lable at ScienceDirect

Journal of Environmental Management 161 (2015) 385e391

Contents lists avai

Journal of Environmental Management

journal homepage: www.elsevier .com/locate/ jenvman

Research article

Performance evaluation for three pollution detection methods usingdata from a real contamination accident

Shuming Liu*, Han Che, Kate Smith, Musuizi Lei, Ruonan LiSchool of Environment, Tsinghua University, Beijing, 100084, China

a r t i c l e i n f o

Article history:Received 21 May 2015Received in revised form9 July 2015Accepted 10 July 2015Available online xxx

Keywords:Contamination detectionPearson correlationEuclidean distanceLinear prediction filter

* Corresponding author.E-mail address: shumingliu@tsinghua.edu.cn (S. L

http://dx.doi.org/10.1016/j.jenvman.2015.07.0260301-4797/© 2015 Elsevier Ltd. All rights reserved.

a b s t r a c t

Early warning systems have been widely deployed to safeguard water security. Many contaminationdetection methods have been developed and evaluated in the past decades. Although encouragingdetection performance has been obtained and reported, these evaluations mainly used artificial or lab-oratory data. The evaluation of detection performance with data from real contamination accidents hasrarely been conducted. Implementation of contamination event methods without full assessment usingfield data might lead to failure of an early warning system. In this paper, the detection performance ofthree contamination detection methods, a Pearson correlation Euclidean distance (PE) based detectionmethod, a multivariate Euclidean distance (MED) method and a linear prediction filter (LPF) method, wasevaluated using data from a real contamination accident. Results improve understanding of the imple-mentation of detection methods to field situations and show that all methods are prone to yieldingworse detection performance when applied to data from a real contamination accident. They alsorevealed that the Pearson correlation Euclidean distance based method is more capable of differentiatingbetween equipment noise and presence of contamination and has greater potential to be used in realfield situations than the MED and LPF methods.

© 2015 Elsevier Ltd. All rights reserved.

1. Introduction

Protection of drinking water systems from accidental andintentional contamination events has increased in importance inrecent years due to security concerns (Liu et al., 2014; Yang et al.,2009). Between 1992 and 2006, an average of 1906 contamina-tion accidents occurred per year in China (Yang et al., 2010). Forexample, the Songhua River was contaminated by nitrobenzenefrom a chemical plant explosion in 2005, which resulted in a 4 daysuspension of water supply to Harbin, China (Wang et al., 2012).One approach for avoiding or mitigating the impact of contami-nation is to establish an early warning system (EWS).

A key part of an EWS is the detection algorithm, which utilizesdata from online sensors to evaluate water quality and detect thepresence of contamination. Many studies have been conducted todevelop detection algorithms using signals from conventionalwater quality sensors. As summarized by McKenna et al. (2008),there are two approaches to developing and testing event detectionmethods using water quality sensor signals. First, laboratory and

iu).

test-loop evaluation of sensors and associated event detection al-gorithms provides direct measurement of chemical changes inbackground water quality caused by specific contaminants (Hallet al., 2007; Yang et al., 2009). Results from these physical experi-ments can be used to quantify which deviations from backgroundwater quality signals are indicative of contamination events. Theseresponses can then be integrated into event detection methods. Forexample, Yang et al. (2009) proposed a real-time event adaptivedetection, identification and warning (READiw) methodology in adrinking water pipe. The suggested adaptive transformation ofsensory measurements reduced background noise and enhancedcontaminant signals.

