Ad

JI

a

ARR2A

KACOPQT

1

utebbgiCtlndtcssls

t

1d

Biochemical Engineering Journal 53 (2010) 38–43

Contents lists available at ScienceDirect

Biochemical Engineering Journal

journa l homepage: www.e lsev ier .com/ locate /be j

chemometric tool to monitor high-rate anaerobic granular sludge reactorsuring load and toxic disturbances

.C. Costa, M.M. Alves, E.C. Ferreira ∗

BB – Institute for Biotechnology and Bioengineering, Centre of Biological Engineering, University of Minho, Campus de Gualtar, 4710-057 Braga, Portugal

r t i c l e i n f o

rticle history:eceived 1 October 2008eceived in revised form7 November 2008ccepted 12 December 2008

a b s t r a c t

Knowing that wide fluctuations in flow rate and presence of toxic compounds can damage the high effi-ciency of high-rate anaerobic granular sludge reactors, the use of Principal Component Analysis (PCA)to detect organic and toxic disturbances was tested. As earlier these disturbances are detected, moreaccurate would be the corrective actions, and less damage will be caused to the microorganisms involvedin the process. The PCA determined a latent variable, combining a weighted sum of operational, physi-ological, and morphological data, which showed high sensitivity to recognize the operational problems

eywords:naerobic digestionontrolrganic loading disturbancerincipal Component Analysisuantitative image analysis

occurred when four organic loading disturbances (OLDs) and three toxic shock loads (TSLs) were appliedto Expanded Granular Sludge Bed (EGSB) reactors. The high loadings/weights linked with the morpho-logical parameters, specially the aggregates size distribution (>0.1, >1), obtained using quantitative imageanalysis techniques, demonstrate the usefulness of monitor the anaerobic granular sludge structuralchanges. The application of PCA chemometric tool to dataset gathering information from all disturbances

n bess of

oxic shock load

allowed the differentiatioeffects caused by each cla

. Introduction

The development of high-rate reactors, based in anaerobic gran-lar sludge, was the key feature that allowed for a great increase inhe use of anaerobic technology for the treatment of a growing vari-ty of industrial wastewaters [1]. Anaerobic granules are particulateiofilms, formed spontaneously by self-immobilization of anaero-ic bacteria in the absence of a support material [1]. Hence, eachranule is a functional unit comprising all the different microorgan-sms necessary for methanogenic degradation of organic matter.onsequently, uncoupling the hydraulic retention time (HRT) fromhe solids retention time allowed the application of high organicoading rates (OLRs), and therefore the use of compact and eco-omical wastewater treatment plants. However, these systems areesigned with reference to a nominal operating condition, in whichhe OLR is assumed to be constant in time. Also, some compoundsan have inhibitory or toxic effects to the microbial populations,uch as detergents and solvents. These facts, coupled with the longtart-up periods imply the need to monitor the anaerobic granu-

ar sludge stability in order to achieve an appropriate control andustainability of the process.

The recognition of parameters that could be used for monitoringhe process is important to efficient control of those processes. It is

∗ Corresponding author. Tel.: +351 253604407; fax: +351 253678986.E-mail address: ecferreira@deb.uminho.pt (E.C. Ferreira).

369-703X/$ – see front matter © 2008 Elsevier B.V. All rights reserved.oi:10.1016/j.bej.2008.12.006

tween organic loading and toxic shock disturbances, as well as the maindisturbance.

© 2008 Elsevier B.V. All rights reserved.

equally feasible to obtain values of parameters measured in solid,liquid or gaseous phases. However, parameters involved in reactorscontrol had been limited to indicators of the liquid and the gaseousphases [2], due to difficulties in obtain and inaccuracy associatedwith morphological parameters.

With the rapid development of instrumental methods theamount of diverse data generated in an environmental processmonitoring and/or control is increasingly drastically [3–5]. Thisadvance guide analysts and researchers to gathering further moremultivariate data. Concurrently, with computer science and tech-nology developments, apply computers and advanced statisticaland mathematical methods to analyse this data became easier.

