25 WSEAS Int. Conf. on DYNAMICAL SYSTEMS nd CONTROL, Venice, Itly, November 2-4, 25 (pp24-246) Modeling nd Optimiztion of Olive Stone Drying Process L. KIRALAKIS, N. C. TSOURVELOUDIS Deprtment of Production Engineering nd Mngement Technicl University of Crete University Cmpus, Chni GREECE Abstrct: - Olive oil cn be extrcted from dried olive stones. Drying of olive stones is procedure difficult to model. The equtions describe het trnsfer inside the drying cylinder re highly non liner nd therefore the mthemticl models used to simulte the drying process re complicted. Here we suggest fuzzy nd neurofuzzy techniques to control the drying procedure. A fuzzy controller is designed bsed on vilble expertise, while neuro-fuzzy controller is built using the Adptive Neuro Inference System () bsed on experimentl dt. Both controllers tested in vrious opertion conditions nd extensive comprtive results re presented. Key-Words: - Rotry dryer, Olive stones, logic control, Process control Introduction Dryers cn be used to remove wter from solid substnces primrily by introducing hot gses into drying chmber. Among vrious dryer types, rotry dryers re the most commonly used in minerls nd food industry. Rotry dryers consist of horizontlly inclined rotting cylinder. The mteril, which is fed t one end nd dischrged t the other end, is dried by contct with heted ir, while being trnsported long the interior of the cylinder. The rotting cylinder cts simultneously s the conveying device nd stirrer, s shown in Fig.. Hot ir Wet olive stone Dry olive stone Exhust ir Fig. : Schemtic representtion of cross-flow rotry dryer In this pper we present novel pproch for the control of the rotry drying process pplied to olive stones. It is known tht the mthemticl modelling of rotry drying is rther complicted nd the dynmics involved re non-liner [], [6]. The proposed pproch is bsed on fuzzy logic nd neuro-fuzzy techniques. logic nd neurofuzzy techniques hve been proven tht cn ccommodte these kinds of procedures []-[4]. A fuzzy logic controller is designed bsed on the knowledge cquired from humn experts. A neurofuzzy controller is designed bsed on numericl dt of the rel system. These dt sets were used to trin Sugeno-type fuzzy system through the Adptive Neuro Inference System () [7]. The pper is orgnized s follows. Section 2 describes the drying process of olive stones in rel fctory. Section 3 describes the mthemticl modeling of the drying process. In this section two different pproches for the controller design re described. The fuzzy pproch bsed on Mmdni type controllers nd the neuro-fuzzy pproch bsed on Tkgi-Sugeno type controllers. In section 4 experimentl nd comprison results re described. Finlly, section 5 comments on the proposed pproch nd suggests potentil future reserch topics. 2 Olive Stone Drying Process Wet mss of olive stones is vilble in lrge quntities in olive oil mills fter the first extrction of oil. Olive stones still contin oil, which cn be chemiclly subtrcted from the dehydrted/dried stones. The olive stones re dried by contct with heted ir, while being trnsported long the interior of rotting cylinder (Fig. ), with the rotting shell cting s the conveying device nd stirrer. Moisture of olive stone t the input of drying cylinder my vry from 4% to 54%. The drying process hs to reduce this moisture to %. The reduction of moisture is criticl becuse ffects the qulity of the
