Static and dynamic neural network models for estimating biomass concentration during thermophilic, lactic acid bacteria batch cultures

Neural networks were used to elaborate static and dynamic models for the on-line estimation of biomass concentration during batch cultures of Streptococcus salivarius ssp. therrmophilus 404 and Lactobacillus delbrueckii ssp. bulgaricus 398 conducted at controlled pH and temperature. Four static models with different structures and input variables were tested. The model relating the increase of lactic acid concentration and the working conditions (pH and temperature) to the increase of biomass was the most appropriate. Nevertheless, all the static models could furnish biased estimations when initial values of biomass were erroneous or when lactic acid measurements were perturbed or noisy. To overcome these drawbacks, recurrent neural networks were used to model the dynamic behaviour of fermentations. These dynamic models, when acting as estimators, performed just as well as the static models but offered more stable responses, due to an implicit corrective action arising from the training methodology and the associated method for biomass estimation.