In this framework, image analysis techniques appear as apromising tool to provide quantitative parameters of the solidphase evolution. Chemometrics-based techniques, such as PrincipalComponent Analysis (PCA), can be useful to detect groups, trends,correlations, and outliers in datasets gathering vast amounts ofinformation. This method allows identifying patterns in data, andexpressing them in order to highlight their similarities and differ-ences. PCA is a projection method for analyse data and reduce itfrom an n-dimensional space to few latent/hidden variables, whilekeeping information on its variability.

Chemometric tools have been proved to be able to monitorwastewater treatment reactors [6–8]. Also, multivariate statisticalanalysis has been used together with image analysis techniques topattern recognition, such as discriminant analysis, neural networks,and decision trees [9]. The relationships between morpholog-

J.C. Costa et al. / Biochemical Engineering Journal 53 (2010) 38–43 39

Table 1Organic loading disturbances (OLDs) and toxic shock loads (TSLs) conditions.

Disturbance OLD1 OLD2 OLD3 OLD4 TSL1 TSL2 TSL3

Ethanol (gCOD/L) 5 1.5 15 15 1.5 1.5 1.5HRT (h) 8 2.5 8 8 8 8 8Toxic – – – – Detergent Detergent Solvent[ER

itr

nibma

2

2

Eesip

fii

licoteT

2

s

TL

V

OCEpV<>>SSLVTv

B

Toxic] (mg/L) – – –xposure phase (h) 72 72 72ecovery phase (d) 7 7 7

cal parameters and biomass properties in aerobic wastewaterreatment processes were also assessed by partial least squaresegression [10] and PCA [11].

The objective of this work was to apply the chemometric tech-ique PCA in order to recognize fluctuations and respective effects

n high-rate anaerobic granular sludge reactors performance causedy organic loading and toxic disturbances. The role of quantitativeorphological parameters in the potential fault detection was also

ssessed.

. Materials and methods

.1. Datasets

Four organic loading disturbances (OLDs) [12] were applied to anxpanded Granular Sludge Bed (EGSB) reactor fed with 1.5 gCOD-thanol/L and HRT of 8 h, in steady-state conditions. Also, three toxichock loads (TSL) were applied in EGSB reactors operating in sim-lar conditions [13,14]. Summary of the disturbances applied areresented in Table 1.

Three programmes previously developed [15] were used as thenal step of a procedure [13] to obtain quantitative morphological

nformation from anaerobic granular sludge.Three datasets were created gathering morphological, physio-

ogical, and reactors performance information. Datasets 1 and 2ncluded observations of OLD and TSL, respectively. The objectiveonsisted in examine the sensitivity of the latent variables to rec-gnize the disturbances. Dataset 3 encompassed all observationso study the differentiation of the OLD from the TSL, and respectiveffects. The variables used during the experiments are defined inable 2.

.2. Principal Component Analysis

PCA aims at finding and interpreting hidden complex, and pos-ibly causally determined, relationships between features in a

able 2oadings/weights associated with the first (t[1]) and second (t[2]) latent variable in organ

ariable OLD TSL

t[1] t[2] t[1]

LR −0.40 0.06 0.12sd – – −0.34ff 0.40 0.02 −0.20H 0.32 −0.12 0.00SS −0.40 −0.12 −0.050.1 −0.12 −0.29 −0.200.1 −0.35 0.19 −0.431 0.36 −0.16 0.44AA 0.33 0.13 −0.26HMA 0.04 −0.41 0.18fA 0.13 0.55 0.24SS/TA 0.16 −0.18 0.17L/VSS 0.05 0.55 0.13sed – – 0.46

old values are marked the loadings with higher influence in each score (higher than 0.30

– 1.6 3.1 40384 56 222 222

7 14 12 7

dataset. Correlating features are converted to the so-called factorswhich are themselves noncorrelated [16].

