Non-invasive fermentation analysis using an artificial neural network algorithm for processing near infrared spectra

The feasibility of using Artificial Neural Networks (ANN) to improve the pe rformance of a fibre optic, near infrared (NIR) spectroscopic probe for mon itoring fermentation processes was investigated. A miniature diode array sp ectrometer, operating between 1100 and 1450 nm, was used for on-line, in si tu fermentation monitoring by placing a bifurcated fibre bundle inside a fe rmentation vessel. Non-linearities in the spectral response were complicate d by the combined effects of spectral variations in the OH vibrational band s due to temperature fluctuations and pH variation. As the fermentation pro ceeds interferences from other absorbing molecules and cell masses increase at an exponential rate. The feasibility of accurately predicting both gluc ose and ethanol concentrations simultaneously during a fermentation process were assessed by applying Partial Least Squares (PLS) and Artificial Neura l Networks (ANN) to the NIR spectral data. For a 5% glucose fermentation, a PLS model was able to predict "online" concentrations with standard errors of prediction (SEP) of 0.19% for glucose and 0.11% for ethanol, Three sepa rate on-line fermentation experiments were performed in order to determine the possibility of using a model developed from one experiment to predict t he concentrations of another experiment. PLS models produced an average SEP of 0.21% when used to predict different fermentation experiments. The ANN algorithm produced an average SEP of 0.13% on the same data.