Naslov
Intelligent control of yeast cultivation

Autori
Beluhan, Damir ; Beluhan, Sunčica

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
BIOTECHNOLOGY 2000. The World Congress on Biotechnology Book of Abstract / - Berlin, 2000, 224 - 225

Skup
BIOTECHNOLOGY 2000. The World Congress on Biotechnology

Mjesto i datum
Berlin, Njemačka, 03.09.2000. - 08.09.2000

Vrsta sudjelovanja
Poster

Vrsta recenzije
Međunarodna recenzija

Sažetak
The recognition of characteristic bioprocess metabolic phases by a neuro-fuzzy-expert system made direct inverse training of system outputs to inputs, by artificial neural networks (ANNs), easier and more realistic. An internal model control structure (IMC)1, further sophisticated this hybrid modeling approach that integrates all available a priori knowledge of the baker's yeast production process. The identification of four bioprocess physiological states (lag/log/O2-limitation/maturation) was based on the empirical operator's fuzzy reasoning (if-then rules), and membership functions automatically tuned by a supervised backpropagation learning procedure2. This neural-fuzzy classification model consisted of four applied linguistic rules on two state variables: oxygen uptake rate and liquid volume. For each phases separate inverse ANN models, trained by the off-line generalized learning approach3, were used for estimation of the current molasses feed rate FEED(t). This mapping was based on measurements of the 'desired' respiratory quotient, RQ(t+1), with sampling period of 1 minute, the past molasses feed rate, FEED(t-1,...t-T), the current and past oxygen uptake rate, OUR(t, t-1,...t-T), carbon dioxide evolution rate, CER(t, t-1,...t-T), concentration of ethanol, EtOH(t, t-1,...t-T) and volume, V(t, t-1,...t-T). The IMC structure was necessary because of inaccuracies of the inverse ANN model, unmeasured process environmental disturbances and internal instability4. Therefore, the negative feedback signal, a difference between the measured process output and noise free prediction estimated by a second feedforward ANN model (applied in parallel connection to a process), was superimposed to the input signal (the desired RQ at the next time step) of the driving inverse ANN controller. The applied neural networks had to capture information from patterns that exist over time and hence, static neural networks were extended to dynamic networks that have short-term memory structures, such as tapped delay line memory structure or a Gamma memory that provides a recursive memory of the input signals past. This hybrid control algorithm of fed-batch yeast cultivation process was successfully realized on a laboratory scale bioreactor under a commercial 'Simatic M7-400' control system.

Izvorni jezik
Engleski

Znanstvena područja
Elektrotehnika, Prehrambena tehnologija

POVEZANOST RADA