Journal of Food Engineering 116 (2013) 86–96

Contents lists available at SciVerse ScienceDirect

Journal of Food Engineering

journal homepage: www.elsevier .com/locate / j foodeng

Towards an online milk concentration sensor using ERT: Correlationof conductivity, temperature and composition

Mohadeseh Sharifi, Brent Young ⇑Department of Chemical and Materials Engineering, University of Auckland, Private Bag 92019, Auckland 1142, New Zealand

a r t i c l e i n f o

Article history:Received 3 August 2012Received in revised form 20 November 2012Accepted 20 November 2012Available online 29 November 2012

Keywords:Skim milkWhole milkConductivityTotal solids content (TS)Electrical resistance tomography (ERT)

0260-8774/$ - see front matter � 2012 Elsevier Ltd. Ahttp://dx.doi.org/10.1016/j.jfoodeng.2012.11.010

⇑ Corresponding author. Tel.: +64 9 9235606; fax: +E-mail address: b.young@auckland.ac.nz (B. Young

a b s t r a c t

The milk powder production process is a multistage process in which the function of each stage affectsthe final product quality. Having accurate and multidimensional knowledge of the milk solution at differ-ent stages of the process is crucial in order to maintain satisfactory control of the process. Electrical resis-tance tomography (ERT) is a novel, robust and high speed method of process imaging. It has the potentialto overcome the lack of online concentration sensors and provide multidimensional concentration infor-mation at various stages of the milk powder production process. In order to achieve this goal an accuratemodel is required to correlate the ERT’s multidimensional conductivity measurements to milk concentra-tion/total solids content.

This work focuses on obtaining a correlation of milk conductivity, temperature and composition(including powder protein, lactose and fat content on a dry basis) to the total solids content for wholeand skim milk in the process to produce milk powder. For this purpose, the Design of Experiments(DOE) methodology was applied to plan and conduct a Central Composite Design (CCD) experiment.The resulting Response Surface, correlating the factors to whole and skim milk concentration, providedan accuracy of over 93%. The results were also compared to the results of a multiple linear regressionmodel correlating only milk conductivity and temperature to its total solids content. Multiple linearregression models presented better accuracy for lower concentrations while the CCD model presentedbetter accuracy for higher concentrations.

� 2012 Elsevier Ltd. All rights reserved.

1. Introduction

In the milk powder production process, manufacturing of prod-ucts which fulfill all qualitative requirements (Pisecky, 1997), to-gether with decreased energy consumption, reduced waste andminimized costs, can be achieved through automatic process con-trol which in turn requires accurate and multidimensional knowl-edge of the process states at different stages of the process. Thisissue becomes especially important when operating conditionschange away from the initial norm. The need for such detailedinformation is due to the fact that the milk powder production pro-cess is a multistage process in which the function of each stage af-fects the final product quality.

In order to achieve an acceptable final product quality in themilk processing industry, strict regulations on the monitoringand control of various aspects of milk such as temperature, pres-sure, flow, concentration, composition, hygiene, taste and smellhave been imposed (Bylund, 1995). Although at present, sensorsand instrumentation for in-line and online measurement of physi-cal and physical–chemical properties of products, such as

ll rights reserved.

64 9 373 7463.).

temperature, pressure, flow and levels in tanks, are commerciallyavailable and applied in the food industry, sensors to determineproduct composition and concentration are not yet much devel-oped (TeGiffel, 2006). Various methods for the measurement ofmilk concentration or total solids content (TS) have been proposed,but most have disadvantages including imprecision and samplingproblems and they do not have the possibility of providing a dy-namic, spatially-distributed image of the data variation. Therefore,the need for multidimensional monitoring of milk concentration,especially in inhomogeneous situations, still remains.

Lately, the measurement of process variables has been revolu-tionized through a new technique called ‘‘Process Tomography’’.This technique offers fresh methods for studying the interior ofan industrial system. One of the most common modalities of Pro-cess Tomography which has significantly progressed since it wasinvented in the 1980s is electrical resistance tomography (ERT).This technique injects current to the various sensors arrangedaround the region of interest and collects the resulting voltagefrom all the other sensors. The collected voltage data are thenreconstructed into multidimensional conductivity data of the re-gion of interest. Such multidimensional conductivity data can thenbe further correlated to various other useful information of theprocess under investigation. ERT has become a favored system in

Table 1Factor variations during standardized milk concentration process to produce milkpowder at Te Rapa-Fonterra Co. Also the ranges applied in the model developmentexperiments in this work.

