Advanced monitoring of high-rate anaerobic reactors through quantitative image analysis of granular sludge and multivariate statistical analysis.

Abstract

Four organic loading disturbances were performed in lab-scale EGSB reactors fed with ethanol. In load disturbance 1 (LD1) and 2 (LD2), the organic loading rate (OLR) was increased between 5 and 18.5 kg COD m(-3) day(-1), through the influent ethanol concentration increase, and the hydraulic retention time decrease from 7.8 to 2.5 h, respectively. Load disturbances 3 (LD3) and 4 (LD4) were applied by increasing the OLR to 50 kg COD m(-3) day(-1) during 3 days and 16 days, respectively. The granular sludge morphology was quantified by image analysis and was related to the reactor performance, including effluent volatile suspended solids, indicator of washout events. In general, it was observed the selective washout of filamentous forms associated to granules erosion/fragmentation and to a decrease in the specific acetoclastic activity. These phenomena induced the transitory deterioration of reactor performance in LD2, LD3, and LD4, but not in LD1. Extending the exposure time in LD4 promoted acetogenesis inhibition after 144 h. The application of Principal Components Analysis determined a latent variable that encompasses a weighted sum of performance, physiological and morphological information. This new variable was highly sensitive to reactor efficiency deterioration, enclosing variations between 27% and 268% in the first hours of disturbances. The high loadings raised by image analysis parameters, especially filaments length per aggregates area (LfA), revealed that morphological changes of granular sludge, should be considered to monitor and control load disturbances in high rate anaerobic (granular) sludge bed digesters.

Full text

Page 1

ARTICLE
Advanced Monitoring of High-Rate Anaerobic
Reactors Through Quantitative Image Analysis
of Granular Sludge and Multivariate
Statistical Analysis
J.C. Costa, I. Moita, A.A. Abreu, E.C. Ferreira, M.M. Alves
Institute for Biotechnology and Bioengineering (IBB), Centre of Biological Engineering,
University of Minho, 4710-057 Braga, Portugal; telephone: þ351-253-604-400;
fax: þ351-253-678-986; e-mail: madalena.alves@deb.uminho.pt
Received 2 May 2008; revision received 17 July 2008; accepted 21 July 2008
Published online 5 August 2008 in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/bit.22071
ABSTRACT: Four organic loading disturbances were per-
formed in lab-scale EGSB reactors fed with ethanol. In load
disturbance 1 (LD1) and 2 (LD2), the organic loading rate
(OLR)wasincreasedbetween5and18.5kgCODm?3day?1,
through the influent ethanol concentration increase, and the
hydraulic retention time decrease from 7.8 to 2.5 h, respec-
tively. Load disturbances 3 (LD3) and 4 (LD4) were applied
by increasing the OLR to 50 kgCODm?3day?1during
3 days and 16 days, respectively. The granular sludge mor-
phology was quantified by image analysis and was related to
the reactor performance, including effluent volatile sus-
pended solids, indicator of washout events. In general, it
was observed the selective washout of filamentous forms
associated to granules erosion/fragmentation and to a de-
crease in the specific acetoclastic activity. These phenomena
induced the transitory deterioration of reactor performance
in LD2, LD3, and LD4, but not in LD1. Extending the
exposure time in LD4 promoted acetogenesis inhibition
after 144 h. The application of Principal Components
Analysis determined a latent variable that encompasses a
weighted sum of performance, physiological and morpho-
logicalinformation.Thisnewvariablewashighlysensitiveto
reactorefficiency deterioration,
between 27% and 268% in the first hours of disturbances.
The high loadings raised by image analysis parameters,
especially filaments length per aggregates area (LfA), reveal-
ed that morphological changes of granular sludge, should be
considered to monitor and control load disturbances in high
rate anaerobic (granular) sludge bed digesters.
Biotechnol. Bioeng. 2009;102: 445–456.
? 2008 Wiley Periodicals, Inc.
KEYWORDS: anaerobic granular sludge; quantitative image
analysis; organic loading disturbances; principal compo-
nent analysis
enclosingvariations
Introduction
Anaerobic granules are particulate biofilms,
spontaneously by self-immobilization of anaerobic bacteria
in the absence of a support material (Lettinga, 1995). Hence,
each granule is a functional unit comprising of all the
different microorganisms necessary for methanogenic
degradation of organic matter (Sekiguchi et al., 1998).
Over the last three decades, numerous works, theories and
models have been developed to explain and/or stimulate the
anaerobic granulation process (Liu et al., 2003).
Biological wastewater treatment plants are normally
designed with reference to a nominal operating condition,
in which the loading rate is assumed to be constant in time.
However, in practice this steady-state assumption is seldom
met and fluctuations, both in flow and influent concentra-
tion occur, which often result in performance degradation
or even digester failure.
Anaerobic digesters run frequently at organic loading
rates (OLR) below the maximum process capacity. Stable
operation of high-rate anaerobic processes is in practice, a
difficult task. Due to the slow growth of syntrophic and
methanogenic bacteria, anaerobic digestion is quite often
unstable during influent fluctuations with volatile fatty
acids accumulation and pH decrease, leading sometimes to
process failure (Voolapalli and Stuckey, 1998). Depending
on the monitoring plan and operator’s expertise, the
influent disturbances can exhibit different configurations:
either as a short/transient overload, which only lasts few
hours, or as a longer step change of days to weeks.
Parameters involved in reactor monitoring and control,
have been limited to indicators of the liquid and the gaseous
phases and have mostly been limited to the assessment of
operational performance measurements. The impact on
microbial community structure has rarely been investigated
(McMahon et al., 2004), although the combination of
physical and biological methods for process monitoring
formed
Correspondence to: M.M. Alves
Contract grant sponsor: Fundac ¸a ˜o para a Cie ˆncia e a Tecnologia (Portugal)
Contract grant number: SFRH/BD/13317/2003; POCI/AMB/60141/2004; POCTI/BIO/
37934/2001
? 2008 Wiley Periodicals, Inc.
Biotechnology and Bioengineering, Vol. 102, No. 2, February 1, 2009 445

