issn 1451 - 9372(print) issn 2217 - 7434(online) july … no3_p369-464... · 2017. 2. 3. ·...

ISSN 1451 - 9372(Print)ISSN 2217 - 7434(Online)JULY-SEPTEMBER 2015Vol.21, Number 3, 369-464

www.ache.org.rs/ciceq

Journal of the Association of Chemical Engineers of Serbia, Belgrade, Serbia

EDITOR-In-Chief Vlada B. Veljković

Faculty of Technology, University of Niš, Leskovac, Serbia E-mail: [email protected]

ASSOCIATE EDITORS Jonjaua Ranogajec

Faculty of Technology, University of Novi Sad, Novi Sad, Serbia

Srđan PejanovićDepartment of Chemical Engineering, Faculty of Technology and Metallurgy,

University of Belgrade, Belgrade, Serbia

Milan Jakšić ICEHT/FORTH, University of Patras,

Patras, Greece

EDITORIAL BOARD (Serbia) Đorđe Janaćković, Sanja Podunavac-Kuzmanović, Viktor Nedović, Sandra Konstantinović, Ivanka Popović

Siniša Dodić, Zoran Todorović, Olivera Stamenković, Marija Tasić, Jelena Avramović

ADVISORY BOARD (International)

Dragomir Bukur Texas A&M University,

College Station, TX, USA Milorad Dudukovic

Washington University, St. Luis, MO, USA

Jiri Hanika Institute of Chemical Process Fundamentals, Academy of Sciences

of the Czech Republic, Prague, Czech Republic Maria Jose Cocero

University of Valladolid, Valladolid, Spain Tajalli Keshavarz

University of Westminster, London, UK Zeljko Knez

University of Maribor, Maribor, Slovenia

Igor Lacik Polymer Intitute of the Slovak Academy of Sciences,

Bratislava, Slovakia Denis Poncelet

ENITIAA, Nantes, France

Ljubisa Radovic Pen State University,

PA, USA Peter Raspor

University of Ljubljana, Ljubljana, Slovenia

Constantinos Vayenas University of Patras,

Patras, Greece Xenophon Verykios University of Patras,

Patras, Greece Ronnie Willaert

Vrije Universiteit, Brussel, Belgium

Gordana Vunjak Novakovic Columbia University,

New York, USA Dimitrios P. Tassios

National Technical University of Athens, Athens, Greece

Hui Liu China University of Geosciences, Wuhan, China

FORMER EDITOR (2005-2007) Professor Dejan Skala

University of Belgrade, Faculty of Technology and Metallurgy, Belgrade, Serbia

Journal of the Association of Chemical Engineers of Serbia, Belgrade, Serbia

Vol. 21 Belgrade, July-September 2015 No. 3

Chemical Industry & Chemical EngineeringQuarterly (ISSN 1451-9372) is published

quarterly by the Association of ChemicalEngineers of Serbia, Kneza Miloša 9/I,

11000 Belgrade, Serbia

Editor:Vlada B. Veljković

[email protected]

Editorial Office:Kneza Miloša 9/I, 11000 Belgrade, Serbia

Phone/Fax: +381 (0)11 3240 018E-mail: [email protected]

www.ache.org.rs

All the manuscripts are not to be returned

For publisher:Tatijana Duduković

Secretary of the Editorial Office:Slavica Desnica

Marketing and advertising:AChE Marketing Office

Kneza Miloša 9/I, 11000 Belgrade, SerbiaPhone/Fax: +381 (0)11 3240 018

Publication of this Journal is supported by theMinistry of Education and Science of the

Republic of Serbia

Subscription and advertisements make payableto the account of the Association of Chemical

Engineers of Serbia, Belgrade, No. 205-2172-71, Komercijalna banka a.d., Beograd

Computer typeface and paging:Vladimir Panić

Printed by:Faculty of Technology and Metallurgy,

Research and Development Centre of PrintingTechnology, Karnegijeva 4, P. O. Box 3503,

11120 Belgrade, Serbia

Abstracting/Indexing:Articles published in this Journal are indexed inThompson Reuters products: Science Citation

Index - ExpandedTM - access via Web of Science®, part of ISI Web of KnowledgeSM

CONTENTS

Nenad D. Nikolić, Đorđe P. Medarević, Jelena D. Đuriš, Dra-gana D. Vasiljević, Comparison of drug release and mechanical properties of tramadol hydrochloride matrix tablets prepared with selected hydrophilic polymers ............................................................................. 369

Saeid Shokri, Mohammad Taghi Sadeghi, Mahdi Ahmadi Marvast, Shankar Narasimhan, Saeid Minaei, Naser Hamdami, Integrating principal component analysis and vector quantization with support vector regression for sulfur content prediction in hydrodesulfurization process ............................................................................... 379

Mile Veljovic, Sasa Despotovic, Milan Stojanovic, Sonja Pecic, Predrag Vukosavljevic, Miona Belovic, Ida Leskosek-Cukalovic, The fermentation kinetics and physicochemical properties of special beer with addition of Prokupac grape variety ..................................... 391

Najmi Ahmed Essawet, Dragoljub Cvetković, Aleksandra Velićanski, Jasna Čanadanović-Brunet, Jelena Vulić, Vuk Maksimović, Siniša Markov, Polyphenols and anti-oxidant activities of Kombucha beverage enriched with Coffeeberry® extract ........................................................... 399

Nataša S. Tomović, Kata T. Trifković, Marko P. Rakin, Marica B. Rakin, Branko M. Bugarski, Influence of com-pression speed and deformation percentage on mechanical properties of calcium alginate particles ........... 411

Radojica Pešić, Tatjana Kaluđerović Radoičić, Nevenka Boš-ković-Vragolović, Zorana Arsenijević, Željko Grbavčić, Pressure drop in packed beds of spherical particles at ambient and elevated air temperatures .............................. 419

Marija A. Ilić, Franz-Hubert Haegel, Vesna M. Pavelkić, Sne-zana J. Zlatanović, Zoran S. Marković, Aleksandar S. Cvjetić, Unusually sluggish microemulsion system with water, toluene and a technical branched alkyl poly-ethoxylate ........................................................................... 429

Shifeng Li, Yanming Shen, Dongbing Liu, Lihui Fan, Zhe Tan, Zhigang Zhang, Wenxiu Li, Wenpeng Li, Experi-mental study of concentration of tomato juice by CO2 hydrate formation................................................................ 441

Aleksandar R. Mladenović, Milka B. Jadranin, Aleksandar D. Pavlović, Slobodan D. Petrović, Saša Ž. Drmanić, Mil-ka L. Avramov Ivić, Dušan Ž. Mijin, Liquid chroma-tography and liquid chromatography-mass spectrometry analysis of donepezil degradation products ......................... 447

Contents continued Jiajia Dai, Benfang H. Ruan, Ying Zhu, Xianrui Liang, Feng Su, Weike Su, Preparation of nanosized fluticasone propionate nasal spray with improved stability and uniformity ............................................................................ 457

Chemical Industry & Chemical Engineering Quarterly

Available on line at Association of the Chemical Engineers of Serbia AChE www.ache.org.rs/CICEQ

Chem. Ind. Chem. Eng. Q. 21 (3) 369−378 (2015) CI&CEQ

369

NENAD D. NIKOLIĆ1 ĐORĐE P. MEDAREVIĆ2

JELENA D. ĐURIŠ2

DRAGANA D. VASILJEVIĆ2 1Hemofarm a.d., Vršac, Serbia

2Faculty of Pharmacy, Department of Pharmaceutical Technology,

University of Belgrade, Belgrade, Serbia

SCIENTIFIC PAPER

UDC 615.453.6:661.12

DOI 10.2298/CICEQ140613040N

COMPARISON OF DRUG RELEASE AND MECHANICAL PROPERTIES OF TRAMADOL HYDROCHLORIDE MATRIX TABLETS PREPARED WITH SELECTED HYDROPHILIC POLYMERS

Article Highlights • HPMC and HPC were used for the preparation of sustained release matrix tablets • Formulation and process parameters effects on tablet quality were evaluated • Simulation of compaction profiles of large scale rotary tablet presses was used • Proportion of tramadol HCl in tablet exhibits the most important influence on the drug

release • The type of filler had the most pronounced effect on tablet mechanical properties Abstract

This study investigates the use of high molecular weight hydrophilic polymers, hypromellose and hydroxypropylcellulose, for the preparation of sustained release matrix tablets containing a high-dose highly soluble drug, tramadol HCl. The proportion of polymer, type of insoluble filler, proportion of tramadol HCl, amount of drug in the tablet and compression pressure were recognized as critical formulation and process parameters and their influence on drug release and tablet mechanical properties was evaluated. Tensile strength was used as an indicator of the mechanical properties of the tablets. Experiments were performed with utilization of a compaction simulator, a device that simul-ates compaction profiles of large scale rotary tablet presses. In formulations with both polymers, the proportion of tramadol HCl was the most critical form-ulation parameter, wherein increasing the proportion of tramadol HCl inc-reased its release rate in the early stages of drug release. Regarding the tablet mechanical characteristics, the filler type had the most pronounced effect in formulations with both polymers. Higher tensile strengths were obtained with Avicel PH 102 as the filler in formulations with both HPMC and HPC.

Keywords: tramadol HCl, matrix tablets, hypromellose, hydroxypropyl-cellulose, drug release, tensile strength.

