1

A Low-Cost Sensor Network for Real Time

Monitoring and Contamination Detection in
Drinking Water Distribution Systems
Theofanis P. Lambrou, Member, IEEE, Christos C. Anastasiou , Christos G. Panayiotou , Senior Member, IEEE,
and Marios M. Polycarpou, Fellow, IEEE

Abstract This paper presents a low cost and holistic approach develop operational response and do not provide a level of
to the water quality monitoring problem for drinking water public health protection in real time. Rapid detection (and
distribution systems as well as for consumer sites. Our approach response) to instances of contamination is critical due to the
is based on the development of low cost sensor nodes for real time
and in pipe monitoring and assessment of water quality on the fly. potentially severe consequences to human health. Breaches in
The main sensor node consists of several in-pipe electrochemical physical and hydraulic integrity of a water distribution system
and optical sensors and emphasis is given on low cost, lightweight can lead to the influx of contaminants across pipe walls,
implementation and reliable long time operation. Such implemen- through breaks, and via cross connections. Drinking water
tation is suitable for large scale deployments enabling a sensor can serve as a transmission vehicle for a variety of hazardous
network approach for providing spatiotemporally rich data to
water consumers, water companies and authorities. Extensive (biological, chemical, etc) agents which can produce adverse
literature and market research is performed to identify low cost effects in humans due to multiple routes (ingestion, inhalation
sensors that can reliably monitor several parameters which can and dermal) of exposure.
be used to infer the water quality. Based on selected parameters Traditional methods of water quality control involve the
a sensor array is developed along with several microsystems manual collection of water samples at various locations and
for analog signal conditioning, processing, logging, and remote
presentation of data. Finally, algorithms for fusing on-line multi at different times, followed by laboratory analytical techniques
sensor measurements at local level are developed to assess the in order to characterize the water quality. Such approaches are
water contamination risk. Experiments are performed to evaluate no longer considered efficient [1], [2], [3], [4], [5]. Although,
and validate these algorithms on intentional contamination events the current methodology allows a thorough analysis including
of various concentrations of escherichia coli bacteria and heavy chemical and biological agents, it has several drawbacks: a) the
metals (arsenic). Experimental results indicate that this inexpen-
sive system is capable of detecting these high impact contaminants lack of real-time water quality information to enable critical
at fairly low concentrations. The results demonstrate that this decisions for public health protection (long time gaps between
system satisfies the on-line, in-pipe, low deployment-operation sampling and detection of contamination) b) poor spatiotem-
cost and good detection accuracy criteria of an ideal early poral coverage (small number locations are sampled) c) it is
warning system. labor intensive and has relatively high costs (labor, operation
Index Terms On line, In Pipe, Water Monitoring, Flat surface and equipment). Therefore, there is a clear need for continuous
Sensors, Turbidity Sensor, Multi-Sensor System, Wireless Sensor on-line water quality monitoring with efficient spatio-temporal
networks, Arsenic - Bacteria Contamination Detection. resolution. US Environmental Protection Agency (USEPA)
has carried out an extensive experimental evaluation [6] of
I. I NTRODUCTION water quality sensors to assess their performance on several
Clean drinking water is a critical resource, important for the contaminations. The main conclusion was that many of the
health and well-being of all humans. Drinking water utilities chemical and biological contaminants used have an effect on
are facing new challenges in their real-time operation because many water parameters monitored including Turbidity (TU),
of limited water resources, intensive budget requirements, Oxidation Reduction Potential (ORP), Electrical Conductivity
growing population, ageing infrastructure, increasingly strin- (EC) and pH. Thus, it is feasible to monitor and infer the water
gent regulations and increased attention towards safeguarding quality by detecting changes in such parameters.
water supplies from accidental or deliberate contamination. Water quality monitoring systems should provide good bal-
There is a need for better on-line water monitoring systems ance between cost and ease of implementation-maintenance,
given that existing laboratory-based methods are too slow to however to ensure their survival must demonstrate operational
benefits (such as better water quality, decreased operating
T. Lambrou, C. Panayiotou and M. Polycarpou are with the KIOS Research costs or reduced customer complaints). Given the absence
Center for Intelligent Systems and Networks and the Department of Electrical of reliable, in-line, continuous and inexpensive sensors for
and Computer Engineering, University of Cyprus, Nicosia, Cyprus.
E-mail: {faniseng,christosp,mpolycar}@ucy.ac.cy monitoring all possible biological and chemical contaminants,
Christos C. Anastasiou is with the Department of Civil Engineering, our approach is to measure physicochemical water parameters
Frederick University, Nicosia, Cyprus. that can be reliably monitored with low cost sensors and
E-mail: [email protected]
An earlier version of this paper was presented at the IEEE SENSORS 2012. develop low cost networked embedded systems (sensor nodes)
Manuscript received December 19, 2013. as well as contamination detection algorithms to fuse these
 2

