2014 3rd author contamination event detection using multiple types of conventional water quality...

EnvironmentalScienceProcesses & Impacts

PAPER

Publ

ishe

d on

16

May

201

4. D

ownl

oade

d by

Tsi

nghu

a U

nive

rsity

on

26/0

1/20

15 0

5:52

:38.

View Article OnlineView Journal | View Issue

Contamination e

School of Environment, Tsinghua Unive

[email protected]; Tel: +86 1062

† Electronic supplementary informa10.1039/c4em00188e

Cite this: Environ. Sci.: ProcessesImpacts, 2014, 16, 2028

Received 31st March 2014Accepted 16th May 2014

DOI: 10.1039/c4em00188e

rsc.li/process-impacts

2028 | Environ. Sci.: Processes Impacts

vent detection using multipletypes of conventional water quality sensors insource water†

Shuming Liu,* Han Che, Kate Smith and Lei Chen

Early warning systems are often used to detect deliberate and accidental contamination events in a water

system. Conventional methods normally detect a contamination event by comparing the predicted and

observed water quality values from one sensor. This paper proposes a new method for event detection

by exploring the correlative relationships between multiple types of conventional water quality sensors.

The performance of the proposed method was evaluated using data from contaminant injection

experiments in a laboratory. Results from these experiments demonstrated the correlative responses of

multiple types of sensors. It was observed that the proposed method could detect a contamination

event 9 minutes after the introduction of lead nitrate solution with a concentration of 0.01 mg L�1. The

proposed method employs three parameters. Their impact on the detection performance was also

analyzed. The initial analysis showed that the correlative response is contaminant-specific, which implies

that it can be utilized not only for contamination detection, but also for contaminant identification.

Environmental impact

Early warning systems are oen used to detect deliberate and accidental contamination events in a water system. Conventional methods normally detect acontamination event by comparing the predicted and observed water quality values from one sensor, which suffer from high false positive and false negativerates. This paper proposes a new method for event detection by exploring the correlative relationships between multiple conventional water quality sensors,which could detect a contamination event 9 minutes aer the introduction of lead nitrate solution (0.01 mg L�1). By implementing the proposed method, theaccuracy and efficiency of an early warning system can be improved.

1. Introduction

Water systems are vulnerable to contamination accidents andbioterrorism attacks because they are relatively unprotected,accessible, and oen isolated.1–3 In 2005, for example, theSonghua River was contaminated by nitrobenzene from achemical plant explosion, which resulted in a 4- day suspensionof water supply service to Harbin, China.4 Besides this, about1906 contamination accidents occurred per year in Chinabetween 1992 and 2006.5 An emerging area of water securityresearch involves developing methods to minimize the publichealth and economic impact of a large-scale accident or attack.An intense effort is currently underway to improve analyticalmonitoring and detection of biological, chemical, and radio-logical contaminants in drinking water systems as part of theoverall aim of securing drinking water supplies.6 One approachfor avoiding or mitigating the impact of contamination is to

rsity, Beijing 100084, China. E-mail:

787964

tion (ESI) available. See DOI:

, 2014, 16, 2028–2038

establish an Early Warning System (EWS). The EWS shouldprovide a fast and accurate means to distinguish betweennormal variations and contamination events.7 Ideally, it shouldbe inexpensive, easy to maintain and integrate into networkoperations and reliable, with few false positives and negatives.8

A key part of an EWS is the detection module, which utilizesonline sensors to evaluate water quality and detect the presenceof contamination. Generally, there are two types of online waterquality sensors. The rst type refers to non-compound specicor conventional water quality sensors, which are normally usedfor checking routine water quality parameters, including pH,chlorine, total organic carbon (TOC), oxidation reductionpotential (ORP), conductivity and temperature. Many commer-cially available technologies for these parameters provide reli-able means of detecting anomalies within water systems. Thesecond type is compound specic water quality sensors oradvanced sensors, which are capable of conrmative detectionat low concentrations for a specic component and are mainlybased on emerging detection technologies. Examples are anAlgae Toximeter for detection of the presence of toxicsubstances,9 a Daphnia Toximeter for pesticides,10 a FishActivity Monitoring System for toxins,11 biological sensors

This journal is © The Royal Society of Chemistry 2014

http://crossmark.crossref.org/dialog/?doi=10.1039/c4em00188e&domain=pdf&date_stamp=2014-07-22

http://dx.doi.org/10.1039/c4em00188e

http://pubs.rsc.org/en/journals/journal/EM

http://pubs.rsc.org/en/journals/journal/EM?issueid=EM016008

Paper Environmental Science: Processes & Impacts

Publ

ishe

d on

16

May

201

4. D

ownl

oade

d by

Tsi

nghu

a U

nive

rsity

on

26/0

1/20

15 0

5:52

:38.

View Article Online

relying on the detection of specic biomolecules (includingadenosine triphosphate (ATP), enzymes, immunoassay andpolymerase chain reaction (PCR) techniques), evaporative lightscattering, refractive index measurement, uorescence, andRaman spectroscopy.12–14

Although compound specic sensors are capable of con-rmative detection for contaminants at low concentration, theirapplication in EWS for quick contamination screening anddetection is not popular since the type of potential contaminantwas unknown at the time of sensor selection. It is difficult todetermine which contaminant must be tested at such an earlystage. In the past, conventional water quality sensors weregenerally used by operators in process control and regulatorycompliance. In recent years, they have played a growing role inEWS. These sensors are advantageous in operationaleconomics.15 This type of ‘dual-use’ is practically attractive. Forexample, in the Water Security Initiative program in the UnitedStates, pH, turbidity, temperature, conductivity, TOC, andchlorine were chosen on the basis of their sustainability forlong-term operation, and to provide ‘dual-use’ benets todrinking water utilities, such as improved water qualitymanagement.16,17

For an EWS with conventional water quality sensors,performance is highly dependent on the contamination detec-tion method. Numerous publications have been devoted todiscussing different ways of event detection using data fromconventional water quality sensors. These mainly includestatistical, articial intelligence and data mining methods. Forexample, Hart et al.18 reported two-state estimation basedalgorithms for event detection, a linear prediction coefficientlter (LPCF) and a multivariate nearest-neighbor (MVNN)algorithm. These algorithms process the water quality data ateach time step to identify periods of anomalous water qualityand provide the probability of a water quality event existing atthat time step. The LPCF method predicts the water quality at afuture time step and evaluates the residual between predictedand observed water quality values. The MVNN approachprovides a measure of the similarity between the sampled waterquality and the previously measured samples contained in thehistory window. Klise and McKenna19 developed an algorithmto classify the current measurement as normal or anomalous bycalculating the multivariate Euclidean distance. Bucak andKalik20 and Bouamar and Ladjal21 utilized articial neuralnetworks (ANN) and support vector machines (SVM) to classifywater quality data into normal and anomalous classes. Acommon feature of these methods is to compare observed andpredicted responses from time series data for one sensor.Meanwhile, several papers were published on optimal locationof sensors for event detection and response actions.22–25 In mostof these studies, an ideal sensor is assumed. In fact, routineoperation or equipment noise could result in uctuation, whichmight lead to a high false alarm rate.

