2015 3rd author contaminant classification using cosine distances based on multiple conventional...

EnvironmentalScienceProcesses & Impacts

PAPER

Publ

ishe

d on

09

Dec

embe

r 20

14. D

ownl

oade

d on

26/

01/2

015

01:3

4:19

.

View Article OnlineView Journal

Contaminant cla

School of Environment, Tsinghua Unive

[email protected]; Tel: +86 1062

† Electronic supplementary informa10.1039/c4em00580e

Cite this: DOI: 10.1039/c4em00580e

Received 29th October 2014Accepted 9th December 2014

DOI: 10.1039/c4em00580e

rsc.li/process-impacts

This journal is © The Royal Society of

ssification using cosine distancesbased on multiple conventional sensors†

Shuming Liu,* Han Che, Kate Smith and Tian Chang

Emergent contamination events have a significant impact on water systems. After contamination detection,

it is important to classify the type of contaminant quickly to provide support for remediation attempts.

Conventional methods generally either rely on laboratory-based analysis, which requires a long analysis

time, or on multivariable-based geometry analysis and sequence analysis, which is prone to being

affected by the contaminant concentration. This paper proposes a new contaminant classification

method, which discriminates contaminants in a real time manner independent of the contaminant

concentration. The proposed method quantifies the similarities or dissimilarities between sensors'

responses to different types of contaminants. The performance of the proposed method was evaluated

using data from contaminant injection experiments in a laboratory and compared with a Euclidean

distance-based method. The robustness of the proposed method was evaluated using an uncertainty

analysis. The results show that the proposed method performed better in identifying the type of

contaminant than the Euclidean distance based method and that it could classify the type of

contaminant in minutes without significantly compromising the correct classification rate (CCR).

Environmental impact

Emergent contamination events have a signicant impact on water systems. Aer contamination detection, it is important to classify the type of contaminantquickly to provide support for remediation attempts. Conventional methods commonly either rely on laboratory-based analysis, which requires a long analysistime, or multivariable based geometry analysis and sequence analysis, which is prone to being affected by the contaminant concentration. This paper proposes anew contaminant classication method, which discriminates contaminants in a real time manner independent of the contaminant concentration. Theproposed method can identify glyphosate and cadmium nitrate 1 and 4 minutes aer detection with the true positive rate of 91.8% and 97.5%. Compared tolaboratory-based methods, classication in minutes with no signicant compromise on the CCR is an advantage.

Introduction

Water systems are vulnerable to contamination accidents.1,2 Forexample, in April 2014, crude oil leaked from a petrochemicalpipeline in Lanzhou, China, contaminating the water source ofa local water plant and introducing hazardous levels of benzeneinto the city's tap water. Water supply to Lanzhou city wassuspended as a result. An intense effort is currently underway toimprove analytical monitoring and detection of biological,chemical, and radiological contaminants in water systems. Oneapproach for avoiding or mitigating the impact of contamina-tion is to establish an Early Warning System (EWS). The EWSshould provide a fast and accurate means of distinguishingbetween normal variations and contamination events, andshould be able to classify the type of contaminant.3

rsity, Beijing, 100084, China. E-mail:

787964

tion (ESI) available. See DOI:

Chemistry 2015

Aer an EWS detects the presence of contamination, the nextimportant issue is to classify the type of contaminant. The mostcommonly used method for contaminant classication islaboratory-based analysis, e.g. ICP-MS. The advantage of thistype of analysis is that it can accurately qualify and quantify thecontaminant. The disadvantage is that it is time-consuming. Inthe event of emergent contamination, the key to all remediationattempts is time. Therefore, methods for fast classication ofcontaminants are in great demand. One possible solution isonline compound-specic sensors, which need less time thanlaboratory-based methods.4–8 However, compound specicsensors can normally only identify one type or a small group ofcontaminants. In this case, low efficiency or failure of contam-inant classication can be expected.