The second approach to event detection is based on signalprocessing and data-driven techniques (McKenna et al., 2008). Forexample, Kroll (2006) reported the Hach HST approach usingmultiple sensors for event detection and contaminant identifica-tion. Hart et al. (2007) reported a linear prediction filter (LPF). TheLPF method predicts the water quality at a future time step andevaluates the residual between predicted and observed waterquality values. Klise and McKenna (2006) developed an algorithmto classify the current measurement as normal or anomalous bycalculating the multivariate Euclidean distance (MED). The MED

S. Liu et al. / Journal of Environmental Management 161 (2015) 385e391386

approach provides a measure of the distance between the sampledwater quality and the previously measured samples contained inthe history window. Liu et al. (2015c) presented a new detectionmethod that identifies the existence of contamination bycomparing Euclidean distance of correlation indicators, which arederived from the correlation coefficients of multiple water qualitysensors. Allgeier et al. (2005) and Raciti et al. (2012) utilized arti-ficial neural networks (ANN) and support vectormachines (SVM) toclassify water quality data into normal and anomalous classes aftersupervised learning. Perelman et al. (2012) and Arad et al. (2013)reported a general framework that integrates a data-driven esti-mation model with sequential probability updating to detectquality faults in water distribution systems using multivariatewater quality time series. In general, these algorithms process thewater quality data at each time step and compare this data with apreset threshold. If the deviation is greater than the presetthreshold value, an alarm is triggered.

Researchers have attempted to evaluate the performance of thesemethods. The first group of methods has generally been evaluatedusing data from laboratory contamination injection experiments(Hall et al., 2007; Kroll, 2006; Liu et al., 2015a,b; Yang et al., 2009). Asargued by McKenna et al. (2008), a drawback of the laboratory andtest-loop results and the resulting algorithms is that variation of thebackground water quality in these systems may be considerably lessthan the variation observed in actual water systems. Evaluation ofthe performance of the second group of methods has mainly usedartificial data or data from injection experiments in laboratory. Theartificial data normally contains actual background data and artificialevent data. For example, in work by McKenna et al. (2008), waterquality data collected in awater utility in theUnited Stateswere usedto represent background water quality conditions. Simulatedanomalouswater quality events (or spikes) were then added to thesedata. Using observed hydraulics data from CANARY and simulatedcontamination event data, Perelman et al. (2012) reported that anANN based detectionmethod yielded a true positive rate of 90%withthree false alarms. The READiw method developed by Yang et al.(2009) was capable of correctly detecting all contamination eventsfor the experimental data under discussion. In a study by Liu et al.(2015c), the Pearson correlation Euclidean distance based methodwas applied to data from an injection experiment and it detected 95%of contamination events correctly with a 2% of false alarm rate. Ingeneral, the performance of these approaches is encouraging.However, these evaluations were conducted using only artificialwater quality data or laboratory data. It is unclear how these ap-proaches would perform in real contamination situations, in whichwater quality data contains much more background noise andfluctuations.

To understand the applicability of contamination detectionmethods, evaluation of these methods using data from actualcontamination accidents is necessary. The objective of this paperwas to evaluate and compare the performance of three detectionmethods using data from an actual contamination accident in awater source.

2. Methods and materials

The three methods evaluated in this study were Pearson cor-relation Euclidean distance (PE) based detection method, multi-variate Euclidean distance (MED) method and linear predictionfilter (LPF) method. These three methods are briefly introducedhere.

2.1. The PE method

In a parallel study, Liu et al. (2015c) proposed the PE method,

which includes three steps: calculation of Pearson correlation co-efficients, calculation of correlation indicators and calculation ofEuclidean distances.

Step 1: Pearson correlation coefficients for multiple sensorsignals are calculated. In a previous study, Liu et al. (2014) re-ported that multiple water quality sensors could respond to acontamination event simultaneously. This is defined as acorrelative response and is utilized in this study for eventdetection. Step 1 involves quantifying the extent of correlationusing Pearson correlation coefficients, r, which are calculated asfollows

rXY ¼Pn

i¼1�xi � X

��yi � YÞffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPn

i¼1�xi � X

�2q*

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPni¼1

�yi � Y

�2q (1)

in which X and Y refer to signal series from two separate waterquality sensors (e.g. pH and ORP). xi and yi are the ith numbers inthe signal series. X and Y stand for mathematical expectation. Thenumber of data or window size is given by n. The window size is thenumber of past observations used to calculate the Pearson corre-lation coefficient.

