| Abstract [eng] |
Biotechnological processes are complex nonlinear systems whose effective control depends on the accurate estimation of process state variables. In practice, some important variables, such as biomass concentration, cannot be measured directly in real time because their determination requires laboratory analysis. For this reason, it is relevant to apply soft sensors based on data-driven methods, which allow difficult-to-measure process parameters to be indirectly estimated using other variables recorded in real time. This work investigates the estimation of state variables in a biotechnological GFP protein synthesis process using neural networks. The study used experimental cultivation process data, which were processed before neural network training: the data were combined according to laboratory measurement times, supplemented with derived variables, normalized, and incorrect or atypical experiments were removed. In total, data from 78 experiments were used for neural network training. In this work, autoassociative neural networks of different structures were developed and analyzed for the reconstruction of the biomass variable. The Levenberg–Marquardt algorithm was applied for network training, while model accuracy was evaluated using the average sum of absolute deviations, the average absolute deviation, and the absolute percentage error. The best results for the autoassociative neural network were obtained using a 4:2:2:2:4 structure: the average sum of absolute deviations was 4485.5 g, the average absolute deviation was 10.86 g, and the absolute percentage error was 9.05 %. In order to evaluate the effectiveness of the obtained results, the autoassociative neural network was compared with a feedforward neural network. The best results for the feedforward neural network were obtained using a 3:2:1 structure: the average sum of absolute deviations was 3097.0 g, the average absolute deviation was 7.49 g, and the absolute percentage error was 6.26 %. The comparison of both methods showed that the feedforward neural network was more accurate for the specific biomass prediction task, while the autoassociative neural network can also be applied for the reconstruction of biotechnological process state variables. |