To overcome this drawback, several researchers haveattempted to develop real-time contaminant classicationmethods. Kroll9 reported the Hach HST approach usingmultiple types of sensors for event detection and contaminantclassication. In the Hach HST approach, signals from 5 sepa-rate orthogonal measurements of water quality (pH, conduc-tivity, turbidity, chlorine residual, and TOC) were processed

Environ. Sci.: Processes Impacts

http://crossmark.crossref.org/dialog/?doi=10.1039/c4em00580e&domain=pdf&date_stamp=2014-12-20

http://dx.doi.org/10.1039/c4em00580e

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

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

Environmental Science: Processes & Impacts Paper

Publ

ishe

d on

09

Dec

embe

r 20

14. D

ownl

oade

d on

26/

01/2

015

01:3

4:19

. View Article Online

from a 5-parameter measure into a single scalar trigger signal.The deviation signal was compared to a preset threshold level. Ifthe signal exceeded the threshold, the trigger was activated.9

The deviation vector was then used for further classication ofthe cause of the contamination. The direction of the deviationvector relates to the agent's characteristics. Seeing that this isthe case, laboratory agent data can be used to build a threatagent library of deviation vectors. A deviation vector from themonitor can be compared to agent vectors in the threat agentlibrary to see whether there is a match within a given tolerancelevel. This system can be used to classify what caused the triggerevent. Yang et al.10 reported a real-time event adaptive detection,classication and warning (READiw) method for event detectionand contamination classication. In this method, fourdiscrimination systems were developed to differentiate the 11tested contaminants according to the various responses ofsensors. The classication process was more based on geometryanalysis. The similarity or dissimilarity between examples andclasses was not quantitatively evaluated. Oliker and Oseld11

developed a contamination event detection method for waterdistribution systems, which comprised a weighted supportvector machine for the detection of outliers, and subsequentsequence analysis for the classication of contaminationevents. It was noticed that either geometry analysis or sequenceanalysis was prone to being affected by the magnitude of sensorresponses, which were normally related to contaminantconcentrations. This could then lead to misclassication.

Although effort has been put into developing methods forcontaminant classication in recent years, more attention isnecessary. Therefore, the objectives of this study are (1) todevelop a classication method which is independent of thecontaminant concentration; (2) to compare the performance ofthe proposed method with a Euclidean distance-based method.

Materials and methodsData collection

In order to collect contamination data, a pilot-scale contami-nant injection experiment (CIE) platform was developed. Aprocess ow schematic of the CIE platform is shown in Fig. 1.The water tank is approximately 85 cm high with a diameter of70 cm, 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. Eight types of sensors developed by Hach HomelandSecurity Technologies were utilized in this study. They canmeasure the following 8 parameters simultaneously andcontinuously: temperature, pH, turbidity, conductivity, oxida-tion reduction potential (ORP), UV-254, nitrate-nitrogen andphosphate. The CIE platform was operated in recirculationmode for baseline establishment. Generally, the process ofestablishing baseline takes 4–6 hours before any contaminantexperiments can be carried out. When operating in single-passcontaminant mode, the contaminant is injected into the pipeconnecting the tank and sensors via another peristaltic pump. Itis injected at a rate of 2–20 mL per minute depending on theconcentration requirement. For more information about the


CIE platform and the injection experiment, the readers couldrefer to Liu et al.12

Contaminants investigated

Specic quantities of various contaminants were injected intothe system simulator. The contaminants investigated weredetermined according to statistical reports on water pollutionincidents in urban water supply systems in China over the past20 years. Three groups of the most common six pollutants wereselected: atrazine, glyphosate, cadmium nitrate, nickel nitrate,sodium uoride and sodium nitrate. They were also selectedbased on China's national standards regarding source waterquality GB3838-2002 and drinking water quality GB5749-2006.The concentration ranges of tested contaminants are providedin the ESI (Table T1†) and were decided using the concentrationlimit given in the above national standards.