Factors Whole milk Skim milk

�1 +1 �1 +1

Total solids content (%) 13 52 9 52Temperature (C) 13 45 13 45Protein content-dry powder based (%) 24 27 33 39Lactose content-dry powder based (%) 38 50 51 58Fat content-dry powder based (%) 28 30 1 2

M. Sharifi, B. Young / Journal of Food Engineering 116 (2013) 86–96 87

observation and examination of various flows, due to its variedbenefits such as high speed, low cost, no radiation hazard, andnon-intrusiveness. ERT may have potential applications in multidi-mensional monitoring of concentration in various stages of indus-trial milk processing. This could be at any point from the collectionand reception step to the concentrate feed to the spray dryer. Cur-rently, for concentrated products, the concentrate flow rate, dryerfeed rate and energy input are controlled to obtain the desired totalsolids content. For milk powders, the total solids content of thespray dryer feed is controlled to obtain a desired quality of product.In reconstitution tanks in the recombined milk product industry,there is also the need to control the total solids content of thereconstituted milk solution (Tamime, 2009).

There is a growing body of evidence on the applicability of ERTto process characterization and diagnostics (Holden et al., 1999;Mann et al., 1999; Deng et al., 2001; Wang et al., 2003; Lee andBennington, 2007; Pakzad et al., 2008; Zhao et al., 2008; Giguèreet al., 2009, Stephenson et al., 2009, Tan and Dong, 2009). However,Sharifi et al. studied the application of ERT in the milk processingindustry in general, including milk concentration measurement,for the first time in 2010 (Sharifi and Young, 2010). Later in2011, we further expanded our studies and also examined theapplication of ERT to milk fat content measurement (Sharifi andYoung, 2012a) and 3-dimensional monitoring of milk mixing tanks(Sharifi and Young, 2011). In 2012, we applied ERT to single phaseflow analysis by developing a novel calculation methodology forthis purpose as previous methods were not applicable (Sharifiand Young, Submitted for publication) and later a methodologywas developed for accurate multidimensional monitoring andvelocity profile measurement of various milk solutions flowing ina pipe (Sharifi and Young, 2012b). In these studies we identifiedthe need for correlating ERT conductivity measurements to milkconcentration or total solids content information. Our study fo-cused on correlating the conductivity and temperature measure-ments of reconstituted whole and skim milk solutions up to 30%totals solids content and 60 �C to total solids content data usingmultiple linear regression. Concentrations above 30% TS in the milkconcentration process to produce milk powder and also the effectof compositional process variations were not examined.

The aim of this work is to obtain accurate correlations betweenERT multidimensional conductivity measurements with milk con-centration. To address this aim and in order to systematically planthe required experiments so that appropriate data will be obtainedthat can be analysed by statistical methods, to yield valid, mean-ingful and objective conclusions, the Design of Experiments(DOE) methodology was applied. DOE is a very comprehensive to-pic including various techniques, designs and principles which arediscussed in detail in various books including ‘‘Design and Analysisof Experiments’’ by Montgomery (Montgomery, 1997). Some ofthese designs and techniques applied in this work are introducedbelow.

1.1. Factor Screening

During the milk concentration process to produce milk powder,standardized milk faces variations in terms of concentration, tem-perature, protein, lactose and fat content. The variations of each ofthe 5 stated factors during the standard/typical milk concentrationprocess will be tested for each factor (Table 1). Design of Experi-ments (DOE) will be applied to plan the experiments. The firststage of the experiments will aim at identifying factors that havelarge effects on the response among the several initially recognizedfactors (factor screening experiments). Factorial Designs are themost efficient technique for this type of experiment in which oneor more process variables or factors are changed in order to ob-serve the effect of the changes on one or more response variables.

Factors showing very little effect on the response when changedcan be omitted from the correlation equation. In this work due tothe relatively large number of factors (k) which require a largenumber of runs outgrowing time and material resources, a one-halffraction of the 2k design or a 2k�1 design (fractional factorial de-signs) will be conducted in order to eliminate inactive factors.(Montgomery, 1997).