25 WSEAS Int. Conf. on DYNAMICAL SYSTEMS nd CONTROL, Venice, Itly, November 2-4, 25 (pp24-246) finl product s well s the sfety of the plnt. Vlues of finl moisture bove % re highly ssocited with hexne retention in the finl product, which is dngerous for public helth. On the other hnd, low (below %) moisture levels increse the chnce of fire inside the rotry dryer. The drying procedure is described briefly s follows. The olive stone is fed into the rotry dryer in rte up to 7 tons per hour. The temperture in the dryer, which is usully mde from steel, my exceed 7 o K. The flow of the ir inside hs the sme direction with the dried mteril. Olive stones re dried minly by contct with heted ir nd surfces inside the rotting cylinder. The rotry dryer except from drying the stone is lso trnsferring it, s shown in Figure. The wings re used to stir the stone inside the dryer (see dryer s cross section in Figure ). The product is moving becuse of the rottion of the dryer nd its slight inclintion (bout two degrees). The out going dry stone is crried from the dryer for dditionl processing. 3 Modeling nd Control Rotry drying cn be mthemticlly described by generl differentil eqution, in which moisture is function of time nd dimension s follows ( ) ( ) xlt, xlt, + vt () = f( xlt,,) () t l where, x represents the moisture of the olive stone, l is the xil coordinte of the dryer, v: is the liner velocity of the olive stones in the dryer nd t is drying time. It is known tht if () is used under relistic ssumptions, (such s, vrying drying ir velocity, unknown size distribution of olive stones nd not constnt wter evportion long the dryer) leds to complex time vrying dynmic model which involves prmeters tht is difficult to be ccurtely mesured []. The gol of the control system is to chieve the desired moisture in the dried product. A prcticl wy to mesure the percentge of moisture in the finl product, or simply, the finl moisture x f, is weight of remining wter x f =, (2) dried product weight The finl moisture x f my be ssocited with physicl prmeters of the drying process s in the following eqution: x f T = f( x ) + f 2 ( A) + f 3 ( ), (3) l where, x α is the initil moisture of the olive s stone, A represents the quntity of olive stone enters the rotry dryer, T is the temperture in the drying cylinder where l is its length. Nottion fi () i is used to represent generic function. According to eqution (3), the vlue of finl moisture depends on the initil moisture (which is given nd cnnot be controlled), the quntity of product in the rotting cylinder nd the het trnsfer rte inside the cylinder. In order to control the system, the prmeters tht ffect the drying process must be identified. For the system under study, these prmeters re the temperture through the rotting dryer nd the quntity of the olive s stone tht is entering the system (feed rte) t given time intervl. The control methodology presented in the next prgrph, mkes djustments to the control prmeters (nmely, temperture nd feed rte) bsed on the observtion of the following input vribles:. The difference (error) between the desired (trget) moisture of the system nd the current moisture. 2. The initil moisture tht is the moisture of mteril entering the system. 3. Logic Approch logic is widely used to fcilitte problems of controlling rotry dryers [3], [4]. Most of the rotry dryers re controlled mnully bsed on the experience of the opertor. Tody it is known tht fuzzy logic offers the mthemticl frmework tht llows for simple knowledge representtion of the production control principles in terms of IF-THEN rules. The IF-prt describes conditions under which the rule is pplicble nd forms the composition of the inputs. The consequent (THENprt) gives the response or conclusion tht should be tken under these conditions. A two-input (ntecedent) rule of the Mmdni type hs the form: IF X is A AND Y is B THEN Z is C, where X, Y re the input nd Z is the output vrible, nd A, B nd C their linguistic vritions, respectively, tht re fuzzy sets with certin membership functions [5]. The crisp control ction is obtined through defuzzifiction method, which in most pplictions, clcultes the centroid of the output fuzzy set. The controller designed here djusts the temperture of the fumes nd the quntity of the mteril entering the system. An exmple of the expert knowledge describes the control objective cn be summrized in the following sttement: If the difference between the desired moisture of olive