Page 2

and control, has been proposed as a way to more effectively
minimize problems of process imbalance in anaerobic
digesters during load disturbance events (Voolapalli and
Stuckey, 1998). There are evidences that microbial and
structural changes of biological aggregates can be detected
before the detection of significant process performance
deterioration after a disturbance (Abreu et al., 2007; Amaral
et al., 2004).
The link between digester performance, and the physiolo-
gical and structural characteristics of the anaerobic sludge is
particularly relevant in up-flow anaerobic sludge blanket
(UASB), expanded granular sludge bed (EGSB), or internal
circulation (IC) reactors (McMahon et al., 2004). Notwith-
standing, limited knowledge about physiological and mor-
phological changes of anaerobic granular sludge under
unsteadystateoperatingconditionsisavailableintheliterature.
The use of quantitative image analysis techniques for
monitoring wastewater treatment processes has been
initiated by Sezgin et al. (1978), who found a well-defined
relationship between the sludge volume index (SVI) and the
filament quantity in activated sludge. Recently, Li et al.
(2007) compared several characteristics of activated sludge
flocs and granules by using different techniques including
image analysis. The statistical tool principal component
analysis(PCA)wasusedtorelateimageanalysisinformation
with the sludge settleability in terms of SVI (Jenne ´ et al.,
2006). The first attempts to use digital image analysis in
high-rate anaerobic digestion processes were limited to size
measurements and number counting (Dudley et al., 1993;
Jeison and Chamy, 1998). Several subsequent works
demonstrated the application of quantitative image analysis
tools to monitor granulation and granules deterioration
processes (Amaral et al., 2004; Araya-Kroff et al., 2004), to
evaluate the effect of detergents in the structure and
microbialactivityofgranularsludge(Costaetal.,2007)orin
thedetectionofstructuralandmicrobialchangesofgranular
sludge during a process of acetoclastic activity recovery
(Abreu et al., 2007).
Because the experimental approach of integrating
reactor performance, physiological and morphological data
may produce correlated and redundant data, a statistical
instrumentshouldbeappliedinordertoextracttheessential
information for process monitoring and fault detection
applications. Data reduction and interpretation can be
approachedthroughtheapplication
statistical methods, such as PCA (Wise and Gallagher,
1996). This method allows identifying patterns in data, and
expressing them in order to highlight their similarities and
differences. PCA is a projection method for analyze data and
reduce it from an n-dimensional space to few latent/hidden
variables (Lee et al., 2006), while keeping information on its
variability.Ithasbeensuccessfullyappliedtothemonitoring
of industrial processes (Li et al., 2000) and wastewater
treatment processes (Lee et al., 2004).
This work aims at studying the effects of organic loading
disturbances on EGSB reactors performance, and on
anaerobic granular sludge morphology and physiology
of multivariate
using quantitative image analysis techniques and specific
methanogenic activity measurements. Then, multivariate
statistical tools were applied to integrate the operational,
physiological, and morphological information in order to
achieve a timely monitoring of the process, with early
recognition of operational problems and recovery states.
Material and Methods
Experimental Set-Up
Four Plexiglas EGSB reactors (Fig. 1), 1.95 m height and
21 mm internal diameter were used in the experiment. The
working volume was 1.15 L and the superficial velocity was
4.0mh?1.Temperaturewaskeptat 37?18Cbymeansofan
external jacket for water circulation.
Inoculum and Substrate
Four hundred milliliters of granular sludge from a lab-scale
EGSB reactor, treating a synthetic effluent with ethanol as
sole organic carbon source, was used as the inoculum of the
four EGSB reactors used in these experiments. The biomass
was characterized in terms of specific methanogenic activity
(SMA) with acetate, and H2/CO2as substrates, morphology
by quantitative image analysis, and volatile suspended solids
(VSS) (Table I). The reactors were fed with ethanol at a
Figure 1.
Expanded granular sludge bed reactor.
446
Biotechnology and Bioengineering, Vol. 102, No. 2, February 1, 2009

