Inteligência computacional aplicada à automação de biorreator para produção de Penicilina G Acilase (PGA)

Abstract

Biotechnology has presented, in the last years, a rapidly growing development. New biotechnological industrial processes are constantly introduced using different microorganisms and/or enzymes. In this context, the application of process control and optimization techniques has become a need for biotechnological based industry. A technological approach was developed in this dissertation to bioreactor monitoring and control operating in fed-batch model to produce the penicillin enzyme G acilase (PGA) by means of wild cepa of the microorganism Bacillus megaterium. This enzyme is of great industrial importance, being used in the manufacture of semi-synthetic β-lactamic antibiotics. This case study presents the main difficulties faced in the control of biological processes in general: the variability of kinetic parameters and the limited availability of on-line information. In order overcome these difficulties, a unconventional architecture was proposed for a dynamic and adaptive controller using filters, some of them developed in this work, and applying Computational Intelligence (CI) methodologies both in direct and hybrid form. The inference for the microbial concentration (Cx) state variable, a very relevant objective for the logic of the controller, was performed by a softsensor that had as input the filtered values of the sensors signals of the molar fractions of CO2 (yco2) and O2 (yo2) in the effluent gases, of the air feeding flow and of agitation velocity. The respiratory quotient (RQ), calculated from these data, was also used by the algorithms of the software developed here. For the Cx inference, the softsensor employed a hybrid intelligent system (HIS) composed by a neural networks ensemble (RNE) and a fuzzy rule based system (FRBS). These techniques were structured to complement each other such that the RNE infers the microbial concentration (Cx) capturing real-time process data (empirical knowledge) and the FRBS corrects this inferred value using phenomenological based knowledge. The obtained results demonstrated a more robust inference by using this architecture, even supporting some degree of extrapolation. Another important operational parameter is the definition of the initial and final instants of feeding flow of supplemental. In order to meet this goal, a logic was employed that is able to accurately predict this moments, using the CO2 (yco2) molar fraction signal, filters and adaptive fuzzy sets.