Modeling of soft sensor for pilot scale polyhydroxyalkanoates fermentation using neural network

Periodical sampling and laboratory analysis is the current practice for determination of biomass concentration with optical density (OD) and wet cell weight (WCW) as the product quality indicator in a pilot scale fermentation process for polyhydroxyalkanoates (PHA) production. This study aims to develop a soft-sensor model that provide estimation of the hard-to-measure variables using online sensor measurements of easy-to-measure process variables as an alternative to offline measurement of biomass that has measurement delay, and increases the risk of contamination. The accuracy of soft sensor models developed using data-driven method such as Feedforward Neural Network (FFNN) and Support Vector for Regression (SVR) relies on the use of good quality and informative data as well as the network configuration. Modeling of soft sensor using real historical data of a large-scale bioprocess have the challenges of data imperfections due to limited data availability, batch variations and different sampling time for online sensor measurement dan offline laboratory analysis data. To address this, down-sampling and PCHIP interpolation method was performed to resample the data so that the sampling interval is synchronized without loss of information. It was observed from statistical analysis, heatmap and box plot comparison that information was retained despite a large reduction of the online data sample. Piecewise Cubic Hermite Interpolation Polynomial (PCHIP) interpolation method for offline data increases the sample size but still preserve the original curve and trend of data without excessive oscillation. Next, variable selection by Cosine Amplitude Method (CAM) was carried out to select input variables with high strength of relationship to OD and WCW. Meanwhile, Pearson’s Correlation Coefficient (PCC) detects collinearity between input variables. Data from batches that showed high input variables influence on output variables were chosen as dataset for soft-sensor modeling. Statistical analysis results for performance of soft-sensor models for OD and WCW and inverse models for agitation speed predictions showed that Multilayer Perceptron Neural Network (MLPNN) models are superior to Radial Basis Neural Network (RBFNN) and SVR. MLPNN method achieved mean square error (MSE) of lower than 0.003 for training, lower than 0.009 for testing, correlation coefficient (R) of at least 0.97 for training and 0.91 for testing data, a-20 index of higher than 0.8, variance accounted for (VAF) higher than 93, mean absolute error (MAE) lower than 0.04 and mean bias error (MBE) lower than 0.009 for all models. With good prediction accuracy of OD and WCW values as those obtained through laboratory analysis, MLPNN soft sensor has the potential for use in monitoring of PHA fermentation.

Text FK 2023 34 - Declaration Form.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (675kB) Text FK 2023 34 - Full Text.pdf Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (5MB) Text FK 2023 34.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (5MB) Official URL or Download Paper: http://ethesis.upm.edu.my/id/eprint/18954

Actions (login required)

View Item