Page 3

concentration of 1.5 gCODL?1. Sodium bicarbonate was
added as the alkalinity source (3 gL?1) and macro- and
micronutrients were added according to Zehnder et al.
(1980). When the reactors were operating in steady-state,
four organic load disturbances were applied (Table I). The
organic overload was performed by increasing the substrate
concentration (LD1, LD3, and LD4), and by decreasing
the HRT (LD2).
Routine Analysis
COD and VSS were determined according to Standard
Methods (APHA, 1989). Biogas flow rate was measured by a
Ritter Milligascounter (Dr. Ing. Ritter Apparatebau GmbH,
Bochum, Germany). Methane content of biogas was
determined by gas chromatography using a Porapack Q
(100–180mesh)column,withHeliumasthecarriergasat30
mLmin?1and thermal conductivity detector. Temperatures
of the detector, injector and oven were 110, 110, and 358C,
respectively. Volatile fatty acids (VFA) and ethanol were
determined by high performance liquid chromatography
using an HPLC (Jasco, Japan) with a Chrompack column
(6.5?30 mm2); sulfuric acid (0.01 N) at a flow rate of
0.7 mLmin?1
was used as mobile phase. Column
temperature was set at 608C. Detection of VFA and ethanol
wasmadesequentiallywithanUVdetectorat210nmandan
RI detector, respectively.
Sludge Sampling and Dilution
A sampler specially designed to minimize the turbulence
inside the reactor was used to remove sludge samples. It
consisted of a wide bore tube connected to a 100 mL syringe.
It was introduced at the top of the reactor and biomass was
collected to the tube, along the expanded granular sludge
bed, avoiding mechanical stress. All the sludge samples were
characterized by image analysis, SMA assays, and VSS
content.
Biomass samples must be diluted for image analysis, but
anoptimalvalueshouldbefound.Whentherearenoobjects
in the image due to an excessive dilution, that image should
be registered and considered for calculations. However the
observer may unconsciously search objects, over estimating
them, and, if the dilution is insufficient, the objects will be
overlaid. The optimal dilution value was determined as the
lowest dilution that enabled the maximum percentage of
objects to be recognized. The percentage of recognition is
the ratio between the area of objects that are completely
inside the image and the total area of objects in the image
including those that are cut-off by the borders of the images
and cannot be completely recognized. The dilution 1:5 was
determined as the optimal one.
Specific Methanogenic Activity Assays
The specific acetoclastic activity (SAA) was measured in
presence of acetate (30 mM) and the specific hydrogeno-
trophic methanogenic activity (SHMA) was measured in the
presence of H2/CO280:20 (v/v), at 1 bar, according to the
previously reported pressure transducer method (Alves
et al., 2001). No nutrients were added. Methane was
measured by gas chromatography with helium as the carrier
gas and a TCD detector. All assays were performed in
triplicate.
Image Acquisition, Processing, and Analysis
For the acquisition of filaments and micro-aggregates
(Deq<0.2 mm) images, a volume of 35 mL from the diluted
sample was distributed on a slide and covered with a
20 mm?20 mm cover slip for visualization and image. This
volume was exactly covered by the cover slip. Each image
corresponded to a volume of 0.0445 mL. Then, more than
120 images were acquired. Image acquisition was obtained
by dividing the cover slip into 42 identicalfields and taking a
Table I.
Inocula characterization and load disturbance conditions.
LD1 LD2LD3LD4
Inocula characterization
Specific methanogenic activity with individual substrates (mLCH4@STPgVSS?1day?1)
Acetate234?15
H2/CO2
1,427?26
Morphology
LfA (mm?1) 20
TL/VSS (mgVSS?1) 1,099
VSS/TA (gm?2)18.4
Deqmacroflocs (mm) 1.06?0.58
VSS (gL?1) 14.78
Disturbance load conditions
Ethanol (gCODL?1)5
HRT (h)8
OLR (kgCODm?3day?1)15
Exposure time (h) 72
Recovery phase (h) 163
367?20
1,686?52
234?15
1,427?26
328?12
1,720?44
18 2018
1,238
14.5
1,099
18.4
1,238
14.5
0.91?0.67
33.57
1.06?0.58
14.78
0.91?0.67
33.57
1.6
2.5
15
72
160
15
8
50
72
163
15
8
50
384
168
Costa et al.: Monitoring of High Rate Anaerobic Reactors
447
Biotechnology and Bioengineering

Page 4

photo in each square. At least three slides were examined to
minimize sampling errors.
Concerning to macro-aggregates (Deq?0.2 mm) images,
an arbitrary volume was transferred to a Petri dish for
visualization and image acquisition. All the aggregates
present in that volume were captured in more than
120 images. Then, the VSS content in the Petri dish was
measured.
Images used to quantify filaments and micro-aggregates
were acquired through phase contrast and bright field,
respectively, on a Nikon Diaphot 300 microscope (Nikon
Corporation, Tokyo, Japan) with 100? magnification.
Images used to quantify macro-aggregates were acquired
through visualization on a Olympus SZ 40 stereo micro-
scope (Olympus, Tokyo, Japan) with 15? magnification. All
the images were digitized and saved with the help of a CCD
AVC D5CE Sony gray scale video camera (Sony, Tokyo,
Japan) and a DT 3155 Data Translation frame grabber (Data
Translation, Marlboro, MA) with 768?576 pixel size in 8
bits (256 gray levels) by Image Pro Plus (Media Cybernetics,
Silver Spring, MD) software package.
The metric unit dimensions were further calibrated to
pixel using a micrometer slide, for each magnification.
Image processing and analysis was accomplished by
means of three programs developed in Matlab (The
Mathworks, Inc., Natick, MA), for filaments, micro, and
macro-aggregates (Amaral, 2003). In the next paragraphs a
brief description of the programs are presented.
The first step of all programs consists in divide the gray-
scale image by the background image in order to remove
background light differences.
Filaments Program
A bottom hat filter (Russ, 1995) was applied to enhance
the filaments and small aggregates that have low gray levels.
The larger aggregates, which have high gray levels, were
subsequently identified on the image resulting from the
background elimination step by applying a 10-order closing
(to enhance the aggregates), a segmentation at a fixed
threshold value, a filling of the resulting binary image (to
remove the inner holes in the aggregates) and a erosion-
reconstruction step to eliminate the debris. Filaments and
small aggregates were then isolated by segmentation at a
fixed threshold and by logic subtraction of the mask binary
image containing the large aggregates. Then, the small
aggregates were eliminated by deleting all the objects smaller
than 32 pixels (in area) or with a gyration radius below
1.2 (Pons and Vivier, 1999). The final image contained only
filamentsthatwerecharacterizedintermsoftheirlengthand
number.
Micro- and Macro-Aggregates Programs
Firstly, histogram equalization was performed to enhance
the micro-aggregates (this step only occurs in the micro-
aggregates program). Afterwards, the image was smoothed
by a Wiener filtering (Glasbey and Horgan, 1995). Then, it
segmented in black (background) and white (aggregates) by
the simultaneous use of a boundary based segmentation
and a user chosen or automatically determined threshold
segmentation. The objects smaller than 3?3 pixels
(small debris) were then removed and small gaps (6?
6 pixels or less) were filled on the remaining objects.
Subsequently, in order to remove filaments, all the objects
smaller than 2,000 pixels in area and with a gyration radius
above 1.2 (Pons and Vivier, 1999) were deleted (micro-
aggregates program). Finally, all the objects cut off by the
image boundaries were removed, and the morphological
characterization of the micro- and macro-aggregates were
performed.
Morphological parameters. The filaments image was
skeletonized and pruned (Russ, 1995). The filaments length
was determined by:
L ¼ N ? 1:122 ? Fcal
(1)
where N is the number of pixels of the skeletonized filament
and Fcalis the calibration factor (mmpixel?1). The factor
1.122 is used in order to homogenize the different angles of
the filaments (Walsby and Avery, 1996).
The specific total filament length was calculated as:
Lspec¼
L
Vfield
(2)
where Vfieldis the volume (in mm3) corresponding to the
field of view (i.e., the image).
Filaments are not only the dispersed bulk filaments, but
include also those that are attached to an aggregate and still
have one free extremity (protruding filaments).
The aggregates program is used to determine:
The aggregate Area from which the Equivalent Diameter
(Deq) is calculated by:
Deq¼ 2Fcal
ffiffiffiffiffiffiffiffiffiffi
p
Area
r
(3)
And, the total area (AT) occupied by aggregates in each
image from which the specific area occupied by aggregates
(Aspec) is calculated:
Aspec¼
AT
Vfield
(4)
Finally, morphological parameters representing the
dynamic evolution of filaments and aggregates inside the
reactor were calculated as:
TL
VSS¼Lspec
VSS
(5)
VSS
TA¼
VSS
Aspecð<0:2mmÞþ Aspecð?0:2mmÞ
(6)
448
Biotechnology and Bioengineering, Vol. 102, No. 2, February 1, 2009

