Automatic real-time uncertainty estimation for onlinemeasurements: a case study on water turbidity

Joonas Kahiluoto & Jukka Hirvonen &

Teemu Näykki

Received: 21 August 2018 /Accepted: 7 March 2019 /Published online: 2 April 2019# The Author(s) 2019

Abstract Continuous sensor measurements are becom-ing an important tool in environmental monitoring. How-ever, the reliability of field measurements is still too oftenunknown, evaluated only through comparisons with lab-oratory methods or based on sometimes unrealistic infor-mation from the measuring device manufacturers. Awa-ter turbidity measurement system with automatic refer-ence sample measurement and measurement uncertaintyestimation was constructed and operated in laboratoryconditions to test an approach that utilizes validationand quality control data for automatic measurement un-certainty estimation. Using validation and quality controldata formeasurement uncertainty estimation is a commonpractice in laboratories and, if applied to field measure-ments, could be a way to enhance the usability of fieldsensor measurements. The measurement system investi-gated performed replicate measurements of turbidity inriver water and measured synthetic turbidity referencesolutions at given intervals during the testing period.Measurement uncertainties were calculated for the resultsusing AutoMUkit software and uncertainties were at-tached to appropriate results. The measurement resultscorrelated well (R2 = 0.99) with laboratory results and the

calculated measurement uncertainties were 0.8–2.1formazin nephelometric units (FNU) (k = 2) for 1.2–5FNU range and 11–27% (k = 2) for 5–40 FNU range.Themeasurement uncertainty estimation settings (such asmeasurement range selected and a number of replicates)provided by the user have a significant effect on thecalculated measurement uncertainties. More research isneeded especially on finding suitable measurement un-certainty estimation intervals for different field condi-tions. The approach presented is also applicable for otheronline measurements besides turbidity within limits setby available measurement devices and stable referencesolutions. Potentially interesting areas of applicationcould be the measurement of conductivity, pH, chemicaloxygen demand (COD)/total organic carbon (TOC), ormetals.

Keywords Measurement uncertainty . Fieldmeasurement . The Nordtest approach .Water qualitymonitoring . Quality control . Turbidity

Introduction

The need for reliable information about the environmentis becoming more and more evident as mankind’s impacton the planet has significantly increased (Rockström et al.2009). Measurements are needed to monitor and distin-guish changes in the environment, to define if thesechanges are natural or an outcome of the human activity,and to evaluate the effects of policies which aim to keepour impact to the environment at an acceptable level

Environ Monit Assess (2019) 191: 259https://doi.org/10.1007/s10661-019-7374-7

J. Kahiluoto (*) : T. NäykkiEnvironmental Measurement and Testing Laboratory, FinnishEnvironment Institute, Ultramariinikuja 4, 00430 Helsinki,Finlande-mail: joonas.kahiluoto@ymparisto.fi

J. HirvonenEnvironmental Measurement and Testing Laboratory, FinnishEnvironment Institute, Yliopistokatu 7, 80100 Joensuu, Finland

(Elliott 2014). Long-term monitoring with comparablemethods and known data quality is essential in utilizingthe measurement data (Ellingsen et al. 2017). The chem-ical and physical state of natural waters is normallymonitored by measuring different water quality parame-ters. Knowing the uncertainties of these measurements iscritical as decisions are made based on the measurementresults. Decisions made based on inaccurate data canhave severe consequences. Surface water monitoringhas traditionally been carried out using the combinationof manual sampling and laboratory analysis, but there is aglobal trend shifting towards field measurements, remotesensing, and citizen science (Giles 2013; Nilssen et al.2015; Dunbabin andMarques 2012; Conrad and Hilchey2010). The quality of results produced in traditionalenvironmental laboratories is ensured by using standard-ized methods, available guides and accreditation (JointCommittee for Guides in Metrology 2008; Eurachem/CITAC Guide CG 4 2012; International Organizationfor Standardization 2012; Magnusson et al. 2017; Inter-national Organization for Standardization 2017). For fieldmeasurements, the quality control procedures are not wellenough established and the measurement uncertainties,including method and laboratory bias, are unknown orevaluated based on sometimes too ideal information pro-vided by measuring device manufacturers (Björklöf et al.2016a, b).

There is a need to improve the reliability, i.e.,knowledge of measurement uncertainty, of these newmonitoring methods in order to validate the produceddata for decision making (European Commission2009; Lewis and Edwards 2016). Continuous fieldmeasurements have a huge potential providing anunrivaled temporal resolution with better efficiencyand lower costs per sample compared to samplingand laboratory analysis. Even classifying water bodieswith conventional techniques can be questioned, be-cause of the poor sampling density, when normalsampling intervals and laboratory analyses are used(Skeffington et al. 2015). Also, the greatest source ofuncertainty is often caused by sampling, sample trans-port, and storage, which are at least partly eliminatedin field measurements (Moser and Wegscheider 2001;Björklöf et al. 2016b). All of the necessary waterquality parameters cannot be measured with onlineinstruments and sensors at the moment (Näykki andVäisänen 2016), but technology is evolving and thelist of parameters is growing all the time (Blombergvon der Geest et al. 2012).