Several researchers have reported the phenomenon ofcorrelative responses. For example, Hall et al.26 reported asensor response experiment for 9 types of contaminants andrealized that more than one sensor responded to each testedcontaminant. Aer noticing this phenomenon, researchers


have attempted to develop contaminant detection methodsusing responses from multiple types of sensors. Yang et al.15

explored a real-time event adaptive detection, identication andwarning (READiw) methodology in a drinking water pipe. Thesuggested adaptive transformation of sensory measurementsreduced background noise and enhanced contaminant signals.In the method employed by Yang et al.,15 the relative value ofconcentrations of free and total chlorine, pH and ORP wereused for contaminant classication. This allowed for contami-nant detection and further classication based on chlorinekinetics. Kroll27 reported the Hach HST approach usingmultiple sensors for event detection and contaminant identi-cation. In the Hach HST approach, signals from 5 separateorthogonal measurements of water quality (pH, conductivity,turbidity, chlorine residual, TOC) were processed from a5-paramater measure into a single scalar trigger signal. Thedeviation signal was then compared to a preset threshold level.If the signal exceeded the threshold, the trigger was activated.28

Murray et al.29 used Bayesian belief networks to detect acontamination event based on data from multiple sensors.Aliker and A. Ostfeld30 reported the application of SVM to detectcontamination based on data from multiple sensors. Perelmanet al.31 and Arad et al.32 reported a general framework thatintegrates a data-driven estimation model with sequentialprobability updating to detect quality faults in water distribu-tion systems using multivariate water quality time series. Inparticular, in Arad et al.'s work, univariate event probabilitiesare fused to give a unied multivariate event probability. Themultivariate probability reects the likelihood of a contamina-tion event based on all data analyzed from all parameters.32 Acommon feature of these methods is that efforts based onmultivariate water quality measurements consider correlationsbetween parameters, even if a correlation coefficient is some-times not calculated explicitly.

In this paper, we will describe a new method for real-timecontamination detection using multiple types of conventionalwater quality sensors for source water. The proposed methodaims to achieve contamination detection by exploring thecorrelative relationship between responses from multiplesensors for the same type of contaminant. The proposedmethod is tested using data from contaminant dosing experi-ments in a laboratory.

2. Materials and methods2.1 Pilot-scale contaminant injection and monitoringsystem

The pilot-scale system used in this study is a recirculatingsystem simulator in the School of Environment Laboratory atTsinghua University, Beijing, China. A process ow schematicof the pilot-scale system used for baseline establishment andsingle-pass contaminant tests is shown in Fig. 1 and thephotograph of the experimental setup is shown in P1, ESI.† Thewater tank is approximately 85 cm high with a diameter of 70cm, and has a total capacity of 300 L. The tank is linked withonline water quality sensors via a peristaltic pump at 0.5 L perminute.

Environ. Sci.: Processes Impacts, 2014, 16, 2028–2038 | 2029


Fig. 1 A process flow schematic of the pilot-scale system.

Environmental Science: Processes & Impacts Paper

Publ

ishe

d on

16

May

201

4. D

ownl

oade

d by

Tsi

nghu

a U

nive

rsity

on

26/0

1/20

15 0

5:52

:38.

View Article Online

The system was operated in recirculation mode for baselineestablishment. In this mode, 300 L source water ows throughthe multi-sensors and back to the tank. The characteristic of thesource water is shown in Table 1. The entire volume of water inthe loop is replaced every 72 hours if no contaminant test isconducted. Generally, the process of establishing baseline takes2–3 hours before any contaminant experiments can be carriedout. When operating in single-pass contaminant mode, thetarget contaminant is injected into the pipe connecting the tankand sensors via another peristaltic pump.33 It is injected at arate of 2–20 mL per minute depending on the concentrationrequirement. The water combined with contaminant owsthrough the sensors directly into a specic waste liquid bucket,avoiding pollution of the water in the tank.

2.2 Sensors investigated

8 sensors developed by Hach Homeland Security Technologieswere utilized in this study. They can measure the following 8parameters simultaneously and continuously: temperature, pH,turbidity, conductivity, oxidation reduction potential (ORP), UV-254, nitrate-nitrogen and phosphate. Table 2 shows a list of the

Table 1 Characterization of source water in current study

Parameter Concentration Parameter Concentration

Temperature 10 �C pH 7.31DO 12.77 mg L�1 Turbidity 1.39 NTUCOD 4 mg L�1 BOD5 <2 mg L�1

Conductivity 690 ms cm�1 NH3–N 0.03 mg L�1

ORP 350 mV NOx–N 3.36 mg L�1

Sulfate 170 mg L�1 Total phosphorus 0.06 mg L�1

Chloride 28 mg L�1 Sulde <0.02 mg L�1

2030 | Environ. Sci.: Processes Impacts, 2014, 16, 2028–2038

parameters and the detailed information of their associatedsensors.

2.3 Contaminants investigated

Specic quantities of various contaminants were injected intothe system simulator. The contaminants were determinedaccording to statistical reports on water pollution incidents inurban water supply systems in China for the past 20 years andincluded three types of the most commonly seen pollutants:herbicides (glyphosate), pesticides (atrazine) and heavy metals(lead nitrate, cadmium nitrate, nickel nitrate and trivalentchromium). They were also selected based on China's nationalstandards regarding source water quality GB3838-2002. Theconcentration ranges were decided by the concentration limitgiven in the standards (Table 3).

