Degree of hydrolysis from mid-infrared spectra

The use of empirical modelling, namely, feed-forward neural network, is proven to be an efficient alternative for the control of bovine haemoglobin proteolysis from mid-infrared (MIR) spectra recording by the mean of an ATR probe. Six batches were experimented and sampled. The situation is challenging since the results of the self-organising maps (SOMs) analysis definitely show clusters of the spectral dataset. Consequently, strong generalisation capabilities are to be foreseen and it is typical situation when one deals with enzymatic processes. Supervised learning is thus used in order to predict an unknown batch from the knowledge provided by the five others. The RMSEP results are estimated around 0.3% in an analytical range of (0, 8.7) of the hydrolysis degree. Moreover, the analysis of the data treatments enable to emphasise on the biochemical information, from one hand, and on the role of each of the hidden nodes from another hand.