Näykki et al. (2015) presented a way to apply theNordtest method based on quality control and vali-dation data (Magnusson et al. 2017; Hovind et al.2011; International Organization for Standardization2012) for Breal-time^ uncertainty estimations in on-line measurements, in which measurement uncer-tainty is broken down into within-laboratory repro-ducibility uRw and method/laboratory bias ub. Thesetwo components can be estimated from the dataproduced by the online measurement system, morespecifically from routine sample replicate measure-ments and synthetic control sample measurements(Hovind et al. 2011). The idea of this approach forreliability estimation is fundamentally differentwhen compared to a more traditional approach,where the reliability is defined through correlationswith laboratory results. Comparisons with currentlyused methods, e.g., laboratory measurements, arerequired for detection of systematic differences be-tween the methods, but for measurement uncertaintyestimations, a continuous automated uncertainty cal-culation procedure is definitely able to reflect thecurrent state of the measurement device better thanan uncertainty value estimated once per year (oreven lifetime) for an instrument. Requirements forthe automated measurement system and automateddata processing were laid out in the paper by Näykkiet al. 2015. In this paper, it is reported how aremotely controlled automated measurement systemand automated data processing system for uncertain-ty calculations was set up and tested in laboratoryconditions. This will further clarify the practicalissues arising from operating such a system andreveal possible problems with the approach and sub-jects for further research.

Water turbidity was selected as the research case forseveral reasons. First of all, the definition of turbidityaccording to ISO 7027-1 (International Organization forStandardization 2016) is Breduction of transparency of aliquid caused by the presence of undissolved matter,^which means that turbidity of water can be heteroge-neous within a sample, as it is caused by undissolvedmatter with different particle sizes and densities(Horowitz 2013). Turbidity in itself is an importantwater quality parameter and in 2012, over 30,000 tur-bidity measurements from natural waters were carriedout in Finnish laboratories (Näykki et al. 2014). Turbid-ity can also be used as a surrogate parameter for examplein continuous suspended solid or nutrient monitoring

259 Page 2 of 14 Environ Monit Assess (2019) 191: 259

when site- and instrument-specific relationships are welldefined (Rymszewicz et al. 2017; Horowitz 2013;Caradot et al. 2015; Bilotta and Brazier 2008). Turbiditymeasurements are divided into two categories, nephe-lometry and turbidimetry, according to the measurementprinciple used. In nephelometry, diffused radiation in a90° angle from the light source is measured whereas inturbidimetry attenuated radiation in a 0° degree angle ismeasured. Results gained with these two different prin-ciples have different units and are not comparable(Joannis et al. 2008). The measurement devices in thispaper are nephelometric and the results are expressed informazin nephelometric units (FNU). The primary ref-erence solutions for turbidity measurements areformazin suspensions synthetized from hexamethylene-tetramine (C6H12N4) and hydrazine sulfate (N2H6SO4)of which hydrazine sulfate is poisonous and may becarcinogenic (International Organization forStandardization 2016).

Requirements for continuous field measurementswith automated measurement uncertaintyestimation

The requirements for automated measurement un-certainty estimations can be divided into two cate-gories: hardware and software. The physical mea-surement station has to be able to perform replicatemeasurements from the sample water and also mea-sure synthetic reference solutions at given intervals.The synthetic reference solutions have to be col-lected as waste in many cases, depending on thereference material used. For turbidity, this is thecase because of the poisonous hydrazine sulfate.The software side is used to control the measure-ment station remotely, store the data into a data-base, and perform the uncertainty calculations withgiven settings from the produced measurement re-sults. The reference solutions with certified refer-ence values have to be distinguishable from routinesamples and from other reference solutions withdifferent certified values in a measurement seriesfor the measurement uncertainty calculations. Thecalculated measurement uncertainties are then auto-matically attached to appropriate results. Transfer-ring the results to different databases or presentingthe results graphically can be implemented withsuitable interfaces.

Design and construction of the measurement station(hardware)

A water turbidity monitoring system (Fig. 1) was de-signed and constructed around a 1-m3 cylindrical tanksimulating a river. Three pumps were installed into thetank to circulate and lift the water enabling the mixing ofsynthetic river waters with approximately known tur-bidities for testing purposes. The synthetic river waterwas prepared by mixing sediment from river Vantaa intotap water according to a defined correlation betweensediment mass, water volume, and turbidity. The de-vices used in the measurement system (Fig. 1) are listedand described in Table 1.

The liquid flows in the system were controlled with12 relays towhich the pumps and valves were connectedto. Two different voltage levels were used (convertedfrom the same power source) because the valves usedare 24-V direct current (VDC) and pumps 12 VDC.Additional flow meter for volume flow informationand control was installed, because the ABB turbidimeterspecifications require a flow between 0.5 and 1.5 l/min,and the flow meter data can be used for quality control.Also, level indicators were installed to the referencesolution containers and the waste container, to preventrunning out of reference solutions and for avoidingwaste overflow situations.