25 WSEAS Int. Conf. on DYNAMICAL SYSTEMS nd CONTROL, Venice, Itly, November 2-4, 25 (pp24-246) stone nd its current moisture is AND the moisture of olive s stone entering the system is THEN the temperture of fumes should be AND the feed rtes Reltively. The bove knowledge my be more formlly represented by fuzzy rules of the following form: IF error is LA (k) AND x is LB (k) THEN T is LC (k) AND A is LD (k) where k is the rule number, LA is linguistic vlue of the vrible error with term set A= {Dngerous,, Reltive, Perfect,, Reltive }, LB is linguistic vlue of the vrible x with term set B= {, Medium, }, LC is linguistic vlue of the vrible T (temperture) with term set C= {Very,, Reltive, Medium, Reltive,, Very } nd LD is linguistic vlue of the vrible A (quntity) with term set D= {Very,, Reltive, Medium, Reltive,, Very }. The temperture T t given time instnt is Tµ C IS error x µ C T= f (, ) = ( T), (4) ( T) while the quntity A tht enters (feed rte) the dryer is Tµ D IS error x µ D ( A) A= f (, ) =, (5) ( A) where f IS( error, x ) represents fuzzy inference system, tht tkes s inputs the out coming moisture error nd the initil moisture xα, of the incoming olive s stone. The membership functions µ C ( T ) nd µ ( A ) which re given by D µ C ( T) = mx min[ µ AND ( error, x ), µ ( k) ( error, x, T)] (6) error, x µ D ( A) = mx min[ µ AND ( error, x ), µ () k ( error, x, A)] (7) error, x where µ AND ( error, x ) is the membership function of the conjunction of the inputs while µ ( k )( error, x, T ) nd µ ( k )( error, x, A) re the membership functions of the k-th ctivted rule. Tht is µ AND ( error, x ) =µ A ( error ) µ B ( x ) nd µ () k ( error, x, T) = f [ () k ( error ), () k ( x), µ () k ( T) ] µ µ, () LA LB LC µ () k ( error, x, A) = f [ () k ( error), () k ( x), () k ( A)] µ µ µ. () LA LB LD () In equtions (), (), (), µ A( error ) represents the membership function of the moisture devition, tht is e e t, nd µ ( x ) is the membership trget ( ) function of the initil moisture. The membership functions tht re used s inputs vribles re presented in Fig. 2 nd 2b. Fig. 2 models the vrible moisture error nd Fig. 2b presents the moisture of olive stones entering the system. Fig. 2c nd 2d show the membership functions of the drying ir temperture nd quntity of mteril entering the system, respectively.. Dngerous Reltive Perfect -45-4 -35-3 -25-2 -5 - -5 5 Error. Very Reltive Reltive B. Medium 4 4 5 5 52 53 54 initil moisture ) b) Medium Reltive Very 35 4 45 5 55 6 65 Temperture. Very Reltive Medium Reltive Very 2 3 4 5 6 7 Feed Rte c) d) Fig. 2: Membership functions of vribles: ) Moisture error, b) Initil moisture, c) Temperture nd d) Feed Rte. 3.2 Neuro- Approch The correct choice of membership function is by no mens trivil nd plys crucil role in the success of fuzzy control pplictions. A point of criticism for the fuzzy controller presented in the previous prgrph, is tht membership functions re heuristiclly selected, bsed on tril nd error experimenttion. In this prgrph, we utilize well known systemtic procedure, the Adptive Neuro Inference System () [7], for the design of the fuzzy controller. A similr ttempt for lbortory size rotry dryer is described in []. utomticlly constructed fuzzy controller using set of bout input-output dt. Experiments were conducted t the industril size rotry dryer of ABEA S.A., Chni, Crete, Greece. The drying cylinder length is 22m, its dimeter is 2.5m nd rottes bout 3.5 times per minute. Neuro-fuzzy controllers were tested for vrious shpes of membership functions. In ll cses presented in Tble nd Tble 2, the trining lsted the sme time epochs. All controllers tested hd the sme number of membership functions per input/output vrible.
