<p>Fokus ove doktorske disertacije je razvijanje modela<br />zasnovanog na konceptu veštačkih neuronskih mreža<br />za predviđanje i projektovanje mikrofiltracije<br />kultivacionih tečnosti preko ispitivanja mogućnosti<br />primene ovog koncepta za modelovanje fluksa<br />permeata pri različitim uslovim a mikrofiltracij e, u<br />sistemima sa i bez primene hidrodinamičkih metoda<br />poboljšanja fluksa permeata i njihove kombinacije,<br />kao i razvoj modela kojim će se objediniti<br />eksperimentalni rezultati u cilju dobijanja jedne<br />jedinstvene neuronske mreže za simulaciju svih<br />metoda poboljšanja fluksa. Dodatan cilj predstavlja<br />razvoj modela za procenu poboljšanja fluksa u<br />stacionarnim uslovim a usled primene metoda<br />poboljšanja fluksa permeata čija će se adekvatnost<br />proveriti sa energetskog stanovišta.<br />Eksperimentalna ispitivanja su obuhvatila razvoj i<br />validaciju deset različitih modela neuronskih mreža<br />kod kojih su nezavisne ulazne promenljive i njihovi<br />rasponi (transmembranski pritisak, protok suspenzije<br />i protok vazduha) utvrđeni Box-Behnken-ovim<br />eksperimentalnim planom uz dodatne parametre<br />vreme trajanja mikrofiltracije i temperature koji su<br />varirani u zavisnosti od uslova izvođenja postupka<br />mikrofiltracije. Nasuprot tome, za razvoj dinamičkog<br />modela u svojstvu zavisno promenljive veličine<br />razmatran je pad fluksa permeata sa vremenom, dok<br />je za razvoj modela procene efikasnosti primenjenih<br />metoda poboljšanja fluksa permeata razmatran fluks i<br />specifična potrošnja energije u stacionarnim<br />uslovima.<br />Normalizacijom eksperimentalnih podataka izbegla<br />se velika razlika u specifičnim težinskim<br />koeficijentim a pojedinih ulaznih promenljivih i predupredila opasnost da te promenljive pokažu veći<br />uticaj nego što ga imaju u realnosti, a balansiranje<br />efekata nekontrolisanih faktora na izlaznu<br />promenljivu izvedeno je randomizacijom na grupu za<br />obučavanje (70% podataka), grupu za validaciju<br />(15% podataka) i grupu za testiranje (15% podataka).<br />Nestacionarnosti koje utiču na efikasnost algoritma<br />obuke i arhitekture neuronskih mreža izbegnute su<br />ispitivanjem m odela sa pet algoritama obuke<br />(Levenberg-M arkuardt-ov algoritam obuke<br />(trainlm), Bayes-ova regularizacija (trainbr), model<br />rezilientnog povratnog prostiranja (trainrp), model<br />skaliranog konjugovanog gradijenta (trainscg) i<br />model jednostepenog sekantnog povratnog<br />prostiranja greške unazad (trainoss)) i dve<br />sigmoidalne aktivacione funkcije u skrivenom sloju<br />(logistička i hiperbolična tangensna), dok je u<br />izlaznom sloju korišćena linearna aktivaciona<br />funkcija. Svi modeli su optimizovani primenom<br />metode probe i greške sa osnovnim ciljem dobiti što<br />jednostavniju mrežu, odnosno mrežu sa minimalnim<br />brojem skrivenih neurona koja pokazuje najbolju<br />sposobnost generalizacije. Kao indikatori nivoa<br />generalizacije i parametara učinka obuke neuronske<br />mreže ispitivani su koeficijent determinacije (R2) i<br />srednja kvadratna greška (MSE), a koeficijent<br />korelacije (r) je odabran kao dodatni parametar<br />adekvatnosti fitovanja vrednosti utvrđenog i<br />neuronskom mrežom procenjenog fluksa permeata.<br />Najbolju sposobnost generalizacije i predikcije<br />pokazao je model neuronske mreže obučavan<br />Levenberg-M arkuardt-ovim algoritmom. Optimalan<br />broj neurona u skrivenom sloju se kretao od 7 do 13<br />što ukazuje na znatnu kom pleksnost mehanizama<br />koji utiču na fluks permeata kako je i procenjeno<br />postavljanjem hipoteze ove doktorske disertacije.<br />Analiza apsolutne relativne greške pokazala je veoma<br />dobro predviđanje pošto je u rasponu od 81% do<br />100 % podataka imalo grešku manju od 10%, a<br />koeficijent determinacije u rasponu od 0,98091 do<br />0,99976 ukazuje da mreža ne može da objasni manje<br />od 2% varijacija u sistemu. Vrednosti koeficijenta<br />korelacije se kreću u rasponu od 0,99041 do 0,99988<br />što sugeriše na dobru linearnu korelaciju između<br />eksperimentalnih podataka i podataka predviđenih<br />neuronskom mrežom. Pored primene koncepta<br />fitovanja podataka ispitana je i mogućnost procene<br />uticaja pojedinih eksperimentalnih promenljivih na<br />fluks permeata primenom jednačine Garsona, a<br />komparativnom analizom dobijenih simulacionih rezultata na eksperimentalim podacima koji nisu bili predstavljeni neuronskoj mreži potvrđen je<br />generalizacijski kapacitet modela neuronske mreže.