Automation and remote management with SykeEnviCal Manager (software)

A cloud service based on open source solutions wasprogrammed to control the measurement station. SykeEnviCal Manager consists of three modules: users, in-strument platform, and data. The user module handlesthe user management and user rights. The instrumentplatform module enables sequential control of pumps,relays, and measurement devices as defined by the userand is used to perform sample water and syntheticcontrol sample measurements automatically. The instru-ment platform also has features like real-time graphicalmonitoring, email alarms, and an alarm history log. Thedata module consists of a database, data analysis tools,and visualization tools. AutoMUkit software is imple-mented as a data analysis into the cloud for measure-ment uncertainty calculations. AutoMUkit calculatesthe measurement uncertainty for results within a speci-fied time interval and concentration range(s) and

Environ Monit Assess (2019) 191: 259 Page 3 of 14 259

combines the measurement uncertainty information toappropriate results either as absolute uncertainty (withthe same unit as the measurement results) or relativeuncertainty (percentage) with a specified coverage fac-tor. Everything is controlled from a web user interface(UI) that also supports mobile devices.

Automated uncertainty calculations

The automated Breal-time^ uncertainty calculationprocedure applying the Nordtest approach presentedby Magnusson et al. (2017) was introduced byNäykki et al. (2015) and is described in more detail

Fig. 1 Measurement system process and instrumentation diagram (PID)

Table 1 Device list

Device Symbol in Fig. 1 Description

Reference solution containers 1–2 Reference solution 1–2 30 l conical

Waste container Waste container 125 l cylindrical

Pumps for sample and reference solutions P1-P3 Solinst 410

Reference solution mixing pumps P4-P5 Biltema Art. 259750

Valves 1–7 V1-V7 Danfoss EV220B base with 24 VDC magnetic coils

Valves 8–9 V8-V9 Manual ball valve

Turbidimeter ABB TM ABB 7998 sensor with 4690 analyzer

Ultrasonic level indicator LIA1-LIA2 DFRobot SEN0204

Waste container level indicator LIA3 12″ eTape

Ultrasonic flow indicator FI1 Cynergy3 UF08B100

Controlling computer – Raspberry Pi 3 Model B

I/O module – Arduino Mega

Relay card (× 2) – 8× Songle SRD-05VDC-SL-C per card

4G modem – ZTE MF823

Power source – XP Power DNR480PS24-I

24 VDC/12 VDC converter – Biltema Art. 38-123

259 Page 4 of 14 Environ Monit Assess (2019) 191: 259

in their paper. In brief, the method utilizes the stan-dard deviation from routine sample replicate measure-ments (Eq. (1)) and the standard deviation from ref-erence solution measurements (Eq. (2)) to estimaterandom error, i.e., within-laboratory reproducibilitycomponent (Eq. (3)).

ur;range ¼∑nr

i¼1 c ið Þmax−c ið Þmin

� �

nr � d

ur;range% ¼

∑nri¼1 100%� c ið Þmax−c ið Þmin

c ið Þmax þ c ið Þmin

2

0

B@

1

CA

nr � d

ð1Þ

where cmax is the maximum and cmin is the minimumconcentration in a replicate series, nr is the number ofreplicate series, and d is a conversion factor from meandifference to the standard deviation (depends on thenumber of replicate series).

uRw;stand ¼ SRw;stand

uRw;stand% ¼ 100%� SRw;standCavg

ð2Þ

SRw,stand is the standard deviation of control samplemeasurement results and cavg the average of controlsample measurement results.

uRw ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiu2r;range þ u2Rw;stand

q

uRw% ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiu2r;range% þ u2Rw;stand%

qð3Þ

Bias is estimated from the difference between refer-ence solution measurement results and certified refer-ence value (Eq. (4)). The reference solution has to havea known certified reference value and a stated uncertain-ty for this value.

ub ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

b2 þ sbffiffiffin

p� �2

þ u2cref

s

ð4Þ

where b is the difference between the control sampleaverage and the actual reference value, sb is the standarddeviation of the control sample sensor measurements, nis the number of sensor measurement results of the

control sample, and uCref is the standard uncertainty ofthe reference value.

uc ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiu2Rw þ u2b

qð5Þ

Combined standard uncertainty is then calculated fromthe reproducibility and bias components (Eq. (5)). Expand-ed uncertainty is calculated by multiplying the combinedstandard uncertaintywith a coverage factor k (usually k = 2for 95% confidence level). With this method, part of therepeatability component is included twice, but this is con-sidered to be small compared to between-days variation(Magnusson et al. 2017). There are a lot of user-definedsettings that can have a significant effect on the uncertaintyestimations. These settings include dividing the estimationrange depending on the behavior of the results as a func-tion of the concentration, number of replicates, and suffi-cient time between different replicate sensor measure-ments. The settings include also the consideration of theamount of time (affecting the number of results) for whichthe measurement uncertainty is calculated for.

As presented in Fig. 2, the relative standard deviationwithin replicate series is stable only at concentrationshigher than 5 FNU. Therefore, the measurement uncer-tainty estimation range should be divided into absoluteand relative ranges at around 5 FNU. The uncertaintycalculations should be performed in absolute units(FNU) from the limit of quantification up to 5 FNUand as relative for concentrations above 5 FNU(Magnusson et al. 2017; Kahiluoto 2017).