Page 5

where VSS are the volatile suspended solids present in each
sample, and Aspec(<0.2 mm)and Aspec(?0.2 mm)are the specific
aggregate area ratio for aggregates of equivalent diameter
<0.2 and ?0.2 mm, respectively.
A morphological parameter based on the ratio of specific
totalfilamentlength(Lspec)tototalaggregatesprojectedarea
(LfA) is determined by:
LfA ¼
Lspec
Aspecð<0:2mmÞþ Aspecð?0:2mmÞ
(7)
Principal Components Analysis
Variables summarizing the morphological, physiological
and performance data obtained during load disturbances
were grouped to create the dataset used to perform the PCA.
Four datasets were created, one for each load disturbance.
Principal components analysis aims at finding and
interpreting hidden complex, and possibly causally deter-
mined, relationships between features in a data set.
Correlating features are converted to the so-called factors
which are themselves noncorrelated (Einax et al., 1997).
PCA modeling, that is, the approximation of a matrix by a
model, defined by variables and a relatively small number of
outer vector products, shows the correlation structure of a
data matrix X, approximating it by a matrix product of
lower dimension (T?P0), called the principal components
(PC), plus a matrix of residuals (E):
X ¼ 1 ? x0þ T ? P0þ E(8)
where the term 1 ? x0represents the variable averages. The
second term, the matrix product T?P0, models the
structure, and the third term, E, is a matrix of residuals,
containing the deviations between the original values and
the projections, that is, contains the noise. T is a matrix of
scores that summarizes the X-variables (Scores), and P is a
matrix of loadings showing the influence of the variables on
each score. Geometrically, it corresponds to fitting a line, a
plane or a hyper plane to the data in the multidimensional
space, with the variables as axes. The scaling of the variables
specifies the length of the axes of this space.
SIMCA-P (Umetrics AB) software package was used to
perform the Principal Components Analysis. The first step
of the analysis consists in the pre-treatment of data by
standardization of the variables, that is, guarantee that each
individual variable has about the same range, avoiding that
some variables would be more important than others
because of scale effects. During this work each variable was
autoscaled by:
Zij¼xij? xj
sj
(9)
where xijis the value of the variable j in the sample i, xjand sj
are the mean and the standard deviation of the variable j,
respectively, and, zijis the autoscaled value of xij. At the end
of this standardization,each variable has mean zero and unit
standard deviation.
Subsequently, SIMCA-P (Umetrics AB) iteratively com-
putes one principal componentata time, comprisingascore
vector taand a loading vector pa. The score vectors contain
information on how the samples relate to each other.
Otherwise,theloadingvectorsdefinethereduceddimension
space and contain informationon how thevariablesrelate to
each other. Usually, few principal components (2 or 3) can
express most of the variability in the dataset when there is a
high degree of correlation among data.
Thecriterionused todeterminethemodel dimensionality
(number of significant components) is cross validation
(CV).Partofdataiskeptoutofthemodeldevelopment,and
then are predicted by the model and compared with the
actual values. The prediction error sum of squares (PRESS)
is the squared differences between observed and predicted
values for the data kept out of the model fitting. This
procedure is repeated several times until data element has
been kept out once and only once. Therefore, the final
PRESShascontributions fromalldata. Foreverydimension,
SIMCA computes the overall PRESS/SS, where SS is the
residual sum of squares of the previous dimension. A
component is considered significant if PRESS/SS is
statistically smaller than 1.0.
Results and Discussion
Reactor Performance
Four load disturbances (LD) were applied when the EGSB
reactors were operating in steady state with organic loading
rates (OLR) of roughly 5 kgCODm?3day?1and hydraulic
retention time (HRT) of 8 h. The COD removal efficiencies
were in the range 80–90% (Fig. 2). In LD1 and LD2 the
organic loading rates (OLR) were increased to 18.5
kgCODm?3day?1, through the substrate (ethanol) con-
centration increase and the hydraulic retention time (HRT)
decrease to 2.5 h, respectively. The LD3 and LD4 were
performed by increasing the substrate concentration (OLR
of50kgCODm?3day?1)during 3and16days,respectively.
The load disturbance LD1 did not cause deterioration in
the reactor performance. After a temporary decrease, the
COD removal efficiency increased to more than 95%,
suggesting that the reactor was operating, in the pre-
disturbance conditions, far from its maximum capacity. In
contrast, when the flow rate increased in LD2, the COD
removal efficiency decreased and stabilized around 73%,
until the end of the disturbance. This led to a temporary
accumulation of acetate in the effluent that peaked to values
of 300 mgCODL?1(data not shown).
When the OLR increased to 50 kgCODm?3day?1(LD3
and LD4), the COD removal efficiency decreased linearly
Costa et al.: Monitoring of High Rate Anaerobic Reactors
449
Biotechnology and Bioengineering