multi-sensor data in order to infer possible contamination important that the spatio-temporal sampling should be signif-
events. Even though this approach may suffer from some false icantly increased in order to collect water quality samples
alarms, it can be compensated/eliminated by the large scale at significantly more locations (if possible at all consumer
deployment and the possibility of correlating the decisions sites). Therefore, the main challenge is to develop reliable
from various sensor nodes which is the topic of our future and low-cost sensing systems for accurate and continuous
work. in-pipe water quality monitoring. The system developed is
A limited number of on-line, reagent-free water moni- intended to be used as a part (sensor node) of a low cost
toring systems are commercially available [7] (e.g. Hach water sensor network to provide water quality information
HST GuardianBlue [8], J-MAR BioSentry [9], etc), but these to consumers, water companies and authorities. The spatio-
systems are bulky (sensors are installed in flow cells located temporal data provided by such network can support complex
in cabinets) and remain cost prohibitive for large scale de- decisions concerning the quality of drinking water, including
ployments (cost tens of thousands of dollars per unit). It is the detection of the location/source of hazardous agents and
worth mentioning that cost is mostly attributed not to sensing pathogens, raising awareness, and encouraging better water-
probes but to instrumentation-automation controllers (analyz- handling and management.
ers) and panels. Such systems can take frequent samples The main contribution of this paper is twofold. First,
of the water quality at a very limited number of locations. it proposes the need for a shift in the current monitoring
However, substantial proportion of contamination problems is paradigm and propose the idea of monitoring the quality of
attributable to problems within distribution systems and due water delivered to consumers, using low cost, low power
to the limited spatio-temporal sampling, it is impossible for and tiny sensors. We argue that this approach can achieve
the water companies and consumers to know the quality of more reliable quality monitoring due to the large spatially
potable water delivered to consumer households. distributed deployment and the possibility of correlating the
A number of bare multi-parametric sensor arrays have been quality measurements from various sensors. Second, it presents
developed and presented in the literature based on various sen- the first step towards this goal which is the design and
sor technologies. A recent review on multi-parametric solid- development of a low cost system that can be used at the
state sensors for water quality is given [3]. A chemical sensor premises of consumers. The embedded systems developed can
array for water quality monitoring based on thick-film technol- also be used in a consumer-oriented manner to continuously
ogy is presented in [29], [30] and [31], [32], these sensors are monitor qualitative water parameters and fuse multi-parametric
very low cost, though they have limited lifetime (few months) sensor response in order to assess the water consumption risk
and require a conventional glass reference electrode to operate at consumer level, locally and independently from other con-
accurately. Along similar lines, a multi-parametric sensor sumer measurements. In particular, the contributions regarding
array based on semiconductor ruthenium oxide nanostructures the low cost system is the design and development of low
is presented in [33]. In addition, several water monitoring cost networked embedded systems as well as optical sensors
microsystems (sensor nodes) have been developed for large (turbidity) for water quality monitoring, the development of
scale water monitoring based on wireless sensor networks event detection algorithms using fusion techniques and the
(WSNs) technology. In [35] a sensor node (CSIRO Fleck-3) experimental evaluation and validation of system performance
is developed for monitoring salinity in ground waters as well in various concentrations of microbiologically (E.coli) and
as the water temperature in surface waters. Finally, a multi- chemically (Arsenic) contaminated drinking water.
sensor system for monitoring coastal water quality in rivers The remaining of this paper is organized as follows. Sec-
and lakes is presented in [34]. tion II presents the methodology and justification for the selec-
Apart from the on going research towards the design and tion of water quality parameters to be monitored. Section III
development of sensors and microsystems another parallel presents the experimental implementation of the hardware
research direction is that of the development of software and and software modules and validates the performance of the
algorithms for the detection of water quality anomalies and developed system. Finally the paper ends with the conclusion.
contamination events. A thorough survey on recent advances
in this area is provided in [36]. A limited number of event
II. M ETHODS
detection software is commercially available (Hach Event
Monitor [8], BlueBox [11]). A currently freely available tool Drinking water quality standards are determined according
is CANARY software [12] developed at Sandia National to World Health Organization (WHO) [13] guidelines for
Laboratories in collaboration with the USEPA. CANARY drinking-water quality as well as other pertinent organizations
indicates possible contamination events by using a range of (i.e. EU [14], USEPA [15]). These organizations set the stan-
mathematical and statistical techniques to identify the onset dards for drinking water quality parameters and indicate which
of anomalous water quality incidents from online raw sensor microbiological, chemical and indicator parameters must be
data. Other event detection and data validation methodologies monitored and tested regularly in order to protect the health
are given in [37] and references therein. However, additional of the consumers and to make sure the water is wholesome
event detection methods can be developed based on multi- and clean.
sensor data fusion methods [38]. For the developed system, the selection of the physico-
Given the size of the water distribution networks (pipe chemical parameters to be monitored was based on extensive
length) and the number of households served, we consider scientific literature review [6], [18], [19], [20], [21], [22], [23]
 3

on the relation between certain physicochemical parameters ORP measurements.