Step 2: The value of rXY is between�1 and 1. If the value of rXY isclose to 0, the correlation between X and Y is deemed to beweak. In this study, a correlation indicator CXY is used to denotewhether two vectors are closely related. The value of CXY iseither 0 or 1, which is obtained, as shown in Equation (2), bycomparing rXY with a pre-set indicator threshold C*.

�CXY ¼ 0 if jrXY j<C* or X ¼ YCXY ¼ 1 if C* � jrXY j � 1

(2)

Step 3: For the case of s sensors, the correlation coefficient formsan s x smatrix, as does the correlation indicator. The correlationindicators above the diagonal are taken to construct a 1 � mdimension vector V, which is called the correlation indicatorvector (Equations (3) and (4)).

26666664

1 C12 C13 / C1s�1 C1s1 C23 / C2s�1 C2s

1 / C3s�1 C3s/ / /

1 Cs�1s1

37777775

(3)

½C12C13/C1s�1C1sC23/ C2s�1C2s/C3s�1C3s/Cs�1s�|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}m

(4)

m is determined by

m ¼Xs�1

i¼1

i (5)

The Euclidean distance of the correlation indicator vector fromthe origin point, dPE, is calculated using

S. Liu et al. / Journal of Environmental Management 161 (2015) 385e391 387

dPEðtÞ ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXmk¼1

½Vk � O�2vuut (6)

in which Vk is the kth item in the correlation indicator vector V andO refers to the point of origin. The items in vector V are the valuesCXY above the diagonal in the correlation indicator matrix.

A contamination alarm will be triggered if

dPEðtÞ � d*PE (7)

in which d*PE is a detection threshold.

2.2. The MED method

The MED method considers changes in water quality bycomparing two successive distances in a multivariate space definedby the water quality signal (Klise and McKenna, 2006; McKennaet al., 2008). The distances compared are (1) the Euclidean dis-tance between the water quality measurement at the current timestep and the mean water quality value over the previous PE timesteps and (2) the Euclidean distance between the water qualityvalue of the previous time step and the mean value.

The distance measure, dMED, is the difference between theEuclidean distances of successive points to the mean of previoustime steps (PMED) in the n dimensional space. PMED denotes thenumber of sensor readings that are taken to calculate the mean.dMED is calculated using

dMEDðtÞ ¼ absðDMEDðtÞ � DMEDðt � 1ÞÞ (8)

in which

DMEDðtÞ ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXnj¼1

hZðtÞj � ZðmÞj

i2vuut

DMEDðt � 1Þ ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXnj¼1

hZðt � 1Þj � ZðmÞj

i2vuut

where n is the number of water quality parameters (or the numberof dimensions of the space), Z(t)j is the jth water quality parameterin the current time step (i.e. t), Z(t � 1)j is the jth water qualityparameter in the previous time step (i.e. t-1), and Z(m)j is the meanvalue of the jth water quality parameter. A constant detectionthreshold, d*MED, is applied to determine if an event has occurred.The MED method identifies a contamination event when dMED isabove this threshold. Otherwise, measurements are consideredbackground.

2.3. Linear prediction filters (LPF) method

A linear predictor estimates the current value of a time series asa linear combination of previous samples. This method is alsoknown as the autoregressive moving average (ARMA) predictor(McKenna et al., 2008; Zetterqvist, 1991), and the most commonrepresentation is

Z*ðtÞ ¼ �XPLPFi¼1

aiZðPLPF � iÞ (9)

inwhich Z*(t) is the predicted water quality value, Z(PLPF� i) are theprevious observed values, ai are the predictor coefficients, and PLPFis the order of the prediction filter polynomial.