2.4 Experimental procedure

Sensors were calibrated in accordance with the manufacturer'srecommendations and were veried with a calibration checkstandard. Before the introduction of contaminants, the experi-mental system was kept running to establish a baseline. Sensordata were collected continuously and archived electronically toestablish stable baseline conditions and to record sensorresponses to injected contaminants. Data from the ORP, nitrate,temperature, pH, conductivity, turbidity and UV sensors weremonitored and recorded every 1 minute during the test period,while the phosphate sensor was recorded every 5 minutes. Aerthe baseline was established, a specic concentration ofcontaminant was injected. Each contaminant injection lastedfor over 30 minutes to reach a stabilized reading. The sensorswere then supplied with uncontaminated raw water and theresponses returned to the baseline. Another different



Table 2 Detailed information of the parameters and sensors

Parameter Sensor name Measuring range Sensitivity Measuring interval

Temperature DPD1R1-WDMP �10–50 �C �0.01 �C 1 minpH DPD1R1-WDMP �2.00–14.00 �0.01 1 minTurbidity LXV423.99.10100 0.001–4000 NTU �0.001 NTU 1 minConductivity D3725E2T-WDMP 0–2 000 000 ms cm�1 �1 ms cm�1 1 minORP DRD1R5-WDMP �1500–1500 mV �0.5 mV 1 minUV-254 LXG418.99.20000 0.01–60 m�1 �0.01 m�1 1 minNitrate–nitrogen LXG.717.99.50000 0.1–100.0 mg L�1 �0.1 mg L�1 1 minPhosphate LXV422.99.20102 0.05–15 mg L�1 �0.05 mg L�1 5 min

Table 3 Concentration limits in GB3838-2002

Contaminant Concentration limit (mg L�1)

Glyphosate 0.7Lead 0.01Atrazine 0.003Nickel 0.02Chromium 0.01Cadmium 0.001


Publ

ishe

d on

16

May

201

4. D

ownl

oade

d by

Tsi

nghu

a U

nive

rsity

on

26/0

1/20

15 0

5:52

:38.

View Article Online

concentration of the same contaminant was injected aersensor responses had returned to the baseline following theprevious test.

2.5 Detection method

In this research, it is assumed that multiple water qualitysensors can respond to a contaminant simultaneously. Theproposed method detects contamination by exploring thecorrelative relationship between responses from multiple waterquality sensors. This relationship is evaluated using the Pearsoncorrelation coefficient. The window size is the number of pastobservations used to calculate the Pearson correlation coeffi-cient. For each sensor, a new observation enters the slidingwindow at every time step t and the oldest observation exits (i.e.,rst in rst out).

The value of rxy is between �1 and 1. In this study, a corre-lation indicator Cxy is calculated using�

Cxy ¼ 0 if��rxy��\thresholdindicator or x ¼ y

Cxy ¼ 1 if thresholdindicator #��rxy��\1

(1)

A contamination alarm will be trigged if

Xcx

Xcy

Cxy $ thresholdalarm (2)

in which x s y, x ˛ (pH, ORP, UV.), y ˛ (pH, ORP, UV.). Thevalue of thresholdindicator and thresholdalarm can be determinedbased on experimental and practical analyses.

The performance of the detection method is measuredthrough detection time (DT), true positive rate (TPR) and falsepositive rate (FPR). TPR and FPR can be calculated by

TPR ¼ TP

TPþ FN(3)


FPR ¼ FP

FPþ TN(4)

where TP (true positive) is the detection of an actual event(alarm on); FP (false positive) refers to a routine operation beingincorrectly classied as a contamination event (alarm on); TN(true negative) refers to a routine operation correctly beingclassied as such (alarm off); FN (false negative) is the situationthat an actual event is not detected (alarm off). A greater TPRmeans themethod is more capable to detect a real event, while asmaller FPR implies the method is less likely to classify aroutine operation as an event.

DT is dened as the time difference between a contaminationevent taking place and when it is detected, and is evaluated by

DT ¼ T1 � T0 (5)

where T0 is the time when the contamination event occurs andT1 is the time when the contamination event is detected. Asmaller DT means the detection method is more effective andcan detect contamination within a shorter time frame.

In this study, the calculation is based on a 1 minute step. Acontaminant injection with period of t is assumed to be tcontamination events. Then TPR is used to evaluate theperformance. For example, for a contamination injection withperiod of 30 minutes, if a contamination event is rst detectedat the 10th minute and 4 contamination events are detectedwithin the remaining 20 minutes, DT and TPR will be 10minutes and

ð1þ 4Þ30

¼ 0:17:

2.6 Evaluation of reproducibility

In this research, the reproducibility of the proposed detectionmethod is evaluated using a concordance correlation coefficientmethod.34 Taking the Pearson correlation coefficients as anexample, the concordance correlation coefficient method isbriey introduced here. Let us assume that pairs of Pearsoncorrelation coefficients of one type of sensor to the others (Yil,Yi2), i ¼ 1, 2, ., n, are calculated from two independentcontaminant injection experiments with means of m1 and m2

and covariance matrix �s1

2 s12

s12 s22

�(6)




Publ

ishe

d on

16

May

201

4. D

ownl

oade

d by

Tsi

nghu

a U

nive

rsity

on

26/0

1/20

15 0

5:52

:38.

View Article Online

The concordance correlation coefficient is calculated by

rc ¼2b1s2

2

ðs12 þ s2

2Þ þ ½b0 þ ðb1 � 1Þm2�2(7)

in which b1 ¼ (s1/s2)r and b0 ¼ m1 � b1m2 represent theregression slope and intercept, respectively. r is the Pearsoncorrelation coefficient. If rc ¼ 1, two groups of Pearson corre-lation coefficients of one sensor from two experiments are inperfect agreement or in perfect reversed (if rc ¼ �1) agreement.For more information about the concordance correlation coef-cient method, the readers can refer to Lin's work.34