1.2. Response surface methodology

After elimination of the low influence factors, the next stepwould be to fit a response curve/surface to the levels of the quan-titative factors using response surface methodology so that theexperimenter has an equation that relates the response to the fac-tors (Montgomery, 1997). For this purpose the Sequential Strategyof experimentation will be applied. In this strategy the 2k designaugmented with centre points needs to be conducted to fit afirst-order model. Then, in experiments which curvature is signifi-cant and lack-of-fit is exhibited, the two-level design can be aug-mented with axial points to allow quadratic terms to beincorporated and obtain a central composite design (Montgomery,1997).

Therefore, the aim of this work, in more detail, is to develop amodel for whole and skim milk concentration/total solids content(response) in relation to the active factors among solution conduc-tivity, temperature, and dry powder composition; protein, lactoseand fat content (factors), during the concentration process to pro-duce milk powder (i.e. milk concentration is a function of activefactors among milk conductivity, temperature, fat, protein, and lac-tose content).

This model will provide the ability to extract accurate concen-tration distribution information from various milk solutions inthe dairy industry based on multidimensional conductivity mea-surements obtained from the application of ERT, and other knowninformation such as temperature and composition.

This paper is organized as follows. The experimental materialsand equipment used are described. Then the experimental proce-dure is described including the milk solution preparation, factorscreening experiments, response surface development experi-ments, multiple linear regression model experiments, and modeltesting experiments. Finally the results are presented, discussedand conclusions drawn.

2. Materials and methods

2.1. Milk solution preparation

The whole and skim milk concentration, temperature, and drypowder compositional variations observed at the Te Rapa site ofFonterra Cooperative Group Ltd during the concentration processof standardized milk to produce milk powder are the ranges beingtested in this work and are provided in Table 1. These are typical ofmilk concentration processes generally (Tamime, 2009).

88 M. Sharifi, B. Young / Journal of Food Engineering 116 (2013) 86–96

The milk samples prepared for the model developmentexperiments were 200 ml samples prepared by reconstitutionof whole and skim milk powder and also Whey Protein Isolate(WPI) as the protein source, lactose powder as the lactosesource and commercial full fat cream as the source for fat.For testing the models developed, 30 kg each of 13%, 24.55%and 47% (weight percentage) Total Solids (TS) contentreconstituted skim milk and whole milk solutions were pre-pared commensurate with raw milk and concentrated milksolutions after the second and third evaporation effects, respec-tively. These ranges were chosen to demonstrate the accuracyof the model over the entire concentration range of the milkconcentrating process in the industry. Milk powders, WPI andlactose powder were kindly provided by Fonterra CooperativeGroup Ltd. Commercial full fat cream was purchased from thesupermarket.

As explained in the authors’ previous studies (Sharifi andYoung, 2011, 2012b), a Silverson high shear batch mixer with asquare hole screen was used for the reconstitution of milk pow-ders, and a METTLER TOLEDO SevenEasy local conductivity probewas used for the conductivity measurement of the samplesolutions in the model development experiments and to gatherreference conductivity and temperature data in the model testingexperiments.

All milk solutions throughout the experiments were prepared asfully homogenized and hydrated milk solutions (Tamime, 2009),which best resemble fresh milk solutions in the various stages ofprocessing using the procedure described in the authors’ previousstudies (Sharifi and Young, 2011, 2012b).

2.2. Experimental rig

The experimental setup of the flow rig used for the purpose oftesting the concentration models developed was the same set upas explained in the authors’ previous study (Sharifi and Young,2012b). This setup is illustrated in Fig. 1.

Fig. 1. Schematic diagram of th

A P2000 ERT system (Sensor and Data Acquisition System) waspurchased from Industrial Tomography Systems plc. for the pur-pose of the model validating experiments. All experiments wereconducted at an optimized injection frequency and current ampli-tude of 9600 Hz and 20–30 mA, respectively. The small amplitudeof the injection current assures no impact on the milk componentsand safety.

Design-Expert 8.0.6 was used to assist with planning of experi-ments and analysis of results.

2.3. Factor screening experiments

As explained in the Introduction section, the aim of this work isto find:

TS ¼ f ðr; T; P; L; FÞ ð1Þ

It is not possible to use conductivity as a variable factor andconcentration as the response in the sample preparation andexperimental stages. In order to make sample preparation possible,conductivity was used as the response while concentration wasmodified as a variable factor:

r ¼ f 0ðTS;T; P; L; FÞ ð2Þ

For Eqs. (1) and (2) r is the Solution conductivity, (mS/cm); TS,Solution total solids content (%); T, Solution temperature (�C); P,Dry based protein content (%); L, Dry based lactose content (%);F, Dry based fat content (%); f, f0, Mathematical functions.