Classication method

Clustering or cluster analysis is the process of grouping a set ofobjects into classes of similar objects. Objects in any one clustershare similar features. Although denitions of similarity varyfrom one clustering model to another, in most of these modelsthe concept of similarity is based on distances, e.g., Euclideandistance and cosine distance.13–15

In cluster analysis, similar objects are assumed to have closevalues. If the distance of an object to a particular class is shorterthan the distances to other classes, the object is deemed asbelonging to that class (Fig. 2). In this way, cluster analysis canbe used to identify the type of contaminant. An object can be anexample or instance of the class. In this study, the term instancerefers to the object in a pre-dened class, while the examplerefers to the object to be classied. Both instances and examplesare vectors consisting of features. The features are extracted andderived from the sensor responses for contaminants.

Fig. 3 shows the responses to cadmium nitrate and atrazineat time t1 and t2 for 8 types of sensors. If the sensor reading is

This journal is © The Royal Society of Chemistry 2015


Fig. 2 Schematic graphs of class and instance.

Fig. 3 Four instances of features of cadmium nitrate and atrazine.

Paper Environmental Science: Processes & Impacts

Publ

ishe

d on

09

Dec

embe

r 20

14. D

ownl

oade

d on

26/

01/2

015

01:3

4:19


taken as the feature, pt1, pt2, qt1 and qt2 are 8-dimensionalvectors. As shown in Fig. 3, the graphs for pt1 and pt2 are clearlysimilar to each other, while the graph for qt1 is closer to thegraph for qt2. An essential task of this study is to quantify thesimilarity or dissimilarity between two vectors, which is thenused for contaminant identication.

Similarity measure

There are several methods of measuring the similarity betweentwo objects (i.e. two l-dimensional vectors). In this study, cosinesimilarity was adopted. Cosine similarity is a measure of simi-larity between two vectors of an inner product space thatmeasures the cosine of the angle between them.16,17 The cosineof two vectors can be derived by using the Euclidean dot productformula.

p$q ¼ kpkkqk cos q (1)

Given two vectors of attributes, p and q, the cosine similarity,cos(q), is represented using

Similarityðp; qÞ ¼ cos q ¼ p$q

kpkkqk ¼

Xn

i¼1

piqiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn

i¼1

pi2

s ffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn

i¼1

qi2

s (2)

in which n is the dimension of vectors p and q.


This function gives a similarity measure in the sense that thecosine value gets larger as the two vectors become more parallelto each other in the l-dimensional space. Or, in other words, asthe two data segments become more similar, their cosinesimilarity approaches 1.0 and their distance approaches 0.0.Therefore, cosine similarity can be used as a distance metric inthe following way:

D(p,q) ¼ 1 � similarity(p,q) (3)

Since the cosine similarity reects the magnitude of theangle between two vectors in the l-dimensional space, it is amany-to-one function. Compared with the other distancemeasures, like the Euclidean distance, the cosine similarityignores the magnitude difference between the two vectors, i.e.

SimilarityðAp; qÞ ¼

Xn

i¼1

ApiqiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn

i¼1

ðApiÞ2s ffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn

i¼1

qi2

s ¼ similarityðp; qÞ (4)

Therefore, when the cosine distance is used for contaminantidentication, the variation range of sensor data need not bepredetermined.

Contaminant classication

The distance from a point p to a class c is given by:

D(p,C) ¼ 1 � similarity(p,mc) (5)

in which, D(p,C) is the distance from a point to a class and mc isthe mean of all instances in class C.

The type of contaminant is identied by comparing thedistances from examples to classes. Assuming there are n typesof contaminants, C1, C2, ., Cn, (or n classes), each classcontains many vectors (i.e. instance of class). For any example pto be identied, if there exists

D(p,Ci) < D(p,Cj), j ¼ 1, 2, ., n, i s j (6)

then it is deemed that p ˛ Ci.