In laboratory measurements, the concept of a repli-cate measurement is well defined and includes the rep-lication of all the analytical steps up to the result(Eurachem Guide 2011). For continuous measurements,there is no clear definition for a replicate measurementand data can be collected at very short intervals down to1 ms for this system. Replicates should be measuredfrom the same sample and at intervals where the randomvariation within the sample and the measurement systemnoise are representative. In this work, an approachwhere the interval is estimated based on sample flowrate and the internal volume of the measurement systemwas selected. This way, the samples can be seen asseparate sub-samples, but the time for the sampled waterbody to change is minimized. In practice for our mea-surement system with around 1-l/min flow rate and a0.15-dm3 cuvette volume, this lead to 10-s intervalsbetween replicate measurements.

Environ Monit Assess (2019) 191: 259 Page 5 of 14 259

Another important aspect to consider is the time periodfor which the measurement uncertainty is calculated for.There has to be enough data for the uncertainty estima-tion, which would favor a long period of data collection,but on the other hand, the measurement uncertainty willlikely change with time as biofouling and instrument driftaffect the results. Biofouling is an acute problem in openmeasurement systems (sondes etc.), possibly affecting theresults only after several days, but the problem also existsin flow-through systems (Delauney et al. 2010). Biofoul-ing is also highly site-dependent, environmental condi-tion–dependent, and temperature-dependent, which com-plicates the situation. The reference solutions can serve asa way to evaluate the drift caused by biofouling providedthat the reference solutions can be reproducibly measuredduring the operating period. The authors of Näykki et al.(2015) suggest that at least 30 measurement results over aperiod of 10 days are to be used for the estimation, whichcan serve as a starting point together with Hovind et al.2011, which states that at least 5% of measurements inlaboratories should be quality control measurements. InISO 11352 (2012), at least six certified reference materialmeasurement batches are recommended for the estima-tion of bias.

Stability and mixing of formazin reference solutions

One week of autonomous operation can be considered aminimum requirement for a cost-effective measurementstation and hence, the minimum stability requirementfor the reference solutions. According to the manufac-turer, the stability of formazin reference solutions is1 month, when the concentration is between 20 and400 FNU and only 12–24 h, when the concentration is2–20 FNU (Sadar 2003). The stability of formazin so-lutions diluted from 4000 FNU stock solution (ThermoScientific, Orion AC45FZ) was studied in Hach 2100AN IS sample cuvettes (30 ml) and for selected turbid-ities in the actual high-density polyethylene (HDPE)reference solution containers. The laboratory measure-ments were conducted with a Hach 2100 AN IS turbi-dimeter (calibrated with a HACH Stablcal® calibrationkit before the experiments) according to ISO 7027-1(2016). Measurement uncertainty (k = 2) for laboratorymeasurements was estimated to be 0.12 FNU in theconcentration range 0–5 FNU and 8.9% in the concen-tration range 5–40 FNU, calculated with MUkit soft-ware according to Magnusson et al. (2017). In thecuvette scale test, seven solutions per concentration

2

4

6

8

0 10 20 30 40

Turbidity (FNU)

Ave

rag

e s

tan

da

rd

devia

tio

n w

ith

in a

re

plica

te s

erie

s %

Instrument

ABB

Optoseven

Fig. 2 Standard deviation within replicate series as a function of turbidity with two different instruments

259 Page 6 of 14 Environ Monit Assess (2019) 191: 259

3.0

3.5

4.0

4.5

5.0

Nov 06 Nov 13 Nov 20 Nov 27

Time

Tu

rb

idity (

FN

U)

4 FNU Formazin solution stability test

Fig. 3 4-FNU formazin solution stability test results

10

15

20

25

Oct 02 Oct 09 Oct 16 Oct 23

Time

Tu

rb

idity (

FN

U)

20 FNU Formazin solution stability test

Fig. 4 20-FNU formazin solution stability test results

Environ Monit Assess (2019) 191: 259 Page 7 of 14 259

level (1, 5, and 20 FNU) were prepared and monitoredfor alteration of turbidity values during a 33-day testperiod. The turbidities of the solutions remained ade-quately stable for 1 to 3 weeks depending on the turbid-ity, with a higher turbidity leading to a better stability.

Because of the possible effects caused by differencesin volume, container material, and exposure to air, theturbidities of 4 FNU and 20 FNU formazin solutionswere also studied with a 10-l batch volume in the actualreference solution containers. The solutions were thor-oughly mixed by manually shaking and tilting the con-tainer and with a mixer attached to a drill to achievehomogeneity before sampling. The samples were col-lected with a pipette from around six random locationsfrom the top half of the container and turbidities weremeasured immediately. The results show that the stabil-ity of the diluted formazin solutions fulfill the minimumrequirement of 1-week stability, with no detectable de-cline in turbidity for the 20 FNU solution and only aslight decrease for the 4 FNU solution during the 4-week test period. The results are shown below in Figs. 3and 4, where the measurement uncertainties presenteddo not include uncertainty caused by sampling.