Eq. (2) is the general format of the model which will be obtainedthrough DOE experimentation and Eq. (1) is the general format ofthe model or response surface of interest which will be obtainedfrom rearrangement of Eq. (2) or the DOE obtained model.

As the first step, a factor screening experiment was designedand conducted. For this purpose and due to resource restrictionsa one-half fraction of the 25 factorial design or in other words25�1 design with 2 replicates was conducted for each milk type,whole and skim. Milk solutions were prepared according to this

e experimental rig/setup.

Fig. 2. (a) Normal plot and (b) pareto chart for whole milk screening experiments demonstrating factor effects.

M. Sharifi, B. Young / Journal of Food Engineering 116 (2013) 86–96 89

Fig. 3. (a) Normal plot and (b) pareto chart for skim milk screening experiments demonstrating factor effects.

90 M. Sharifi, B. Young / Journal of Food Engineering 116 (2013) 86–96

Table 2Anova summary of results for testing the first-order model against curvature for (a)whole milk; and (b) skim milk.

AdjustedF-value

Modelp-value

UnadjustedF-value

Modelp-value

a: ANOVA summaryModel 18.86 0.0003 1.07 0.4778Curvature 133.70 <0.0001Lack of fit 1.14 0.3903 47.14 0.0004

b: ANOVA summaryModel 59.99 <0.0001 6.52 0.0017Curvature 132.11 <0.0001Lack of fit 68.17 0.0001 611.77 <0.0001

M. Sharifi, B. Young / Journal of Food Engineering 116 (2013) 86–96 91

design and as explained previously, and duplicate conductivitymeasurements were made with the average measurementrecorded.

2.4. Response surface methodology experiments

After elimination of low/no-influence factor(s), the next step ofexperiments was conducted on the remaining factors using thesequential strategy of experimentation. Estimation of the experi-mental error will be possible using the replicates of the centrepoints.

As quadratic effects were found to be significant, further axialruns were required for the aim of developing a quadratic responsesurface. Due to the fact that the factor ranges applied to the exper-iments were the complete ranges for each factor and predictionsoutside these ranges were not of interest for the aim of this work,an a (the distance of the axial points from the centre point) of 0.5was used to produce axial runs which are at the median of the twofactorial levels. The axial runs were all replicated twice for the esti-mation of experimental error.

2.5. Multiple linear regression experiments

In the authors’ previous study (Sharifi and Young, 2012a), amultiple linear regression model was developed correlating wholeand skim milk conductivity and temperature to concentration forsolutions of up to 30% TS (weight percentage). In the concentrationprocess to produce milk powder, the concentrate TS reaches up to�52% TS (weight percentage). Therefore, the same methodologywas applied in this work to obtain a correlation for whole and skimmilk solutions with concentrations in the range of 30–50% TS(weight percentage).

2.6. RSM and MLR model testing

Finally, in order to test the models developed using the DOEtechnique and the multiple linear regression for TS predictionaccuracy across the complete concentration range of interest inthe milk concentrating process in the industry, reconstituted skimmilk and whole milk solutions were prepared (13%, 24% and 47% TS(weight percentage) commensurate with raw milk and concen-trated milk solutions after the second and third evaporation ef-fects, respectively). These solutions were put to test in theexperimental rig in which multidimensional conductivity datawere obtained using ERT. These conductivity measurements werethen converted into TS information using the response surfacedeveloped using the CCD technique and the correlation modeldeveloped using Multiple Linear Regression (MLR). The results ofthese techniques were compared to the real TS as used in the prep-aration of samples for accuracy.

3. Results and discussions

3.1. Factor screening analysis

Selected results from the Design-Expert software analysis of thefactor screening experiments for whole and skim milk are shownin Figs. 2 and 3.