Hydrophilic polymers, such as hypromellose, polyethylene oxide, hydroxypropylcellulose, guar gum, carrageenan, xanthan gum, etc., are widely used in the formulation of hydrophilic matrices with controlled drug release. Formation of a gel layer in contact with surrounding media allows sustained drug release from such systems. Once the original protect-ive gel layer is formed, it controls the further penet- Correspondence: Đ.P. Medarević, Faculty of Pharmacy, Depart-ment of Pharmaceutical Technology, Vojvode Stepe 450, 11221 Belgrade, Serbia. E-mail: [email protected] Paper received: 13 June, 2014 Paper revised: 24 October, 2014 Paper accepted: 29 October, 2014

ration of water into the tablet and release of drug [1]. The basic processes that determine drug release from hydrophilic matrix tablets are drug diffusion through the gel layer, as well as erosion of the swollen gel layer [2]. The solubility of the drug can significantly affect the mechanism of drug release. Release of a highly water-soluble drug from hydro-philic matrix system is mainly controlled by the dif-fusion through the swollen gel layer, whereas the release of poorly soluble drugs is predominantly con-trolled by the polymeric matrix erosion [3]. For pro-longed release of high-dose, highly soluble drugs, it is necessary to use hydrophilic polymers of very high viscosity [4,5]. Quality by design (QbD) includes pro-

N.D. NIKOLIĆ et al.: COMPARISON OF DRUG RELEASE AND MECHANICAL… Chem. Ind. Chem. Eng. Q. 21 (3) 369−378 (2015)

370

cesses of determining critical quality attributes (CQA) of the drug product, as well as selecting the types and amounts of excipients to deliver drug product of the specified quality [6]. The most critical quality attribute in the development of sustained release hydrophilic matrix tablets is the drug release rate, which is pre-dominantly determined by the properties of the poly-mer (type, concentration, solubility and viscosity) and drug (solubility and concentration). High proportions of both drug and matrix polymer can significantly affect mechanical properties that are considered as an important quality attribute of the hydrophilic matrix tablets. Therefore, in the formulation of hydrophilic matrix tablets, both the ability of the polymer to pro-vide sustained drug release and the mechanical characteristics of the prepared tablets should be eval-uated. The ability of the polymer to affect the release rate characteristics of high-dose, highly soluble drug can be evaluated in two ways: as the ability to prevent premature drug release in early stage of drug release, as well as the ability to maintain drug release with predictable kinetics in time points over a prolonged period of time. Mechanical properties of powder mix-tures and compacts can be assessed through eval-uation of compressibility (solid fraction vs. compaction pressure), compactibility (tensile strength vs. solid fraction) and tabletability (tensile strength vs. com-paction pressure). Compression and compaction pro-perties of hypromellose were evaluated with the inf-luence of particle size, moisture content, compression force, compression speed, viscosity grade and sub-stitution level of free OH-groups [7]. Mechanical pro-perties and drug release characteristics of directly compressible controlled release matrix systems con-taining high viscosity hydroxypropylcellulose and highly or sparingly soluble model drug were evaluated for different types of fillers [5]. Tablet geometry plays also important role in drug release kinetics for dif-fusion-controlled systems and it has been studied in detail by Siepmann et al. [8]. Effect of a tablet sur-face/tablet volume ratio (SA/Vol) on the drug release from hydrophilic matrices was investigated for hypro-mellose matrix tablets [9] and for hydroxypropylcel-lulose matrix tablets [10]. Most of the studies that investigated mechanical characteristics of hydrophilic matrix tablets were conducted on either excenter tablet presses or instrumented small rotary presses. Using compaction simulator enables thorough studies of compaction characteristics of materials, as well as evaluation of the influence of different process vari-ables of the compaction phase on tablet properties, determination of scale-up parameters, and creating database of the compaction properties of new active

pharmaceutical ingredients (APIs) or excipients [11]. Compaction simulators are useful as the tools for evaluation and comparison of powder mechanical properties in simulated production conditions. Mathe-matical models, such as force-time, force-distance, and die-wall force parameters can be used to des-cribe work of compaction, elasticity, plasticity, and time dependent deformation behavior of pharma-ceuticals. Parameters such as the bonding index, brittle fracture index and strain index can be used to predict compaction related problems [12]. This study investigates the performance of two selected high viscous hydrophilic polymers, hypromellose and hyd-roxypropylcellulose in the formulation of hydrophilic matrix tablets with tramadol-hydrochloride, as a model of high-dose, highly soluble drug. The influ-ence of the selected formulation and process vari-ables on drug release characteristics and tablet mechanical properties was evaluated.

EXPERIMENTAL

Materials

The following materials were used: hypromel-lose, type 2208 (Metolose 90SH–100000, Shin-Etsu Chemical Co., Ltd, Tokyo, Japan), hydroxypropyl-cellulose (Klucel HXF, Hercules Incorporated Aqualon Division, Wilmington, DE, USA), microcrystalline cel-lulose (Avicel PH 102, FMC Biopolymer, USA), par-tially pregelatinised maize starch (Starch 1500, Color-con, Dartford, UK), colloidal silicon dioxide (Aerosil 200 Pharma, Degussa), magnesium stearate, (Mallin-ckrodt, St. Louis, MO, USA). Tramadol HCl (Hemo-farm A.D., Vršac, Serbia) was used as model of high dose, highly water-soluble drug.

Experimental design

In the first part, experiments were conducted according to 25-2 fractional factorial design (Table 1) for each polymer, in order to evaluate the influence of different formulation (proportion of polymer, type of insoluble filler, proportion of tramadol HCl, amount of drug in tablet) and process input variables (com-pression pressure) on tramadol HCl release from compressed matrices. Percentages of drug released from matrix tablets after 30 and 240 min were sel-ected as responses. Using this design type it is pos-sible only to recognize factors that have the highest impact on the response variable, having in mind that effects that are small relative to the noise of the pro-cess will be confounded. Therefore, the applied design type was used for screening purposes in order


371

to select factor that have pronounced influence on the selected response variable.

The effects of the input variables on responses were compared for each polymer. In order to eli-minate the influence of tablets geometry on drug rel-ease, normalization of the percentage of drug rel-eased after 30 and 240 min, was performed by div-iding obtained drug released percentages with sur-face area per volume ratio (SA/Vol) of tablets.

In the second part of the study, 23 full factorial design was used to evaluate the influence of the mat-rix polymer proportion, type of filler and drug pro-portion on the tablet mechanical properties (Table 2). From this point forward, two experimental sets according to 23 design was applied, for each polymer separately. Tensile strength (σt) was used as an indi-cator of tablets mechanical properties, enabling com-parison of mechanical properties of tablets with differ-ent dimensions. Tensile strength was calculated according to the following equation:

σπ

=t2Fdh

(1)

where F is the crushing force, d is the tablet diameter and h is the tablet thickness.

Statistical software Design Expert 7.0 (StatEase, Inc.) was used throughout the study.

Tablets preparation

Powder mixtures for compression were pre-pared by using a Turbula® shaker-mixer (Glen Mills Inc., Clifton, NJ, USA). Tablets were compressed with direct compression method by using of PressterTM single station compacting simulator (Metropolitan Computing Corporation, East Hanover, NJ, USA). Simulation of the rotary tablet press Korsch PH336 was used, with simulated die table speed of 30 rpm which conforms to 65,000 tablets per hour and dwell time of 20 ms. Tablets were prepared using punches with diameters of 7, 10 and 13 mm, while tablet masses were 180, 360 and 720 mg, respectively, according to the experimental design. Compression pressure was calculated from the measured com-paction force per cross-sectional area of tablets. Tablets were compressed on different compaction pressures in the range of 100 to 500 MPa with simul-

Table 1. Experimental matrix according to 25-2 experimental design with obtained responses

Formul-ation

Proportion of polymer, % Filler type

a Proportion of

tramadol HCl, %Compression

pressure, MPaTramadol HCl per tablet, mg

HPMC matrix tablets HPC matrix tablets

SA/Vol Q30b Q240c SA/Vol Q30b Q240c

F1 25 Starch 1500 55.6 300 200 0.894 27.5 78.7 0.883 28.5 82.6

F2 35 Starch 1500 55.6 150 100 1.047 28.4 78.7 1.037 28.7 83.3

F3 25 Avicel PH 102 55.6 150 200 0.896 28.0 83.0 0.891 28.1 82.8

F4 35 Avicel PH 102 55.6 300 100 1.056 30.3 82.2 1.048 26.7 80.8

F5 25 Starch 1500 27.8 300 100 0.896 26.2 75.5 0.898 26.2 73.3

F6 35 Starch 1500 27.8 150 200 0.709 20.8 69.9 0.710 23.1 70.5

F7 25 Avicel PH 102 27.8 150 100 0.930 22.5 71.9 0.920 26.8 72.4

F8 35 Avicel PH 102 27.8 300 200 0.750 20.3 65.8 0.736 21.9 68.1aStarch 1500 - partially pregelatinized maize starch; Avicel PH 102 - microcrystalline cellulose; bnormalized percentage of drug release after 30 min; cnormalized percentage of drug release after 240 min

Table 2. Experimental matrix according to 23 experimental design with obtained responses

Formulation Proportion of polymera, %

Filler typeb Proportion of

tramadol-HCl, %

σt / N cm-2 HPMC matrix tablets HPC matrix tablets

120 MPa 250 MPa 120 MPa 250 MPa

F1 25 Starch 1500 55.6 75 105 95 130

F2 35 Starch 1500 55.6 115 180 107 140

F3 25 Avicel PH 102 55.6 145 195 145 201

F4 35 Avicel PH 102 55.6 170 213 130 182

F5 25 Starch 1500 27.8 60 95 75 130

F6 35 Starch 1500 27.8 85 130 95 135

F7 25 Avicel PH 102 27.8 295 360 253 305

F8 35 Avicel PH 102 27.8 245 330 205 265 aProportion of polymer refers to a quantity of HPMC or HPC depends on experimental design; bStarch 1500 - partially pregelatinised maize starch; Avicel PH 102 - microcrystalline celullose


372

ation of the compaction profile of the Korch PH336 rotary tablet press. Tensile strength was calculated from measured hardness and tablet dimensions. Inf-luence of the selected input variables on the tablet tensile strength was analyzed.

Drug release testing

Drug release test was performed using rotating basket apparatus (Sotax, Allschwil, Switzerland) during 8 h (rotational speed 75 rpm, medium volume 600 ml, temperature 37 °C). Drug release testing procedure was developed as a part of in-house test-ing. Samples were taken after predefined time inter-vals: 30, 120, 180, 240, 360 and 480 min and the amount of dissolved tramadol HCl was determined spectrophotometrically at λ = 271 nm. The data col-lected up to 240 min were selected for modeling of drug release. During the drug release testing, the pH of the medium was changed by adding buffer sol-utions as follows: in the first 30 min, pH 1.2; from 30- –120 min, pH 2.3; from 120-180 min, pH 6.8; from 180-240 min, pH 7.2. Artificial gastric juice, pH 1.2, as well as two buffer solutions, were used for changing the pH during the dissolution test:

Buffer 1: K2HPO4, 63 g/100 ml (6 ml at 31 min) Buffer 2: NaOH 15 g/100 ml (6 ml at 121 min

and 3 ml at 181 min).