The error (or distance) generated by this estimate is

DLPFjðtÞ ¼ Z*ðtÞ � ZðtÞ (10)

in which Z(t) is the observed true signal value, DLPF(t) is the dif-ference between prediction Z*(t) and observation Z(t), and j denotesthe jth water quality parameter.

dLPFðtÞ ¼1n

Xnj¼1

abshDLPFjðtÞ

i(11)

Each water quality signal is treated separately, and then allwater quality signals are combined by taking the average absolutevalue of the differences across all signals (Equation (11)). Theaverage difference or distance is then compared to the detectionthreshold d*LPF to determine whether the difference is significant. Ifthe difference is greater than the detection threshold, an alarmwillbe sounded.

2.4. Data standardization

In the MED and LPF methods, values from or derived fromdifferent sensors are accumulated. These data are in different units.Therefore, they need to be standardized to a common scale. In thisstudy, standardization is achieved by subtracting the dataset mean,m, from each water quality signal, X(t), and then dividing this dif-ference by the standard deviation, s, of the dataset:

ZðtÞ ¼ XðtÞ � m

s(12)

The m refers to themean of the dataset in the previous time stepswith a window size denoted by nml. The performance of the MEDand LPF methods is also evaluated using the ROC curve. It should benoted that data for the PE method do not need to be standardizedbecause the Pearson coefficient is dimensionless.

2.5. Performance evaluation

The accuracy of the event detection method is assessed by itsability to place the current state of water quality into one of twoclasses: background and event. Evaluation of the event detectionmethod requires that the results be examined on a common scale toassess the tradeoffs between false positive (FP) and false negative(FN) decisions as a function of the sensitivity of the detectionmethod. In this study, the received operating characteristic (ROC)curve is adopted as an evaluation tool (McKenna et al., 2008). Thiscurve has been used in several other studies (Debon et al., 2010;Wang et al., 2015).

The ROC curve defines the probability of the detection (PD) thatcan be obtained as a function of the corresponding false alarm rate(FAR). FAR is equal to the number of FPs divided by the total numberof values that are actually below the detection threshold. The PD isdefined as the number of true positives (TPs) divided by all eventsthat actually exceed the detection threshold. The area under theROC curve is the preferred single-valued measure of accuracy of thetechnique being evaluated (Swets, 1988). The maximum value ofthe area under the ROC curve is 1. The closer the area to 1, the betterthe performance yielded by the detection method.

FAR ¼ FPFP þ TN

(13)

S. Liu et al. / Journal of Environmental Management 161 (2015) 385e391388

PD ¼ TPTP þ FN

(14)

in which TN is true negative.The points on the ROC curve represent the PDs and FARs the

detection method can achieve for all possible threshold values(parameter values in this study). Each point on the ROC curverepresents a trade-off between FAR and PD. Evaluation of an eventdetection method requires that the results be examined on acommon scale to assess the tradeoffs between FP and FN decisions.To facilitate this, the false alarm rate (FAR) at a given probability ofdetection (PD) (in this case PD ¼ 0.95) is used in this study. Asmaller value of FAR at a PD of 0.95 suggests the performance of thedetection method is better.

2.6. Optimal parameter values

The three methods discussed in this paper all contain threeparameters: PE e n (window size), C* (indicator threshold) and d*PE(distance threshold); MED e nml (window size), PMED (number ofprevious time step) and d*MED (detection threshold); LPF e nml

(window size), PLPF (the order of the prediction filter polynomial)and d*LPF (detection threshold). In this study, the optimal valueswere obtained using an optimization analysis. The objective func-tion is configured to be:

Objective function ¼ maxðarea under the ROC curveÞ (15)

In the optimization analysis, a range was pre-set for eachparameter according to the problem under discussion. Taking thePE method as an example, all possible values of n, C* and d*PE fromthe ranges were taken to calculate FARs and PDs, which form theROC curves. The ROC curve with maximum area is denoted as theoptimal ROC curve. By iterating all possible values for all parame-ters, the optimal ROC curve (i.e. the curve that satisfies the objectivefunction) can be obtained. For more information about the opti-mization method, refer to Liu et al. (2015c).