and chemical or biological contaminations that present in Temperature (T) measurement is important because bio-
water. Table I enumerates the suggested parameters to be logical and chemical activities are heavily influenced by
monitored from high to low correlation significance when the temperature of water. Moreover the measurements
interpreting water contaminations (assess hazard). Table I also of many water quality sensors (e.g. pH, EC) depend
presents the measurement cost (for purchase and maintenance) on temperature and therefore temperature compensation
associated with these parameters based on recent review [24] is needed to obtain accurate results. The most widely
of measurement and instrumentation methods, compensation used sensor in water quality monitoring is the platinum
and calibration procedures and probe lifetime concerning these (Pt) Resistive Temperature Detector (RTD) because is the
parameters. Therefore, the parameters selected to be monitored most accurate, linear and stable over time and temperature
are the following: 1) Turbidity, 2) Oxidation Reduction Poten- sensor. A constant current must be passed through the
tial (instead of Free Chlorine), 3) Temperature 4) pH, and 5) RTD and the change of voltage determines the tempera-
Electrical Conductivity. ture.
pH (pH) is the measure of the acidity or alkalinity of the
Parameter Units Quality Range Meas. Cost water and is defined as the negative decimal logarithm of
1 Turbidity NTU 05 Medium the hydrogen ion activity (concentration) in a water so-
2 Free Residual Chlorine mg/L 0.2 2 High lution. The minimum and maximum allowable pH range
3 ORP mV 650 800 Low for portability is 6.58.5. No health effects are associated
4 Nitrates mg/L <10 High
5 Temperature oC Low with pH, unless dealing with extreme cases of alkaline
6 pH pH 6.5 8.5 Low or acidic conditions; however a sudden change of more
7 Electrical Conductivity S/cm 500 1000 Low than 0.5 pH units indicates a possible contamination.
8 Dissolved Oxygen mg/L Medium
Low pH allows a more effective disinfection provided
TABLE I by chlorination. A combined pH glass electrode is used
S UGGESTED PARAMETERS TO BE M ONITORED . Q UALITY RANGE IS for pH measurement and temperature compensation is
SUGGESTED BY WHO [13] GUIDELINES AND EU STANDARDS [14]. needed for accurate measurements.
Electrical Conductivity (EC) is the measure of the ability
of water to conduct electrical current and is an indicator
The importance of the selected parameters as well as their of the concentration of total dissolved solids (TDS) (e.g.
correlation to water quality is briefly described next. ions) in water. Conductivity is determined by applying
Turbidity (TU) is a measure of cloudiness of the water an AC voltage (in the low kHz region) and measuring
sample caused by suspended particles and microscopic the resulting current on the electrode. Conductivity is
organisms. Turbidity is determined by measuring the expressed in S/cm. Drinking water conductivity should
intensity of infrared light scattered at 90o angle to the be between 500 to 800 S/cm (max 1050 S/cm). There
traversing beam (ISO 7027 method) and is expressed in are basically two types of sensors for measuring water
Nephelometric Turbidity Units (NTU). Turbidity is an conductivity: conductive sensors and toroidal - inductive
important water quality indicator as excessive turbidity in sensors. Conductive sensors are more sensitive and low
potable water can provide food and shelter to microbes cost but more prone to fouling. Inductive sensors use a
and thus pathogens are more likely to be present in highly magnetic field (coils) to sense conductivity and toroidal
turbid waters. Turbidity in distribution system should be coils are not in contact with water which makes the
maintained under 5 NTU. sensor virtually immune to fouling, however these sensors
Oxidation Reduction Potential (ORP) is a measure of are more expensive and lack the sensitivity of contact-
the oxidizing or reducing power of the water and is ing sensors. Both types of sensors require temperature
measured in mV (i.e. is the tendency of a solution compensation for accurate measurements. The presence
to acquire electrons). Oxidation refers to any chemical of minerals in drinking water (hardness or high EC) is
action in which electrons are transferred between atoms. preferred not only because of the health benefits, but also
Oxidation and reduction always occur together. ORP the flavor. On the other hand, soft water (low EC) tastes
is a reliable and cost effective method to measure the salty and is sometimes not suitable for drinking. On the
effectiveness of water disinfection sanitizers (chlorine, other hand, excess water hardness often causes deposits
sodium hypochlorite, bromine and ozone) in real time. in pipes and in household appliances, which can, in time,
As the measured value of ORP increases, the solution harbor microbial communities.
has more potential to oxidize and thus to destroy any
unwanted organisms. WHO adopted an ORP standard It is noted that Free Chlorine concentration (HOCl) can be
for drinking water disinfection of 650 mV. Research has approximated based on the ORP, pH and temperature measure-
shown that at 650-700 mV of ORP, bacteria such as E. ments. Free chlorine monitoring using amperometric sensors
coli and Salmonella are killed on contact, while tougher is expensive because it is very sensitive in the pH, temperature,
organisms (listeria, yeasts, molds) may require 750 mV or flow and pressure of the sample. Therefore accurate free chlo-
higher in order to be killed. A combined ORP electrode rine measurements require a flow cell with additional pH and
is used and temperature compensation is not needed for temperature sensors for compensation. Furthermore, sensors
 4