Mechanical characterization of tablets

Tablet hardness was measured using an 8M tablet hardness tester (Dr. Schleuniger Pharmatron, Thun, Switzerland) and tablet dimensions were measured with digital caliper. Tensile strength was calculated from dimensions of tablets and measured hardness, according to Eq. (1).

RESULTS AND DISCUSSION

Evaluation of drug release from hypromellose (HPMC) and hydroxypropylcellulose (HPC) matrices

Tramadol HCl release profiles from HPMC mat-rices and analogue formulations with HPC (formul-ation F1-F8) are presented in Figure 1a and b, res-pectively.

Since in the hydrophilic matrices with high visc-osity polymers and highly water-soluble drug diffusion release mechanism is predominant in respect to erosion, drug release process follows Higuchi’s model. There is linear relationship between drug release and the square root of time for both HPMC and HPC matrices (Figure 2a and b, respectively) in certain time intervals and it is related to the geometry characteristics of the matrix tablets. The calculated determination coefficients values (R2) for all formul-

ations were above 0.95, indicating good fitting to the Higuchi model.

One of the prerequisites for formulation of matrix tablets with extended drug release is that premature drug release should not happen. Percentages of tra-madol HCl released after 30 and 240 min were sel-ected as indicator of polymer ability to prevent pre-mature drug release and enable sustained release of drug (Table 1). Furthermore, both time points are within the interval in which drug release follows Higuchi’s model. Normalized values of drug released percentages were used in order to eliminate the inf-luence of geometrical characteristics on drug release rate. Normalization was performed by dividing the percentage of drug released with surface area per volume ratio (SA/Vol) of the matrix tablets containing HPMC or HPC as the release rate control polymer (Tables 3 and 4).

The influence of selected formulation and pro-cess parameters (proportion of polymer, type of insol-uble filler, proportion of tramadol HCl, amount of drug in tablet and compression pressure) on normalized percentage of drug released at 30 and 240 min was analyzed and calculated factor effects are summar-ized in Table 3.

From the obtained results it can be concluded that in formulations containing HPMC as release rate control polymer, only the proportion of drug in tablet exhibits statistically significant effect on the percent-age of drug release after 30 min (p < 0.05). By increasing the proportion of drug in the tablet, the drug release rate is increased. Other formulation and process variables in selected range have no statistical significant influence on tramadol HCl released rate. For the drug release at a later time point, 240 min, none of the examined factors exhibit statistically sig-nificant influence on drug release.

Contrary to HPMC matrix tablets, statistically significant effects of the proportion of tramadol HCl, proportion of polymer and amount of tramadol HCl in the tablet on the normalized percentage of drug released after 30 min were observed for HPC matrix tablets. The proportion of tramadol HCl exhibits the highest effect on the drug release in the early stage (30 min), and with its increase, the drug release rate increased. The proportion of the HPC polymer has the opposite effect on the drug release, wherein its increase led to slower drug release rate. Quantity of drug per tablet also showed a statistically significant effect for the early stage of drug release, indicating that there could be difference in drug release between tablets containing different amounts of the drug. For the normalized value of drug released at 240 min, the

N.D. NIKOLIĆ et al.: COMPARISON OF DRUG RELEASE AND MECHANICAL… CI&CEQ 21 (3) 369−378 (2015)

373

Figure 1. Tramadol HCl release profiles from HPMC (a) and HPC (b) matrix tablets.

same formulation and process variables including type of filler had statistically significant effects. It was found that the proportion of tramadol HCl has the largest effect on the normalized percentage of drug release after 240 min. Increase in proportion of tra-madol HCl results in the increase in percentage of drug release after 240 min, the same as after 30 min. Other input variables did not have statistically sig-nificant influence on the release of tramadol HCl.

Mechanical characterization of matrices with HPMC and HPC

Full factorial design was performed in order to evaluate the influence of input variables (proportion of polymer, type of filler and proportion of tramadol HCl per tablet) on mechanical characteristics of matrix tablets. Due to different tablet dimensions, tensile strength was used for evaluation of mechanical char-acteristics of matrix tablets. Profiles of tensile strength


374

(a)

(b)

Figure 2. Tramadol HCl released per square root of time for HPMC (a) and HPC (b) matrix tablets.

Table 3. Calculated effects of the formulation and process parameters on normalized percentage of drug release after 30 and 240 min; * – statistically significant effects

Parameter HPMC matrix tablets HPC matrix tablets

Q30a Q240

b Q30a Q240

b

Proportion of polymer -0.9099 -1.1403 -7.1839* -7.6050*

Filler type -0.3721 0.0091 -2.3426 -5.0700*

Proportion of Tramadol HCl 5.0441* 3.6032 10.9322* 40.9222*

Compression Pressure -0.9509 0.1186 2.6549 3.8025

Tramadol HCl per tablet 2.2326 0.9943 5.3099* 5.2511* aNormalized percentage of drug release after 30 min; bnormalized percentage of drug release after 240 min

versus compaction pressure (tabletability profiles) for the matrix tablets with HPMC and HPC are presented in Figure 3a and b, respectively.

Compression of tablet mixture was performed with compression force up to 40 kN, since with punch face diameter of 13 mm and compression pressure

ofabout 300 MPa compression the force is near to the maximum possible force that could be achieved on the compaction simulator. Tensile strengths of matrix tablets on compaction pressures of 120 and 250 MPa were analyzed and compared as output variables (Table 2).


375

Figure 3. Profiles of tensile strength versus compaction pressure for HPMC (a) and HPC (b) matrix tablets.

The values of tensile strength were extracted from trend lines of tabletability profiles for matrix tablets with HPMC and HPC (Figure 3a and b, res-pectively). Values for compaction pressure of 120 MPa correspond to values of tensile strength in the ascending part of the tabletability profiles diagrams in all trials, while the compression pressure of 250 MPa corresponds to the part where the profile reaches a plateau. The higher values of tensile strength were obtained when Avicel PH 102 was used as the filler compared to Starch 1500 in formulations with both polymers. In the range of the lower compression pres-sure (120 MPa) for matrix tablets with both polymers, none of the evaluated formulation variables including

interaction between them have a statistically signific-ant influence on the output variable, i.e., variation of either of the evaluated variables in the evaluated range did not exhibit statistically significant effects on the tensile strength (Table 4).

On the contrary, in the range of the higher com-pression pressures, ∼250 MPa, for matrices formul-ated with HPMC all formulation variables including interaction between them except proportion of HPMC polymer and interaction between proportion of HPMC polymer and proportion of Tramadol HCl had signific-ant effects on the tensile strength. The most pro-nounced effect is achieved with variation of the type of filler, where matrix tablets containing Avicel PH102


376

had much higher tensile strengths compared to mat-rices formulated with Starch 1500 as the filler. Inc-reasing of the proportion of Tramadol HCl had a neg-ative influence on the tensile strength, as well as the combination of using Starch 1500 with increasing pro-portion of Tramadol HCl, and the combination of the same filler with increasing proportion of HPMC poly-mer.

In the case of matrix tablets formulated with HPC and with higher compression pressure, only the

type of filler had a statistically significant effect on the tensile strength. Same as with HPMC as a polymer, matrix tablets containing Avicel PH102 has much higher tensile strengths compared to matrices for-mulated with Starch 1500 as filler.

Influence of proportions of polymer, filler and drug (comprising 100% of tablet) on tensile strength for matrix tablets with HPMC polymer, is presented in Figure 4a and b for Avicel PH 102 and Starch 1500, respectively, as well as for the matrix tablets with

Table 4. Calculated effects of the formulation and process parameters on tablet tensile strength; * – statistically significant effects

Parameter HPMC matrix tablets HPC matrix tablets

120 MPa 250 MPa 120 MPa 250 MPa

Proportion of polymer 0.6667 12.2500* -0.7561 -2.7500

Filler type 8.6667 73.5000* 8.8049 26.1250*

Proportion of Tramadol HCl -3.0000 -27.7500* -3.6829 -11.3750

Proportion of polymer×Filler type -1.5000 -15.2500* -2.3171 -4.6250

Proportion of polymer×Proportion of tramadol HCl 1.5000 11.0000 0.6098 1.6250

Filler type×Proportion of tramadol HCl -4.5000 -42.7500* -5.2439 -12.0000

(a)

(b)

Figure 4. Influence of proportions of polymer (HPMC), filler and drug (comprising 100% of tablet) on tensile strength of tablets prepared with Avicel PH 102 (a) and Starch 1500 (b).


377

HPC polymer in Figure 5a and b for both fillers in the same order.

CONCLUSION

A comparison of hydrophilic polymers as matrix forming materials was made with respect to drug release modification abilities, as well as mechanical properties with the selected high-dose highly soluble model drug, tramadol HCl. Formulations with both polymers, HPMC and HPC, had similar character-istics with respect to sensitivity of drug release rate on the variation of proportion of the tramadol HCl in early

stages of drug release. Increasing of the proportion of tramadol HCl in formulation resulted in higher amounts of drug released. Regarding the mechanical properties of tablets, the type of filler had the most critical effect on the powder mixture tabletability in for-mulations with both polymers. Mechanical properties of matrix tablets were significantly better with mic-rocrystalline cellulose compared to partially pregel-atinised starch. These findings could be useful in selection of the polymer and optimization of for-mulation of sustained release matrix tablets with high dose, highly soluble model drug.

(a)

(b)

Figure 5. Influence of proportions of polymer (HPC), filler and drug (comprising 100% of tablet) on tensile strength of tablets prepared with Avicel PH 102 (a) and Starch 1500 (b).


378

Acknowledgements

This work was done under the project No. TR 34007, supported by the Ministry of Education, Sci-ence and Technological Development, Republic of Serbia.