Evaluation of classication performance

The performance of the classicationmethod is evaluated usingthe correct classication rate (CCR). CCR can be calculated by

CCR ¼ CC

CCþ IC� 100% (7)

where CC refers to the correct classication of a contaminant,IC is the incorrect classication of a contaminant as anothertype of contaminant. A greater CCR means that the method ismore capable of contamination identication.

Robustness of the proposed method

The proposed method relies on the readings of online waterquality sensors. Inevitably, uctuations exist in online readings,




Publ

ishe

d on

09

Dec

embe

r 20

14. D

ownl

oade

d on

26/

01/2

015

01:3

4:19


which might come from equipment noise or ambient varia-tions. An important issue for a contaminant classicationmethod is how robust it is when dealing with uctuations inreadings. To evaluate the robustness of the proposed method,articial uncertainties were added to the raw readings. It isassumed that the uncertainty obeys Gaussian distributions. Theuncertainty quantication is achieved through a sampling-based method, the Latin hypercube sampling (LHS) technique.In LHS,18 values of stochastic tested vectors are generated in arandom, yet constrained way. First, the values of variables in theoriginal tested vectors are taken as the mean (i.e. raw readingsof sensors) and the standard deviation is equal to 1% of themean value (i.e., coefficient of variation Cv ¼ 0.01, for example).The range of each vector variable can be calculated using aGaussian distribution equation, which is then divided into Ns

non-overlapping intervals on the basis of equal probability.Aer that, a single random value is selected from each interval.This process is repeated for all variables in a feature vector.Once that is done, the Ns values obtained for the rst vectorvariable are paired in a randommanner with Ns values obtainedfor the second vector variable and so on. Ns feature vectors aregenerated from the original feature vector. By repeating thesame process, feature vectors and the associated uncertaintycan be obtained for all time steps. The CCRs for feature vectorswith uncertainty can then be obtained. Finally, the robustness isevaluated using eqn (8).

Robustness ¼ CCRconfidence limit

CCRo

(8)

in which, CCRcondence limit is the 95% condence limit of theCCRs with uncertainty and CCRo is the original CCR. Forexample, if the 95% condence limit of the CCRs with uncer-tainty is 0.8 and the original CCR is 1, then the robustness valueis 0.8/1¼ 0.8. A higher robustness value means that the methodis more robust.

Experiments and resultsFormation of classes of contaminants

In this study, features were extracted to facilitate the quantita-tive evaluation of similarity or dissimilarity between different

Fig. 4 The demonstration of feature vectors at glyphosate concen-trations of 1.4, 2.8, 7, and 14 mg l�1.


types of contaminants. For all sensors in this study, the sensorresponses obtained at each time step were adopted to form afeature vector (8 dimensions). For instance, the vector at the 1st

minute for glyphosate was [1.32, 7.06, 757.67 10.76, 276.96,3.16, 9.42, 0.08] with the vector sequence being turbidity, pH,conductivity, temperature, ORP, nitrate, UV and phosphate.Fig. 4 shows the corresponding feature vectors at differentconcentrations. As shown in Fig. 4, the extracted features sharesome similarity, but dissimilarity also exists. By extracting suchdata from all time steps, the class for glyphosate was estab-lished. The same procedure was repeated for the othercontaminants examined in this study and a library containing 6classes was obtained.

Contaminant classication

Glyphosate and cadmium nitrate were chosen to demonstratethe performance of the contaminant classication method. Theconcentrations for glyphosate were 1.4 mg l�1, 2.8 mg l�1,7.0 mg l�1 and 14.0 mg l�1. For cadmium nitrate, the concen-trations were 0.004 mg l�1, 0.008 mg l�1, 0.016 mg l�1 and0.032 mg l�1. A new group of contaminant injection experi-ments were conducted to produce data for contaminant clas-sication. The raw experimental data contained sensorresponses for baseline and the presence of contaminant at4 concentrations. As reported by Liu et al.,12 the contaminationevents were detected 1 minute aer the introduction ofcontaminants. The sensor response data aer detection wereseparated from the raw data and used in the classication. Theywere treated using the procedure above to obtain examplefeature vectors. In total, there were 110 glyphosate and200 cadmium nitrate example data to be tested.