directly installed in-pipe with no flow cell to condition the Conventionally, ultrasonic, brush, water-jet, or chemical type
sample provide erratic responses as the readings are both flow of automatic cleaners [25] are used to remove coatings from
and pressure dependent. In addition, amperometric sensors are the sensor probes. Recently, several alternative cost effec-
more prone to fouling due to their specific construction and tive methods have been proposed that can either actively
this results in increased cleaning and calibration frequency. remove fouling (e.g. electrolysis) or passively prevent fouling
In [16], a new low cost ruthenium oxide thin film chlorine (copper mesh or CuO2 doped materials). In this work, flat
ion measurement system was presented, however the sensor is measuring surface probe method [28] is used because is
again sensitive to pH, temperature and time. the most cost effective, passive self-cleaning method and is
Nitrates, though considered as an important parameter for based on the mechanical package and design of the probe.
human health is not selected because measurement methods When the electrodes flat measuring surface is exposed to
are subjected to failures (Ion-Selective electrodes) or are cost turbulent flow, the resulting scrubbing action provides a self-
prohibitive (UV spectrophotometric method). In [17], a new cleaning effect in most applications under medium range
promising method is presented based on a PCB planar electro- flows. The flat sensing surface virtually eliminates deposits
magnetic sensor. Finally, dissolved oxygen is not selected due that can foul the electrode and significantly reduces necessary
to several compensations and frequent membrane replacement maintenance. This simple, but effective method has no moving
needed for accurate measurements. parts, requires no power and also prolongs electrode life
Convectional combined electrodes (for ORP and pH) have and eliminates breakage. Additional antifouling technologies
been widely used due to their good sensitivity, selectivity, have be proposed for solid-state and optical sensors based on
stability and long lifetime. However, convectional pH glass nano-scale materials possessing super-hydrophobic properties
electrodes have several disadvantages due to the intrinsic [26]. Moreover, super-hydrophobic surfaces have found to also
nature of the glass membrane. For example, they have limited prevent biofouling [27].
pressure tolerance, exhibit a sluggish response, require a high
input impedance signal conditioning circuits and it is difficult III. R ESULTS
to miniaturize based on current manufacturing technologies. A. System and Sensors Development and Integration
Therefore, a number of emerging-alternative sensor technolo- A modular but holistic approach is adopted for the design
gies in various stages of research and development have been and development of the system. Modularity enables the simul-
proposed in the literature. taneous instrumentation and sampling of all parameters and the
Thick film chemical sensor arrays developments show that it decomposition of several operations like calibration, upgrades
is possible to develop a single miniaturized multi-parametric and repair or replacement of faulty parts.
sensor probe in a cost effective manner, however thick film
chemical sensors have limited lifetime (few months), suffer
from electrode drift (due to salt loss) and the development of
a stable reference electrode is not possible so far [30].
ISFET based microsensors (developed using MOSFET
semiconductor technology) offer advantages such small size
(mass fabrication and compact probes), robustness (no glass
membranes), low output impedance and rapid response, how-
ever they have several limitations as they require a glass refer-
ence electrode to operate robustly (REFET designs suffer from
short lifetime and residual ionic sensitivity) and encapsulation
is difficult, which increases dramatically the final cost of the
sensors [39].
Nano-sensors based on nanostructures of noble metals
and their oxides (like Pt, Ru, Ir) is a recent - promising
concept however developments so far suffer from several
drawbacks like temperature dependent delay response and non-
deterministic potential drift (electrolysis of water on oxide
surfaces and unpredictable temperature dependence) [33].
Therefore, despite the recent advances in sensor develop-
ment technologies, the reliability and performance of convec- Fig. 1. System architecture.
tional glass electrodes is still unsurpassed for continuous wa-
ter quality monitoring [7]. Therefore, convectional (pH,ORP) The overall system architecture under discussion in pre-
glass electrodes and solid-state sensors (TU, EC, T) are used sented in Fig. 1 and is comprised of the following three
in this work as they provide the most reliable technology. subsystems: a central measurement node (PIC32 MCU based
In line water sensors illustrate the need for efficient and board) that collects water quality measurements from sensors,
periodic probe cleaning to maintain reliable measurements. implements the algorithm to assess water quality and trans-
Cleaning mechanisms constitute an important cost parameter mits data to other nodes, a control node (ARM/Linux web-
which can consume as high as 50% of the operational budgets. server based platform) that stores measurement data received
 5

from the central measurement node in a local database and

provides gateway to the internet, visualize data (charts), and
sends email/sms alerts and finally a tiny notification node(s)
(PIC MCU based board) that receives information from the
central measurement node through an interconnected ZigBee
RF transceiver and provides local near-tap notifications to the
user (water consumer) via several interfaced peripherals (LED,
LCD, Buzzer).
It should be noted that the central measurement node serves
as the sensor node. The idea is to installed these sensor nodes
in many consumer sites in a spatially-distributed manner to
form a WSN that will monitor the drinking water quality in
the water distribution system from the source to the tap.

Fig. 4. Complete System Photo.

turbidity sensor for continuous in pipe turbidity monitoring

in water distribution systems using commercial off-the self-
components. The turbidity sensor development was based
on the ratio turbidimeter design (see Fig. 5(a)) where both
transmitted and scattered light intensities are measured to
. eliminate errors (interferences) due to IR emitter intensity drift
(a) Central measurement node. (b) Notification node. and sample absorption characteristics. An infrared (860nm)
Fig. 2. Nodes photos. narrow beam LED emits light through an optical gap to the
water sample and two IR photodiodes separated around 1cm
A photo of the central measurement and notification node from the emitter receive simultaneously the 90o scattered
prototypes are presented in Fig. 2. The control node is tiny and 0o transmitted light. The photodiodes spectral sensitivity
ARM/Linux based computer (Raspberry Pi). Fig.3 presents are selected to fit with that of the IR light source. The
the multi-parameter sensor array probes comprised of the instrumentation and analog signal conditioning of the sensor
Turbidity(TU), ORP, pH, Electrical Conductivity(EC) and is as follows: The IR emitter is pulsed at 1kHz with a square
Temperature(T) sensors. The in-pipe Turbidity sensor is con- wave signal and the photodiodes convert the light directly
structed from scratch based on our previous work [1] while the into electrical current, then a high-gain, low-noise CMOS
other sensor probes obtained from SensoreX Corpr . The pH (Complementary metal-oxide-semiconductor) transimpedance
sensor embeds an RTD sensor which is used for temperature amplifier with background light rejection is used to convert
sensing and temperature compensation of pH and Electrical the each photocurrent to voltage output. The ac output of each
Conductivity measurements. TU, ORP, pH and toroidal EC transimpedance amplifier is then converted to a dc signal using
sensors have flat measuring surfaces for cost effective self- a precision active peak detector. Finally the 90o scattered dc
cleaning. The complete system photo, with TU, ORP, pH, EC signal is further conditioned by an instrumentation amplifier
and T sensors as well as a rotor-flow sensor mounted in a for 0 NTU offset nulling and additional amplification. The
plastic pipe, is shown in Fig 4. conditioned voltage outputs are then sampled by a 10 bit A/D
converter with reference voltage of 1.1V and the sensor output
V90o
voltage V = c.V 0o
is given as the signal ratio of the scattered
V90o to the transmitted V0o voltage, c is calibration coefficient.
An indirect method for the sensor calibration was employed,
in order to avoid the use of the carcinogen and expensive
chemical formazin solutions. Therefore, a number of samples
were created and the turbidity of each sample is measured both
by the turbidity sensor under calibration and by a laboratory
Fig. 3. Sensor photos. TU, ORP, pH and inductive EC have flat sensing turbidimeter (Lutron TU-2016) used as reference. Then the
surfaces for cost effective self-cleaning. relationship between turbidity (in NTU) and the voltage output
(in mV) of the turbidity sensor is extracted and given by
Turbidity Sensor Development: Although there is plenty
T U = 0.1035V 0.292 (1)
of turbidity measuring instruments available on the market
at the moment, most of them are expensive and not di- The sensor generates an output voltage proportional to the
rectly compatible with in-pipe, in-line requirements as well turbidity or suspended particles and has a linear response in
as wireless sensor networks technology. Therefore, the goal the range of 0-100 NTU with 0.1 NTU resolution. Finally, as
is to develop a low cost, easy to use and accurate enough shown in Fig. 5 the turbidity sensor probe was mounted in a
 6