REFERENCES

[1] Using METHOCEL Cellulose Ethers for Controlled Rel-ease of Drugs in Hydrophilic Matrix Systems, product brochure, http://www.colorcon.com/literature/marketing/ /mr/Extended%20Release/METHOCEL/English/hydroph_matrix_broch.pdf (accessed 2 April 2014)

[2] P. Colombo, R. Bettini, P. Santi, A. De Ascentiis, N.A. Peppas, J. Controlled Release 39 (1996) 231-237

[3] R. Bettini, PL. Catellani, P. Santi, G. Massimo, NA. Peppas, P. Colombo, J. Controlled Release 70 (2001) 383-391

[4] T. Cabelka, A. Faham, H. Bernthal, A. Rajabi-Siahboomi, Application of Quality by Design Principles to the For-mulation of a Hydrophilic Matrix Tablet of a High Dose/High Solubility Drug, in: AAPS Annual Meeting, Seattle, WA, 2009, Poster T2282

[5] T. Dürig, G. Skinner, W. Harcum, Compression and Drug Release Characteristics of Directly Compressible KLUCEL® Hydroxypropylcellulose Controlled Release Matrix Systems, Pharmaceutical Technology Report PTR-019-1, Ashland, 2002

[6] I. Robertson, S. Tiwari, T. Cabelka, Pharm. Technol. 36 (2012) 106-116

[7] A. Nokhodchi, J. Ford, P. Rowe, M. Rubinstein, Int. J. Pharm. 129 (1996) 21-31

[8] J. Siepmann, H. Kranz, N.A. Peppas, R. Bodmeier, Int. J. Pharm. 201 (2000) 151-164

[9] T.D. Reynolds, S.H. Gehrke, A.S. Hussain, L.S. She-nouda, J. Pharm. Sci. 87 (1998) 1115-1123

[10] M.K. Divi, T. Dürig, W. Harcum, Drug Release from Hydroxypropylcellulose Sustained Release Matrix Tab-lets: Implications of Tablet Surface Area/Volume Ratio, MW and Drug Solubility, Aqualon Pharmaceutical Tech-nology Report PTR-068, 2008

[11] L. Yang, K. Venkatesh, R. Fassihi, J. Pham. Sci. 85 (1996) 1085-1090

[12] S. Patel, A.M. Kaushal, A.K. Bansal, Crit. Rev. Ther. Drug. Carrier Sys. 23 (2006) 1-65.

NENAD D. NIKOLIĆ1 ĐORĐE P. MEDAREVIĆ2

JELENA D. ĐURIŠ2

DRAGANA D. VASILJEVIĆ2 1Hemofarm a.d., Beogradski drum b.b.,

Vršac, Srbija 2Katedra za farmaceutsku tehnologiju i

kozmetologiju, Farmaceutski fakultet, Univerzitet u Beogradu, Vojvode Stepe

450, 11221 Beograd, Srbija

NAUČNI RAD

POREĐENJE OSLOBAĐANJA TRAMADOL- -HIDROHLORIDA I MEHANIČKIH KARAKTERISTIKA MATRIKS TABLETA IZRAĐENIH SA ODABRANIM HIDROFILNIM POLIMERIMA

U ovom radu ispitivana je mogućnost primene hipromeloze i hidroksipropilceluloze, kao hidrofilnih polimera velike molekulske mase, za izradu matriks tableta sa produženim oslo-bađanjem, sa visoko rastvorljivom, visoko doziranom lekovitom supstancom tramadol-hid-rohloridom. Udeo hidrofilnog polimera, vrsta nerastvorljivog sredstva za dopunjavanje, udeo tramadol-hidrohlorida, količina lekovite supstance u pojedinačnoj tableti i pritisak kompresije su prepoznati kao kritični parametri formulacije i procesa i u radu je ispitivan njihov uticaj na oslobađanje lekovite supstance i mehaničke karakteristike izrađenih tab-leta. Zatezna čvrstina tableta je korišćena kao indikator mehaničkih karakteristika tableta. Svi eksperimenti su vršeni korišćenjem simulatora kompakcije, koji omogućava simuliranje profila kompakcije rotacionih tablet mašina velikog kapaciteta. Kod formulacija izrađenih sa obe vrste polimera, udeo tramadol-hidrohlorida se pokazao kao najkritičniji faktor formu-lacije, pri čemu je povećanje udela tramadol-hidrohlorida dovelo do povećanja brzine oslo-bađanja ove lekovite supstance u početnim fazama oslobađanja lekovite supstance. Vrsta sredstva za dopunjavanje je pokazala najveći uticaj na mehaničke karakteristike izrađenih tableta, kod formulacija izrađenih sa oba tipa hidrofilnih polimera. Više vrednosti zatezne čvrstine tableta su postignute kod formulacija izrađenih korišćenjem Avicel PH 102 kao sredstva za dopunjavanje, bez obzira da li je u sastav tableta kao matriks polimer ulazila hipromeloza ili hidroksipropilceluloza.

Ključne reči: tramadol-hidrohlorid, matriks tablete, hipromeloza, hidroksipro-pilceluloza, brzina oslobađanja lekovite supstance, zatezna čvrstina.




379

SAEID SHOKRI1

MOHAMMAD TAGHI SADEGHI1

MAHDI AHMADI MARVAST2

SHANKAR NARASIMHAN3 1Department of Chemical

Engineering, Iran University of Science and Technology (IUST),

Tehran, Iran 2Process & Equipment Technology

Development Division, Research Institute of Petroleum Industry

(RIPI), Tehran, Iran 3Department of Chemical

Engineering, IIT Madras, Chennai, India

SCIENTIFIC PAPER

UDC 66.094.522:004.932

DOI 10.2298/CICEQ140418039S

INTEGRATING PRINCIPAL COMPONENT ANALYSIS AND VECTOR QUANTIZATION WITH SUPPORT VECTOR REGRESSION FOR SULFUR CONTENT PREDICTION IN HYDRODESULFURIZATION PROCESS

Article Highlights • Designing of a reliable data-driven soft sensor to predict sulfur content in HDS Process • Integrating Principal Component Analysis (PCA) and Vector Quantization (VQ) with

SVR • Comparing between PCA and VQ methods on prediction accuracy of support vector

regression model • Improving prediction accuracy and computation time of SVR model by proposed

approach Abstract

An accurate prediction of sulfur content is very important for the proper oper-ation and product quality control in hydrodesulfurization (HDS) process. For this purpose, a reliable data-driven soft sensor utilizing Support Vector Regres-sion (SVR) was developed and the effects of integrating Vector Quantization (VQ) with Principle Component Analysis (PCA) were studied in the assess-ment of this soft sensor. First, in the pre-processing step the PCA and VQ techniques were used to reduce dimensions of the original input datasets. Then, the compressed datasets were used as input variables for the SVR model. Experimental data from the HDS setup were employed to validate the proposed integrated model. The integration of VQ/PCA techniques with SVR model was able to increase the prediction accuracy of SVR. The obtained results show that integrated technique (VQ-SVR) was better than (PCA-SVR) in prediction accuracy. Also, VQ decreased the sum of the training and test time of SVR model in comparison with PCA. For further evaluation, the per-formance of VQ-SVR model was also compared to that of SVR. The obtained results indicated that VQ-SVR model delivered the best satisfactory predicting performance (AARE = 0.0668 and R2 = 0.995) in comparison with investigated models.

Keywords: Principal Component Analysis (PCA), Vector Quantization (VQ), Support Vector Regression (SVR), soft sensor, hydrodesulfuri-zation (HDS) process.

HDS process is one of the key catalytic units that play a major role in most petroleum refineries. High concentration of sulfur in HDS product has a negative impact on the refining processes, health and

Correspondence: M.T. Sadeghi, Department of Chemical Eng-ineering, Iran University of Science and Technology (IUST), Tehran, Iran. E-mail: [email protected] Paper received: 18 April, 2014 Paper revised: 7 September, 2014 Paper accepted: 29 October, 2014

the environment. To increase process performance, it would be necessary to manage the ultra low sulfur content in final product of this process. Hence, sulfur content prediction is very important for the proper operation of HDS units [1]. Furthermore, more strin-gent environmental restrictions give remarkable importance to have an accurate prediction of sulfur content in the HDS process [2]. For this purpose, either hardware on-line analyzers or analytical labor-atory tests are used in HDS units. Hardware on-line

S. SHOKRI et al.: INTEGRATING PRINCIPAL COMPONENT ANALYSIS… Chem. Ind. Chem. Eng. Q. 21 (3) 379−390 (2015)

380

analyzers are too expensive. Moreover, operators and engineers find many problems such as calibration necessity, insufficient accuracy and long dead time when hardware sensors are used. Moreover, anal-ytical laboratory tests are tedious and unreliable. Soft sensors (soft analyzers) are key technologies for managing high quality products when hardware pro-cess analyzers are not available. Soft sensors can also be applied as an alternative to laboratory tests.

A soft sensor is a predictive model that des-cribes the relationship between the predicted process variables and the measured variables. A soft sensor model can be developed by using either model-driven approaches or data-driven approaches. Model-driven soft sensors are based on first principle mathematical models. First principle models describe the physical and chemical background of the process. They are often not available due to complexity of industrial processes. These models, obtained from the fundam-ental process knowledge, require a lot of process expert knowledge, effort and time to be developed. Data-driven models operate based on actual data measured within the operational plants, and describe the real process conditions [3]. Furthermore, process data have become widely available in modern indus-trial plants.

Data-driven soft sensors can be applied to the online estimation of product indices using process measurement data because they have become widely available in many chemical plants [4,5]. Unlike phys-ical sensors, which directly measure the value of a variable, data driven-soft sensors measure the pro-cess variables whose direct measurements are asso-ciated with some technical problems. Therefore, soft sensors use the frequently sampled process variables such as temperature, pressure, flow rate, etc. to measure these hard to measure variables. In these processes, machine learning techniques are fre-quently used.

Support vector machines (SVM) are efficient machine learning techniques derived from statistical learning theory by Vapnik [6]. Compared with artificial neural networks (ANN), an SVM provides more reli-able and better performance [7]. The SVM is used for classification and regression tasks. When applied for regression tasks, SVM is also called SVR [8].