2.7. Dataset

Excessive phenol was detected in Yangtze River water inFebruary 2012 due to an illegal chemical spill from a cargo ship.This accident was recorded downstream by a water quality stationequipped with 10 conventional water quality sensors. At 20:00 onFebruary 5, laboratory analysis of water samples taken from thewater quality station showed that phenol concentration (0.008 mg/l) exceeded the national limit (0.005 mg/l). The concentration ofphenol in water sampled from the Yangtze River continued toexceed the standard for another 44 h. By 16:00 February 7, phenollevels had decreased to below the standard. The data from the 10water quality sensors were recorded at a frequency of every two

Table 1Detailed information of the parameters and sensors.

Parameter Sensor type

Temperature DPD1R1-WDMPpH DPD1R1-WDMPTurbidity LXV423.99.10100Conductivity D3725E2T-WDMPChemical oxygen demand (COD) COD max plus SCFluoride FBM-1601Dissolved oxygen (DO) Zullig S-14Total organic carbon (TOC) BiotechorNitrate (NH3eN) LXG.717.99.50000Total phosphate (TP) LXV422.99.20102

hours. Table 1 shows a list of the parameters and the detailed in-formation of their associated sensors.

Sensor signals before, during and after the contamination acci-dent were taken to evaluate the performance of the detectionmethods. The dataset covered a total of 314 h and included 157individual data for each sensor. To facilitate analysis, the datasequence rather than the actual time was used. In Fig. 1, forexample, the first point refers to 14:00 January 29 and the 90th datapoint refers to 20:00 February 5 (contamination starting time).Periods A and C represent periods with no contamination events,while period B denotes the contamination period. Each time step isdefined as a contamination event or a background situation. Intotal, there are 22 contamination events and 135 backgroundsituations.

3. Results and discussions

Using the method introduced in Section 2.3, the optimal valuesof parameters were obtained and are shown in Table 2. Meanwhile,Fig. 2 displays the ROC curves obtained in the parameter optimi-zation process. The optimal ROC curve of each method is high-lighted in red. The values of areas under the optimal ROC curves inFig. 2 are listed in Table 3. As shown in Table 3, the areas under theROC curves for the PE, MED and LPFmethods are 0.83, 0.52 and 0.68respectively. This suggests that the PE method has a better detec-tion performance than the MED and LPF methods.

Meanwhile, the FAR at a PD of 0.95 for the PE method is 0.33(also shown in Fig. 2), which is lower than the FARs for theMED andLPF methods (0.88 for the MED and 0.82 for the LPF). This meansthat when the PE method detects 95% of contamination eventscorrectly, the FAR rate is compromised so that 33% of backgroundvalues are incorrectly classified as events. To achieve the same levelof probability of detection, the MEDmethod wrongly classified 88%of background as events, and the LPF method wrongly classified82% of background as events. This suggests that the detectionperformance of the MED and LPF methods is less satisfactory thanthat of the PE method. A high FAR could lead to the unnecessarydeployment of manpower for on-site water sampling (i.e. toconfirm whether or not an event really occurred) and wouldsignificantly decrease confidence in the early warning system. Theresults for areas under the ROC curves and FARs at a PD of 0.95show that the PE method demonstrated a better performance thanthe MED and LPF methods when applied with data from a fieldcontamination accident.

Mckenna et al. (2008) evaluated the performance of LPF andMED methods using artificial event data, which were obtained byinserting simulated event data into observed background data. Aspike spanning a single time step represented the simulated event.Spike strength values were in the units of the standard deviation ofobserved background data and ranged from 1.0 (no change in thewater quality) to 3.5 (a change of 3.5 standard deviation away from

Measuring range Sensitivity

�10e50 �C ±0.01 �C�2.00e14.00 ±0.010.001e4000 NTU ±0.001 NTU0e2000000 ms/cm ±1 ms/cm10e500 mg/l ±0.3 mg/l0e99.9 mg/l ±0.5 mg/l0e15 mg/L ±0.05 mg/L0e1250 mg/L ±0.3 mg/L0.1e100.0 mg/L ±0.1 mg/L0.05e15 mg/L ±0.05 mg/L

Fig. 1. Sensor readings from a real contamination accident.