3. Experiments and results3.1 Correlative responses

As an example, the results from the experiment involving leadnitrate are shown in Fig. 2. The experimental results forglyphosate, atrazine, cadmium nitrate, nickel nitrate andtrivalent chromium can be found in the ESI† (see Fig. F1–F5). Inthe experiment, lead nitrate solutions with concentrations of0.01 mg L�1, 0.02 mg L�1, 0.04 mg L�1 and 0.08 mg L�1 wereadded in sequence. The concentrations are at the sensors, notin the contaminant tank. This is illustrated using solid greenbars at the top of Fig. 2. As shown in Fig. 2, ORP and nitrateincrease due to the presence of lead nitrate, while pH decreases.Sensor responses show correlative relationships, especially forpH, nitrate, ORP and temperature. This suggests that thecorrelative response is caused by the introduction of contami-nant and implies that this type of phenomenon can be utilizedfor detection of the presence of contamination. Themagnitudesof the sensors' responses were related to contaminant concen-trations. In order to justify the applicability of the proposedmethod, the lead nitrate injection experiment with the sameconguration was repeated with drinking water. The sensors'responses are shown in Fig. F6 (ESI†). As shown in Fig. F6,† theresponses of sensors are similar to the case of source water. Thissuggests that the proposed method can also be used for

Fig. 2 Sensor responses for lead nitrate (concentrations: 0.01, 0.02, 0.0


contamination detection in drinking water. By comparing withresults from other types of contaminants (see Fig. F1–F6, ESI†),it is clear that the response curves are contaminant-specic,which implies that the correlative response could be utilized notonly for contamination detection, but also for contaminantidentication.

Table 4 summarizes the responding sensors for differentcontaminants. As shown in Table 4, the nitrate, conductivityand UV sensors have correlative responses to the introductionof atrazine. For ORP, pH, conductivity and nitrate also respondsimultaneously to the presence of cadmium nitrate. ORP showsresponse to atrazine, glyphosate and lead nitrate. Other studieshave also revealed a similar phenomenon. For example, Hallet al.26 conducted a sensor response experiment for 9 types ofcontaminants and realized that there was more than one sensorresponding to each tested contaminant. Yang et al.15 alsoreported similar ndings from an experiment for 11 contami-nants. Drinking water was used in the studies by Hall et al. andYang et al., while source water is utilized in the current study.The sensor arrays adopted differed and manufacturers werealso different in these three experiments. Therefore, compari-sons are difficult to make. However, as shown in Table 4, for alltested contaminants, it was found that multiple types of sensorsrespond to the introduction of a contaminant and a sensor canrespond to different types of contaminants. This veried thecorrectness of the assumption of the proposed method.

3.2 Contamination detection

Taking lead nitrate as an example, the implementation of theproposed method is demonstrated here. Fig. 3 presents theresponses of 8 types of sensors before and aer introduction oflead nitrate solution with concentration of 0.01 mg L�1 (the 1st

lead nitrate experiment). As shown in Fig. 3, from the 1st to the60th minute, the system was running on baseline and nocontaminant was added. The lead nitrate solution was intro-duced at the 61st minute and lasted for 37 minutes. The values

4, 0.08 mg L�1).



Table 4 The responding sensors to different contaminantsa

Contaminant

Responding sensors

Hall et al. (2007) Yang et al. (2009) Current studySensor array Sensor array Sensor arrayA, B, D, E, G, H, I A, B, C, D, F D, F, G, I, J, K, L, MWater type Water type Water typeDrinking water Drinking water Source water

Aldicarb A, B, D, E, I A, B, C, DArsenic trioxide A, B, D, G, H, IAtrazine D, G, I, J, K, L, MCadmium nitrate D, F, G, I, J, K, MColchicine A, B, C, DDicamba D, FE. coli A, B, E, H, I A, B, C, D, FGlyphosate A, C, D, E A, B, C, D, F D, F, G, I, J, MK ferricyanide B, C, DLead nitrate D, F, G, I, J, MMalathion A, D, E, IMercuric chloride A, C, D, FNickel nitrate D, F, G, I, JNicotine A, B, D, E, G, H A, B, D, FNutrient broth A, B, C, DPotassium ferricyanide A, C, D, E, G, H,Terric broth A, B, D, E, I A, B, D, FTrivalent chromium D, F, G, I, JTyptic soy broth A, B, C, D

a Note: A – free chlorine; B – total chlorine; C – chloride; D – ORP; E � TOC; F – pH; G – nitrate–nitrogen; H – ammonia-nitrogen; I – turbidity; J –temperature; K – conductivity; L – UV; M – Phosphate.

Fig. 3 Sensor responses for lead nitrate (concentration: 0.01 mg L�1).


Publ

ishe

d on

16

May

201

4. D

ownl

oade

d by

Tsi

nghu

a U

nive

rsity

on

26/0

1/20

15 0

5:52

:38.

View Article Online

of thresholdindicator, thresholdalarm and window size adoptedwere 0.8, 6 and 30 respectively. The Pearson correlation coeffi-cients for each couple of sensors and the correlation indicatorsfor ‘no contamination’ and ‘contamination’ scenarios werecalculated and are listed in Tables 5 and 6 respectively. The ‘nocontamination’ and ‘contamination’ scenarios represent twoextreme situations. In the ‘no contamination’ scenario, thedataset containing observations from the 31st to 60th minute(the baseline) were used, while in the ‘contamination’ scenario,


data from the 69th to 98th minutes (injection of lead nitrate)were adopted.

As shown in Table 5, the relationships between sensors'responses in the ‘no contamination’ scenario are weak. Onlyturbidity and pH show a moderate negative relationship with acoefficient of �0.61. All Pearson correlation coefficient valuesare smaller than the value of thresholdindicator, and the value ofthe correlation indicator is 0. Therefore, no contaminationalarm was trigged off for the ‘no contamination’ scenario.



Table 5 Pearson correlation coefficients and correlation indicators (no contamination)a

Turb. pH Cond. Temp ORP Nitra. UV Phos.