The linear dependency between different vari-ables in the dataset influences the generalization abi-lity of the SVR model [9]. Moreover, due to encount-ering large datasets in process industries, the training time increases for SVR model. To tackle these prob-lems, this paper uses the data compression tech-niques such as PCA or VQ. These methods can be

used to generate a smaller training set with greatly reduce training time. Furthermore, integration of data compression techniques with SVR can strengthen the generalization ability of the SVR model and therefore increase prediction performance.

PCA has been used as a pre-processing step during SVR modeling to reduce dimensionality of the original multivariable dataset. The integrated PCA and machine learning methods showed good perfor-mance in various prediction fields, such as assess-ment of coronary artery diseases [10], forecasting greenhouse gas emissions [11] and predicting gaso-line homogeneous charge compression ignition com-bustion behavior during transient operation [12]. VQ is a data compression method based on the principle of block coding. The purpose of using this technique is to simplify the training set and to reduce training time [13]. A few studies have been carried out that inves-tigate the effect of using compression techniques with a number of machine learning algorithms. Moreover, no comparison has been carried out between the effects of VQ and PCA on the SVR model in literature. In this work PCA and VQ have been integrated with SVR to predict the sulfur content of the treated gas oil. To train and test SVR model, data collection is carried out from a HDS setup.

The main objectives of the present study are: 1) designing an accurate and reliable data-driven pre-diction model for sulfur content prediction in HDS process; 2) aplying a novel integrated technique using VQ/PCA and SVR model to increase the prediction performance; 3) comparing prediction accuracy and CT of both integrated models (VQ-SVR & PCA-SVR).

MATERIALS AND METHODS

Experimental setup

The main control index in product quality of hydrodesulfurization (HDS) process is the sulfur con-tent. Accurate prediction of sulfur content plays an important role in this process. In this study, a HDS setup was used to obtain experimental datasets. The schematic diagram of the system is depicted in Figure 1. Gas oil stream is entered into the preheated sec-tion via a dozing pump. Then, preheated stream passes through a fixed bed reactor (trickle-bed), in which hydrogenation occurs. The selected catalyst is Co-Mo/Al2O3. The output liquid from the bottom of the reactor enters into a high pressure separator in which it is separated to treated gas oil and H2S. H2S is then absorbed by NaOH solution in caustic column vessel. The operating conditions and catalyst specifications of the mentioned setup are shown in Table 1.


381

In all experiments, the gas oil with total sulfur content of 7200 ppm was used for the feed stream. In this setup, the reactor diameter and the reactor length were 0.0127 and 0.63 m, respectively. Also, catalyst bed length of 0.11 m was selected.

Table 1. Setup specification

Value Component

0.0127

0.63 0.11

3.4 13.6

235 0.53 200 99

0.72

Reactor: Reactor diameter, m Reactor length, m Catalyst bed length, m Catalyst: Chemical composition, wt%, dry basis Cobalt Molybdenum Physical properties: Surface area, m2/g Pore volume, cm3/g (H2O) Flat plate crush strength, N/cm (lb/mm) Attrition index Compacted bulk density, g/cm3

In order to train and test the SVR model, a set of experiments were designed for the pilot. Experimental design for wide range of sulfur content was done. The minimum and maximum sulfur content in the treated gas oil products were found to be 10 and 4900 ppm, respectively. The HDS conversion varies with changes in the operating conditions. The independent oper-ating parameters that affect the hydrogenation proce-dure are: inlet reactor temperature (T), reactor pres-sure (p), H2/oil ratio and liquid flow rate (Q). In this work, these operating conditions are considered according to real condition of the refineries.

Therefore, the effect of these parameters was studied using the following range of values: T, 320, 337, 353 and 370 °C; p, 50, 60 and 70 bar; H2/oil ratio, 85, 100, 120, 140 and 170 nm3/m3; liquid flow rate, 0.2, 0.23, 0.26, 0.29 and 0.32 cm3/min. Only one of the above parameters was allowed to change in every test. The samples are collected based on 4 h of operation under nearly steady state conditions. A time interval of 4 h was required to reach the next steady state experimental condition. Change of any oper-ating conditions was carried out continuously and inc-

Figure 1. A schematic diagram of the diesel oil HDS setup.


382

rementally with moderate slope within time intervals of 4 h to avoid a shock to the system and prevent catalyst deactivation.

The operating data was recorded minute-by-minute over this time periods. Therefore, the datasets were obtained with 240 data for each sensor (QH2, Qgas oil, Tpreheat, Tin, Tout and p). While the sulfur content in the product stream was obtained from laboratory tests once every 4 h. Hence, in twenty four hours, only 3 treated gas oil samples are collected. There-fore, 294 samples for 98 working days were collected. As there was only one data for sulfur in time intervals of 4 h, the values of other parameters were averaged within the time intervals and finally 294 datasets were selected.

The location of these measurements is shown in Figure 1 as follows: hydrogen flow rates (1); feed diesel flow rate (4); preheat temperature (6); tempe-rature of the inlet stream to reactor (10); temperature of the outlet stream from reactor (11); reactor pres-sure (12).

In order to determine total sulfur content in HDS product after reaching steady state condition, the treated gas oil samples were analyzed by two methods: 1) ASTM D4294 and 2) ASTM D5453.

The ASTM D4294 method is a standard test method to determine the sulfur content in petroleum product by energy dispersive X-ray fluorescence spectrometry. It is capable of determining sulfur over a wide range of concentrations. This test method pro-vides precise measurement of total sulfur in petro-leum product with a minimum of a sample prepar-ation. The applicable concentration range for this method is 0.015 to 5 mass%. To determine the sulfur content of less than this range the ASTM D5453 method is used. The ASTM D5453 is an ultra low sulfur analysis that uses ultraviolet fluorescence to determine the sulfur content in ultra low sulfur diesel. Liquid samples were collected for analysis every four hours. A typical analysis time is 5 min per sample after calibration and standardization.

Principal Component Analysis (PCA)

PCA is one of the multivariate statistical methods that are widely used to find a low dimensional repre-sentation of data matrix [14-16]. The PCA method is designed to transform a large set of interrelated inde-pendent variables into, uncorrelated new variables (axes), also known as principal components (PCs) [17,18]. In this method, the information of input vari-ables will be presented without much loss of infor-mation [19,20]. This approach can be implemented as follows:

For a given p-dimensional dataset X, the m principal axes T1,T2,…,Tm, where 1 ≤ m ≤ p, generally, T1,T2,…,Tm can be determined by the m leading eigenvectors of the sample covariance matrix:

( ) ( )μ μ=

= − −1

1 N Ti i

i

S x xN

(1)

where xi ∈ X, μ is mean of the sample, N is the num-ber of data points in the sample, so that:

λ= ∈, 1,2,...,i i iST T i m (2)

where λi is the i-th largest associated eigenvalue of the sample covariance matrix. The m principal com-ponents of a given observation vector xi ∈ X is given by:

= = = 11 2 2, ,..., , ,...,T T T T

m my y y y T x T x T x T x (3)

where y is the vector containing the m principal com-ponents of x. Thus, each original data vector can be represented by its principal component vector with dimensionality m. The eigenvectors with the highest eigenvalues are projected into space. In mathematical terms, PCA involve the following major steps:

1) Standardization of the variables X1,X2,...,Xp by the Z matrix:

( )−= ij jij

j

x xZ

s for i = 1,2,...,n and j = 1,2,..,q (4)

where jx and js are, respectively, the mean and standard deviation of the generic variable jx .

2) Calculation of the Kaisere-Meyere-Olkin (KMO): KMO index can be between 0 and 1. The index value 0.5 or more is considered determines the PCA can act efficiently [21]:

=+

2

2 2ij

ij ij

rKMO

r a (5)

where ijr is the correlation coefficient between vari-ables i and j and ija is the partial correlation coefficient between them.

3) Calculation of the difference between the sample values and the means of input data set.

4) Calculation of the variance–covariance matrix. 5) Calculation of the eigenvalues and eigen-

vectors using covariance matrix. 6) Determination of principal components (PCs).

Vector quantization (VQ) method

Quantization is a survey process from an unlim-ited set of scalar or vector quantities by a limited set of scalar or vector quantities. Two types of quanti-


383

zation techniques exist: scalar quantization (SQ) and VQ. SQ deals with the quantization of samples on a sample by sample basis, while VQ deals with quant-izing the samples in groups called vectors. VQ is a data compression method based on the principle of block coding. Using VQ, the training time for input parameters of predictive model is greatly reduced. The most influential gains here are the robustness of such systems [22].

The prediction speed is very important in soft sensors design. Therefore, In order to speed up the training time and reliability prediction of SVR model, the VQ technique is applied for data compression. The main goal of this method is to simplify the training set and increasing the prediction accuracy.

VQ method reduces the size of the training data-set. In Vector quantization the data is quantized in the form of contiguous blocks called vectors rather than individual samples. The VQ maps a vector x of K-dim-ensional in the vector space Rk to another vector y of K-dimensional that belongs to a finite set C (code book) of output vectors (code words) [23]. In this method K-dimensional input vectors are derived from input data {X} = {xi: i = 1,2,...,N}. Data vectors are quantized into a finite set of code words {Y} = {yj: j = = 1,2,…,K}. Each vector yj is called a code vector or a codeword and the set of all the code words is called a code book where the overall distortion of the system should be minimized. The purpose of the generated code book is to provide a set of vectors which gener-ate minimal distortion between the original vector and the quantized vector.

Therefore, VQ comprises of three stages: 1) code book generation, 2) vector encoding and 3) vec-tor decoding. It works by encoding values from a mul-tidimensional vector space into a finite set of values from a discrete subspace of lower dimension.

The generation of code book is the most impor-tant process that decides the performance of vector quantization. The aim of code book generation is to find code vectors (code book) for a given set of train-ing vectors by minimizing the average pair-wise dis-tance between the training vectors and their corres-ponding code words [24].