SIMCA-P (Umetrics AB) software package was used to performthe PCA. The first step of the analysis consists in the pre-treatmentof data by standardization of the variables, i.e. guarantee that eachindividual variable has about the same range, avoiding that somevariables would be more important than others because of scaleeffects. During this work each variable was autoscaled so that eachvariable has mean zero and unit standard deviation.

Subsequently, the software iteratively computes one PrincipalComponent (PC) at a time, comprising a score vector ta and a load-ing vector pa. The score vectors contain information on how thesamples relate to each other. Otherwise, the loading vectors definethe reduced dimension space and contain information on how thevariables relate to each other. Usually, few PCs (2 or 3) can expressmost of the variability in the dataset when there is a high degree ofcorrelation among data.

The criterion used to determine the model dimensionality(number of significant components) was cross-validation (CV). Partof data is kept out of the model development, and then are predictedby the model and compared with the actual values. The predictionerror sum of squares (PRESSs) is the squared differences betweenobserved and predicted values for the data kept out of the modelfitting. This procedure is repeated several times until data elementhas been kept out once and only once. Therefore, the final PRESShas contributions from all data. For every dimension, SIMCA com-putes the overall PRESS/SS, where SS is the residual sum of squaresof the previous dimension. A component is considered significantif PRESS/SS is statistically smaller than 1.0.

3. Results and discussion

3.1. Summary of the disturbances effects

Since the main objective of this study was the rapiddetection of potential problems, emphasis should be given to

ic loading disturbances (OLD) and toxic shock loads (TSL).

Name

t[2]

−0.20 Organic loading rate−0.07 Toxic (detergent or solvent) concentration

0.30 COD removal efficiency0.38 pH

−0.28 Effluent volatile suspended solids0.30 % of Aggregates projected area with Deq < 0.1 mm

−0.09 % of Aggregates projected area with 0.1 ≤ Deq (mm) < 10.07 % of Aggregates projected area with Deq ≥ 1 mm0.30 Specific acetoclastic activity

−0.04 Specific hydrogenotrophic methanogenic activity0.38 Total filaments length/total aggregates projected area

−0.33 VSS/total aggregates projected area0.44 Total filaments length/volatile suspended solids0.01 Settling velocity

).

4 ngineering Journal 53 (2010) 38–43

ca

atoeoV(cfafoco

edeewIouHd

3

ihto

OgwbSlveitHm

Fl

0 J.C. Costa et al. / Biochemical E

hanges that occurred immediately after the disturbances werepplied.

In OLD1, although the COD removal efficiency was unaffected,change in granules size distribution was observed. Consequently

he effluent volatile suspended solids (VSSs) increased after 24 hf disturbance [12]. The hydraulic shock load (OLD2) caused anfficiency decrease from 90 to 73%. Filaments release was alsobserved, with consequent increase of total filaments length perSS (TL/VSS), total filaments length to total aggregates area ratio

LfA) and VSS parameters [12]. OLD3 and OLD4 caused reactors effi-iency decreases, from 90 to 30%. In these disturbances granulesragmentation was observed, with the % of aggregates projectedrea with equivalent diameter (Deq) higher than 1 mm decrease,rom 90 to 58%. Filaments release and biomass washout was alsobserved. The exposure time increase, in OLD4, caused a signifi-ant decrease in the specific acetoclastic activity (SAA) after 140 hf disturbance [12].

Relatively to the two detergent shock loads [13], the reactorfficiency was unaffected with 1.6 mg/L (TSL1), but a significantecrease in the COD removal efficiency was observed after 80 h ofxposure to 3.1 mg/L (TSL2). Increases in LfA and in VSS param-ters were observed in both shock loads. However, the washoutas detected only 100 and 270 h after the disturbance began.

n the TSL2 the biomass activity was inhibited. The main effectf the solvent shock load (TSL3) was detected by the gran-les size distribution, because of the fragmentation phenomenon.owever, this disintegration was gradually increasing during theisturbance [14].