Figs. 2a and 3a show the normal plot of residuals for whole andskim milk, respectively. The normal plot is used to highlight activefactors or factors with significant effect on the response. The ideaof this approach is that if none of the factors is active, the variationin the estimates of effect will be purely due to random variation, sothat a normal plot of the estimates will be roughly linear. Other-wise, the points not on the underlying linear pattern show signifi-cant positive and negative effects depending on which side of theline they are located at. Figs. 2b and 3b show the Pareto chart oft-values of the effects for whole and skim milk, respectively. ThePareto chart is used to give a picture of the relative sizes of the dif-ferent effects. Pareto charts are usually used in conjunction withthe normal plot for active factor selection and help prevent over-selection of factors from the normal plot. We can see that all inde-pendent factors of Whole milk (TS, temperature, protein, lactoseand fat) show significant effects, while for skim milk fat contentdoes not show a significant effect on conductivity and thereforewas eliminated in the following experiments. The insignificant ef-fect of fat content on skim milk conductivity is due to the verysmall range in which it appears in skim milk contents. For bothtypes of milk some interactions also show significant effects, espe-cially for whole milk.

3.2. CCD response surface analysis

From the factor screening experiments’ analysis, as the effect offat content for skim milk was found to be insignificant, thus thisfactor was eliminated from the response surface experiments.Therefore whole milk remained with 5 and skim milk with 4 signif-icant factors to be tested for response surface development. A 25�1

design for whole milk and a 24 design for skim milk with 6 centrepoints were conducted to allow for checking of the first-ordermodel.

The Anova results of the first-order model (Table 2) showed sig-nificant curvature for both milk types, whole and skim. Therefore itwas concluded that the effect of curvature in the response is signif-icant and needs to be considered for an accurate model. In order toconsider the curvature and consequently quadratic effect of themodel, extra axial runs with the aim of developing a quadratic re-sponse surface were conducted.

Figs. 4 and 5 show selected quadratic surface analysis resultsfrom design-expert for whole and skim milk, respectively.

Figs. 4a and 5a show the normal plot of residuals which shouldideally be a straight line, and in this case for whole and skim milkthey line up adequately. Figs. 4b and 5b show the predicted versusactual plots for whole and skim milk, respectively. These plotsshow how the model predicts over the range of data and should ex-hibit random scatter about the 45 degree line. Clusters of pointsabove or below the line indicate problems of over and under pre-dicting. Therefore there is no significant problem with the predic-tion of the developed response surfaces for whole and skim milk,although better prediction can be seen in the case of Skim milk.This is due to the fact that for whole milk, due to the fat content,keeping the samples in a homogenous situation for conductivitymeasurement was found to be difficult to maintain in the case ofhigher fat content samples.

Fig. 4. Selected design-expert results for whole milk central composite design; (a) normal plot of residuals; (b) predicted vs. actual plot b:

92 M. Sharifi, B. Young / Journal of Food Engineering 116 (2013) 86–96

Table 3 shows the resulting quadratic surface models developedfor each milk type, and the various values demonstrating the accu-racy of the developed models.

It is again obvious that the model developed for skim milk ismuch better than the whole milk model which could be causedby the fat content causing inhomogeneity.

Fig. 5. Selected design-expert results for skim milk central composite design; (a) normal plot of residuals; (b) predicted vs. actual plot.

M. Sharifi, B. Young / Journal of Food Engineering 116 (2013) 86–96 93

Table 3(a) Quadratic response surface for whole milk; (b) quadratic response surface for skimmilk.

Final equation in terms of coded factors

a: Conductivity =+2.48+0.34 *A+0.32 *B�0.24 *C�0.55 *D�0.13 *E�0.29 *A*C�.19 *B*C�0.19 *B*D�0.21 *C*D�0.23 *D*E+3.75 *D2

�2.43 *E2

Std. Dev. 0.43 R-squared 0.8849Mean 3.12 Adj R-squared 0.8191C.V.% 13.89 Pred R-squared 0.5194PRESS 16.44 Adeq precision 14.910

b: Ln(Conductivity) =+1.93+0.20 *A+0.095 *B�0.25 *C�0.24 *D+0.028 *A*B+0.019 *A*C+0.057 *A*D�0.23 *C*D�0.48 *A2

Std. Dev. 0.030 R-squared 0.9963Mean 1.69 Adj R-squared 0.9949C.V.% 1.76 Pred R-squared 0.9891PRESS 0.060 Adeq precision 101.747

* A: Total Solids Content; B: Temperature; C: Protein; D: Lactose; E: Fat.