For glyphosate, the cosine distances to all classes for eachexample (or 1 minute time step) were calculated using eqn (4)and are shown in Fig. 5. The green dots show the distancebetween examples and the glyphosate class. For all time steps(from 1 to 110), it can be noted that, although the concentrationvaries, the cosine distances from the examples to the glyphosateclass are rather stable and small. They are mostly in the range of0–0.02. The distances to the other classes are much greater. Forexample, the distance to chromium nitrate is around 0.16 (the

Fig. 5 The cosine distance of glyphosate to 6 classes.



Table 1 Averaged cosine distances from examples to classes

Tested example

Cosine distances from examples to classes

Atrazine Glyphosate Cadmium nitrate Nickel nitrate Sodium uoride Sodium nitrate

Glyphosate Mean 0.0190 0.0105 0.1240 0.1555 0.0148 0.0470Standard deviation 0.0063 0.0048 0.0065 0.0065 0.0063 0.0065

Cadmium nitrate Mean 0.0781 0.0880 0.0277 0.0593 0.0819 0.0494Standard deviation 0.0065 0.0066 0.0065 0.0065 0.0065 0.0066


Publ

ishe

d on

09

Dec

embe

r 20

14. D

ownl

oade

d on

26/

01/2

015

01:3

4:19


blue circles in Fig. 5). This is shown in Table 1, along with meanand standard deviation values of the distances. The proposedmethod classied the type of contaminant by comparing thecosine distance. The one with the closest distance is deemed tobe the correct class. Table 1 and Fig. 5 reveal that the examplesare closer to the glyphosate class. On the basis of eqn (6), thefeature vectors of the example are more similar to the ones forglyphosate. Therefore, it can be concluded that the contami-nant is glyphosate. Using eqn (7), the CCR of the classicationwas calculated to be 0.918, which suggests that the testedcontaminant is correctly classied in 91.8% of situations in thisstudy.

For cadmium nitrate, Fig. 6 shows the cosine distances todifferent classes, in which the red dots indicate the distances tothe cadmium nitrate class. For all time steps, the distances tothe cadmium nitrate class were in the range of 0.01 to 0.04 withthe mean of 0.0277 (Table 1). It is obvious that the distances tocadmium nitrate are smaller than the ones to other classes inmost cases in this study. The CCR was calculated to be 0.975.

In terms of the time needed for classication, once acontamination event is detected by an EWS, the contaminantclassication module will be activated. Theoretically, the type ofcontaminant can be classied within 1 minute (i.e. the sensorreporting step). However, in practice, the time might be a bitlonger since the sensor responses to the presence of contami-nant might sometimes need to stabilize. As shown in Fig. 5, thecontaminant was classied correctly to be glyphosate 1 minuteaer the contamination event alarm. This means that thedistance to the correct class was the smallest from the 1st

Fig. 6 The cosine distance of cadmium nitrate to 6 classes.


minute onwards. For the case of cadmium nitrate, the proposedmethod can classify correctly 6 minutes aer activation. In therst 5 minutes, the tested examples were incorrectly classied.The key strength of the proposed method is that it classies thetype of contaminant in a real time manner. Compared tolaboratory-based methods, classication in 6 minutes with nosignicant compromise of the CCR is an advantage.