Cf

flat surface PTFE housing and sealed in a hydraulic Tee fitting Rf

Rf
for inline installation. EC Probe

Rl
2 R 2
TU Probe
1 1
LED D Vout Vout
3 3
Vexc

(a) Turbidity preamplifier (b) Conductivity preamplifier

Precision current
source <1mA
R1 R2 R1 R2
Vout

2 2 Temp Probe
ORP Probe pH Probe NTC or RTD
1 1
Vout Vout
(a) Ratio turbidimeter measurement principle R 3 R 3

(c) ORP preamplifier (d) pH preamplifier (e) Temperature

Fig. 6. The first stage of analog signal conditioning circuitry for each
parameter

(b) Probe board with (c) Probe mounted in (d) Housing sealed
optoelectronic com- a flat surface PTFE in a hydraulic Tee
ponents (teflon) housing fitting for inline in-
stallation

Fig. 5. Turbidity sensor

Apart from TU sensor, analog signal conditioning circuits,

calibration and compensation procedures were developed for
pH, OPR, RTD and conductive/inductive EC sensors. Con-
siderable attention is given to acquire linear response, reduce
noise and attain high resolution and accuracy. A dedicated PIC
based microsystem is developed for each parameter to accom-
plish this task. The first stage of analog signal conditioning
circuitry for each parameter is presented in Fig. 6 while Table Fig. 7. Software Platform.
II shows the results regarding laboratory evaluation (using
standard buffer solutions and reference instruments) of each
parameter along with the quality range suggested by WHO detection algorithm enables the system to act as an early
guidelines and EU standards. The overall power consumption warning system for possible potable water quality deteriora-
of the central measurement sensor node with the on board tion at the point of installation (e.g. homes). A flowchart of
LEDs off and the RF Xbee transceiver module sending water the two algorithms is illustrated in Fig. 9. Both algorithms are
quality data every 5s is about 50mA at 5V operating voltage, based on normalized sensor outputs given by
however further improvements are planned to minimize the |Si i |
power consumption. Ni = (2)
i i
The software platform developed for the control node is
illustrated in Fig. 7. This platform enables real time measure- where Si is the current measurement of parameter i
ment charts of monitored parameters, real time assessment {T U, ORP, pH, EC}, i , i are the mean and standard de-
of water quality and sensor calibration instructions through viation over a moving time window w and i is a sensor
a Graphical User Interface (GUI). It also logs sensor data in based parameter associated with measurement accuracy of
a local database and posts data to web using Pachube open each parameter i. Normalized sensor outputs Ni are used to
source web platform. Using Pachube scripts the user can setup filter baseline (i.e mean) fluctuations.
various thresholds for sending notifications via sms or email. The objective of the event detection algorithms is to activate
Fig. 8 illustrates the main window of the internet platform. a trigger event (alarm) when normalized sensor outputs exhibit
sudden and significant changes, given that these changes are
bounded within the quality ranges suggested by drinking water
B. Contamination Event Detection Algorithms quality standards (see Table II, quality range). The detection
Two event detection algorithms were developed to fuse on- of water quality changes that are outside the expected quality
line multi-sensor measurements in order to assess the water ranges (min/max violations) is easier and can be done by
contamination risk when anomalies are detected. An event a weighted multi-parameter cost function in the form of
 7

Parameter Measurement principle Units Range Resolution Accuracy Quality Range

Turbidity Optical/infrared scattering NTU 0 100 0.1 0.5 05
ORP Galvanic cell, platinum electrode mV -2000 2000 2 10 600 800
pH Galvanic cell, glass electrode pH 0 14 0.05 0.1 6.5 8.5
Conductivity Conductive cell S/cm 100 20000 10 5% 500 1000
Conductivity Inductive cell S/cm 200 3000 10 5% 500 1000
Temperature RTD resistance oC -5 100 0.1 0.1
Flow Magnetic rotor, hall effect sensor L/min 1115 0.0015 15%
TABLE II
S PECIFICATIONS AND ACCOMPLISHED PERFORMANCE FOR EACH MONITORED PARAMETER .