Each vector is compared with a collection of representative code vectors, = cˆ , 1,...,iX i N taken from a previously generated code book. Best match code vector is chosen using a minimum distortion rule. To minimize the distortion, the following formula is used to determine the distance between two code words:

=

= − 21

1ˆ ˆ( , ) ( )n

i ii

d X X x xN

(6)

where ˆ( , )d X X denotes the distortion incurred in rep-lacing the original vector X with the code vector X̂ .

Support Vector Regression (SVR)

The SVR is a computational tool that has rec-ently received much attention in soft sensor design due to its successes in building nonlinear data-driven models [26]. SVR has more popularity over ANN due to having many attractive features and promising empirical performance. The main difference between conventional ANN and SVR lies in the risk minimiz-ation principle. Conventional ANN implement the empirical risk minimization (ERM) principle to mini-mize the error on the training data, while SVR models are based on the Structural Risk Minimization (SRM) principle which equips them with greater potential to generalize. Therefore the most important features of SVR are: 1) excellent generalization capability, 2) solving the high-dimension problems, 3) avoiding from local minima and the over fitting phenomenon, 4) does not require to determine network topology in advance and 5) needs fewer a priori-determined para-meters than ANN. These aspects of SVR make it a more generalizable tool, permitting more robust pre-diction despite a small number of learning samples [27-29].

SVR is utilized to determine a nonlinear relation of the form y = f(x) between the vectors of observation x and the desired y from a given set of training samples. A number of cost functions such as the Laplacian, Huber’s, Gaussian, and ε-insensitive can be used for the SVR formulation. Among these, the robust ε-insensitive loss function (Lε) is more common [30, 31]:

εε ε − − − ≥− =

( ) ( )( ( ) )

0 Otherwise

f x y f x yL f x y (7)

where ε is a precision parameter representing the radius of the tube located around the regression func-tion, f(x) (Figure 2).The goal in using the ε-insensitive loss function is to find a function that fits the current training data with a deviation less than or equal to ε. C and ε are user-defined parameters in the empirical analysis. A penalty parameter C > 0 is a parameter determining the trade-off between generalization abil-ity and accuracy in the training data, while the para-meter ε defines the degree of tolerance to errors. The optimization problem can be reformulated as:


384

( )ξ ξ− +=

+ +21

1Min

2

l

i ii

w C (8)

That is subject to the constraints given below:

( )( )

ε ξ

ε ξ

ξ ξ

− + ≤ += + − ≤ +

≥

*

*

,

,

, 0

i i i

i i i

i i

y w x b

y w x b y

The positive slack variables ξ and ξ* represent the distance from actual values to the corresponding boundary values of the ε-tube. The nonlinear model involves the inner products of feature vectors in a high dimensional feature space. For SVR based models, four different kernel functions, including linear, quadratic, Gaussian, and polynomial are used. Generally, the application of Gaussian function is shown to yield a better prediction performance [32]:

σ−

= −

2

2( , ) exp( )2i jx yK x y (9)

In order to build a SVR model efficiently, the SVR parameters must be specified carefully. These parameters include the kernel function, regularization parameter C, bandwidth of the kernel function (σ2) and the tube size of ε-insensitive loss function (ε).

Model development

Figure 3 represents the structure of the pro-posed method. The proposed system consists of two stages. One is the development a pre-processing step based on the VQ or PCA. The other is the SVR model and estimation the trained model.

In the first stage, experimental data taken from the setup were divided into two distinct sets including

training and testing data. From a total of 294 experi-mental data, 240 data were used for training, and 54 for the testing set. Then, the PCA/VQ technique was implemented on the training and testing data.

In the second stage, the k fold cross-validation technique was employed to solve the over fitting prob-lem of the training data. The training dataset was randomly partitioned into k subsets (folds) of approx-imately equal size. Next, k–1 subsets were used for training the model with the selected set of parameters while the model performance was measured by the only remaining subset (validation dataset). This pro-cedure was repeated k times in a way that each subset was used as a validation subset once the others performed the role of training dataset. Finally, the overall model generalization ability for each set of parameters was estimated by averaging the perfor-mance values obtained over the k trails. In this paper, AARE is selected as the performance criteria whereas a 5-fold cross-validation is utilized to eval-uate the performance of the model. Meanwhile, the input space is transformed into feature space by means of the Radial Basis Function (RBF) kernel.

In this study, the LIBSVM package was emp-loyed for developing SVR model. The implementation was carried out in MATLAB 7.10 simulation software [34]. The experimental results were obtained using a personal computer equipped with Intel (R) Core (TM) 2 CPU (3.0 GHz) and 3.25 GB of RAM.

The main steps of model development were as follows:

1. Dividing data into two parts, namely training and test data;

2. Applying cross validation technique (the 5-fold cross validation technique was used);

3. Selecting support vector machine (the ε-SVR model was used);

Figure 2. A schematic diagram of SVR using ε-sensitive loss function (with permission from the publisher [33]).


385

4. Selecting the type of core kernel (Gaussian kernel was used);

5. Optimizing model parameters (C and g (1/2σ2) using pattern search (PS) algorithm;

6. Validating of model and prediction of the results.

Model comparison criteria

The performance criteria and their calculations for comparison of different approaches including ave-rage absolute relative error (AARE) and squared correlation coefficient (R2) were applied:

( )=

−= 1

ˆ1 n i ii

i

y yAARE

n y (10)

( )( )

==

=

−= − =

−

2

2 12 1

1

ˆ 11 ,

nni ii

in iii

y yR y y

ny y (11)

RESULTS AND DISCUSSIONS

Tuning parameters of SVR model

For the SVR model, the Grid Search Method (GSM) is the most common method to determine appropriate values of hyper parameters [35]. The GSM does not consider all the values for parameters in parameters space. Hence, it is unable to converge to the global optimum. The accuracy of the GSM depends on the parameter range in combination with

the chosen interval size. Therefore, this method is quite time-consuming and depends on the selection of boundary parameters. Sensitivity analysis of the hyper parameters of the SVR model is shown in Figure 4.

In this figure, the x and y-axes are log2 C and log2 g, respectively. The z-axis is the AARE. In this study, a typical wide range is selected that could cover a broad range of analysis with adequate inc-remental size which is not too big to reduce the accu-racy and not too small to make CT too long. log2 C and log2 g varied within [-5,20] and [-30,20], res-pectively, with incremental sizes of 3 and 2. It can be seen that by changing the hyper parameters (C,g), the AARE varies in a wide range. Since ε has little effect on ARRE it is assumed to be 0.01. It shows that tuning of the hyper parameters greatly influences the prediction accuracy.

Since the accuracy of SVR model depends on a proper setting of SVR hyper-parameters [36], pattern search (PS) has been used in conjunction with SVR to find an optimum set of hyper parameters for SVR model [37].

The PS method is a class of direct search method to solve nonlinear optimization problems. The algorithm calculates the function values of a pattern and tries to find the minimum. For hyper-parameter optimization with PS algorithm the procedures was summarized as follows:

1. Parameters setting, set iteration i = 0.

Figure 3. Block diagram of the proposed integrated model and SVR.


386

2. Set iteration i = i+1 3. Model training: Hyper-parameter optimiza-

tion. 4. Fitness definition and evaluation. 5. Termination: the evolutionary process pro-

ceeds until stopping criteria (maximum iterations predefined or the error accuracy of the fitness func-tion is met). Otherwise, go to step 2.

The PS optimization procedure is a relatively fast alternative for the time consuming grid search approach. For choosing an overall optimal hyper-parameter, the AARE criterion was least for the test set. The optimal values of the C and g (1/2σ2) were obtained to be 20 and 0.017, respectively.

Integrated model results

This paper proposes a novel soft sensor model by integrating a data reduction technique (VQ or PCA) with SVR. The main idea is based on appli-cation of VQ or PCA for generation of a smaller train-ing dataset. In our approach, these two techniques were used in the pre-processing step to make SVR

model more effective. The major advantage of this approach is to train the model using the extracted low-dimensional datasets.

Before utilizing the PCA technique, The KMO index was applied to find out the applicability of PCA. The obtained value of the KMO index was 0.6978, which was above 0.5 and hence PCA implementation was feasible.

Table 2 shows the variance distribution of PCs (PC1-PC6). It is clear from this table that the cumul-ative variance of PC1 to PC4 is 99.98%. In this study, the first four principal components have got more than 95% of the total cumulative variance. Therefore the first four principal components will be sufficient to develop the model and therefore they were selected as the main model parameters. It was observed that only PC5 and PC6 were insignificant within all vari-ables. Moreover, Figure 5 shows Pareto Chart in MATLAB software in which only the first four PCs that have more than 95% of the cumulative distribution were depicted.

Figure 4. Sensitivity analysis of the hyper-parameters for SVR model.

Table 2. Descriptive statistics of the created PCs

PCs Variance Variance proportion, % Cumulative variance proportion, %

1 0.4549 47.2042 47.2042

2 0.2276 23.6176 70.8218

3 0.1549 16.0770 86.8988

4 0.1261 13.0878 99.9866

5 0.0001 0.0099 99.9965

6 0.0000 0.0034 100


387

The main contribution in this work was to imple-ment VQ technique to reduce the dimension feature vector of the training dataset. This technique can be used to reduce the computation time (CT) for soft sensor model, thus soft sensor model was trained on a low-dimensional dense datasets. The aim of apply-ing this method was to reduce the dimension of the training dataset in order to reduce the training time. Therefore, using VQ technique leads to higher accu-racy for sulfur content prediction.

Typical results by integrated VQ-SVR method are shown in the Table 3. As can be seen in Table 3, integration of VQ with SVR model has good prediction

accuracy for the treated gas-oil sulfur content in a wide range.

The proposed integrated approaches were com-pared with SVR. The validity of these methods was evaluated by the statistical parameters (AARE and R2) in Table 4. Moreover, the parity plots for different optimization algorithms integrated with the SVR model are shown in Figure 6.

Comparison of three methods shows that the VQ-SVR model is more accurate and faster than PCA-SVR in predicting sulfur content. Also both approaches (VQ-SVR and PCA-SVR) exhibit high performance in accuracy and CT compared to the

Figure 5. Pareto chart for the first four PCs (PC1–PC4).

