Abstract:

Neural networks with nonlinearity correctly describe the dynamic process of microorganism cultivation. In this paper, the biomass of Bacillus cereus DM423 during a batch cultivation was measured by a soft-sensor based on the partial recurrent neural network （PRNN） , and a PRNN with the topology of 11-5-1 was constructed, in which the pH value, the temperature, the dissolved oxygen content, the glucose concentration at two previous times, as well as the delays and feedbacks of estimated biomass concentration at three previous times, were used as the input variables, the current biomass concentration was used as the output variable, and the BPTT algorithm was employed. The results show that the constructed network is of good generalization and that a mean square error of 0. 56× 10^-3 is attained. It is also found that the the constructed network is robust in resisting low Gaussian noise, and is suitable for the accurate multi-step prediction of biomass of Bacillus cereus DM423 during a batch cultivation.

Key words: recurrent neural network, biomass, soft-sensing, prediction