Behavior comparison for biomass observers in batch processes

Abstract

On-line estimation of biomass concentration in batch biotechnological processes is an active area of research because normally, the biomass is the desired process product output, and also because it is necessary for control purposes to replace the unavailable biomass concentration measurements with reliable and robust on-line estimations. This work presents five different alternatives to face the problem of biomass estimation in a particular batch bioprocess (δ-endotoxins production of Bacillus thuringiensis), namely: a phenomenological estimator based on dissolved oxygen balance, an extended Kalman filter estimator, a Gaussian process regression-based estimator, an artificial neural networks-based estimator, and finally, an estimator based on information fusion by a decentralized Kalman filter. Each proposed biomass estimation method has its own advantages and drawbacks according to their ability to take into account the model uncertainties and the measurement errors. First, the design techniques of these five biomass estimators are exposed, and finally, the behavior of each estimation method is compared. The availability of efficient biomass estimators is of great importance for engineers because, on the one hand, it allows developing new control strategies for other bioprocess variables such as for instance: the growth rate of the microorganism, the dissolved oxygen concentration, and so on. On the other hand, it is also important to improve the performance of the bioprocess optimization procedure. This work also aims to show the evolution on biomass estimation techniques from classical to more contemporary approaches, such as the design based on neural networks and Gaussian processes regression.