Table 3. Typical input and output data for proposed integrated VQ-SVR method

Test No. Qgas oil / 10-6 kg s–1 QH2 / 10-5 m3 h-1 p / kPa TPreheat / °C Tin / °C Tout / °C Sexp / ppm Spre / ppm

1 4.56 3.55 5000 170.0 370.0 374.0 151 160

2 2.85 3.55 5000 120.8 320.0 324.3 3669 3392

3 3.71 3.55 5000 119.5 320.0 322.3 4362 3904

4 3.71 6.91 5000 145.6 345.0 347.0 1393 1176

5 3.71 11.3 5000 169.2 370.0 376.5 64 74

6 4.56 3.55 5000 120.3 320.0 322.2 4845 4469

7 4.56 3.55 5000 144.3 345.0 348.6 1948 1524

8 2.85 3.55 7000 119.4 320.0 324.2 3738 3644

9 2.85 3.55 7000 144.2 345.0 348.7 924 843

10 4.56 11.3 7000 144.4 345.0 347.4 2116 2063

11 4.56 11.3 7000 169.8 370.0 372.3 211 307

12 3.71 11.3 7000 119.7 320.0 323.2 4382 4360

13 3.71 6.91 7000 145.0 345.0 348.2 1546 1287

14 4.56 3.55 7000 145.9 345.0 347.7 2113 2473

15 2.85 11.3 7000 119.5 320.0 320.4 3681 3650


388

conventional algorithm based on SVR. The PCA red-uced the dataset used as an input to SVR method and increased the accuracy. Obviously, the reduction of the input vector dimensions has resulted in the reduction of SVR size and hence has decreased the training and testing time. However, VQ performed greatly in accuracy prediction of sulfur content and training time of datasets for SVR model.

CONCLUSION

Determination of ultra low sulfur content of treated gas oil in HDS process is highly important to increase the productivity and efficiency of refinery operations. Sulfur content prediction by online hard-ware analyzers is mostly expensive with high main-tenance cost. SVR is a new model developed for soft

Table 4. Effect of VQ and PCA on predictive performance of SVR model

No. Method R2 AARE / % CT(s)

Training data Test data Training data Test data

1 SVR 0.978 0.970 16.74 22.32 243

2 PCA-SVR 0.989 0.988 9.57 11.86 126

3 VQ-SVR 0.997 0.995 5.32 6.68 85

Figure 6. The parity plots for different models.


389

sensor applications in process engineering. However, the industrial applications have different approaches. Due to the huge size of industrial data sets used by soft sensors, training and validation time of SVR model has become a challenging issue. In this paper, an accurate and reliable data-driven soft sensor has been developed by means of a SVR integrated with a data compression technique (VQ/PCA) to predict the sulfur content in an industrial HDS process. The pro-posed integrated technique incorporated two stages: 1) the data compression stage and 2) the prediction stage. First, the PCA and VQ were applied to reduce the dimensionality of the dataset and then a SVR model was developed. In order to evaluate the per-formance of the proposed approach, a wide range of experimental data according to real condition of the refineries were taken from a HDS setup. Therefore, the results are generalizable to real processes in the refinery.

The performance of both VQ–SVR and PCA- -SVR methods were compared with that of SVR. Some statistical criteria (AARE, R2) were used to evaluate the prediction performance of models. Com-parison of the results indicated that the best predict-ion accuracy could be obtained using VQ–SVR method, i.e., highest R2 (0.995) and lowest AARE (0.0668). Moreover, it was observed that VQ-SVR had a shorter CT in comparison with PCA-SVR.

Abbreviations

AARE Average absolute relative error ASTM American Society for Testing and Materials KMO Kaisere-Meyere-Olkin CT Computation time SVR Support vector regression PCA Principal Component Analysis HDS Hydrodesulfurization PS Pattern Search GSM Grid search method VQ Vector quantizaion RBF Radial basis function

Nomenclature

ija Partial correlation coefficient X̂ Code vector ˆiy The predicted value

iy The observed values ijr Correlation coefficient

K(x,y) Kernel function Sexp Sulfur content from laboratory tests Spre Predicted sulfur content C Regularization parameter (hyper-parameter) k Subsets (folds) expi Actual values

prei Predicted values Lε Loss function PCs Principal component W Weight vector R2 Squared correlation coefficient Tm m principal axes b0 Intercept QH2 Hydrogen flowrate QGas oil Gas oil flowrate Greek symbols and subscripts

μ Mean of the sample σ Width of kernel of radial basis function ε Precision parameter (hyper-parameter) ξ ξ *,i i Slack variables λi i-th largest associated eigenvalue

REFERENCES

[1] E.A. Medina, J.I.P. Paredes, Math. Comput. Modell. 49 (2009) 207-214

[2] F.S. Mederos, J. Ancheyta, Appl. Catal., A 332 (2007) 8- -21

[3] P. Kadlec, R. Grbic, B. Gabrys, Comput. Chem. Eng. 35 (2011) 1-24

[4] H. Kaneko, K. Funatsu, Ind. Eng. Chem. Res. 50 (2011) 10643-10651

[5] J. Ji, H. Wang, K. Chen, Y. Liu, N. Zhang, J. Yan, J. Taiwan Inst. Chem. Eng. 43 (2012) 67-76

[6] V.N. Vapnik, The nature of statistical learning theory, Springer, New York, 1995, p.93

[7] G. Liu, D. Zhou, H. Xu, C. Mei, Expert. Syst. Appl. 37 (2010) 2708-2713

[8] P. Niu, W. Zhang, Neurocomputing 78 (2012) 64-71 [9] H. Son, Ch. Kim, Ch. Kim, Automat. Constr. 27 (2012) 60-

-66

[10] I. Babaoglu, O. Findik, M. Bayrak, Expert. Syst. Appl. 37 (2010) 2182-2185

[11] D.Z. Antanasijevic, M.D. Ristic, A.A.P. Grujic, V.V. Pocajt, Int. J. Greenhouse Gas Con. 20 (2014) 244-253

[12] V.M. Janakiraman, X. Nguyen, D. Assanis, Appl. Soft Comput. 13 (2013) 2375-2389

[13] G.R. Lloyd, R.G. Brereton, R. Faria, J.C. Duncan, J. Chem. Inf. Model. 47 (2007) 1553-1563

[14] L.I. Smith, A tutorial on principle component analysis, http://www.cs.otago.ac.nz/cosc453/student_tutorials/principal_components.pdf (2002)

[15] K. Polat, S. Gunes, Expert Syst. Appl. 34 (2008) 2039- -2048

[16] M. Aminghafari, N. Cheze, J. Poggi, Comput. Stat. Data Anal. 50 (2006) 2381-2398

[17] I.T. Jolliffe, Principal Component Analysis, 2nd ed., Sprin-ger, Berlin, 2002, p.29

[18] C. Sarbu, H. F. Pop, Talanta 65 (2005) 1215-1220


390

[19] J.D. Wu, C.T. Liu, Expert Syst. Appl. 38 (2011) 14284- -14289

[20] I. Lindsay, A. Smith, A tutorial on principal components analysis, http://kybele.psych.cornell.edu/ edelman/Psych- -465-Spring-2003/PCA-tutorial (2002)

[21] S. Shrestha, F. Kazama, Environ. Modell. Software 22 (2007) 464-475

[22] K. Sayood, Introduction to Data Compression, Fourth ed., Morgan Kaufmann, University of Nebraska, 2012, p.295

[23] A. Gersho, R.M. Gray, Vector Quantization and Signal Compression, Kluwer Academic Publishers, Boston, MA, 1992, p.309

[24] M.H. Horng, Expert Syst. Appl. 39 (2012) 1078-1091 [25] J. Yin, Procedia Environ. Sci. 8 (2011) 173-178 [26] S.B. Chitralekha, S.L. Shah, Can. J. Chem. Eng. 88

(2010) 696-709

[27] K.Y. Chen, Reliab. Eng. Syst. Saf. 92 (2007) 423-432 [28] V. Vapnik, S. Golowich, A. Smola, Advances in Neural

Information Processing Systems, Cambridge, MA, Vol. 9, 1997, pp. 281-287

[29] C. Bergeron, F. Cheriet, J. Ronsky, R. Zernicke, H. Labelle, Eng. Appl. Artif. Intel. 18 (2005) 973-983

[30] F. Si, C.E. Romero, Z. Ya, Z. Xu, R.L. Morey, B.N. Liebowitz, Fuel Process. Technol. 90 (2009) 56-66

[31] S. Zaidi, Chem. Eng. Sci. 69 (2012) 514-521 [32] N. Cristianini, J. Shawe-Taylor, An Introduction to Sup-

port Vector Machines and Other Kernel-based Learning Methods, Cambridge University Press, Cambridge, 2000, p. 112

[33] S. Nandi, Y. Badhe, J. Lonari, U. Sridevi, B.S. Rao, S.S. Tambe, B.D. Kulkarni, Chem. Eng. J. 97 (2004) 115–129

[34] C.C. Chang, C.J. Lin, LIBSVM: A Library for Support Vector Machines, Software,http://www. csie.ntu.edu.tw/ /cjlin/libsvm/, (2001)

[35] B. Scholkopf, A.J. Smola, Learning with kernels, MIT Press, Cambridge, 2002, p.187

[36] S.K. Lahiri, N.M. Khalfe, Chem. Ind. Chem. Eng. Q. 15 (2009) 175-187

[37] M. Momma, K.P. Bennett, in Proceedings of the Second SIAM International Conference on Data Mining, Arlington, VA, USA (2002), pp. 261-274.