.2. Recognition of organic load and toxic disturbances

The PCA expressed the importance of the proposed morpholog-cal parameters to recognize, possible problematic, disturbances toigh-rate anaerobic granular sludge reactors. The latent variables

[1] and t[2] showed high percentages of variation in the first hoursf exposure in every disturbances applied (Figs. 1 and 2).

As showed in Table 1, OLD3 and OLD4 were the most severeLDs. Therefore, its higher variation in t[1] is visible in Fig. 1, sug-esting that the potential problems recognition is directly linkedith the disturbance extent/magnitude. This was essentially caused

y the granules fragmentation, biomass washout, and decrease inAA and COD removal efficiency. These variables had the higheroadings in PC1 (Table 2). However, with the objective of identify theariables with higher influence in the potential problem detection

mphasis should be given to the variable contribution plots showedn Fig. 3. This plot was used to understand why the first observa-ion, after the disturbances were applied, differs from the inoculum.ere, is confirmed that the variables associated with granules frag-entation (percentages of aggregates projected area with Deq > 1,

ig. 1. First latent variable (t[1]) evolution during organic loading disturbances: LD,oad disturbance phase; R, recovery phase.

Fig. 2. First latent variable (t[1]) evolution during TSLs: TSL, toxic shock load phase;R, recovery phase.

and in the range 1 > Deq > 0.1) were the most valuables for the distur-bances recognition, especially in OLD1, 3 and 4 (Fig. 3a–d). However,in OLD2 the variable LfA was identified as the one with higherweight in the disturbance recognition (Fig. 3b). In fact, althoughthe COD removal efficiency was unaffected, granule fragmentationwas observed in OLD1. In OLD2, while no changes in granules sizedistribution occurred, a severe filaments release was observed. Forthat reason the TL/VSS and LfA parameters had high loadings in PC2(Table 2), and the latent variable t[2] show the higher variation inOLD2 (data not shown).

Relatively to TSLs it was observed that no significant effects onreactor performance and biomass characteristics occurred in deter-gent TSL1. These were detected by the PCA, since the latent variablet[1] does not show large variation during this TSL (Fig. 2). The deter-gent TSL2 caused the biggest decrease in reactor performance, and,the solvent TSL3 caused granules fragmentation and continuousbiomass and reactor performance deterioration. Observing the t[1]evolution in Fig. 2, an immediate deviation between the inoculumobservation in TSL2 and the first observation during shock load(after 8 h) is clearly visible. This deviation was originated by theoperational problems caused by the detergent. In the same figure,a constant decrease of t[1] is visible during TSL3. In the TSLs, con-stant filaments release was observed. This morphological changewas responsible by the variability detected in PC2 (data not shown)as showed by the high loadings of TL/VSS associated with the latentvariable t[2] (Table 2).

Looking at the variable contribution for the shock load recogni-tion (Fig. 4) is visible that the morphological variables VSS per totalaggregates area (VSS/TA) and % of aggregates projected area withDeq < 0.1, had the higher weights associated to TSL1 (Fig. 4a) andTSL2 (Fig. 4b) detection, respectively. In TSL3, since the effects weregradual and no significant effects occurred during the first 24 h,it was impossible to clearly distinguish the variables with higherinfluence in the variance between inoculum and first observationafter the disturbance (Fig. 4c and d), besides the imposed variablesToxic concentration and OLR. However, looking at the weights asso-ciated to the latent variables t[1] and t[2] (Table 2) is visible thatthe morphological parameters suffer the main effects caused bythe toxics.

The high loadings/weights associated with morphological vari-ables, enhances the need/usefulness of biomass morphologymonitoring to control the reactors. It may be suggested that, from

the liquid and gaseous phase, the pH can be an important parame-ter to monitor/control the process, as reported by others [17]. Theresults show the adequacy of use the chemometric technique PCA torecognize disturbances in high-rate anaerobic reactors and detectthe respective effects.