Turb. 1.00(0) �0.61(0) 0.34(0) 0.04(0) 0.09(0) �0.12(0) �0.19(0) 0.00(0)pH �0.61(0) 1.00(0) �0.45(0) �0.25(0) 0.14(0) 0.49(0) 0.07(0) 0.00(0)Cond. 0.34(0) �0.45(0) 1.00(0) 0.16(0) 0.08(0) �0.13(0) 0.05(0) 0.00(0)Temp. 0.04(0) �0.25(0) 0.16(0) 1.00(0) 0.28(0) �0.35(0) 0.20(0) 0.00(0)ORP 0.09(0) 0.14(0) 0.08(0) 0.28(0) 1.00(0) 0.09(0) 0.17(0) 0.00(0)Nitra. �0.12(0) 0.49(0) �0.13(0) �0.35(0) 0.09(0) 1.00(0) �0.35(0) 0.00(0)UV �0.19(0) 0.07(0) 0.05(0) 0.20(0) 0.17(0) �0.35(0) 1.00(0) 0.00(0)Phos. 0.00(0) 0.00(0) 0.00(0) 0.00(0) 0.00(0) 0.00(0) 0.00(0) 1.00(0)Sum of correlation indicator: 0

a Note: numbers outside brackets are Pearson correlation coefficients; numbers within brackets are correlation indicators.

Table 6 Pearson correlation coefficients and correlation indicators (lead nitrate, the 1st experiment)a


Turb. 1.00(0) 0.59(0) �0.14(0) �0.68(0) �0.75(0) �0.50(0) �0.38(0) 0.73(0)pH 0.59(0) 1.00(0) �0.26(0) �0.98(1) �0.89(1) �0.99(1) �0.48(0) 0.82(1)Cond. �0.14(0) �0.26(0) 1.00(0) 0.32(0) 0.28(0) 0.25(0) �0.07(0) �0.27(0)Temp. �0.68(0) �0.98(1) 0.32(0) 1.00(0) 0.93(1) 0.95(1) 0.38(0) �0.84(1)ORP �0.75(0) �0.89(1) 0.28(0) 0.93(1) 1.00(0) 0.84(1) 0.43(0) �0.84(1)Nitra. �0.50(0) �0.99(1) 0.25(0) 0.95(1) 0.84(1) 1.00(0) 0.49(0) �0.76(0)UV �0.38(0) �0.48(0) �0.07(0) 0.38(0) 0.43(0) 0.49(0) 1.00(0) �0.44(0)Phos. 0.73(0) 0.82(1) �0.27(0) �0.84(1) �0.84(1) �0.76(0) �0.44(0) 1.00(0)Sum of correlation indicator: 18

a Note: numbers outside brackets are Pearson correlation coefficients; numbers within brackets are correlation indicators.

Table 7 The correlation indicator and detection timea

Time(m)

Sumof Cxy

DT(m)

Time(m)

Sumof Cxy

DT(m)

Time(m)

Sumof Cxy

DT(m)

30 2 N/T 53 0 N/T 76 6 1631 2 N/T 54 0 N/T 77 6 1732 2 N/T 55 0 N/T 78 6 1833 2 N/T 56 0 N/T 79 6 1934 2 N/T 57 0 N/T 80 6 2035 2 N/T 58 0 N/T 81 10 2136 2 N/T 59 0 N/T 82 10 2237 2 N/T 60 0 N/T 83 10 2338 2 N/T 61 0 N/T 84 8 2439 2 N/T 62 0 N/T 85 8 2540 2 N/T 63 0 N/T 86 10 2641 2 N/T 64 2 N/T 87 10 2742 2 N/T 65 2 N/T 88 12 2843 2 N/T 66 2 N/T 89 12 2944 2 N/T 67 4 N/T 90 12 3045 2 N/T 68 2 N/T 91 12 3146 2 N/T 69 6 9 92 12 3247 2 N/T 70 6 10 93 12 3348 0 N/T 71 6 11 94 12 3449 0 N/T 72 6 12 95 12 3550 0 N/T 73 6 13 96 14 3651 0 N/T 74 6 14 97 16 3752 0 N/T 75 6 15 98 18 38

a Note: N/T means ‘not detected’.



Publ

ishe

d on

16

May

201

4. D

ownl

oade

d by

Tsi

nghu

a U

nive

rsity

on

26/0

1/20

15 0

5:52

:38.

View Article Online

In the ‘contamination’ scenario, as shown in Table 6, ORPand nitrate show strong negative relationships with pH. ThePearson correlation coefficients are �0.88 (ORP) and �0.99(nitrate) respectively. ORP shows strong positive relationshipswith nitrate (0.84) and negative relationships with phosphate(�0.84) and pH (�0.89). The value of the correlation indicatorwas calculated to be 18. Based on this evaluation, a contami-nation event was conrmed at the 99th minute.

Fig. 3 shows that graphs for turbidity, conductivity and UVhave a number of peaks and troughs. No signicant differencesbefore and aer introduction of lead nitrate (the le and rightparts of each graph) can be observed. The peaks and troughs aremainly due to equipment noises. These noises are independentand not related to other sensors' responses. This is veried bythe weak Pearson correlation coefficients for turbidity,conductivity and UV in Tables 5 and 6. This also suggests thatturbidity, conductivity and UV do not respond to the presence oflead nitrate. If a detection decision were drawn in the light ofthese peaks or troughs, false positive and false negative errorswould be obtained.

A common question for the contamination detectionmethod is how fast the contamination event can be detected orwhat the detection time is. In a practical situation, the proposedmethod will calculate the Pearson correlation coefficients andcorrelation indicators, and make a detection decision at eachtime step (1 minute for the sensors used in this research) oncethe new readings from online sensors are received. As shown bythe rectangle with a dashed line in Fig. 3, the calculation starts



Table 8 Summary of detection performance for differentcontaminants

Contaminant TPR FPRDetectiontime (minute)

Glyphosate 0.80 0.10 2Cadmium 1.00 0.00 1Atrazine 1.00 0.00 1Nickel 0.76 0.00 3Chromium 0.79 0.00 8thresholdindicator ¼ 0.6, thresholdalarm ¼ 8 and window size ¼ 30


Publ

ishe

d on

16

May

201

4. D

ownl

oade

d by

Tsi

nghu

a U

nive

rsity

on

26/0

1/20

15 0

5:52

:38.

View Article Online

from the 30th minute with the time step of 1 minute. The sumsof correlation indicators and detection time are listed inTable 7. It is shown that the proposed method can detect acontamination event 9 minutes aer a 0.01 mg L�1 lead nitratesolution is added to the water with a window size of 30, athresholdindicator value of 0.8 and a thresholdalarm value of 6.Meanwhile, using eqn (3) and (4), the TPR and FPR werecalculated to be 78.95% and 0, respectively.