Table 2The initial range and optimal values of parameters for the three methods.

Detection method Parameter Optimal values

PE n 36C* 0.25d*PE 5.29

MED nml 12PMED 6d*MED 0.49

LPF nml 34PLPF 4d*LPF 0.5

Table 3The performance of the three detection methods.

Method Area under ROC FAR at a PD of 0.95

PE 0.83 0.33MED 0.52 0.88LPF 0.68 0.82

S. Liu et al. / Journal of Environmental Management 161 (2015) 385e391 389

the mean). For each dataset, one spike strength was used to facil-itate investigation of the impact of spike strength (contaminantconcentration) on detection performance. Results from the studyby Mckenna et al. (2008) revealed that the detection performanceof the LPF and ME methods varied with the strength of spikes. Inthis study, the contamination data were not from artificial spikenumbers but from field observation. Therefore, whereas deviationof contamination data from the mean was a constant value inMcKenna et al. (2008), deviation varies from time to time in thepresent study. This makes it difficult to compare the detectionperformance obtained in this study (using field contamination

Fig. 2. ROC curves for the PE

data) with the performance obtained by McKenna et al. (2008)(using artificial contamination data). To overcome this, theaverage spike strength of the standardized water quality data for alltime steps in the contamination period was adopted to representthe strength of contamination. This was achieved by summing theabsolute value of standardized water quality data obtained from 12.The average spike strength was then calculated to be 1.2 times thestandard deviation. When applied to artificial contamination dataand when using a spike strength of 1.2, the FAR values for a PD of0.95 for the LPF and MED method were 0.75 and 0.85 (McKennaet al., 2008). However, as shown in Table 3, when applied to fieldcontamination data, the detection performance of both methodsworsened and the FAR values for a PD of 0.95 were 0.82 and 0.88.

Liu et al. (2015c) evaluated the performance of the PE methodusing data from a contaminant injection experiment. This studyreported that the FAR value for a PD of 0.95 for the PE method was

, MED and LPF methods.

Table 4The calculated distances of the LPF method at three time steps.

Distance pH DO COD TP NH3eN TOC Conductivity Turbidity Temperature Fluoride

DLPF(48) 1.50 0.64 0.64 0.57 0.67 0.60 0.09 0.57 0.59 0.98DLPF(88) 0.01 0.41 0.05 0.08 0.44 0.50 0.31 0.45 0.02 0.44DLPF(99) 0.46 0.02 0.17 0.04 3.29 0.40 2.14 0.86 1.94 0.06dLPF(48) ¼ 0.68; dLPF(88) ¼ 0.27; dLPF(99) ¼ 0.90

Table 5The Euclidean distance and distance measure for three time steps.

Euclidean distance Distance measure

DMED(48) 4.33 DMED(47) 2.17 dMED(48) 2.17DMED(88) 1.53 DMED(87) 2.21 dMED(88) 0.68DMED(99) 4.73 DMED(98) 2.35 dMED(99) 2.38

S. Liu et al. / Journal of Environmental Management 161 (2015) 385e391390

0.02. This is significantly lower than the FAR obtained from the fieldcontamination data (0.33, Table 3). Therefore, it is obvious that theperformance of all three methods decreases when applied to datafrom a real contamination accident. This is understandable becausefield conditions are more complicated and field data contain moreuncertainty and background noise, which compromises detectionaccuracy.