DiscussionComparison to the Euclidean distance based method

In previous studies, Liu et al.12 reported that the magnitudes ofsensor responses vary with the concentration of the contami-nant (or see Fig. F1, F2, F3, F4 and F5 in the ESI†). This istypically obvious for pH, nitrate, phosphate and ORP. Forexample, the pH and ORP values for the glyphosate concentra-tion of 1.4, 2.8, 7.0, and 14.0 mg l�1 are 6.89, 6.71, 6.41, and 6.10and 277.66, 283.33, 291.67, and 299.29 mV, respectively. Theaim of this study is to establish a method to classify the type ofcontaminant by evaluating the similarity between examples andclasses. The classication method should be independent of orless related to the concentration of the contaminants since thisis not known in advance in a real event. In other words, thedistance evaluationmethod should not be too dependent on thecontaminant concentration. If the distance evaluation is closelyrelated to the magnitude of sensor response, the classicationmethod might fail to differentiate events caused by the sametype of contaminant with different concentrations.

There are several types of evaluation methods for thedistance of vectors. The most commonly used one is theEuclidean distance, which is the “ordinary” distance betweentwo points.19,20 The Euclidean distance between points p and q isthe length of the line segment connecting them, which can becalculated using

Eðq; pÞ ¼ kq� pk ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiðq� pÞ$ðq� pÞ

p(9)

in which E(q,p) is the Euclidean distance between points pand q.

Fig. 7 schematically shows the Euclidean distances andcosine distances between points p1, p2, q1 and q2. Points p1 andp2 are the sensor response vectors of contaminant 1 atconcentrations 1 and 2. Points q1 and q2 are the vectors ofcontaminant 2 at concentrations 1 and 2. As shown in Fig. 7, theEuclidean distance between p1 and p2 is kp1 � p2k and thecosine distance is 0. For p2 and q2, the Euclidean distance is



Fig. 7 Schematic drawing of cosine and Euclidean distances.

Table 3 Statistics of CCR under uncertainty (original CCR: 0.92)

CCR Cv ¼ 0.005 Cv ¼ 0.01 Cv ¼ 0.02 Cv ¼ 0.03

Mean 0.92 0.90 0.82 0.75Standard deviation 0.01 0.02 0.03 0.04Robustness 0.97 0.93 0.82 0.73


Publ

ishe

d on

09

Dec

embe

r 20

14. D

ownl

oade

d on

26/

01/2

015

01:3

4:19


kp2 � q2k and the cosine distance is 1 � cos(q). Therefore, byusing the cosine distance method, p1 and p2 (also q1 and p2)can be classied to the correct class. However, if the Euclideandistance were used, it might group p2 and q2 into the same classbecause kp2 � q2k < kp1 � p2k. To further explain this, thevectors associated with glyphosate and cadmium nitrate atdifferent concentrations were taken as examples to calculate theEuclidean and cosine distances.

Table 2 shows the cosine and Euclidean distances betweenpoints a, b, c and d, in which a is the vector of sensor responsesto glyphosate at a concentration of 1.4 mg l�1, b is the vector forglyphosate at 14.0 mg l�1, c is the vector for cadmium nitrate ata concentration of 0.008 mg l�1 and d is the vector for cadmiumnitrate at a concentration of 0.032 mg l�1. In Table 2, cosinedistances and Euclidean distances are shown. As shown inTable 2, the cosine distances for points from the same class aresmaller than those for points from different classes. Forexample, D(a,b) ¼ 0.0027, while D(a,c) ¼ 0.1091. This explainsthe correctness of the assumption in Fig. 2, i.e. similar objectshave a shorter distance.