Meanwhile, for the contaminants examined in this study,Table 8 summarizes detection performances. As shown in Table8, by taking the default parameter values, the proposed methodhas rather good performance for all contaminants underdiscussion.

Fig. 4 Impact of thresholdindicator on detection performance.

4. Discussion4.1 Discrimination of equipment noise and contamination

In previous studies, a detection alarm was normally set off if thedeviation between the predicted and observed sensor values wasgreater than a preset threshold value. However, this type ofjudgment is greatly dependent on the reliability and stability ofa sensor. For example, as shown in Fig. 3, some peaks andtroughs (e.g. point a in the UV graph) shied signicantly fromthe previous reading due to equipment noise. This type of shiis difficult to predict and a big deviation between prediction andobservation can be expected. If the detection decision is madebased on the deviation, a false positive error would occur.Taking the autoregressive moving average method reported byHou35 as an example, and setting the 1% deviation of predictionand observation as the detection threshold, point a in Fig. 3(also in Fig. F7 in the ESI†) was grouped into a contaminationevent, which is obviously a false positive. Meanwhile, Fig. F7†shows all the deviations between the predictions and observa-tions. The red line is the detection threshold. The TPR and FPRwere calculated to be 0.28 and 0.32. In this study, the proposedmethod overcomes this by exploring the correlative responsebetween sensors. As shown in Fig. 3 and Table 5, althoughconductivity, UV and ORP show obvious uctuations at thesame time period, their correlative relationships are weak,which means the uctuations are more related to the equip-ment noise than external reasons, for example, presence ofcontamination. By exploring the internal relationship proposedin this method, the inuence of equipment reliability andstability on detection can be reduced.


In the case of ORP, as shown in Tables 5 and 6, the Pearsoncorrelation coefficients of ORP with other sensors are 0.09(turbidity), 0.14 (pH), 0.08 (conductivity), 0.09 (nitrate), 0.17(UV)and 0.00 (phosphate) for the le part (no contamination) and�0.75 (turbidity), �0.89 (pH), 0.28 (conductivity), 0.84 (nitrate),0.43 (UV) and�0.84 (phosphate) for the right part (contaminantadded) respectively. This means the uctuations in the le partare mainly due to random equipment noises, while the uctu-ations in the right part are mainly due to the introduction ofcontaminant. The signicant values of Pearson correlationcoefficients in Table 6 also further indicate the correlativeresponses to the introduction of lead nitrate. A key to an effi-cient contamination detection method is being able todiscriminate between these two types of uctuations. As shownin the le part of the ORP graph, the peaks or troughs shisignicantly from other readings. If a detection decision ismade based on the deviation of response from the sensor, afalse positive can be expected. In the proposed method, thedifference between these two types of uctuations is evaluatedand differentiated using the Pearson correlation coefficients.

4.2 Impacts of parameters

In the proposed method, there are three parameters: thresh-oldindicator, thresholdalarm and window size. The values of theseparameters might inuence the performance of the detectionmethod. In order to understand this, their impact on theperformance of the detection method was investigated. Tofacilitate the analysis, the other two parameters were keptconstant when analyzing one parameter. The default values forthresholdindicator, thresholdalarm and window size were 0.8, 6 and30 respectively. The performance of detection was evaluatedusing TPR and FPR.

The data used for this analysis were originally from the rstlead nitrate experiment (Fig. 2) with some arbitrary combina-tions. Four datasets were regrouped. Datasets 1, 2, 3, and 4 arethe data for baseline and for lead nitrate with concentration of0.01 mg L�1, 0.02 mg L�1, 0.04 mg L�1 and 0.08 mg L�1

respectively. For each parameter conguration, TPR and FPRwere calculated with the time step of 1 minute. The totalnumber of calculations for each dataset for 1 parameter equalsthe difference between the length of the dataset and the windowsize. The averaged TPR and FPR over the entire time period were



Fig. 5 Impact of thresholdalarm on detection performance.

Fig. 6 Impact of window size on detection performance.

Table 9 The Pearson correlation coefficients and the concordancecorrelation coefficients rc (lead nitrate, the 2nd experiment)


Turb. 1.00 0.53 �0.09 �0.62 �0.73 �0.45 �0.46 0.71pH 0.53 1.00 �0.18 �0.98 �0.88 �0.99 �0.62 0.74Cond. �0.09 �0.18 1.00 0.24 0.17 0.17 0.00 �0.22Temp. �0.62 �0.98 0.24 1.00 0.92 0.96 0.56 �0.78ORP �0.73 �0.88 0.17 0.92 1.00 0.83 0.61 �0.83Nitra. �0.45 �0.99 0.17 0.96 0.83 1.00 0.61 �0.67UV �0.46 �0.62 0.00 0.56 0.61 0.61 1.00 �0.53Phos. 0.71 0.74 �0.22 �0.78 �0.83 �0.67 �0.53 1.00rc 0.81 0.96 �0.06 0.96 0.86 0.95 0.53 0.88


Publ

ishe

d on

16

May

201

4. D

ownl

oade

d by

Tsi

nghu

a U

nive

rsity

on

26/0

1/20

15 0

5:52

:38.

View Article Online

deemed to be the true positive rate and false positive rateassociated with this parameter conguration. The analysisresults are shown in Fig. 4–6, in which TRP 1, TPR 2, TPR 3 andTPR 4 refer to the TPRs for the concentration of 0.01 mg L�1,0.02 mg L�1, 0.04 mg L�1 and 0.08 mg L�1.

4.2.1 Parameter thresholdindicator. In the analysis forthresholdindicator, its value changes from 0.6 to 0.9 with the stepof 0.1 while window size and thresholdalarm are kept constant. Asshown in Fig. 4, FPR and TPRs decrease with an increase in thethresholdindicator value. This indicates that a small thresh-oldindicator value could incorrectly classify equipment noise as acontamination event. Meanwhile, it also suggests that a highthresholdindicator might result in the overlook of a realcontamination event.

4.2.2 Parameter thresholdalarm. In the analysis for thresh-oldalarm, its value changes from 4 to 12 in step of 1. As shown inFig. 5, when the thresholdalarm increases, both FPR and TPRsdecrease. From thresholdalarm ¼ 5, FPR approaches 0, whichmeans a large thresholdalarm value can signicantly reduce falsepositives. However, a large thresholdalarm can also lead to falsenegatives and result in a low true positive rate, especially in thecase of low concentrations. It also shows that the TPR1 fordataset 1 (lead nitrate concentration at 0.01 mg L�1) dropssignicantly with an increase in the thresholdalarm value. Thismight be because the correlative responses are rather weak atlow concentration. This is consistent with the graphs in Fig. 2.