J.C. Costa et al. / Biochemical Engineering Journal 53 (2010) 38–43 41

Fig. 3. Variable contribution for the latent variable t[1] variation, between inoculum observation and the first observation during organic loading disturbance: (a) OLD1, (b)OLD2, (c) OLD3 and (d) OLD4.

Fig. 4. Variable contribution for the latent variable t[1] variation, between inoculum observation and the first observation during toxic shock load: (a) TSL1, (b) TSL2, (c) TSL3,and (d) variable contribution for the latent variable t[2] variation in TSL3.

42 J.C. Costa et al. / Biochemical Engineering Journal 53 (2010) 38–43

l Com

3

dtohobet1i(

vTpAdodeawg

ido

4

dDewvntIpsTot

[

[

[

Fig. 5. PCA of the first Principal Component vs. the second Principa

.3. Distinguish organic load disturbances from TSLs

The two firsts PCs gathered 51.2% of the total variability inataset 3 (encompassing observations from organic loading andoxic disturbances). However, analysing the score and loading plotsf PC1–PC2 plane (Fig. 5) is possible to visualize the variables withigher influence to distinguish load disturbances from shock loadbservations and respective effects, i.e. the variables more affectedy the disturbances. The inoculum observations are located close toach other in the top-right quadrant (Fig. 5a). The OLDs displacedhe observations in direction of negative scores in PC1 and PC2 (linein Fig. 5a). Concerning to the TSL, the observations were dislocated

n the direction of positive scores in PC1 and negative scores in PC2line 2 in Fig. 5a).

Watching at the direction lines of the exposure phase obser-ations (Fig. 5a) and the loading plot (Fig. 5b), is visible that theSLs affected mostly the morphological variables filaments lengther aggregates area (LfA) and filaments length per VSS (TL/VSS).s stated before a severe filaments release was observed in theseisturbances. So, these variables were responsible for the detectionf operational problems even before the COD removal efficiencyecrease. The OLDs caused VSS increase, and pH and COD removalfficiency decreases. Severe fragmentation phenomenon’s werelso observed, as showed by the % of aggregates projected areaithin the range 0.1 < Deq < 1 mm (>0.1), and consequent % of aggre-

ates projected area with Deq > 1 mm (>1) decrease.The recovery phases observations tend to return to close the

nocula observations (Fig. 5a). Effectively, the reactors regain its pre-isturbances performance few hours after their stop, indicating thatnly temporary inhibitions occurred.

. Conclusions

The morphological parameters associated with aggregates sizeistribution, i.e. the percentages of aggregates projected area witheq < 0.1 mm (<0.1), 0.1 < Deq < 1 mm (>0.1), and Deq > 1 mm (>1),xperienced high variation in the first hours of disturbances, eitherhen an organic or a TSL occurred. Their high weights were rele-

ant for the immediate recognition of deviations occurred in theormal operation of the EGSB reactors. These were detected byhe latent variable t[1], determined by the PCA, especially in OLDs.n the cases where no fragmentation occurred, the morphological

arameters LfA (in hydraulic shock load) and VSS/TA (in detergenthock load) showed high sensitivity to the disturbance recognition.he results enhance the important role that biomass morphol-gy monitoring can have in the recognition of organic loading andoxic disturbances in anaerobic granular sludge reactors. The PCA

[

ponent: (a) score plot t[1] vs. t[2], and (b) loading plot p[1] vs. p[2].

allowed to distinguish organic loading from toxic disturbances, andthe main effects caused by each other. Increase the OLR causedgranules fragmentation and washout (increase in effluent VSS), andCOD removal efficiency decrease. Exposure to detergent and solventcaused the filaments release and consequent increase in the LfAparameter, and specific hydrogenotrophic methanogenic activitydecrease.

Acknowledgements

We grateful acknowledge the financial support to J.C.Costa through the Grant SFRH/BD/13317/2003 and the ProjectPOCTI/AMB/60141/2001 from the Fundacão para a Ciência e a Tec-nologia (Portugal).

References

[1] G. Lettinga, Anaerobic digestion and wastewater treatment systems, Antoin. vanLeeuw. Int. J. 67 (1995) 3–28.