Therefore, the risk indicator RIV DA is given by

{
1 if ||N N0 ||2 > d
RIV DA = (3)
0 otherwise
Note that VDA algorithm requires the normalized control
signal vector N0 as well as a calibration threshold d (obtained
from a learning phase) to execute.
The second event detection algorithm is denoted as Polygon
Area Algorithm (PAA) and the risk indicator RIP AA function
used in this algorithm is estimated based on the ratio of the
polygon area AN formed by the N vector components (when
projected (displayed) on a two-dimensional spider graph with
four (TU, ORP, pH, EC) axes starting from the same point) to
the polygon area A1 formed by the 1 ones vector components
(i.e 1 = [1111]T ). Therefore, the risk indicator RIP AA is given
by {
P AA 1 if A A1 > 1
N
RI = (4)
0 otherwise
Note that PAA algorithm does not require any further infor-
mation to execute.

Fig. 8. On-line Web Platform.

RO = i wOi Ji , where Ji are binary variables that indicate
whether parameter i has been violated and wOi are non-
negative weights which imply the significance of the violation
of each parameter i. If RO = 0 no violation is assumed,
however as RO > 0 increases the water contamination risk
is also increases.
As previously indicated, the objective in this paper is
to detect anomalies when water quality changes are inside
the expected quality ranges by fusing the multi-sensor data.
Therefore a risk indicator RI function is defined that takes a Fig. 9. Risk Assessment Algorithms.
value RI = 1 if a contamination event is detected or RI = 0
otherwise.
The first event detection algorithm is denoted as Vector Dis-
tance Algorithm (VDA) and the risk indicator RIV DA function C. Experimental Validation
used in this algorithm is estimated based on the Euclidian In this section we present the results of the experimen-
distance between the normalized sensor signal vector N and tal trials performed to validate the behavior and evaluate
the normalized control signal vector N0 of pure (clean) water. the performance of the developed hardware and algorithms
 8

on intentional contamination events. The experimental setup E.coli

8.4

ORP (mV) TU (NTU)

consists of the sensor node (central measurement node) that 6.3
4.2
takes samples every 5s from potable water flowing through 2.1
0
a flow cell. Intentional contamination of two important con-
591
taminants (escherichia coli bacteria and arsenic) of various 528
465
concentrations was injected at discrete time intervals and the 402
performance of the event detection algorithms is evaluated on 8.38
8.28
real time. Escherichia coli bacteria and arsenic contamination 8.18

pH
8.08
7.98
in drinking water is very severe problem causing serious 7.88

EC (S/cm)
poisoning to large numbers of people all over the world [23]. 1053
Currently there are no sensors that can detect all biolog- 983
913
ical and chemical contaminants, even when single-specific 843
5x102cfu/mL

5x101cfu/mL

5x100cfu/mL

5x101cfu/mL

5x103cfu/mL5x104cfu/mL1x107cfu/mL

contaminants are concerned and sensors do exist, there is 30.15

T (C)
29.75
an absence of reliable, in-line, continuous and inexpensive 29.35
28.95
sensors. In fact, developing and deploying sensors for all 28.55
0 150 300 450 600 750 900 1050 1200 1350 1500
Time (s)
contaminants would not be feasible (too many contaminants
which translates to very high cost). On the other hand, many (a) Sensors responses to E.coli bacteria contaminated water.
contaminants cause changes in some parameters of the water 5x102cfu/mL 5x101cfu/mL 5x100cfu/mL 5x101cfu/mL 5x103cfu/mL5x104cfu/mL1x107cfu/mL
3.5

that can be easily monitored. Therefore, by monitoring these
changes we can infer possible contamination events, even 3
though this approach may suffer from some false alarms, it
can be compensated/eliminated by the large scale deployment 2.5

and the possibility of correlating the decisions from various

sensor nodes which is the topic of our future work. 2 ||N N0 ||2
AN /A1
1) Microbiologically (E.coli) contaminated drinking water:
I
R

RVI DA
RPI AA
The first experiment considers the case of microbiologically 1.5

(E.coli) contaminated drinking water. Most E. coli strains are

in general harmless to humans, but some types can cause 1

serious food and water poisoning. However, the presence of

E.coli is used to indicate that other pathogenic organisms 0.5

may be present (often of faecal origin). According to WHO

0
guidelines & EU Drinking Water Directive E.coli parametric 0 500 1000 1500
Time (s)
value is 0 CFU/100mL.
Fig. 10(a) presents the measurements received using the (b) E.Coli bacteria contamination detection.
developed sensor node when the following concentrations of Fig. 10. Experiments with E.coli bacteria.
E.coli were injected: 5x102 , 5x101 , 5x100 , 5x101 , 5x103 ,
5x104 , 1x107 CFU/mL. It is evident that TU and EC sensors
responded well when microbiological contaminants injected
in chlorinated potable water. ORP sensors has responded with degenerative processes and causes skin cancer. According to
delay and pH sensor has a spiky type of response. Fig. 10(b) WHO guidelines & EU Drinking Water Directive Arsenic
presents the output signals of the Vector Distance Algorithm parametric value is 10g/L (it was 50g/L in the previous
(VDA) and Polygon Area Algorithm (PAA). The results of directives).
Fig. 10(b) indicate that both algorithms miss the detection of Fig. 11(a) presents the measurements received using the
5x102 CFU/mL because sensors responses were very close to developed sensor node when the following concentrations of
background levels (no anomalies occurred). It should be noted Arsenic were injected: 5, 10, 25, 50, 125, 500, 1000 g/L. Ar-
that the performance of PAA algorithm is better and given that senic solutions created from a standard solution of 1000mg/L
it utilizes less information, PAA algorithm is better than the As. Unfortunately, almost all sensors did not respond at low
VDA algorithm. arsenic contamination. However, at concentrations above 25
2) Chemically (Arsenic) contaminated drinking water: g/L ORP and pH sensors have responded and at higher
The second experiment considers the case of chemically concentrations (above 500 g/L) all sensors responded well.
(Arsenic) contaminated drinking water. Water contamination Fig. 11(b) presents the output signals of the Vector Distance
by toxic heavy metals and especially arsenic contamination Algorithm (VDA) and Polygon Area Algorithm (PAA). The
is a common problem encountered in many countries due results of Fig. 11(b) indicate that both algorithms miss the
to undue deposition of mining, agricultural, industrial and detection of 5 and 10 g/L because sensors responses were
urban wastes in water resources. Arsenic is known to affect very sluggish and close to background levels and that the
negatively the mental and central nervous system function, to VDA algorithm exhibits two false alarms. Therefore, the
damage the blood composition, lungs, kidneys, liver, and other performance of PAA algorithm is again better (sharp response,
vital organs, as well as it contributes to certain neurological no false alarms) than the VDA algorithm.
 9