94 M. Sharifi, B. Young / Journal of Food Engineering 116 (2013) 86–96

3.3. Multiple linear regression analysis

A multiple linear regression model correlation for whole andskim milk concentration to the solution conductivity andtemperature was developed for a 30–50% TS (weight percentage)

Fig. 6. Multiple linear regression for 30

concentration range (Fig. 6) as the authors’ previous work (Sharifiand Young, 2012a) only covered concentrations up to 30% TS(weight percentage). As in the previous study, for concentrationsof 30–50% TS also the interaction effect of temperature (T) andconductivity (C) showed to be statistically significant in theregression analysis.

The points represent the experimental data and the lines rep-resent the regression. The models developed show good accuracyas can be seen through the high R-squared values. The slightlyhigher effect of temperature on whole milk conductivity couldbe due to the presence of higher fat content.

3.4. Model testing results

Using the conductivity–concentration models obtained fromCCD and MLR, ERT multidimensional conductivity measurementsfrom various milk solutions are converted to concentration data.Fig. 7 shows a comparison between the TS prediction of the re-sponse surface developed through CCD, including the effects ofpowder compositional variations, and the simple multiple linearregression (MLR) model focusing only on the effects of conductiv-ity and temperature. In order to collect the most accurate ERTconductivity measurements of the various milk solutions, refer-ence measurements were taken from the same milk solutionbeing tested.

These estimations are compared to the real TS value as usedduring the sample preparations for error evaluation (error(%) = (model predicted TS value � real TS value)/real TSvalue � 100%). MLR shows an average error of 5% in TS estimationwhile using the CCD model this value is slightly higher at 6.8% forboth types of milk. These average errors provide an overall viewon the accuracy of the model while Fig. 7 clearly shows the errorsof the CCD and MLR models for each milk sample. Details on theaccuracy of the CCD model such as standard deviation can befound in Table 3. As can be seen in Fig. 7, the CCD model providesbetter TS estimation in the higher whole and skim milk concen-tration ranges compared to MLR. MLR shows an increase in TS

estimation error with the increase of milk concentration, whilefor CCD better estimation can be seen in higher concentrationranges. Both CCD and MLR predictions show less accuracy forwhole milk TS estimations.

-50% TS whole and skim milk.

Fig. 7. Comparison of CCD and MLR model TS prediction errors.

M. Sharifi, B. Young / Journal of Food Engineering 116 (2013) 86–96 95

4. Conclusions

In order to develop and apply ERT as an online multidimen-sional concentration sensor in the milk powder production indus-try, an accurate model needs to be established to correlate ERTmultidimensional conductivity measurement to milk concentra-tion data. For better applicability and precision the model needsto consider any process variations effecting the correlation. Previ-ously milk concentration was correlated to its conductivity andtemperature only through MLR. Also the developed model fo-cused only on a subsection of the milk concentration process(up to 30% TS). The present work has completed the concentra-tion range and developed an MLR model for the correlation ofmilk concentration from conductivity and temperature for higherconcentration ranges (30–50%). The authors have also developeda response surface model through DOE and CCD that as well asthe effects of conductivity and temperature, takes the effects ofcompositional variations (protein, lactose and fat content) intoconsideration. In the model development procedure the concen-tration, temperature, and compositional variations of the entireconcentration process of the standardized milk has been takeninto account. Both models provide relatively similar accuraciesin the range of 93–95%. The relatively high accuracy of the devel-oped models provides ERT with the potential ability of being ap-plied as an online concentration sensor in the milk powderproduction process.

Acknowledgement

The authors gratefully acknowledge the support of FonterraCooperative Group Ltd., for providing us with the milk powder,WPI and lactose required for experiments.

References

Bylund, G., 1995. Dairy Processing Handbook. Tetra Pak Processing Systems AB.Deng, X., Dong, F., Xu, L.J., Liu, X.P., Xu, L.A., 2001. The design of a dual-plane ert

system for cross correlation measurement of bubbly gas/liquid pipe flow.Measurement Science and Technology 12 (8), 1024–1031. Available from:<http://www.scopus.com/inward/record.url?eid=2-s2.0-0035422515&partnerID=40&md5=87e5b3bf83f4cedfc2acc0756574f0b4>.

Giguere, R., Fradette, L., Mignon, D., Tanguy, P.A., 2009. Analysis of slurry flowregimes downstream of a pipe bend. Chemical Engineering Research and Design87 (7), 943–950.