Results from the simulation experiments

An experiment simulating the intended use of the mea-surement station was set up in laboratory conditions. A1-m3 tank with pumps circulating synthetic river waterprepared from sediment and tap water simulated aflowing water body. Sediment and tap water were addedand the bottom of the tank was stirred manually multipletimes during the experiment. The measurement stationwas programmed to measure 5 replicate measurementswith 10-s intervals from the synthetic river water (re-ferred to as sample) once per hour. The reference solu-tions weremeasured once every 24 h (three results savedwith 1-s intervals). The uncertainties of the diluted ref-erence solutions used in the experiments were estimatedwith GUM workbench pro (version 2.4) to be 2.3% forthe 20 FNU reference solution and 0.1 FNU for the 4FNU solution (k = 2). The measurement uncertainty cal-culations were performed weekly (four calculation runsin total), in order to include a satisfactory number ofreference solution results. Two-week and 1-month cal-culation intervals were also tested for the same data set.A total of three reference solution measurement results

0

10

20

30

40

Jan 22 Jan 29 Feb 05 Feb 12 Feb 19

Time

Tu

rb

idity (

FN

U)

Sample type & measurement method

Sample.ABB

Sample.Lab

Standard solution 1.ABB

Standard solution 1.Lab

Standard solution 2.ABB

Standard solution 2.Lab

Fig. 5 Simulation experiment results. Black line represents measurement results and gray area around the results describes the calculatedmeasurement uncertainty (k = 2) for the online turbidity sensor

259 Page 8 of 14 Environ Monit Assess (2019) 191: 259

Tab

le2

Measurementu

ncertainty

calculationresults

Calculatio

ninterval

Num

berof

replicateseries

in0–5-FN

Urange

Reproducibility

with

in-

laboratory

u Rw(FNU)

Methodand

laboratory

bias

u b(FNU)

Expanded

measurement

uncertaintyfor

the0–5-FN

Urangeexpressed

inFN

U(k=2)

Num

berof

replicate

series

in5–40-FNU

range

Reproducibility

with

in-

laboratory

u Rw%(%

)

Methodand

laboratory

bias

u b(%

)

Expanded

measurement

uncertaintyfor

the5–40-FNU

rangeexpressed

in%

(k=2)

1-weekcalculationintervals

23Jan–30

Jan2018

480.22

0.34

0.81

103

7.4

6.7

19.9

30Jan–6Feb

2018

500.57

0.15

1.18

111

6.2

11.7

26.4

6Feb–13Feb2018

230.41

0.088

0.83

138

5.0

2.7

11.2

13Feb–22

Feb2018

143

0.79

0.16

1.62

525.2

2.9

11.9

2-weekcalculationintervals

23Jan–7Feb

2018

104

0.49

0.15

1.02

230

8.1

7.9

22.5

7Feb–22Feb2018

160

0.65

0.13

1.33

174

5.1

2.6

11.3

1-month

calculationinterval

23Jan–22

Feb2018

264

0.57

0.12

1.2

404

7.9

4.9

18.6

Environ Monit Assess (2019) 191: 259 Page 9 of 14 259

had to be deleted due to detected ABB instrumentmalfunction during the test. The test results are present-ed graphically in Fig. 5 and the calculated measurementuncertainties are tabulated in Table 2.

A total of 45 laboratory samples were taken forcomparison at the same time as the measurement stationperformed measurements. The river water laboratorysamples were taken with a pump through a separate linefrom the same spot in the tank as the measurementsystem intake (± 10 cm). Reference solution laboratorysamples were taken with a pipette straight from thecontainers. The laboratory samples were analyzed im-mediately with a Hach 2100 AN IS turbidimeter. The 45laboratory measurements from this experimentcomplemented by 48 comparison measurements fromearlier experiments are presented in Fig. 6. The labora-tory results and sensor results show a strong correlationacross the measurement range (R2 = 0.99).

The instrument drift during the experiment was studiedby measuring a dry secondary reference after the primarycalibration of the instrument before the experiment andmeasuring the same dry reference again after the experi-ment. The secondary reference yielded a result of 1.333FNU before the experiment and 1.404 FNU after the

experiment resulting in a 5% drift. This can also be seenin Fig. 5, where the difference between the laboratory andsensor measurement results for reference solution 1seemed to increase towards the end of the experiment.

The limit of quantification (LOQ) was studiedfrom the average standard deviation within replicatemeasurement sets measured from ~ 1 FNU river wa-ter. Twenty-seven measurements were performedover a period of 2 days and the LOQ was calculatedmultiplying the average standard deviation withinreplicate measurement sets (0.12 FNU) by 10 leadingto LOQ = 1.2 FNU.

A second similar experiment without laboratory samplecomparison was conducted between 28 March and 2May 2018. Results of the second simulation experimentare presented with 1-week calculation intervals in Table 3and Fig. 7.

Results and discussion

According to the test results, it is evident that theresults of the online measurement system comparewell with the laboratory measurements (R2 = 0.99).

y = 1.06x + 0.05

R2 = 0.99

0

10

20

30

40

0 10 20 30

Laboratory result (FNU)

Sensor R

esult (

FN

U)

Sample matrix

Sample

Standard solution

Fig. 6 Laboratory results compared with sensor results

259 Page 10 of 14 Environ Monit Assess (2019) 191: 259

The estimated expanded measurement uncertainties(k = 2) are close to the recommended ± 20% forturbidities over 5 FNU (being ± 19% if calculatedfor the whole month and even less in the secondexperiment), but the recommended limit of quantifi-cation 0.5 FNU and expanded measurement uncer-tainty of ± 0.2 FNU for low measurement range(0.5–1 FNU) was not reached with this system(Näykki and Väisänen 2016).