Heavy Metal (As) nitrates) and install the system in several locations of the water
4
TU (NTU) distribution network to collect spatiotemporally rich water
3
2
1
0
quality data and characterize system/sensors response in real
field deployments. Moreover, additional risk assessment and
ORP (mV)

237
227
217 anomaly detection algorithms will be investigated as well as
207
197 algorithms for contaminant identification. Finally, we plan to
8.14 investigate a sensor network approach for quality monitoring
7.94
pH

7.74 of drinking water distribution systems. Towards this goal,

7.54
7.34 new fusion algorithms and methodologies to assess the water
EC (S/cm)

630
623 quality over the entire water distribution network (network
616
609 level) will be investigated.
602
5g/L 10g/L 25g/L 50g/L 125g/L 500g/L 1000g/L

21.97 ACKNOWLEDGMENT
T (C)

21.57
21.17
20.77 The authors would like to thank their colleagues Dr. Evdokia
0 150 300 450 600 750 900 1050 1200 1350 1500 1650 1800 1950 2100 2250
Time (s) Kastanos and Dr. Marinos Stylianou for the preparation and
(a) Sensors responses to Heavy metal (As) contaminated water. testing of E.coli and Arsenic contaminated water samples.
This work is partly supported by the European
5g/L 10g/L 25g/L 50g/L 125g/L 500g/L 1000g/L
Project EFFINET (FP7-ICT-2011-8-31855) and by the

3
Cyprus Research Promotion Foundation under grant
EIX/POION/0609/87.
2.5
||N N0 ||2
AN /A1
R EFERENCES
RIV DA
2
RIP AA
[1] T.P. Lambrou, C.C. Anastasiou , C.G. Panayiotou, A Nephelometric
Turbidity System for Monitoring Residential Drinking Water Quality, in
I
R

1.5 Sensor Networks Applications, Experimentation and Logistics, 2009

[2] T.P. Lambrou, C.G. Panayiotou and C.C. Anastasiou, A Low-Cost System
for Real Time Monitoring and Assessment of Potable Water Quality at
1
Consumer Sites, in IEEE Sensors 2012, 28-31 Oct 2012.
[3] S. Zhuiykov, Solid-state sensors monitoring parameters of water quality
0.5 for the next generation of wireless sensor networks, Sensors and Actuators
B: Chemical, Volume 161, Issue 1, pp 1-20, 2012
[4] A. Aisopou, I. Stoianov, N. Graham, In-pipe water quality monitoring in
0
0 500 1000
Time (s)
1500 2000 water supply systems under steady and unsteady state flow conditions: A
quantitative assessment, Water Research, 46(1), pp. 235-246, 2012
[5] S. Panguluri, G. Meiners, J. Hall, J.G. Szabo, Distribution System Water
(b) Heavy metals (As) contamination detection. Quality Monitoring: Sensor Technology Evaluation Methodology and
Results, U.S. Environmental Protection Agency, EPA/600/R-09/076, 2009
Fig. 11. Experiments with Heavy metals (As). [6] J. Hall, A.D. Zaffiro, R.B. Marx, P.C. Kefauver, E.R. Krishnan, R.C.
Haught, and J.G. Herrmann, On-line Water Quality Parameters as Indi-
cators of Distribution System Contamination, Journal AWWA, 2007
[7] M.V. Storey, B. Gaag, B.P. Burns, Advances in on-line drinking water
Finally, it should be noted that the signatures of normalized quality monitoring and early warning systems, Water Research, Volume
sensor outputs can be further processed to minimize false 45, Issue 2, January 2011, Pages 741-747, ISSN 0043-1354
alarms and to identify the type of contaminants, given that [8] Hach HST, GuardianBlue Early Warning System Brochure, 2008
[9] JMAR, BioSentry Contamination Warning System Technical Overview,
a contamination library is available/developed. 2006
[10] i::scan, Multi-parameter spectrophotometer probe, 2012
[11] Whitewater Technologies, BlueBox Intelligent Water Analytics System
IV. C ONCLUSIONS Brochure, 2012
In this article, the design and development of a low cost [12] D. Hart, S. McKenna, K. Klise, V. Cruz and M. Wilson, CANARY: A
Water Quality Event Detection Algorithm Development Tool. in World
sensor node for real time monitoring of drinking water quality Environmental and Water Resources Congress, 2007.
at consumer sites is presented. The proposed sensor node [13] World Health Organization Guidelines for drinking-water quality, fourth
consist of several in-pipe water quality sensors with flat edition, 2011
[14] European Communities Drinking Water Regulations (No.2), 2007
measuring probes. Unlike commercially available analyzers, [15] U.S. Environmental Protection Agency 2012 Edition of the Drinking
the developed system is low cost, low power, lightweight and Water Standards and Health Advisories, EPA 822-S-12-001
capable to process, log, and remotely present data. Moreover, [16] J.C. Chou, C.C. Chen, and C.C. Lee, Development of Microcontroller
Applied to Chlorine Ion Measurement System, IEEE Sensors Journal, vol.
contamination event detection algorithms have been developed 12, no 6, pp. 2215-2221, 2012
and validated to enable these sensor nodes to make decisions [17] M. Yunus and S. Mukhopadhyay, Novel planar electromagnetic sensors
and trigger alarms when anomalies are detected. Such imple- for detection of nitrates and contamination in natural water sources, IEEE
Sensors Journal, vol. 11, no 6, pp. 1440-1447, 2011
mentation is suitable for large deployments enabling a sensor [18] C.N. Haas, M.A. Meyer, and M.S. Paller, Microbial alterations in
network approach for providing spatiotemporally rich data to water distribution systems and their relationship to physical-chemical
water consumers, water companies and authorities. characteristics, in Journal of American Water Works Association , 1983
[19] K.N. Power and L.A. Nagy, Relationship between bacterial regrowth
In the future, we plan to investigate the performance of the and some physical and chemical parameters within Sydneys drinking
event detection algorithms on other types of contaminants (e.g. water distribution system, in Water Research, 1999
 10