Holden, P.J., Wang, M., Mann, R., Dickin, F.J., Edwards, R.B., 1999. On detectingmixing pathologies inside a stirred vessel using electrical resistancetomography. Chemical Engineering Research and Design 77 (8), 709–712.Available from: http://dx.doi.org/10.1205/026387699526845.

Lee, Q.F., Bennington, C.P.J., 2007. Measuring flow velocity and uniformity in amodel batch digester using electrical resistance tomography. Canadian Journalof Chemical Engineering 85 (1), 55–64. Available from: <http://www.scopus.com/inward/record.url?eid=2-s2.0-34548220425&partnerID=40>.

Mann, R., Wang, M., Forrest, A.E., Holden, P.J., Dickin, F.J., Dyakowski, T., Edwards,R.B., 1999. Gas–liquid and miscible liquid mixing in a plant-scale vesselmonitored using electrical resistance tomography. Chemical EngineeringCommunications 175, 39–48. Available from: <http://www.scopus.com/inward/record.url?eid=2-s2.0-0042627897&partnerID=40>.

Montgomery, D.C., 1997. Design and Analysis of Experiments, fifth ed. John Wiley &Sons Inc.

Pakzad, L., Ein-Mozaffari, F., Chan, P., 2008. Measuring mixing time in the agitationof non-newtonian fluids through electrical resistance tomography. ChemicalEngineering and Technology 31 (12), 1838–1845. Available from: <http://www.scopus.com/scopus/inward/record.url?eid=2-s2.0-60649099840&partnerID=40>.

Pisecky, I.J., 1997. Handbook of Milk Powder Manufacture. Niro A/S.Sharifi, M., Young, B., 2010. The potential utilisation of ert for monitoring and

control in milk powder processing. Proceedings of the 6th World Congress ofIndustrial Process Tomography, Beijing, China.

Sharifi, M., Young, B., 2011. 3-Dimensional spatial monitoring of tanks for the milkprocessing industry using electrical resistance tomography. Journal of FoodEngineering 105 (2), 312–319. Available from: <http://www.scopus.com/inward/record.url?eid=2-s2.0-79954414837&partnerID=40&md5=75bc337b230598452df87c19a4a9b22e>.

Sharifi, M., Young, B., Submitted for publication. Electrical resistance tomography(ert) for flow and velocity profile measurement of a single phase liquid in ahorizontal pipe. Chemical Engineering Research and Design.

96 M. Sharifi, B. Young / Journal of Food Engineering 116 (2013) 86–96

Sharifi, M., Young, B., 2012b. Milk total solids and fat content soft sensingvia electrical resistance tomography and temperature measurement.Food and Bioproducts Processing. http://dx.doi.org/10.1016/j.fbp.2012.05.001.

Sharifi, M., Young, B., 2012c. Qualitative visualization and quantitative analysis ofmilk flow using electrical resistance tomography. Journal of Food Engineering.http://dx.doi.org/10.1016/j.jfoodeng.2012.04.008.

Stephenson, D.R., Rodgers, T.L., Mann, R., York, T.A., 2009. Application of three-dimensional electrical impedance tomography to investigate fluid mixing ina stirred vesseled. (Eds.), 13th European Conference on Mixing, London, UK,14–17.

Tamime, A.Y., 2009. Dairy Powders and Concentrated Products. Wiley-Blackwellpublishing Ltd.

Tan, C., Dong, F., 2009. Gas–water two-phase flow regime recognition with data andfeature fusion from a dual-plane ert systemed. (Eds.), 5th InternationalConference on Natural Computation, ICNC 2009, Tianjian, 218–222.

Tegiffel, 2006. Dairy Processing, Improving Quality. Woodhead Publishing Ltd.Wang, M., Jones, T.F., Williams, R.A., 2003. Visualization of asymmetric solids

distribution in horizontal swirling flows using electrical resistance tomography.Chemical Engineering Research and Design 81 (8), 854–861. Available from:http://dx.doi.org/10.1205/026387603322482095.

Zhao, Z.F., Mehrvar, M., Ein-Mozaffari, F., 2008. Mixing time in an agitated multi-lamp cylindrical photoreactor using electrical resistance tomography. Journal ofChemical Technology and Biotechnology 83 (12), 1676–1688. Available from:<http://www.scopus.com/inward/record.url?eid=2-s2.0-61349198707&partnerID=40&md5=d65148b23528106cfea51f216f681446>.