SAEID SHOKRI1

MOHAMMAD TAGHI SADEGHI1

MAHDI AHMADI MARVAST2

SHANKAR NARASIMHAN3 1Department of Chemical Engineering,

Iran University of Science and Technology (IUST), Tehran, Iran

2Process & Equipment Technology Development Division, Research

Institute of Petroleum Industry (RIPI), Tehran, Iran

3Department of Chemical Engineering, IIT Madras, Chennai, India

NAUČNI RAD

PREDVIĐANJE SADRŽAJA SUMPORA U PROCESU HIDROSULFURIZACIJE INTEGRACIJOM ANALIZE GLAVNIH KOMPONENTI I VEKTORSKE KVANTIZACIJE SA PODRŽANOM VEKTORSKOM REGRESIJOM

U procesu hidrodesulfurizacije (HDS) veooma je važno precizno predvideti sadržaj sumpora, kako bi se obezbedili pravilan rad i kontrola kvaliteta proizvoda. U tu svrhu, razvijen je pouzdan soft sensor (virtualni senzor) koji koristi podrzanu vektorsku regresiju (SVR). Proučavan je efekat integrisanja vektorske kvantizacije (VK) i analize glavnih komponenti (PCA) na izvršenje ovog sensora. Kao prvo, u fazi prethodne obrade, PCA i VQ tehnike su korišćene za smanjenje dimenzije originalnih ulaznih podataka. Nakon toga, komprimovani podaci su korišćeni kao ulazne promenljive za SVR model. Eksperimentalni podaci iz HDS koraka su iskorišćeni za validaciju predloženog integrisanog modela. Integracija VQ/PCA tehnike sa SVR modelom povećava tačnost predviđanja samog SVR modela. Dobijeni rezultati pokazuju da je integrisana tehnika (VQ-SVR) bolja od PCA-SVR u predviđanju tačnosti. Takođe, VQ smanjuje ukupno vreme obuke i testiranja SVR modela u poređenju sa PCA. Za dalju procenu, performanse VQ-SVR modela su poređene sa SVR modelom. Dobijeni rezultati ukazuju da VQ-SVR model daje najbolje performanse u predviđanju (AARE= 0,0668 and R2= 0,995) u poređenju sa analiziranim modelima.

Ključne reči: analize glavnih komponenti, vektorska kvantizacija, podrzana vektorska regresija, soft senzor, proces hidrodesulfurizacije.




391

MILE VELJOVIC1

SASA DESPOTOVIC1

MILAN STOJANOVIC2

SONJA PECIC3

PREDRAG VUKOSAVLJEVIC1

MIONA BELOVIC4

IDA LESKOSEK-CUKALOVIC1 1Faculty of Agriculture, University

of Belgrade, Zemun-Belgrade, Serbia

2Agricultural Extension Service Institute Tamiš, Pančevo, Serbia

3Economics Institute, Belgrade, Serbia

4Institute of Food Technology, Novi Sad, Serbia

SCIENTIFIC PAPER

UDC 663.4.098:634.8.076

DOI 10.2298/CICEQ140415041V

THE FERMENTATION KINETICS AND PHYSICOCHEMICAL PROPERTIES OF SPECIAL BEER WITH ADDITION OF PROKUPAC GRAPE VARIETY

Article Highlights • The mixture of wort and grape mash is a more nutritious medium for yeast growth than

pure wort • The grape beer is a better source of natural antioxidants than regular lager beer • The rate of yeast growth in grape beer was higher compared with control beer Abstract

Over the last decade, the market of special beers with improved healthy func-tion and/or with new refreshing taste has significantly increased. One of the possible solutions enables mixing beer with bioactive components in grapes responsible for well-known health-promoting action of red wine. The effects of the addition of the Prokupac grape on the physicochemical properties and the fermentation kinetics of the grape beer were studied and the results were compared with a control lager beer. The effect of grape addition on the activity of yeast was also studied. Original extract, alcohol content, degree of ferment-ation, fermentation rate and yeast growth were significantly higher in beers with grapes as a consequence of higher concentration of simple sugars in grapes compared with pure wort. Based on the CIELab chromatic parameters the color of grape beer samples was yellow with certain proportion of redness, while the control beer was purely yellow. The increase in the concentration of grape mash affects the reduction of lightness and yellowness of beers, while the redness of samples was directly proportional with grape quantity. The phenolic content and antioxidant capacity of grape beers was remarkably higher compared to the control beer, which indicates that the grape beer is a better source of natural antioxidants than regular lager beer.

Keywords: beer, grape, phenolic compounds, antioxidants, yeast growth.

Beer is the most popular alcoholic beverage in the world, and probably one of the oldest fermented beverages, dating back more than 8000 years. Since ancient times, many different types of beer and beer-based beverages have been developed at various countries worldwide [1]. Such diversity is caused by wide variety of raw materials and technologies, which are used in their production. According to the Rein-heitsgebot (beer purity law which governing commer-cial brewing in Germany, firstly introduced into Bava-

Correspondence: M. Veljovic, Faculty of Agriculture, University of Belgrade, Nemanjina 6, 11080 Zemun-Belgrade, Serbia. E-mail: [email protected] Paper received: 15 April, 2014 Paper revised: 15 September, 2014 Paper accepted: 5 November, 2014

ria in 1516), beer could be brewed only from water, malted barley, hops and yeast. However, in other countries laws governing beer production are less stringent and brewers have more flexibility, e.g., in selection of carbohydrate sources (adjuncts). Brewing adjuncts are “any carbohydrate source other than malted barley which contributes sugars to the wort”, where cereals (malted or unmalted) and sugar syrups are the most widely used, usually in conjunction with barley malt [2]. In addition, from early times different fruits have also been used in brewing as sources of fermentable extract and as flavoring agents. Further-more, because the grains do not host naturally occur-ring yeast, many ancient brewers inoculated the wort by adding fruit, wine or honey [3].

M. VELJOVIC et al.: THE FERMENTATION KINETICS… Chem. Ind. Chem. Eng. Q. 21 (3) 391−397 (2015)

392

Such a wide range of raw materials that can be used in the production of beers and beer-based pro-ducts provides a great opportunity for brewers to con-quer new markets and to meet demands of uncon-ventional consumer groups. In recent years, the mar-ket of special beers with improved healthy function and/or with new refreshing taste has significantly inc-reased [4,5]. The utilization of dietary compounds and natural products as potential disease prevention agents in the form of functional foods has become an important task in current health researches [6]. A number of studies supports the hypothesis that mod-erate drinking of any alcoholic beverage, particularly red wine and beer, significantly reduces the risk of cardiovascular diseases [7]. Such effects can be explained by a high content of natural antioxidants, particularly phenolics compounds [8].

The antioxidants have a very important role in brewing due to their ability to delay and prevent oxi-dation reactions. Antioxidant capacity of beer mainly depends on the content of phenolics and Maillard compounds [9]. Beer polyphenols come from barley (malt) (70-80%) and hops (20-30%), which are basic raw materials for its production. However, besides the influence of raw materials, the total antioxidant con-tent of beer significantly depends on the brewing pro-cess used [10,11]. Phenolic compounds, especially flavonoids and stilbenes, exhibit a number of bioact-ive effects, such as anti-inflammatory, antimicrobial, antiallergic, antithrombotic, anticancerogenic, antimu-tagenic, antiaging and vasodilatory activities [12]. Except for a physiological role, phenolics have a sig-nificant affect to sensorial properties of beer, such as appearance, taste, mouth-feel, fragrance, astringency and bitternes [13]. In addition, various antioxidants (sulfites, ascorbic acid etc.) can be added during the brewing process to improve flavor stability of products [14]. However, minimizing the use of additives and increasing the content of antioxidants from natural sources to improve flavor stability and increase the shelf-life of products are growing trend in food and beverage industry [15,16].

In our previous work, sensorial acceptability and phenolic content of special type of beer produced by fermenting wort with different proportion of grape must were investigated [17]. Since the beer is a more flexible category than wine, such a product is usually considered a specialty beer rather than specialty wine. The obtained results indicated that special grape beers have unique sensorial profile completely acceptable for consumers and significantly higher content of phenolic compounds. Today, several craft breweries, mainly in Belgium and the United States,

produce grape beers. The most famous grape beer producers are Dogfish Head Brewery (USA), Allagash Brewing Company (USA), Cantillon Brewery (Belgium), Paeleman (Belgium), Blue Moon Brewing Company (USA), etc.

Dynamic of fermentation is one of the most important parameter in the beer production. Addition of grape in the fermenting medium has a great impact on the rate of fermentation, because it contains a higher content of fermentable sugars, mostly glucose and fructose. The main objective of this study was to investigate the influence of the addition of the Pro-kupac grape on the physicochemical properties and the fermentation kinetics of the grape beer. The effect of grape addition on the activity of yeast was also studied.

EXPERIMENTAL

Prokupac, Serbian autochthonous variety used for making table and top quality rose and red wines, was obtained from experimental school estate “Rad-milovac” of Faculty of Agriculture, Belgrade. The all- -malt wort and a bottom-fermenting industrial yeast strain Saccharomyces pastorianus used in this study were obtained from a local brewery collection.

Chemicals

Gallic acid, Folin-Ciocalteu’s phenol reagent, hydrochloric acid, sodium acetate trihydrate, glacial acetic acid and sodium carbonate (anhydrous) were purchased from Merck (Darmstadt, Germany). 2,4,6- -Trypyridyl-s-triazine (TPTZ), ferric chloride hexahyd-rate, 2,2-diphenyl-1-picrylhydrazyl (DPPH), and 6-hyd-roxy-2,5,7,8-tetramethylchroman-2-carboxylic acid (Trolox) were purchased from Sigma-Aldrich (Stein-heim, Germany). Ammonium hydroxide was pur-chased from Fisher Scientific (Loughborough, UK).

Fermentation

The grape was sorted manually, washed in cold water to remove impurities, and the clean grape was destemmed and crushed by hand. The wort and grape mash were mixed in different proportions (70:30 and 80:20) and the pH of obtained mixtures was adjusted to 5.3 with 2 vol.% solution of amm-onium hydroxide. These wort:grape mash ratio was selected after pre-experimental sensorial testing of grape beers with different grape propotion (10, 20, 30, 40 and 50% of grape mash). The fermentation media (4 L) were poured into 5 L laboratory stainless steel fermenters and seeded with yeast suspension such that the concentration of cells was 17 million yeast cells per milliliter of wort. Pitching was performed at

M. VELJOVIC et al.: THE FERMENTATION KINETICS… Chem. Ind. C

issn 1451 - 9372(print) issn 2217 - 7434(online) july … no3_p369-464... · 2017. 2. 3. ·...

Documents