[20] J.B. Cook, J. F. Byrne, R. C. Daamen and E. A. Roehl, Distribution

System Monitoring Research at Charleston Water System, in 8th Annual
Water Distribution Systems Analysis Symposium, 2006
[21] Y.J. Yang, R.C Haught and J.A Goodrich, Real-time contaminant de-
tection and classification in a drinking water pipe using conventional
water quality sensors: Techniques and experimental results, in Journal of
Environmental Management, 2009
[22] C.C. Anastasiou, P. Grafias, T.P. Lambrou, A. Kalli and C. Onisiphorou,
Use of holding tanks and their effect on the quality of potable water in
households; the case study of Cyprus, in Protection and Restoration of
the Environment XI,PRE-XI, 2012
[23] John DeZuane, Handbook of drinking water quality. Wiley.com, 1997.
[24] T.P. Lambrou, A Technical Report for Water Qualitative and Quantitative
Sensors, University of Cyprus, 2011
[25] A. Whelan and F.Regan, Antifouling strategies for marine and riverine
sensors, Journal of Environmental Monitoring, 2006
[26] J. Genser, K. Efimenko, Recent development in superhydrophobic sur-
faces and their relevance to marine fouling: a review, Biofouling vol.22,
pp. 339360, 2006.
[27] A.J. Scardino, H. Zhang, D.J. Cookson, R.N. Lamb, R. de Nys, The role
of nano-roughness in antifouling, Biofouling vol. 25, pp. 757767, 2009.
[28] SensoreX Corp Flat Surface Operating Principles.
[29] M. Sophocleous, M. Glanc, Monika, J. Atkinson and E. Garcia-Breijo,
The effect on performance of fabrication parameter variations of thick-
film screen printed silver/silver chloride potentiometric reference elec-
trodes, Sensors and Actuators A Physical, 197, 1-8, 2013
[30] J. Atkinson, John, M. Glanc,M. Prakorbjanya, M. Sophocleous, R.
Sion and E. Garcia-Breijo, Thick film screen printed environmental and
chemical sensor array reference electrodes suitable for subterranean and
subaqueous deployments. Microelectronics International, 30, (2), 92-98,
2013.
[31] R. Martinez Manez, J. Soto, E. Garcia Breijo, J. Ibanez Civera, E. Gadea
Morant, System for determining water quality with thick-film multisensor,
IEEE Electron Devices 2005.
[32] R. Martinez-Manez, J. Soto, E. Garca-Breijo, L. Gil, J. Ibanez, E. Gadea,
A multisensor in thick-film technology for water quality control, Sensors
and Actuators A: Physical, Volume 120, Issue 2, 2005.
[33] S. Zhuiykov, D. OBrien, M. Best, Water quality assessment by integrated
multi-sensor based on semiconductor RuO2 nanostructures, Meas. Sci.
Tech, nol. 20, 2009.
[34] B. OFlyrm, R. Martinez, J. Cleary, C. Slater, F. Regan, D. Diamond,
H. Murphy, SmartCoast: A Wireless Sensor Network for Water Quality
Monitoring, in 32nd IEEE Conference on Local Computer Networks,
2007.
[35] P. Corke, T. Wark, R. Jurdak, H. Wen, P. Valencia, D. Moore, Envi-
ronmental Wireless Sensor Networks, Proceedings of the IEEE , vol.98,
no.11, pp.1903-1917, Nov. 2010.
[36] S. McKenna, W. Mark and K. Katherine. Detecting changes in water
quality data, Journal of the American Water Works Association, vol. 100,
no.1,pp.74-85, 2008.
[37] A. Jonathan, M. Housh, L. Perelman and A. Ostfeld. A dynamic
thresholds scheme for contaminant event detection in water distribution
systems, in Water research, vol. 47, no. 5,pp. 1899-1908, 2013.
[38] Harvey B. Mitchell, Multi-Sensor Data Fusion: an introduction.
Springer, 2007.
[39] C. Jimenez-Jorquera, J. Orozco and A. Baldi, ISFET Based Microsensors
for Environmental Monitoring. in Sensors, vol.10, pp. 61-83, 2010.