The different measurement uncertainty calculationintervals presented in Table 2 demonstrate the tempo-ral dimension of measurement uncertainty estimationfor continuous measurement systems. If data from along period of time is used for the estimation, in thiscase, 1 month, which can still be considered as arelatively short period, the resulting measurement un-certainty ends up being an average over the timeperiod. When shorter evaluation periods are used, theestimated measurement uncertainty reflects the currentperformance of the system better, but the estimationbecomes more vulnerable to outliers. An example ofthis can be seen in the second week of Table 2 (30January–6 February), where the reference solutionbatch used for the relative uncertainty estimation hada lower than expected concentration (both sensor andlaboratory results), leading to a high ub and therefore ahigh measurement uncertainty. With the 1-month cal-culation interval, the effect of this one batch of refer-ence solution to the calculated measurement uncertain-ty is significantly smaller. The same effect could becaused also for example by a short period of challeng-ing measuring conditions or rapid biofouling betweenmaintenances. There is a significant variation in bothuRw and ub which would infer that the performance ofthe measurement device (measurement uncertainty)depends on the properties of the measured samples(uRw) and the state of the measuring system (ub). Thehighest measurement uncertainties were calculated forweeks (30 January–6 February and 11 April–18 April)where ub component was high due to differences be-tween the reference solution batches. This highlightsthe importance of trueness and homogeneity of theused reference solutions.

Based on laboratory standards and this limitedamount of data produced in laboratory conditions froma single parameter (turbidity) with only one matrix andone measurement device, the authors suggest the fol-lowing measurement uncertainty calculation settings forpossible further research:T

able3

Measurementu

ncertainty

calculationresults

forsimulationexperiment2

Calculatio

ninterval

Num

berof

replicateseries

in0–5-FN

Urange

Reproducibility

with

in-laboratory

u Rw(FNU)

Methodand

laboratory

bias

u b(FNU)

Expanded

measurement

uncertaintyfor

the0–5-FN

Urangeexpressed

inFNU(k=2)

Num

berof

replicateseries

in5–40-FNU

range

Reproducibility

with

in-laboratory

u Rw%(%

)

Methodand

laboratory

bias

u b(%

)

Expanded

measurement

uncertaintyfor

the5–40-FNU

rangeexpressed

in%

(k=2)

28Mar–4

Apr

2018

330.33

0.13

0.71

121

2.4

5.3

11.7

4Apr–11Apr

2018

0–

––

161

5.1

514.3

11Apr–18Apr

2018

320.59

0.85

2.08

129

5.6

2.5

12.3

18Apr–25Apr

2018

290.42

0.17

0.91

132

5.8

2.2

12.3

25Apr–2

Apr

2018

510.41

0.11

0.84

109

8.4

6.5

21.3

Environ Monit Assess (2019) 191: 259 Page 11 of 14 259

& The results used for a measurement uncertainty es-timation should cover a minimum 1-week timeperiod

& The number of synthetic control sample measurementresults used for a measurement uncertainty estimationshould be at least 6 for each estimation range(International Organization for Standardization 2012)

& The number of replicate sample measurements usedfor a measurement uncertainty estimation should beat least 60 for each estimation range (Magnussonet al. 2018)

& Themeasurement uncertainty estimation could be per-formed with the same interval as the reference solu-tions are measured, e.g., daily (autonomous mode)

During the study, the AutoMUkit software had to berun manually to calculate measurement uncertainties forpast measurement results. In the future, an autonomousmode will be implemented, which enables calculation ofthe measurement uncertainty on a set interval, e.g., dailyusing a defined quantity of historical data, e.g., 100–200results during 1 or 2 weeks. AutoMUkit will then attachthe calculated measurement uncertainties to the future

measurement results until the next uncertainty calcula-tion is performed. This way, the most recent measure-ment uncertainty value represents the current state of thesystem and the measurement uncertainty estimation isclose to real time.

The procedure, described by Näykki et al. (2015)which was tested in this paper, is not limited to onlyturbidity measurements and it can be applied for uncer-tainty estimations in any automated continuous mea-surements in which routine sample replicate measure-ments and reference material measurements are per-formed sequentially. The same uncertainty estimationprocedure could in theory also be utilized with onlinegas analyzers. The real limitations are set by the avail-ability of suitable reference materials and measurementdevices. Most of the field measurement devices andsensors are not so called flow-through model, which isrequired because of the need to measure synthetic refer-ence solutions, but this problem can be solved simply byinstalling the sondes or sensors into a flow-through cell.A more severe problem can be caused by the availabilityof stable reference solutions and reference solutionmixing, because after all, a dilute reference solution

0

10

20

30

40

Apr 02 Apr 09 Apr 16 Apr 23 Apr 30

Time

Tu

rb

idity (

FN

U)

Sample type

Sample

Standard solution 1

Standard solution 2

Fig. 7 Results of the second continuous measurement test. Black line represents measurement results and gray area around the resultsdescribes the calculated measurement uncertainty (k = 2) for the online turbidity sensor

259 Page 12 of 14 Environ Monit Assess (2019) 191: 259

should retain a stable reference value throughout thecontainer volume for long enough time periods.