4.2.3 Parameter window size. The window size denotes thenumber of data involved in the calculation of the Pearsoncorrelation coefficient. In this analysis, the window sizeincreases from 8 to 34 in step of 2. The maximum value of the


window size is based mainly on the contaminant injectionperiod. As shown in Fig. 6, when the window size is smaller than18, TPRs decrease with an increase in the value of the windowsize, while they keep rather at aer 18. This implies that theperformance of the detection method is more sensitive to smallvalues of the window size. Fig. 6 also suggests that FPR reachespeak values with medium values of the window size

From Fig. 4–6, it is concluded that the values of parametershave impacts on the performance of the detection method,which suggests that they should be determined carefully toachieve a better detection performance in practical applica-tions. For example, by taking the value of thresholdindicator ¼0.75, thresholdalarm ¼ 6 and window size ¼ 30, as shown inFig. 4, a 0.01 mg L�1 lead nitrate contamination event can bedetected with a true positive rate of 80% and false positive rateof 2%. The detection time is 9 minutes (shown in Table T1 inthe ESI†). A comprehensive sensitivity analysis would benetthe implementation of the proposed method and the optimalvalues of parameters should be determined for a speciccontaminant through experiment and analysis.

4.3 Reproducibility

Data from two independent lead nitrate injection experimentswere used to evaluate the reproducibility of the proposeddetection method. Fig. F8† shows the sensors' responses fromthe second lead nitrate injection with the sequence of 0.08 mgL�1, 0.04 mg L�1, 0.02 mg L�1 and 0.01 mg L�1 (Fig. 2 is for therst injection experiment). The experiment conditions for thetwo experiments are the same. Similar to Tables 6, Table 9shows the Pearson correlation coefficients for the 2nd experi-ment, which are directly calculated from the experiment data.Each column represents the Pearson correlation coefficients ofother sensors and this sensor. For example, the 2nd column liststhe Pearson correlation coefficients of turbidity and othersensors. Using eqn (7), the concordance correlation coefficientsof data in Tables 6 and 9 were obtained and are shown at thebottom of Table 9. As shown in Table 9, the concordancecorrelation coefficients for turbidity, pH, temperature, ORP,nitrate and phosphate are greater than 0.81, which suggestshigh agreement between the calculation results of the data fromthe 1st and the 2nd lead nitrate experiments. It is also noticedthat the sensors with high concordance correlation coefficientsare consistent with the one having high correlative coefficients



Fig. 7 Radar map of sensor's responses for lead and atrazine.


Publ

ishe

d on

16

May

201

4. D

ownl

oade

d by

Tsi

nghu

a U

nive

rsity

on

26/0

1/20

15 0

5:52

:38.

View Article Online

in Tables 6 and 9. This implies that the responses of thesesensors are mostly due to injection of lead nitrate, rather thanequipment noises. For conductivity and UV, the concordancecorrelation coefficients are low, which are consistent with theirlow values of Pearson correlation coefficients in Tables 6 and 9.This suggests that the responses of conductivity and VU aremostly from equipment noises. Therefore, their reproducibilityis low. This will not affect the reproducibility of the proposedmethod since the low Pearson correlation coefficient values arenot taken into account for the correlation indicator. By usingthe same parameter values (the default ones), the TPR and FPRwere calculated as being 78.95% and 0 respectively, which arethe same as the ones from the rst experiment. In summary, theproposed method has a good reproducibility.

4.4 Future work

Using lead nitrate as an example, this study shows that thedetection performance is highly sensitive to the values ofparameters. Meanwhile, in a real situation, the type ofcontaminant is not known a priori. Therefore, the values of theparameters should not be determined for a specic type ofcontaminant, but for a large group of contaminants. A gener-alized set of values of parameters is more deserved. Therefore,in the future, the optimal determination of the values of theparameters should be conducted.

In an EWS, the question aer detection of contamination ishow to identify the type of contaminant quickly. A commerciallyavailable Hach method relies on correlations betweenresponses of different types of sensors for the contaminantidentication. From this study, as shown in Fig. F1–F5,† thesensors' responses are contaminant dependent. For a clearerview, Fig. 7 shows all sensors' responses for lead nitrate andatrazine using radar map. Obviously, the shapes formed by axesfor lead nitrate and atrazine are different. By utilizing thecontaminant dependent feature, it is possible to differentiatethe types of contaminants. Therefore, in future, more studiesshould be carried out on how to extract features and patterns forcontaminant identication. Possible techniques are datamining and pattern recognition. For example, by comparing the


Euclidean distance between samples and classes for differentcontaminants, it is possible to identify the type of contaminant.

5. Conclusion

EWSs should provide a fast and accurate means to distinguishbetween normal variations and contamination events.Compared with component-specic sensors, conventionalwater quality sensors are still widely used because they are cost-effective and easily maintained. For an EWS with conventionalwater quality sensors, a key issue is how to efficiently detect thepresence of contamination. In this study, a platform with 8types of online water quality sensors was established andutilized for contaminant injection experiment. By analyzing theresults from the experiment, the following conclusions aredrawn.

(1) A contamination detection method utilizing the correla-tions between multiple types of conventional water qualitysensors was proposed in this study. The results from theexperiment and analysis showed that the proposed methodcould detect a lead nitrate contamination 9 minutes aer theintroduction of contaminant at the concentration of 0.01 mgL�1 using responses from online water quality sensors. The TRPand FPR were 80% and 2% respectively.

(2) It was noticed that multiple sensors responded simulta-neously to the presence of contamination. Initial analysisshowed that the correlative response was contaminant-specic,which implied that the correlative response could be utilizednot only for contamination detection, but also for contaminantidentication and even for quantication. Meanwhile, in someprevious studies, an ideal sensor was assumed. In a real situa-tion, this is not always the case. If the sensor fails to functionproperly, a false positive is expected. The proposed method canovercome this by utilizing the correlations between sensors.