25 WSEAS Int. Conf. on DYNAMICAL SYSTEMS nd CONTROL, Venice, Itly, November 2-4, 25 (pp24-246) Experiments were conducted, for vrious temperture men vlues, feed rtes nd initil moisture levels. Tble 2 presents the men control error for vrious membership function shpes. Tble : Test results for initil temperture 5 o K nd feed rte 6 tons per hour. Initil Moisture Initil Moisture Initil Moisture 4-5% 5-52% 52-54% Membership Men Vrince Men Vrince Men Vrince function Type Error Error Error Bell-shped.6.54 -..45 -.34.77 Tringulr.4.56.24.4 -.5.74 Guss.6.274.34.5 -.7.2 Tble 2: Test results for initil temperture 55 o K nd feed rte 7.5 tons per hour. Initil Moisture Initil Moisture Initil Moisture 4-5% 5-52% 52-54% Membership Men Vrince Men Vrince Men Vrince function Type Error Error Error Bell-shped..4 -.42.7 -.335.77 Tringulr.6.5.4.6 -.. Guss.5.32.5.46..3 From the experiments conducted we my mke the following observtions: ) The time needed by the controller to pproch the trget vlue (finl moisture %) is the sme for lmost ll test cses. b) Selection of tringulr membership functions results to less oscilltion t the output, in comprison to the other two membership shpes we tested. In most cses the controller with tringulr membership functions performed better thn the others. In the comprtive results, presented in the next section, ll membership functions of both (pure fuzzy nd neuro-fuzzy) controllers hve tringulr shpe. 4 Results nd Comprisons Two controllers, the one designed bsed on experience (FUZZY), nd the other bsed on experimentl dt (neuro-fuzzy-) re compred in this section. Extensive comprtive study hs been performed for vrious testing conditions. In the first set of experiments, the initil moisture of the olive stone remined the sme (5-52%).Three test cses re grphiclly presented here. Fig. 3 presents the response of the drying system for Test Cse : initil temperture in the drying cylinder is 55 o K nd olive stone feed rte is tons per hour. Fig. 4 presents the response of the drying system for Test Cse 2: initil temperture in the drying cylinder is 4 o K nd olive stone feed rte is 5 tons per hour. Fig. 5 presents the response of the drying system for Test Cse 3: initil temperture in the drying cylinder is 52 o K, while the olive stone feed rte is 6.5 tons per hour. 26 24 22 2 6 4 6 Fig. 3: Finl moisture vrition for Test Cse (initil temperture: 55 o K, nd feed rte: tons/hour). 5 4 3 7 Fig. 4: Finl moisture vrition for Test Cse 2 (initil temperture: 4 o K, nd feed rte: 5 tons/hour). Moitsure (%).5.5.5.5 7.5 FUZZY Fig. 5: Finl moisture vrition for Test Cse 3 (initil temperture: 52 o K, nd feed rte: 6.5 tons/hour). In the second set of experiments, the initil temperture nd the feed rted remined constnt. Tht is: initil temperture 55 o K nd feed rte tons/hour. The controller s performnce is exmined for three initil moisture vrition internls, nmely, 4-5%, 5-52% nd 52-54%. These intervls represent the ctul vritions of the initil moisture
25 WSEAS Int. Conf. on DYNAMICAL SYSTEMS nd CONTROL, Venice, Itly, November 2-4, 25 (pp24-246) of olive stones. Fig. 6 presents the response of the drying system for Test Cse 4: initil moisture rndomly vries from 4 to 5%. In prctice, depending on the mesuring device ccurcy, the vrition of the initil moisture my look like in Fig. 7..5 5 4 3.5 7.5 Fig. 6: Finl moisture vrition for Test Cse 4 (initil moisture vries from 4 to 5%). Initil 5 4. 4.6 4.4 4.2 4 4. 