Conclusions

In the future, where continuous measurements will mostlikely have a significantly larger role in environmentalmonitoring, this approach should be studied in fieldconditions and with different measurements to gainknowledge about suitable parameters and calculationsettings. For all applications, such a heavy quality con-trol system is not needed, but this paper demonstratesthat it is possible to have an automated quality controlsystem for continuous field measurements. The systemis capable of producing measurement results with higherquality and traceability compared to current best avail-able commercial technologies, considering the addition-al information on data quality, but at the cost of highermaintenance requirements. Also, special care should betaken to ensure the quality of reference solution mea-surements as failures to accurately and reproduciblymeasure the synthetic reference solutions will causeoverestimated measurement uncertainties.

Acknowledgements Open access funding provided by FinnishEnvironment Institute (SYKE). Atte Virtanen is acknowledged onhis work with programming the AutoMUkit software. KennethArandia is acknowledged on his workwith the small-scale stabilitytesting of formazin solutions.

Open Access This article is distributed under the terms of theCreative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestrict-ed use, distribution, and reproduction in any medium, providedyou give appropriate credit to the original author(s) and the source,provide a link to the Creative Commons license, and indicate ifchanges were made.

References

Bilotta, G. S., & Brazier, R. E. (2008). Understanding the influ-ence of suspended solids on water quality and aquatic biota.Water Research, 42, 2849–2861.

Björklöf, K., Leivuori, M., Näykki, T., Väisänen, T., Ilmakunnas,M. & Väisänen, R. (2016a). Kenttämittausvertailu 13/2016 -Luonnonvesien happi, lämpötila, pH, sähkön-johtavuus jasameus. Suomen ympäristökeskuksen raportteja 46/2016.Suomen Ympäristökeskus. https://helda.helsinki.fi/handle/10138/170301. Accessed 14 March 2017.

Björklöf, K., Näykki, T. & Väisänen, T. (2016b). Luotettavakenttämittaus edellyttää osaavaa mittaajaa ja riittäväälaadunvarmistusta. Vesitalous 4/2016, 39-42.

Blomberg von der Geest, K., Hyvönen, J. & Laurila, T. (2012).Real-time determination of metal concentrations in liquidflows using microplasma emission spectroscopy. PhotonicsGlobal Conference (PGC), Singapore: IEEE, 1–5.

Caradot, N., Sonnenberg, H., Rouault, P., Gruber, G., Hofer, T.,Torres, A., Pesci, M., & Bertrand-Krajewski, J. (2015).Influence of local calibration on the quality of online wetweather discharge monitoring: feedback from five interna-tional case studies. Water Science & Technology, 71, 45–51.https://doi.org/10.2166/wst.2014.465.

Conrad, C. C., & Hilchey, K. G. (2010). A review of citizenscience and community-based environmental monitoring:issues and opportunities. Environmental MonitoringAssessment, 176, 273–291.

Delauney, L., Compère, C., & Lehaitre, M. (2010). Biofoulingprotection for marine environmental sensors. Ocean Science,6, 503–511.

Dunbabin, M., & Marques, L. (2012). Robotics for environmentalmonitoring. IEEE Robotics and Automation Magazine,19(1), 24–39.

Ellingsen, K. E., Yoccoz, N. G., Tveraa, T., Hewitt, J. E., &Thrush, S. F. (2017). Long-term environmental monitoringfor assessment of change: measurement inconsistencies overtime and potential solutions. Environmental MonitoringAssessment, 189, 595. https://doi.org/10.1007/s10661-017-6317-4.

Elliott, M. (2014). Integrated marine science and management: wad-ing through the morass.Marine Pollution Bulletin, 86, 1–4.

Eurachem Guide TAM (2011). Terminology in analytical mea-surement - introduction to VIM 3. Available from www.eurachem.org. Accessed 17 October 2017.

Eurachem/CITAC Guide CG 4 (2012). Quantifying uncertainty inanalyt ical measurement ht tps: / /www.eurachem.org/images/stories/Guides/pdf/QUAM2012_P1.pdf.Accessed 17 October 2017.

European Commission (2009). Guidance Document No. 19Guidance on surface water chemical monitoring under thewa t e r f r amework d i r e c t i v e . h t t p : / / e c . eu ropa .eu /env i ronmen t /wa te r /wa te r - f r amework / f ac t s_figures/guidance_docs_en.htm. Accessed 4 March 2017.

Giles, C. (2013). Next generation compliance. https://www.epa.gov/sites/production/files/2014-09/documents/giles-next-gen-article-forum-eli-sept-oct-2013.pdf. Accessed 21June 2018.

Horowitz, A. J. (2013). A review of selected inorganic surfacewater quality-monitoring practices: are we really measuringwhat we think, and if so, are we doing it right? EnvironmentalScience & Technology, 47, 2471–2486.

Hovind, H., Magnusson, B., Krysell, M., Lund, & U., Mäkinen, I.(2011). Nordtest Technical Report 569 - Internal QualityControl - Handbook for Chemical laboratories, 4th edition.h t t p : / /www.no rd te s t . i n fo / i ndex .php / t echn i ca l -reports/item/internal-quality-control-handbook-for-chemicallaboratories-trollboken-troll-book-nt-tr-569-english-edition-4.html. Accessed 18 Oct 2017.

International Organization for Standardization (2012). ISO 11352,Water quality—estimation of measurement uncertainty basedon validation and quality control data. Geneva: ISO.