Page 6

from 90% to 30% (Fig. 2c) in the first 40 h of disturbance.
After this period, and extending the exposure time (LD4,
Fig. 2d), theCOD removalefficiency wasrecurrently around
35%. Few hours after the end of the shock exposure period,
all the reactors regained the pre-disturbance efficiencies
(Fig. 2).
Accumulation of acetate was observed in LD3 and LD 4,
achievingatemporarypeakof1,400and1,000mgCODL?1,
respectively. A fast decrease of pH from 7.5 to <6 was
observed in the first hours of disturbances (data not shown).
In LD3, the pH remained near 6, during all the period of
shock exposure, whereas in LD4, a gradual increase in the
pH from 5.8 to 7 was observed in the last 150 h of exposure.
The pH increase was simultaneous with a decrease in acetate
concentration and an increase in the ethanol concentration
up to 5,000 mgCODL?1in LD3 and 9,000 mgCODL?1
in LD4, suggesting an inhibition of acetogenic reaction
between ethanol and acetate.
Morphological, Structural and Physiological Data
Quantitative morphological information retrieved by image
analysis techniques can be related with the operational and
physiological informationinordertointerpret theimpact of
load disturbances in the structure of granular sludge and
consequently its impact in the sludge bed stability.
In the organic shocks imposed by increasing the influent
concentration (LD1,LD3,andLD4)itwasclearly observed a
decrease in aggregates size. The percentage of aggregates
projected area with equivalent diameter (Deq)>1 mm
decreased to 70% in LD1, and to 58% in LD3 and LD4.
Concurrently, the % of projected area of aggregates inside
the range of Deq0.1–1 mm increased to 28% and 40%,
respectively, in LD1 and LD3/4 (Fig. 2a, c, and d). A higher
flow of biogas production through the granular structure,
induced by the increase in the substrate concentration, is
likely responsible for the observed fragmentation, being
expectable that the respective impact could be dependent on
the magnitude of the load disturbance. In the hydraulic
shock LD2, a negligible change in size distribution was
observed (Fig. 3b). However, in this experiment, an increase
in the total filaments length per volatile suspended solids
(TL/VSS) was observed immediately (8 h) after starting the
disturbance period (Fig. 3j), likely induced by the increasing
shear stress inside the reactor, due to the higher up-flow
velocity applied. In LD1 a negligible increase of free/
protruding filaments was observed (Fig. 3i), and in LD3/4 a
peakoffilamentsoccurredonly24hafterincreasing theload
(Fig. 3k and l).
It is interesting to observe that, although the granules
were fragmented, LD1 disturbance did not impact the
reactor performance, but LD2 affected the reactor efficiency,
likely due to the erosion of filamentous bacteria from
the granules, followed by detachment and washout. There-
fore, by evaluating reactor’s performance (Fig. 2a and b), it
can be concluded that stability in the granules size
distribution is of minor importance when compared to
the capacity of filaments retention in the granular microbial
structures.
Figure 2.
Time course of organic loading rate (OLR, —), and COD removal Efficiency (~). a: LD1, (b) LD2, (c) LD3, (d) LD4.
450
Biotechnology and Bioengineering, Vol. 102, No. 2, February 1, 2009

Page 7

Reactor performance deterioration was also subsequent
to a high release of filaments in LD3 and LD4. Although
granular sludge is a complex structure, it is accepted that, in
general, acetoclastic bacteria are located mainly in the
core of a stratifiedmicrobial structure (Harmsen et al., 1996;
Liu et al., 2002a). This fact allows them to be more
protected against operational problems such as overloads,
pH oscillations or presence of toxic compounds (Liu
Figure 3.
0.1 mm?Deq<1 mm;
(effluent VSS, &); and (i–l) total filament length per volatile suspended solids (TL/VSS, ^), and volatile suspended solids per total aggregates projected area (VSS/TA, &). a, e, i:
LD1; (b, f, j) LD2; (c, g, k) LD3; and (d, h, l) LD4.
Time course of morphological parameters: (a–d) percentage of aggregates projected area by equivalent diameter (Deq) ranges (D, Deq?1 mm; *,
, Deq<0.1mm); (e–h) dynamic between total filament length and total aggregates projected area (LfA, *), and, effluent volatile suspended solids
Costa et al.: Monitoring of High Rate Anaerobic Reactors
451
Biotechnology and Bioengineering

Page 8

et al., 2002b). In result of the observed fragmentation, the
inner core of the granules becomes exposed to the bulk
conditions, promoting the release and the selective washout
of methanogenic bacteria. Because filamentous forms
increased in the bulk, in LD2, LD3, and LD4 (Fig. 3j, k,
and l), it is worth to suggest that filamentous forms could
have been selectively washed-out during those load
disturbances, causing a decrease in the COD removal
efficiency (Fig. 2b–d). The observed decrease in the specific
acetoclastic activity (SAA), especially during LD3 and LD4
(Fig. 4), the increase of effluent VSS (Fig. 3f, g, and h) and
the kinetic characteristics of the Methanosaeta genera, which
is a very slow growing archaea may allow to speculate that
Methanosaeta genera could have been selectively washed out
during LD2, LD3, and LD4. Apparently, extending the
exposure time in LD4, the deterioration of reactor
performance was persistent and simultaneous with the
increaseineffluentVSSconcentration.Thiswaslikelydueto
filaments release, detachment and washout, inducing
minimalvalues,bothforfilaments(Fig.3l)andSAA(Fig.4).
It was observed in different previous works (Abreu
et al., 2007; Araya-Kroff et al., 2004) that aggregation and
densification of aggregates was simultaneous with an
increase in the SHMA. In this respect, it is remarkable
that, particularly during the recovery phase, the SHMA
(Fig. 4) and the VSS/TA (Fig. 3) presented similar trends in
LD4.
The LfA parameter quantifies the level of biomass
aggregation by measuring the ratio between the free
filaments length and the projected area of the aggregates.
A high correlation with the total filaments length per VSS
was observed. Amaral (2003) and Costa et al. (2007)
proposed that LfA could be an early alert of washout during
influent disturbances in EGSB reactors, particularly when
toxic substances were accidentally fed. In the present study,
however the effects on the morphology caused by the over-
loads were simultaneous with the increase in the effluent
VSS. Therefore, the LfA parameter followed the trend of
effluent VSS without advance (Fig. 3e–h), forbidding an
earlier detection of the washout phenomenon.
By the end of the exposure periods, the turbulence and
instabilitycausedbytheraiseinbiogasproductioninsidethe
reactors, decreased. The effluent VSS decreased to values
similar to the pre-disturbance conditions (Fig. 3e–h).
During the recovery phase the morphological parameters
stabilized in values close to the initial ones, with the
exception of aggregates size distribution (Fig. 3).
Principal Components Analysis
Theextractionofthreeprincipalcomponentsgatheredmore
than 80% of the total variability in all four datasets. The use
of other PCs did not enhance significantly the correlation
factors.
The score maps (Fig. 5) can be seen as windows in the X
space, displaying the observations as situated on the
projection planes or hyperplanes, and revealing groups,
trends, outliers and similarities between samples. Two
clusters clearly isolate the observations during the load
disturbances and the recovery period, besides the inoculum
sample, which emerge asanisolated observation. Thisshows
that immediately after the load disturbance, a deviation
occurred.
The information that can be obtained about the variables
and the role they play in the principal component (PC),
resides in the so-called loadings and respective loading maps
(Fig. 6). It allows deciding which variables are most
Figure 4.
Time course of specific acetoclastic activity (SAA, ^), and specific hydrogenotrophic methanogenic activity (SHMA, &). a: LD1, (b) LD2, (c) LD3, (d) LD4.
452
Biotechnology and Bioengineering, Vol. 102, No. 2, February 1, 2009