From the analysis above, it is clear that, while all three methodsperform worse when applied to field data, the PE method dem-onstrates more encouraging detection performance than the LPFandMEDmethods. It is instructive to further investigatewhy the PEmethod has an advantage over the LPF and MED methods. The keyto reducing FAR is to correctly identify background according tosensor signals. Online sensors are used for their ability to quicklyprovide water quality information, but they have also been criti-cized for their low stability. A common issue for online sensors isthat their signals always contain device noise. Both device noiseand change in water quality (e.g. due to the presence of contami-nant) can lead to signal variations. A good contamination detectionmethod should be able to differentiate the cause of signal variation.The MED and LPF methods depend either on comparison with theprevious time step or on comparison with predicted and observedvalues for the same time step. Fluctuations caused due to devicenoise and those caused by presence of contaminants are identical tothe MED and LPF methods. The MED and LPF methods might beable to identify sudden changes in sensor readings, but they cannotdifferentiate the reason for such a fluctuation. This introduces falsepositive errors.

To gain more insight into this, the 48th, 88th and 99th timesteps were taken as samples and the detection processes for thesetime steps are presented here. Both the 48th and 88th time stepsrepresent background and the 99th time step is a real contamina-tion event. As shown in Table 4, the average distances at the 48thand the 99th time steps are greater than d*LPF (d*LPF¼0.5, Table 2).Therefore, these two time steps were classified as contaminationevents. The LPF method failed to identify the difference betweenfluctuations at the 48th and the 99th time steps. Given the 48thtime step represents background, this is clearly a false positive.

Table 5 shows the differences between observed values and themean of previous PMED time steps (PMED ¼ 6, Table 2) at these threetime steps, the previous time steps and their distances. The valuesof dMED(48), dMED(88) and dMED(99) are all higher than 0.49, thed*MED. Therefore, all three time steps were diagnosed as contami-nation events, which meant that both the 48th and 88th time stepsyielded false positives. The fluctuations at the 48th and 88th timesteps were mainly caused by device noise, while the ones at the99th time steps were generally responses to the presence of phenol.

From this analysis, it is obvious that the MED method cannotdifferentiate the fluctuations caused by contamination from pres-ence of contamination or device noise. Although the causes for thefluctuations at the 48th and 99th time steps were different, the LPFand MED methods viewed the fluctuations as the same becausethey all represented outliers. This led to the false positive error atthe 48th time step.

The PE method attempts to differentiate the presence ofcontamination from device noise by employing correlative co-efficients between water quality parameters. The Euclidean dis-tances dPE(48)¼ 4.58 and dPE(88)¼ 4.80 were smaller than d*PE(5.29,Table 2) and therefore both the 48th and 88th time steps werecorrectly grouped as background. The Euclidean distancedPE(99) ¼ 5.74 was greater than the threshold value. Therefore, the99th time stepwas classified as a contamination event. A greater dPEsuggests that there are more ones in the correlation indicatorvector (Equation (1)), which implies a greater number of correlativecoefficients are far away from zero. This means that the fluctuationsat the 99th minute are more closely interrelated and have a higherpossibility of being caused by a contamination event. Althoughfluctuations did occur at the 48th and 88th minutes, these weremore likely to be caused by device noise. Therefore, the correlationindicator vectors at these two time steps contain more zeros. Fromthis analysis, it is obvious that the PE method has a better ability todifferentiate between fluctuations caused by presence of contam-inant and those caused by device noise than the conventional MEDand LFP methods.

4. Conclusions

All three detection methods under discussion are prone toyielding worse detection performance when applied to data from areal contamination accident. However, the PE method demon-strated better detection performance than the MED and LPFmethods when applied to data from a real contamination accident.By employing the correlative relationship between multiple waterquality parameters, the PE method was more capable of differen-tiating device noise from the presence of contamination. It isconcluded that the PEmethod has better potential to be used in realfield situations.

Acknowledgements

This work is financially supported by the National Nature andScience Foundation and Beijing Science and Technology Program(Z141100006014048).