</p> / <p>Focus of this doctoral dissertation is to develop<br />a model based on the artificial neural networks<br />concept for predicting and designing cultivation<br />broth microfiltration by examining the<br />feasibility of this concept for modeling<br />permeate flux under different microfiltration<br />conditions, in systems with and without<br />hydrodynamic im provem ent methods, as well<br />the development of a model that will combine<br />the experimental results in order to obtain a<br />single neural network to simulate all methods of<br />flux improvement. An additional goal is the<br />development of a model in quasi steady state in<br />term so fadequacy of flux enhancement methods<br />application, which will be checked from the<br />energy point of view.<br />Experimental tests included the development<br />and validation of ten different models оf neural<br />networks in which the independent input<br />variables and their ranges (transmembrane<br />pressure, suspension flow and air flow) were<br />determined by Box-Behnken's experimental<br />plan with added microfiltration parameters time<br />and temperature, varied depending on the<br />conditions of the microfiltration procedure. In<br />contrast, for the development оf a dynamic<br />model as a dependent variable, the decrease in<br />permeate flux with time was considered, while<br />for the development of a model for evaluating<br />the efficiency оf applied permeate flux<br />im provement methods, flux and specific energy<br />consumption in quasi steady state conditions<br />were considered.<br />Normalization of experimental data avoided a<br />large difference in specific weight coefficients of individual input variables and prevented the<br />danger that these variables show a greater<br />impact than they have in reality, and balancing<br />the effects of uncontrolled factors on the output<br />variable was performed by randomization on the<br />training group (70% o f data), a validation group<br />(15% of data) and a testing group (15% of data).<br />Non-stationarities affecting the efficiency of the<br />training algorithm and neural network<br />architecture were avoided by testing the model<br />with five diferent training algorithms<br />(Levenberg-M arquardt training algorithm<br />(trainlm), Bayesian regularization (trainbr),<br />resilient backpropagation algorithm (trainrp),<br />scaled conjugate gradient method (trainscg) and<br />a one-step secant m ethod (trainoss)) and two<br />sigmoid activation functions in the hidden layer<br />(logistic and hyperbolic tangent), while a linear<br />activation function was used in the output layer.<br />All models are optimized by applying the trial<br />and error method with the basic goal of having<br />the simplest possible network, ie a network with<br />a minimum num ber o f hidden neurons that<br />shows the best ability to generalize.<br />Determ ination coefficient (R2) and mean square<br />error (MSE) were examined as indicators of<br />generalization level and neural network training<br />performance parameters, and correlation<br />coefficient (r) was selected as an additional<br />param eter o f adequacy оf fitting the value of<br />determined and neural network estimated<br />permeate flux.<br />The best ability to generalize and predict was<br />shown by a model of a neural network trained<br />by the Levenberg-M arquardt algorithm. The<br />optimal num ber of neurons in the hidden layer<br />ranged from 7 to 13, which indicates a<br />significant complexity of the mechanisms that<br />affect the permeate flux, as assessed by the<br />hypothesis of this doctoral dissertation.<br />Absolute relative error analysis showed very<br />good prediction as in the range of 81% to 100 %<br />of the data had an error of less than 10 %, and<br />the coefficient of determination in the range of<br />0.98091 to 0.99976 indicates that the network<br />cannot explain less than 2 % variation in the<br />system. The values оf the correlation coefficient<br />range from 0.99041 to 0.99988 suggests a good<br />linear correlation between the experimental data<br />and the data predicted by the neural network. In addition to the application of the concept of data<br />fitting, the relative importance of input variables<br />was also investigated by applying the Garson<br />equation. Comparative analysis of the obtained<br />simulation results on experimental data that<br />were not presented to the neural network<br />confirmed the generalization capacity of the<br />neural network model.</p>
Identifer | oai:union.ndltd.org:uns.ac.rs/oai:CRISUNS:(BISIS)114867 |
Date | 22 October 2020 |
Creators | Nikolić Nevenka |
Contributors | Jokić Aleksandar, Ikonić Bojana, Stamenković Olivera, Grahovac Jovana, Lukić Nataša |
Publisher | Univerzitet u Novom Sadu, Tehnološki fakultet Novi Sad, University of Novi Sad, Faculty of Technology at Novi Sad |
Source Sets | University of Novi Sad |
Language | Serbian |
Detected Language | English |
Type | PhD thesis |
Page generated in 0.0216 seconds