It is also observed that the cosine distance is not ‘sensitive’ tothe magnitude the vector (in other words, the concentration ofthe contaminants). As shown in Fig. 4, the magnitude of sensorresponse vectors at 1.4 mg l�1 and 14 mg l�1 is obviouslydifferent. However, their cosine distances to other points areclose. For example, D(a,c) ¼ 0.1091 and D(b,c) ¼ 0.1081. TheEuclidean distance, on the other hand, is related to the

Table 2 The cosine and Euclidean distancesa

Euclidean

Cosine

Glyphosate

a – 1.4 mg l�1

Glyphosate a – 1.4 mg l�1 0b – 14 mg l�1 95.5981

Cadmium nitrate c – 0.008 mg l�1 92.3888d – 0.032 mg l�1 113.1857

a Note: The numbers above the diagonal are cosine distances, while the o


magnitude of the vector. For example, E(a,c) ¼ 92.3888, whileE(b,c) ¼ 158.4424. Furthermore, the case may arise where theEuclidean distance between points from the same class mightbe greater than that between points from two different classes.This is shown in Table 2. For example, E(a,b) ¼ 95.5981 andE(a,c) ¼ 92.3888. In this case, incorrect classication wouldoccur if Euclidean distances were used for contaminant classi-cation. Point c would be wrongly classied as being in thesame class as point a if the Euclidean distance was adopted.Therefore, it was concluded that the cosine distance is moresuitable than the Euclidean distance for classifying the type ofcontaminant since the evaluation for similarity is more relatedto the contaminant's characteristics rather than the magnitudeof sensor responses.

Robustness

The level of uncertainty is given by the value of Cv. In thisstudy, four values of Cv (0.005, 0.01, 0.02 and 0.03) were used(Table 3). The value of Ns is determined according to theliterature. For a given Cv, by setting Ns ¼ 2000, 220 000 featurevectors with uncertainty were nally generated for glyphosate.These feature vectors were divided into 2000 groups. Eachcontains 110 feature vectors. By feeding the 2000 groups offeature vectors into the proposed contaminant classicationmethod, the CCRs for every group were obtained. The histo-grams of these CCRs are displayed in Fig. 8, which shows thatthe proposed method has a robustness of over 0.82 foruncertainty Cv ¼ 0.005, Cv ¼ 0.01 and Cv ¼ 0.02. This suggeststhat the performance of the proposed contaminant classi-cation method is steady and reliable and can cope well withthe uncertainty from the online sensors. For the case ofCv ¼ 0.03, the performance of the method is less satisfactory.The CCR for this case is 0.75, which is much lower than theoriginal CCR (0.92). It should be noted that the uncertainty

Cadmium nitrate

b – 14 mg l�1 c – 0.008 mg l�1 d – 0.032 mg l�1

0.0027 0.1091 0.13910 0.1081 0.1381

158.4424 0 0.0302166.8406 25.9895 0

nes below are Euclidean distances.



Fig. 8 The histogram of CCRs with uncertainty and robustness.


Publ

ishe

d on

09

Dec

embe

r 20

14. D

ownl

oade

d on

26/

01/2

015

01:3

4:19


examined in this study is assumed to be from equipmentnoise or ambient variations. A change of sensor reading dueto sudden sensor failure or the presence of contaminant is nottreated as noise, but instead as an event, which normallymeans a 1–20% change of sensor reading. Therefore, it isdeemed that the uncertainty levels adopted in this study aresignicant enough.

It is worth noting that previous studies about online sensorsin water supply systems generally assumed perfect sensors,which means that sensors worked in good condition.21–23

Although this assumption has allowed researchers to makesignicant progress in early warning system design, sensorfailures can signicantly impact the reliability of an earlywarning system design. It is commonly known that sensorreadings do contain uncertainties as they are easily affected byambient variations and equipment condition. From an imple-mentation perspective, it is essential that an early warningsystem using online sensor data is robust enough and can copewith uncertainty from sensors. The analysis here shows that theproposed method has good capacity to tolerate uncertainty insensor readings.

Future studies

This study proposed a concentration-independent contami-nant classication method based on conventional waterquality sensors. The basis of this method is that points inone class stay close and are separate from other classes. Inspite of great improvement in recent years, readings fromonline sensors are still affected by noise and ambient varia-tions. For a method based on online sensor readings, it isimportant to understand the impact of uncertainty fromsensor readings on the model output. Although this studydemonstrated the robustness of the proposed method in theevent of sensor uncertainty or ambient variations through aninitial uncertainty analysis, a global sensitivity analysis


would be more helpful to understand the extent of uncer-tainty from each sensor. This should be conducted in futurestudies.