Environ Monit Assess (2019) 191: 259 Page 13 of 14 259

International Organization for Standardization (2016). ISO 7027–1, Water quality—determination of turbidity—part 1: quan-titative methods. Geneva: ISO.

International Organization for Standardization (2017). ISO 17025,General requirements for the competence of testing andcalibration laboratories. Geneva: ISO.

Joannis, C., Ruban, G., Gromaire, M.-C., Bertrand-Krajewski, J.-L., & Chebbo, G. (2008). Reproducibility and uncertainty ofwastewater turbidity measurements. Water Science &Technology, 57, 1667–1673. https://doi.org/10.2166/wst.2008.292.

Joint Committee for Guides in Metrology (2008). Evaluation ofmeasurement data - guide to the expression of uncertainty inm e a s u r e m e n t ( G UM ) . h t t p s : / / w ww. b i p m .org/utils/common/documents/jcgm/JCGM_100_2008_E.pdf.Accessed 17 October 2017.

Kahiluoto, J. (2017). Kenttäkäyttöisen veden sameudenmittauksen automatisointi jatkuvatoimisella mittaustuloksenlaadunvarmistustekniikalla. https://aaltodoc.aalto.fi/handle/123456789/28908. Accessed 21 June 2018.

Lewis, A., & Edwards, P. (2016). Validate personal air-pollutionsensors. Nature, 535, 29–31.

Magnusson B., Näykki T., Hovind H., Krysell M. & Sahlin E.(2017). Handbook for calculation of measurement uncertaintyin environmental laboratories, Nordtest Report TR 537 (ed. 4).Available from www.nordtest.info. Accessed 5 March 2017.

Magnusson, B., Hovind, H., Krysell, M., Lund, U., Mäkinen, I.(2018). Nordtest technical report 569 - internal quality con-trol - handbook for chemical laboratories, 5th edition.http://www.nordtest.info/images/documents/nt-technical-reports/NT_TR_569_ed5_1_Internal_Quality_Control_English.pdf. Accessed 18 January 2019.

Moser, J., &Wegscheider, W. (2001). Quantifying the measurementuncertainty of results from environmental analytical methods.Fresenius’ Journal of Analytical Chemistry, 370, 679–689.

Näykki, T. and Väisänen, T. (2016). Laatusuosituksetympäristöhallinnon vedenlaaturekistereihin vietävälletiedolle. Vesistä tehtävien analyyttien määritysrajat, mittaus-epävarmuudet sekä säilytysajat ja -tavat. - 2. uudistettupainos. Suomen ympäristökeskuksen raportteja 22/2016.Suomen ympäristökeskus. https://helda.helsinki.fi/handle/10138/163532. Accessed 13 September 2017.

Näykki, T., Koponen, S., Väisänen, T., Pyhälahti, T., Toivanen, T.,& Leito, I. (2014). Validation of a new measuring system for

water turbidity field measurements. Accredited QualityAssurance, 19, 175–183.

Näykki, T., Virtanen, A., Kaukonen, L., Magnusson, B., Väisänen,T., & Leito, I. (2015). Application of the Nordtest method forBreal-time^ uncertainty estimation of on-line field measure-ment. Environmental Monitoring Assessment, 187, 630.https://doi.org/10.1007/s10661-015-4856-0.

Nilssen, I., Ødegård, Ø., Sørensen, A., Johnsen, G., Moline, M., &Berge, J. (2015). Integrated environmental mapping andmonitoring, a methodological approach to optimise knowl-edge gathering and sampling strategy. Marine PollutionBulletin, 96, 374–383.

Rockström, J., Steffen, W., Noone, K., Persson, Å., Chapin, F. S.,Lambin, E. F., Lenton, T. F., Scheffer, M., Folke, K.,Schellnhuber, H. J., Nykvist, B., de Witt, C. A., Hughes, T.,van der Leeuw, S., Rodhe, H., Sörlin, S., Snyder, P. K.,Costanza, R., Svedin, U., Falkenmark, M., Karlberg, L.,Corell, R. W., Fabry, V. J., Hansen, J., Walker, B.,Liverman, D., Richardson, K., Crutzen, P., & Foley, J. A.(2009). A safe operating space for humanity. Nature,461(24), 472–475.

Rymszewicz, A., O’Sullivan, J. J., Bruen, M., Turner, J. N.,Lawler, D. M., Conroy, E., & Kelly-Quinn, M. (2017).Measurement differences between turbidity instruments,and their implications for suspended sediment concentrationand load calculations: a sensor inter-comparison study.Journal of Environmental Management, 199, 99–108.

Sadar, M. (2003). Turbidity standards technical information series—booklet no. 12, Hach. https://www.hach.com/asset-get.download-en.jsa?code=61799. Accessed 4 December 2017.

Skeffington, R. A., Halliday, S. J., Wade, A. J., Bowes, M. J., &Loewenthal, M. (2015). Using high-frequency water qualitydata to assess sampling strategies for the EU WaterFramework Directive. Hydrology and Earth SystemSciences, 19, 2491–2504.

Publisher’s note Springer Nature remains neutral with regard tojurisdictional claims in published maps and institutionalaffiliations.

259 Page 14 of 14 Environ Monit Assess (2019) 191: 259