Software sensors for biomass concentration in a SSC process using artificial neural networks and support vector...

Area of Science:

• Biotechnology
• Process Engineering
• Computational Intelligence

Background:

• Fermentation processes often lack direct sensors for critical state variables like biomass concentration.
• Software sensors offer a viable solution to estimate these unmeasured variables.
• Solid substrate cultivation (SSC) is a key fermentation process where biomass monitoring is crucial.

Purpose of the Study:

• To compare the efficacy of different computational intelligence models as software sensors for biomass concentration in SSC.
• To evaluate the performance of NARX-ANN, NARMAX-ANN, NARX-SVM, and NARMAX-SVM models under varying noise conditions.
• To determine the best model for on-line estimation of biomass concentration using easily measurable variables.

Main Methods:

• Development and comparison of four software sensor models: NARX-ANN, NARMAX-ANN, NARX-SVM, and NARMAX-SVM.
• Utilizing CO₂ and O₂ measurements as inputs for the software sensors.
• Evaluating model performance using the SMAPE (Symmetric Mean Absolute Percentage Error) index under 20% amplitude noise.
• Assessing model convergence capabilities under perturbation of initial conditions.

Main Results:

• The NARMAX-SVM model achieved the best performance, with an SMAPE index below 9 under 20% noise.
• NARMAX models generally outperformed NARX models due to the inclusion of prediction errors as inputs, enhancing predictive capabilities.
• NARX models showed better convergence when initial conditions of the autoregressive variable were perturbed.

Conclusions:

• Software sensors based on computational intelligence techniques can reliably estimate difficult-to-measure variables like biomass concentration in fermentation.
• The NARMAX-SVM model is highly effective for on-line biomass estimation in SSC processes.
• This approach enables accurate process monitoring using easily accessible variables such as CO₂ and O₂.