Page 9

important for the differences observed between the samples.
The loading plots can be very useful to detect correlations
between variables. Variables close to each other, that is, with
similar coordinates, represent strongly correlated variables.
Concurrently, variables with symmetric coordinates are
correlated,althoughinverselyproportional.Combinationof
score and loading maps (Figs. 5 and 6) allows for the
visualization of the main effects/problems occurred during
EGSB reactorsoperation. For instance, the inoculum sample
is located in the upper part of the plot, with high score in
PC2 (Fig. 5b,t[2]).The variablesthat most influence thisPC
are >1, >0.1, SHMA, Eff, and VSS (Fig. 6b, p[2]) that
evidenced immediate changes when the disturbance was
applied. The main clusters refer to observations with
negative (disturbance) and positive (recovery) scores in PC1
(Fig. 5b, t[1]). The variables with higher loadings in PC1
(Fig. 6b, p[1]), that is, responsible for grouping disturbance
and recovery phase observations are HRT, pH, SAA, OLR,
LfA, and TL/VSS. The LfA showed the highest variation
caused by the decrease of HRT, that is, the highest loading
(Fig. 6b, p[1]).
Some correlations should be distinguished from the
loading maps. It was observed a high positive correlation
between LfA and TL/VSS parameters in all datasets.
Previously some hypotheses were postulated, regarding
correlations between reactor performance deterioration,
biomass activity, and erosion/fragmentation of granules and
consequent filaments release and washout.
Emphasis should be given to the correlation showed in
LD3 (Fig. 6c) between the apparent granules density (VSS/
TA) and the specific hydrogenotrophic methanogenic
activity (SHMA), supporting the hypothesis presented
earlier, where densification of aggregates was simultaneous
with an increase in the SHMA (Abreu et al., 2007).
Special care should be put in the analysis of the loading
plot of LD4 (Fig. 6d), because extending the period of
disturbance,inducedanew dynamic offilamentsreleaseand
washout. Consequently, the correlations detected were
poorer in this case. To elucidate this new dynamic, in
Figure 6d, it is observed a negative correlation between the
efficiency (Eff) and the % of aggregates with equivalent
diameter within the range 0.1–1 mm (>0.1). The increase of
smaller aggregates suggests the occurrence of erosion and/or
granules fragmentation. This phenomenon could have
destroyed the granules structure and architecture, likely,
exposing the inner core of acetoclastic bacteria to the bulk
conditions. The decrease in the removal efficiency was
probably associated to the selective washout of those
bacteria, evidenced by the negative correlation between the
SAA and the effluent VSS (Fig. 6d).
The increasing number of reports with methods for
monitoring and control of anaerobic digestion processes
Figure 5.
this article, available at www.interscience.wiley.com.]
PCA score plot of the first PC (t[1]) versus the second PC (t[2]), in dataset of: (a) LD1; (b) LD2; (c) LD3; and (d) LD4. [Color figure can be seen in the online version of
Costa et al.: Monitoring of High Rate Anaerobic Reactors
453
Biotechnology and Bioengineering

Page 10

does not consider the integration of quantitative morpho-
logical indicators. In the present work, the application of
PCA to datasets obtained during load disturbances, allowed
theestablishmentofanewlatentvariablethatencompassesa
weighted sum of performance, physiological and morpho-
logical information. This new variable can be used as a
warning indicator of operational problems during load
disturbances as can be clearly observed in Table II—
recognition of load disturbance. The variable t[1] was
calculated in the first hours of operation in the load
disturbances and the corresponding % of variation in t[1]
ranged between 27% and 268%, evidencing the high
sensitivity of the latent variable to detect the disturbance.
In LD1, the percentage of variation after 8 h was the smallest
one (27%) because this was the softer shock applied.
Simultaneously, was observed that this variable, t[1], was
also sensible to detect the recovery of the reactors
(Table II—recognition of recovery state). This was possible
due to the rapid recovery in reactor performance and
morphologicalchangesthatoccurredfewhoursaftertheend
of the load disturbance.
For a better understanding of principal components
analysis and the factors involved, is necessary to remember
that the score (ti) of an observation (i) on a principal
component PCj(tiPCj) is the weighted sum of the original
variables (xi). The weights (piPCj) are called the loadings of
the variables on that PCj. The loading of a variable is related
to its variation (Massart and Vander Heyden, 2005)
tiPCj¼
X
piPCjxi
(10)
Analyzing the loadings associated with the new latent
variable t[1] described previously it is possible to distinguish
Figure 6.
online version of this article, available at www.interscience.wiley.com.]
PCA loading plot of the first and second principal components (p[1] vs. p[2]), from dataset of: (a) LD1; (b) LD2; (c) LD3; and (d) LD4. [Color figure can be seen in the
Table II.
at the end of load disturbance and in the beginning of recovery phase (recognition of recovery state).
Coefficients of the first latent variable t[1] in PC1, at the beginning and after few hours of load disturbance (recognition of load disturbance), and,
LD
Recognition of load disturbance Recognition of recovery state
t¼0 h
?1.343
þ1.482
þ2.132
?3.701
t¼8 h
?0.976
?2.493
?2.364
—
t¼24 h
—
—
—
0.817
% of variationt¼72 h
?2.119
?1.069
?0.875
—
t¼96 h
3.938
2.229
1.344
—
t¼384 h
—
—
—
3.329
t¼417 h
—
—
—
?1.386
% of variation
1
2
3
4
27
268
211
122
286
308
254
142
454
Biotechnology and Bioengineering, Vol. 102, No. 2, February 1, 2009

