Sustainable Municipal Infrastructure, Issue No. 1.1, April 2003.
https://fcm.ca/sites/default/files/documents/resources/guide/infraguide-deterioration-inspection-water-
distribution-systems-mamp.pdf.
Jafar, R., Shahrour, I., and Juran, I. (2010). “Application of artificial neural networks (ANN) to model the failure
of urban water mains.” Mathematical and Computer Modelling, 51(9-10), 1170–1180.
doi:10.1016/j.mcm.2009.12.033.
Kabir, G., Tesfamariam, S., Francisque, A., and Sadiq, R. (2015). “Evaluating risk of water mains failure using
a Bayesian belief network model.” European Journal of Operational Research, 240(1), 220-234.
https://doi.org/10.1016/j.ejor.2014.06.033.
Kirchhoff, D., and Doberstein, B. (2006). “Pipeline risk assessment and risk acceptance criteria in the State of
Sao Paulo, Brazil.” Impact Assessment and Project Appraisal, 24(3), 221-234.
https://doi.org/10.3152/147154606781765156.
Kleiner, Y. (1998). “Risk factors in water distribution systems.” British-Colombia Water and Waste Association
26th Annual Conference, Whistler, British Columbia, 26–29 April 1998.
Ricci, P. F., Sagen, L. A., and Whipple, C. G. (1981). Technological Risk Assessment Series E, Applied Series
no. 81, NATO ASI Series, Erice (Italy).
Sadiq, R., Kleiner, Y., and Rajani, B. (2007). “Water quality failures in distribution networks-risk analysis using
fuzzy logic and evidential reasoning.” Risk Analysis, 27 (5). DOI: 10.1111/j.1539-6924.2007.00972.x.
Sadiq, R., Saint-Martin, E., and Kleiner, Y. (2008). “Predicting risk of water quality failures in distribution
networks under uncertainties using fault-tree analysis.” Urban Water Journal, 5(4), 287–304. DOI:
10.1080/15730620802213504.
Sattar, A., Gharabaghi, B., and McBean, E. (2016). “Prediction of timing of water main failure using gene
expression models.” Water Resour Manage, 30, 1635–1651. DOI 10.1007/s11269-016-1241-x.
Winkler, D., Haltmeier, M., Kleidorfer M., Rauch W., and Tscheikner-Gratl, F. (2018). “Pipe failure modelling
for water distribution networks using boosted decision trees.” Structure and Infrastructure Engineering,
14(10), 1402-1411. DOI:10.1080/15732479.2018.1443145.

37

You might also like