4.6 4.4 4.2 4 Fig. 7: Initil moisture monitoring for Test Cse 4. Fig. presents the response of the drying system for Test Cse 5. In this cse the initil moisture vries from 5% to 52%, s shown in Fig.. System s response for initil wter content between 52 to 54% is shown in Fig. (Test Cse 6). The sttisticl vlues of the finl moisture, for the test cse 4, tht hve been derived from the two controllers nd Neuro-fuzzy re presented in the Tble 3. Tble 3: Controller s Performnce for different initil moisture levels (initil temperture 55 o K nd feed rte tons/hour) FUZZY Initil Men Error Vrince Men Error Vrince moisture 4-5%.6.5.7.3 5-52% -3.77.2.3 52-54% -..465 -.4.4 7 Fig. : Finl moisture for Test Cse 5 (initil moisture from 5 to 52%). Initil 52 5. 5.6 5.4 5.2 5 5. 5 5 5 Fig. : Initil moisture monitoring for Test Cse 5. 6 4 6 Fig. : The finl moisture for two different vrinces of the initil moisture from 52 to 54%. Some observtion bsed on ll test cses nd sttisticl nlysis presented, mybe the following:. The neuro-fuzzy controller gives smller men error of the finl moisture in two out of the three test cses (Tble 3). 2. The pure fuzzy pproch gives better results when the initil moisture is higher (52-54%). Although the men error in the chieved finl moisture is smller, the vrince is lrger. On the other hnd, the vlue of vrince of the
25 WSEAS Int. Conf. on DYNAMICAL SYSTEMS nd CONTROL, Venice, Itly, November 2-4, 25 (pp24-246) finl moisture tht derived from the neurofuzzy controller is smll in ll the cses, which indictes more stble behvior in ll rnges of initil moisture. 3. Trining of the controller is bsed on dt tht might not model the behvior of the drying system, under ll working conditions. In this sense, the results of the pure fuzzy controller my cover more relistic opertion cses. 5 Conclusions The production of olive oil is the desired outcome of the rotry drying process of olive stones. After the collection of olive stone from the oil mills, it s used s rw mteril for the drying process. The dried stone is mixed with hexne, which results to the subtrction of oil from the stone. Controlling of n industril size rotry dryer is not n esy tsk, minly becuse of its size nd the corresponding long trnsporttion times of the prticles, nd the delys between control ction nd observble results due to these ctions. In this pper, two different techniques were used for the control of the drying process of olive stones. A fuzzy logic controller designed bsed on expert knowledge. For the trining of neuro-fuzzy controller the Adptive Neuro Inference System () ws used bsed on dt of the rel system. A set of experiments ws conducted with the neuro-fuzzy controller nd vrious types of membership functions. The shpe of membership functions nd if-then rules prmeters were tuned from dt. Another set of experiments ws conducted to compre the performnce of the two different controllers t the drying process control. The consequences tht derived from most of tests show the supremcy of the controller tht ws designed bsed on dt. In the future, it will be interesting to incorporte more vribles in the control scheme, such s, the rottion speed of the drying cylinder. [3] M. Tyel, M. R. M. Rizk, H. A. Hgrs, A Logic Controller for Dry Rotry Cement Kiln, Sixth IEEE Interntionl Conference on Systems, 7. [4] M. Aklp, A. L. Dominguez, R. Longchmp, Supervisory Control of Rotry Cement Kiln, MELECON' 4 Mediterrnen Electrotechnicl Conference, 4. [5] D. Drinkov, H. Hellendroom, & M. Reinfrnk, An introduction to fuzzy control, Springer, 2 nd revised edition, 6. [6] F. Incroper & D. DeWitt, Fundmentls of Het nd Mss Trnsfer, Wiley, 4 th edition, 6. [7] J-S R. Jng, C-T Sun, E. Mizutni, Neuro- nd Soft Computing, A computtionl pproch to lerning nd mchine intelligence, Prentice Hll, 7. References: [] L. Yliniemi, Advnced Control of Rotry Dryer, Ph.D. Thesis, Act Universits Ouluensis,. [2] L. Yliniemi, J. Koskinen, K. Leivisk, Dtdriven Modeling of Rotry Dryer, Interntionl Journl of Systems Science, Vol. 34, 5 November-5 December 23, pp - 36.
25 WSEAS Int. Conf. on DYNAMICAL SYSTEMS nd CONTROL, Venice, Itly, November 2-4, 25 (pp24-246)