Meanwhile, since the proposed method classies bycomparing the distances to predened classes, incorrect clas-sication error would occur if two (or more) classes overlap eachother. This study involved a limited number of contaminantsand no overlaps were noticed, but the possibility does exist. In afuture study, this has to be addressed. A possible solution tothis is that the classication decision could be made based ondistances from more than one type of features. For example, ifthe features using original sensor responses from two types ofcontaminants overlap, another type of feature (e.g. the deviationbetween real readings and baseline) can be employed todifferentiate these two classes.

Conclusions

By using data from online water quality sensors, this studyproposed a real time and concentration-independent contami-nant classication method. From the analysis, the followingconclusions were drawn.

(1) The proposed method classies the type of the contami-nant by comparing its cosine distances to predened classes.Results from the analysis show that the proposed method canidentify glyphosate and cadmium nitrate 1 and 6 minutes aerdetection with the CCR of 91.8% and 97.5%. Compared tolaboratory-based methods, classication in minutes withoutsignicantly compromising the CCR is an advantage.

(2) The results show that the performance of the proposedmethod was not related to the contaminant concentration.This implies that the proposed method is more suitable thanthe Euclidean distance method for contaminant classicationsince the concentration of the contaminant is not known apriori.




Publ

ishe

d on

09

Dec

embe

r 20

14. D

ownl

oade

d on

26/

01/2

015

01:3

4:19


Acknowledgements

This work is jointly supported by Tsinghua IndependentResearch Program (2011Z01002) and Water Major Program(2012ZX07408-002).

References

1 USEPA, Baseline threat information for vulnerability assessmentsof community water systems, Washington, DC, 2002.

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

3 M. V. Storey, B. van der Gaag and B. P. Burns, Water Res.,2011, 45, 741–747.

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

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

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

7 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.

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


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

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

11 N. Aliker and A. Ostfeld, Water Res., 2014, 51, 234–245.12 S. M. Liu, H. Che, K. Smith and L. Chen, Environ. Sci.:

Processes Impacts, 2014, 16, 2028–2038.13 M. Tabacchi, C. Asensio, I. Pavon, M. Recuero, J. Mir and

M. C. Artal, Appl. Acoust., 2013, 74, 1022–1032.14 Z. P. Zhao, P. Li and X. Z. Xu, Applied Mathematics &

Information Sciences, 2013, 7, 1243–1250.15 K. A. Nguyen, R. A. Stewart and H. Zhang, Environmental

Modelling & Soware, 2013, 47, 108–127.16 A. A. Pascasio, Lin. Algebra. Appl., 2001, 325, 147–159.17 M. W. Sohn, Soc. Network., 2001, 23, 141–165.18 G. Manache and C. S. Melching, J. Water. Resour. Plann.

Manag., 2004, 130, 232–242.19 L. Liberti, C. Lavor, N. Maculan and A. Mucherino, SIAM

Rev., 2014, 56, 63–69.20 C. Stoecker, S. Welter, J. H. Moltz, B. Lassen, J. M. Kuhnigk,

S. Krass and H. O. Peitgen, Med. Phys., 2013, 40, 091912.21 M. Weickgenannt, Z. Kapelan, M. Blokker and D. A. Savic, J.

Water. Resour. Plann. Manag., 2010, 136, 629–636.22 A. Kessler, A. Ostfeld and G. Sinai, J. Water. Resour. Plann.

Manag., 1998, 124, 192–198.23 A. Ostfeld and E. Salomons, J. Water. Resour. Plann. Manag.,

2004, 130(5), 377–385.


ksm

Highlight


2015 3rd author contaminant classification using cosine distances based on multiple conventional...

Documents