Page 11

the variables that mostly influence the early detection of
reactor. The morphological parameters, with emphasis to
the LfA parameter that relates the free filaments length per
aggregates projected area emerges due to its high loadings,
that is, weights (absolute value) in all datasets (Table III).
The results clearly suggest that quantitative morphological
parameters should be considered in monitoring high rate
anaerobic reactors, especially those based on granular
sludge. It is important to mention, in this respect, that
granular sludge is a highly valuable resource used to start-up
industrial UASB reactors, EGSB reactors, or IC reactors. If
available, cost of granular sludge can attain about 300 Euro
per ton, including 2,000 km transportation. Therefore,
physiological quality but also morphological stability should
be assessed when operating in a different reactor or in
different conditions.
The integration of morphological information quantified
by image analysis, reactor performance, specific activity, and
chemometric methods constitutes a package for advanced
monitoringofanaerobic granularsludgebedreactors,which
in being tested at industrial scale with promising results.
Conclusions
The effects of organic load disturbances on the operational,
physiological and morphological properties of anaerobic
granular sludge were studied by combining quantitative
image analysis, methanogenic activity assays and perfor-
mance data.
An increase of OLR from 5 to 18.5 kgCODm?3day?1
caused by the increase in influent concentration did not
affect the reactor efficiency. However, when the same
increase was due to the HRT decrease or the OLR was
increased to 50 kgCODm?3day?1, the COD removal
efficiencydecreasedto72%and30%,respectively.Inthelast
case, increasing the exposure time caused acetogenic
inhibition. Concomitantly, granules fragmentation/erosion
was observed in organic shock loads (LD1, LD3, LD4).
However, in hydraulic shock (LD2), a negligible change in
size distribution was observed although an increase in
filaments was observed. Therefore, by evaluating reactor’s
performance, it can be concluded that stability in the
granules size distribution is of minor importance when
compared to the capacity of filaments retention in the
granular microbial structures.
It was demonstrated that the use of a multivariate
statistical tool such as PCA was appropriate to visualize and
isolate the main effects caused by the transient instabilities.
The proposed morphological parameters proved to be more
sensitive to detect promptly the problems than the normal
operating parameters. The new latent variable t[1], defined
as an weighted sum of all variables included in the dataset,
showed high variations in the first hours of disturbance and
in the recognition of the recovery state, in all datasets.
Concurrently, the high loadings/weights of the morpholo-
gical parameters enhanced the need to monitor the
anaerobic digestion process solid phase in order to achieve
an effective and timely control of the process.
Nomenclature
Reactor Performance Data
OLRorganic loading rate
HRThydraulic retention time (only used in LD2)
Eff chemical oxygen demand removal efficiency
pH pH
VSSeffluent volatile suspended solids
Physiological Data
SAAspecific acetoclastic activity
SHMAspecific hydrogenotrophic methanogenic activity
Morphological Data
LfAtotal filaments length per total aggregates projected area
TL/VSS total filaments length per volatile suspended solids
VSS/TAvolatile suspended solids per total aggregates projected area
(apparent granules density)
>1percentage of aggregates projected area with equivalent diameter
(Deq)?1 mm
percentage of aggregates projected area within the range
0.1?Deq(mm)<1 mm
percentage of aggregates projected area with Deq<0.1 mm
>0.1
<0.1
We thank the financial support to J.C. Costa, I. Moita, and A. Abreu
through the grant SFRH/BD/13317/2003 and the projects POCI/
AMB/60141/2004 and POCTI/BIO/37934/2001, respectively, from
the Fundac ¸a ˜o para a Cie ˆncia e a Tecnologia (Portugal).
References
Abreu AA, Costa JC, Araya-Kroff P, Ferreira EC, Alves MM. 2007. Quan-
titative image analysis as a diagnostic tool for identifying structural
changes during a revival process of anaerobic granular sludge. Water
Res 41(7):1473–1480.
Alves MM, Mota Vieira JA, A´lvares Pereira RM, Pereira MA, Mota M. 2001.
Effects of lipids and oleic acid on biomass development in anaerobic
fixed bed reactors. Part I. Biofilm growth and activity. Water Res
35:255–263.
Table III.
Loadings/weights of the variables in datasets associated to PC1.
VariableLD1LD2LD3LD4Observation
HRT
OLR
Eff
pH
VSS
<0.1
>0.1
>1
SAA
SHMA
LfA
TL/VSS
VSS/TA
—0.390
?0.390
0.245
0.336
?0.232
?0.080
0.111
?0.097
?0.375
?0.018
?0.387
0.151
?0.356
——
0.348
?0.368
?0.219
0.369
0.189
0.295
?0.316
?0.372
?0.102
?0.284
?0.270
?0.186
Controlled variable
Controlled variable
?0.309
?0.134
0.197
?0.036
?0.250
0.293
?0.270
0.211
?0.308
0.419
?0.380
0.407
?0.355
0.320
0.345
?0.341
?0.094
?0.283
0.283
0.284
0.196
?0.314
0.222
?0.318
Morphological variable
Morphological variable
Morphological variable
Morphological variable
Morphological variable
Morphological variable
Costa et al.: Monitoring of High Rate Anaerobic Reactors
455
Biotechnology and Bioengineering