[2] J.B. Van lier, A. Tilche, B.K. Ahring, H. Macarie, R. Moletta, M. Dohanyos, L.W.Hulshoff-Pol, P. Lens, W. Verstraete, New perspectives in anaerobic digestion,Water Sci. Technol. 43 (2001) 1–18.

[3] W. Bourgeois, J.E. Burgess, R.M. Stuetz, On-line monitoring of wastewater qual-ity: a review, J. Chem. Technol. Biot. 76 (2001) 337–348.

[4] K. Schügerl, Progress in monitoring, modeling and control of bioprocesses dur-ing the last 20 years, J. Biotechnol. 85 (2001) 149–173.

[5] H. Spanjers, J.B. van Lier, Instrumentation in anaerobic treatment—research andpractice, Water Sci. Technol. 53 (2006) 63–76.

[6] J.A. Grove, H. Kautola, S. Javadpour, M. Moo-Young, W.A. Anderson, Assessmentof changes in the microorganism community in a biofilter, Biochem. Eng. J. 18(2) (2004) 111–114.

[7] S.G. Kulkarni, A.K. Chaudhary, S. Nandi, S.S. Tambe, B.D. Kulkarni, Modelingand monitoring of batch processes using principal component analysis (PCA)assisted generalized regression neural networks (GRNN), Biochem. Eng. J. 18 (3)(2004) 193–210.

[8] S.H. Hong, M.W. Lee, D.S. Lee, J.M. Park, Monitoring of sequencing batch reactorfor nitrogen and phosphorus removal using neural networks, Biochem. Eng. J.35 (3) (2007) 365–367.

[9] Y.P. Ginoris, A.L. Amaral, A. Nicolau, M.A.Z. Coelho, E.C. Ferreira, Recog-nition of protozoa and metazoa using image analysis tools, discriminantanalysis, neural networks and decision trees, Anal. Chim. Acta 595 (2007)160–169.

10] A.L. Amaral, E.C. Ferreira, Activated sludge monitoring of a wastewater treat-ment plant using image analysis and partial least squares regression, Anal. Chim.Acta 544 (2005) 246–253.

11] R. Jenné, E.N. Banadda, G. Gins, J. Deurinck, I.Y. Smets, A.H. Geeraerd, J.F. VanImpe, Use of image analysis for sludge characterisation: studying the rela-tion between floc shape and sludge settleability, Water Sci. Technol. 54 (2006)167–174.

12] J.C. Costa, I. Moita, A.A. Abreu, E.C. Ferreira, M.M. Alves, Advanced monitoringof high rate anaerobic reactors through quantitative image analysis of granular

sludge and multivariate statistical analysis, Biotechnol. Bioeng. 102 (2) (2009)445–456.

13] J.C. Costa, A.A. Abreu, E.C. Ferreira, M.M. Alves, Quantitative image analysisas a diagnostic tool for monitoring structural changes of anaerobic gran-ular sludge during detergent shock loads, Biotechnol. Bioeng. 98 (2007)60–68.

nginee

[

[

J.C. Costa et al. / Biochemical E

14] J.C. Costa, I. Moita, E.C. Ferreira, M.M. Alves, Morphology and physiology ofanaerobic granular sludge exposed to organic solvents. J. Hazard. Mater. (inpress).

15] A.L. Amaral, Image Analysis in Biotechnological Processes: Application toWastewater Treatment. PhD thesis, University of Minho, Portugal, 2003(http://hdl.handle.net/1822/4506).

[

[

ring Journal 53 (2010) 38–43 43

16] J.W. Einax, H.W. Zwanziger, S. Geiss, Chemometrics in Environmental Analysis,VCH, Weinheim, 1997.

17] J. Liu, G. Olsson, B. Mattiasson, Monitoring and control of an anaerobic upflowfixed-bed reactor for high-loading-rate operation and rejection of disturbances,Biotechnol. Bioeng. 87 (2004) 43–53.