References

Allgeier, S., Murray, R., Mckenna, S., Shalvi, D., 2005. Overview of Event DetectionSystems for WaterSentinel. Environmental Protection Agency, Washington, DC.

Arad, J., Housh, M., Perelman, L., Ostfeld, A., 2013. A dynamic thresholds scheme forcontaminant event detection in water distribution systems. Water Res. 47,1899e1908.

Debon, A., Carrion, A., Cabrera, E., Solano, H., 2010. Comparing risk of failure modelsin water supply networks using ROC curves. Reliab. Eng. Syst. Safe. 95, 43e48.

Hall, J., Zaffiro, A.D., Marx, R.B., Kefauver, P.C., Krishnan, E.R., Herrmann, J.G., 2007.On-line water quality parameters as indicators of distribution system

S. Liu et al. / Journal of Environmental Management 161 (2015) 385e391 391

contamination. J. Am. Water Works Assess. 99, 66e77.Hart, D., McKenna, S.A., Klise, K., Cruz, V., Wilson, M., 2007. In: . CANARY: a Water

Quality Event Detection Algorithm Development Tool, Proceedings of the WorldEnviron and Water Res Congress. ASCE, Reston, pp. 1e9.

Klise, K.A., McKenna, S.A., 2006. Water quality change detection: multivariate al-gorithms. Proc. SPIE 6203, 62030J.

Kroll, D., 2006. Securing Our Water Supply: Protecting a Vulnerable Resource.Pennwell, Oklahoma, USA.

Liu, S., Che, H., Smith, K., Chang, T., 2015a. Contaminant classification using cosinedistances based on multiple conventional sensors. Environ. Sci. Process. Impacts17, 343e350.

Liu, S., Che, H., Smith, K., Chang, T., 2015b. A real time method of contaminantclassification using conventional water quality sensors. J. Environ. Manage. 154,13e21.

Liu, S., Che, H., Smith, K., Chen, L., 2014. Contamination event detection usingmultiple types of conventional water quality sensors in source water. Environ.Sci. Process. Impacts 16, 2028e2038.

Liu, S., Smith, K., Che, H., 2015c. A multivariate based event detection method andperformance comparison with two conventional methods. Water Res. 80,109e118.

McKenna, S.A., Wilson, M., Klise, K.A., 2008. Detecting changes in water quality data.J. Am. Water Works Assess. 100, 74e85.

Perelman, L., Arad, J., Housh, M., Ostfeld, A., 2012. Event detection in water distri-bution systems from multivariate water quality time series. Environ. Sci.Technol. 46, 8212e8219.

Raciti, M., Cucurull, J., Nadjm-Tehrani, S., 2012. Anomaly Detection in Water Man-agement Systems, Critical Infrastructure Protection. Springer, pp. 98e119.

Swets, J.A., 1988. Measuring the accuracy of diagnostic systems. Science 240,1285e1293.

Wang, C., Feng, Y., Zhao, S., Li, B.-L., 2012. A dynamic contaminant fate model oforganic compound: a case study of nitrobenzene pollution in Songhua River,China. Chemosphere 88, 69e76.

Wang, S.J., Li, D., Petrick, N., Sahiner, B., Linguraru, M.G., Summers, R.M., 2015.Optimizing area under the ROC curve using semi-supervised learning. PatternRecogn. 48, 276e287.

Yang, J., Bi, J., Zhang, H., Li, F., Zhou, J., Liu, B., 2010. Evolvement of the relationshipbetween environmental pollution accident and economic growth in China.China Environ. Sci. 30, 571e576 (in Chinese).

Yang, Y.J., Haught, R.C., Goodrich, J.A., 2009. Real-time contaminant detection andclassification in a drinking water pipe using conventional water quality sensors:techniques and experimental results. J. Environ. Manage. 90, 2494e2506.

Zetterqvist, L., 1991. Statistical estimation and interpretation of trends in waterquality time series. Water Resour. Res. 27, 1637e1648.