(3) The proposed method employs three parameters:thresholdindicator, thresholdalarm and window size. From theanalysis, it was concluded that these parameters have an impacton the detection performance. For a specic contaminant, theoptimal values of parameters should be determined throughexperiment and analysis.




Publ

ishe

d on

16

May

201

4. D

ownl

oade

d by

Tsi

nghu

a U

nive

rsity

on

26/0

1/20

15 0

5:52

:38.

View Article Online

(4) The basis of the proposed method is that multiplesensors respond to one type of contaminant simultaneously.This has been veried by experiment in this study and otherstudies. However, it should be envisaged that the types ofcontaminants previously tested are still limited. More experi-ments should be conducted in the future.

Acknowledgements

This work is jointly supported by Beijing High TechnologyResearch (Z121100000312097), Development Program andTsinghua Independent Research Program (2011Z01002) andWater Major Program (2012ZX07408-002).

References

1 USEPA, Baseline threat information for vulnerabilityassessments of community water systems, Washington, DC,2002.

2 USEPA, Planning for and responding to drinking watercontamination threats and incidents, Washington, DC, 2003.

3 W. M. Grayman, R. A. Deninger and R. M. Clark, CE News,2002, 14, 34–38.

4 C. Wang, Y. J. Feng, S. S. Zhao and B. L. Li, Chemosphere,2012, 88(1), 69–76.

5 J. Yang, J. Bi, H. Y. Zhang, F. Y. Li, J. Zhou and B. Liu, ChinaEnviron. Sci., 2010, 30(4), 571–576, in Chinese.

6 J. Hasan, S. States and R. Deininger, Journal of ContemporaryWater Research and Education, 2004, 129, 27–33.

7 M. V. Storey, B. van der Gaag and B. P. Burns, Water Res.,2011, 45(2), 741–747.

8 M. Brussen, On-line Water Quality Monitoring, Sydney WaterReport, Sydney, 2007.

9 C. J. de Hoogh, A. J. Wagenvoort, F. Jonker, J. A. Van Leerdamand A. C. Hogenboom, Environ. Sci. Technol., 2006, 40, 2678–2685.

10 J. Jeon, J. H. Kim, B. C. Lee and S. D. Kim, Sci. Total Environ.,2008, 389, 545–556.

11 B. van der Gaag and J. Volz, Real-time On-line Monitoring ofContaminants in Water: Developing a Research Strategy fromUtility Experiences and Needs, KIWA Water Research,Nieuwegein, 2008.

12 R. K. Henderson, A. Baker, K. R. Murphy, A. KHambly,R. M. Stuetz and S. J. Khan, Water Res., 2009, 43(4), 863–881.

13 P. R. Hawkins, S. Novic, P. Cox, B. A. Neilan, B. P. Burns,G. Shaw, W. Wickramasinghe, Y. Peerapornpisal,W. Ruangyuttikarn, T. Itayama, T. Saitou, M. Mizuochi andY. Inamori, J. Water Supply: Res. Technol.–AQUA, 2005, 54,509–518.


14 C. P. Marshall, S. Leuko, C. M. Coyle, M. R. Walter,B. P. Burns and B. A. Neilan, Astrobiology, 2007, 7, 631–643.

15 J. Y. Yang, R. C. Haught and J. A. Goodrich, J. Environ.Manage., 2009, 90, 2494–2506.

16 USEPA, WaterSentinel System Architecture, U.S.Environmental Protection Agency, Office of Water, Office ofGround Water and Drinking Water, Washington, DC, 2005.

17 USEPA, Water Security Initiative Cincinnati Pilot Postimplementation System Status: Covering the Pilot Period:December 2005 Through December 2007, U.S. EnvironmentalProtection Agency, Office of Water, Office of Ground Waterand Drinking Water, Washington, D.C, 2008.

18 D. Hart, S. A. McKenna, K. Klise, V. Cruz and M. Wilson,Proceedings of the World Environmental and Water ResourcesCongress, ASCE, Reston, VA, 2007, pp. 1–9.

19 K. A. Klise and S. A. McKenna, Proceeding of the InternationalSociety for Optical Engineering, Saito, 2006.

20 I. O. Bucak and B. Kalik, Ekoloji, 2011, 20(78), 75–81.21 M. Bouamar and M. Ladjal, Int. J. Comput. Intell. Appl., 2012,

11(2), 1250013.22 S. Alvisi, M. Franchini, M. Gavanelli and M. Nonato,

J. Hydroinf., 2012, 14(2), 345–365.23 M. Guidorzi, M. Franchini and S. Alvisi, Urban Water, 2008,

6(2), 115–135.24 A. Ostfeld and E. Salomons, J. Water. Res. Pl.–Asce, 2004,

130(5), 377–385.25 A. Preis and A. Ostfeld, J. Hydroinf., 2008, 10(4), 267–274.26 J. Hall, A. D. Zaffiro, R. B. Marx, P. C. Kefauver,

E. R. Krishnan, R. C. Haught and J. G. Herrmann, J. Am.Water Works Assoc., 2007, 99(1), 66–77.

27 D. Kroll, Securing our water supply: Protecting a vulnerableresource, Pennwell, 2006.

28 D. Kroll and K. King, Proceeding of the 8th Annual WaterDistribution Systems Analysis Symposium, ASCE, Reston, VA,2006.

29 S. Murray, M. Ghazali and E. A. McBean, J. Water. Res. Pl.–Asce, 2012, 138(1), 63–70.

30 N. Aliker and A. Ostfeld, Water Res., 2014, 51, 234–245.31 L. Perelman, J. Arad, M. Housh and A. Ostfeld, Environ. Sci.

Technol., 2012, 46, 8212–8219.32 J. Arad, M. Housh, L. Perelman and A. Ostfeld, Water Res.,

2013, 47(5), 1899–1908.33 J. S. Hall, J. Szabo, S. Panguluri and G. Meiners, Distribution

system water quality monitoring: sensor technology evaluationmethodology and results, EPA 600/R-09/076, 2009.

34 L. I. Lin, Biometrics, 1989, 45(1), 255–268.35 D. Hou, X. Song, G. Zhang, H. Zhang and H. Loaiciga,

Environ. Sci. Pollut. Res., 2013, 20, 4496–4508.



2014 3rd author contamination event detection using multiple types of conventional water quality...

Documents