Page 12

AmaralAL.2003.ImageAnalysisinBiotechnologicalProcesses:Application
to Wastewater Treatment. PhD dissertation, Braga: University of
Minho, Portugal. (http://hdl.handle.net/1822/4506).
Amaral AL, Pereira MA, da Motta M, Pons M-N, Mota M, Ferreira EC,
Alves MM. 2004.Development of image analysis techniques as a tool to
detect and quantify morphological changes in anaerobic sludge. II.
Applicationtoagranuledeteriorationprocesstriggeredbycontactwith
oleic acid. Biotechnol Bioeng 87(2):194–199.
APHA, AWWA, WPC. 1989. Standard methods for the examination of
water and wastewater. 17th edn. Washington, DC: American Public
Health Association.
Araya-Kroff P, Amaral AL, Neves L, Ferreira EC, Pons M-N, Mota M, Alves
MM.2004.Developmentofimageanalysistechniquesasatooltodetect
andquantifymorphologicalchangesinanaerobicsludge.I.Application
to granulation process. Biotechnol Bioeng 87(2):184–193.
Costa JC, Abreu AA, Ferreira EC, Alves MM. 2007. Quantitative image
analysis as a diagnostic tool for monitoring structural changes of
anaerobic granular sludge during detergent shock loads. Biotechnol
Bioeng 98(1):60–68.
Dudley BT, Howgrave-Grahm AR, Bruton AG, Wallis FM. 1993. Image
analysis to quantify and measure UASB digester granules. Biotechnol
Bioeng 42(3):279–283.
Einax JW, Zwanziger HW, Geiss S. 1997. Chemometrics in environmental
analysis. Weinheim: VCH.
Glasbey CA, Horgan GW. 1995. Image analysis for the biological sciences.
Chichester: John Wiley and Sons.
Harmsen HJM, Kengen HMP, Akkermans ADL, Stams AJM, de Vos WM.
1996. Detection and localization of syntrophic propionate-oxidizing
bacteria in granular sludge by in situ hybridization using 16S rRNA-
based oligonucleotide probes. Appl Environ Microbiol 62(5):1656–
1663.
Jeison D, Chamy R. 1998. Novel technique for measuring the size distribu-
tion of granules from anaerobic reactors for wastewater treatment.
Biotechnolgy Tech 12(9):659–662.
Jenne ´ R, Banadda EN, Gins G, Deurinck J, Smets IY, Geeraerd AH, Van
Impe JF. 2006. Use of image analysis for sludge characterisation:
Studying the relation between floc shape and sludge settleability. Water
Sci Technol 54(1):167–174.
Lee JM, Yoo CK, Choi SW, Vanrolleghem PA, Lee IB. 2004. Nonlinear
process monitoring using kernel principal component analysis. Chem
Eng Sci 59:223–234.
Lee DS, Lee MW, Woo SH, Kim Y-J, Park JM. 2006. Multivariate online
monitoring of a full-scale biological anaerobic filter process using
kernel-based algorithms. Ind Eng Chem Res 45:4335–4344.
Lettinga G. 1995. Anaerobic digestion and wastewater treatment systems.
Antoine van Leeuwenhoek 67:3–28.
Li W, Yue H, Valle-Cervantes S, Qin SJ. 2000. Recursive PCA for adaptive
process monitoring. J Process Control 10:471–486.
Li J, Garny K, Neu T, He M, Lindenblatt C, Horn H. 2007. Comparison of
some characteristics of aerobic granules and sludge flocs from sequen-
cing batch. Water Sci Technol 55(8–9):403–411.
Liu WT, Chan O-C, Fang HHP. 2002a. Characterization of microbial
community in granular sludge treating brewery wastewater. Water
Res 36:1767–1775.
Liu Y, Xu H-L, Show K-Y, Tay J-H. 2002b. Anaerobic granulation technol-
ogy for wastewater treatment. World J Microbiol Biotechnol 18:99–
113.
Liu Y, Xu H, Yang S, Tay J. 2003. Mechanisms and models for anaerobic
granulation in upflow anaerobic sludge blanket reactor. Water Res
37:661–673.
Massart DL, Vander Heyden Y. 2005. From tables from visuals: Principal
component analysis. Part 2. Practical data handling. LC–GC 18(2):84–
89.
McMahon KD, Zheng D, Stams AJM, Mackie RI, Raskin L. 2004. Microbial
population dynamics during start-up and overload conditions of
anaerobic digesters treating municipal solid waste and sewage sludge.
Biotechnol Bioeng 87(7):823–834.
Pons MN, Vivier H. 1999. Biomass quantification by image analysis. Adv
Biochem Eng/Biotechnol 66:133–184.
Russ CR. 1995. The image processing handbook. Boca Raton, FL: CRC
Press.
Sekiguchi Y, Kamagata Y, Nakamura K, Syutsubo K, Ohashi A, Harada H,
Nakamura K. 1998. Diversity of mesophilic and thermophilic granular
sludge determined by 16S rRNA gene analysis. Microbiology 22:2655–
2665.
Sezgin M, Jenkins D, Parker DS. 1978. A unified theory of filamentous
activated sludge bulking. J Water Pollut Control Fed 50:362–381.
Voolapalli RK, Stuckey DC. 1998. Stability enhancement of anaerobic
digestion through membrane gas extraction under organic shock loads.
J Chem Technol Biotechnol 73:153–161.
Walsby AE, Avery A. 1996. Measurement of filamentous cyanobacteria by
image analysis. J Microbiol Methods 26:11–20.
Wise BM, Gallagher NB. 1996. The process chemometrics approach
to process monitoring and fault detection. J Proc Cont 6(6):329–
348.
Zehnder AJB, Huser BA, Brock TD, Wuhrmann K. 1980. Characterization
of an acetate-decarboxylating, non-hydrogen-oxidizing methane bac-
terium. Arch Microbiol 124:1–11.
456
Biotechnology and Bioengineering, Vol. 102, No. 2, February 1, 2009