Avoiding misleading predictions in fluorescence-based soft sensors using autoencoders

Abstract

Excitation-emission matrix (EEM) fluorescence data can be explored by using deep convolutional neural networks to enhance the predictive performance of bioprocess variables. This article proposes the use of residual neural networks (ResNet) for the prediction of ethanol, glucose, and biomass concentrations of S. cerevisiae cultivations based on fluorescence data collected in situ. A trust screening for unkown samples, based on autoencoder (AE) reconstruction error, is also proposed. Its characteristic of reconstructing the inputs is the key feature to avoid misleading predictions and forecast if a new sample should be trusted as usual or flagged as abnormal. The 83 layers deep ResNet successfully predicted the desired outputs, with R2 higher than 0.98, in the test subset. The best-fitted autoencoder had a 3-layer architecture, with three neurons in the bottleneck and using rectified linear unit (ReLU) activation for the encoder and linear activation for the decoder. The mean reconstruction Root Mean Square Error (RMSE) for the fermentation's EEMs was 4.61 rel. fluorescence intensity units, representing an error smaller than 1% (of the total amplitude of change). To evaluate the AE capability to work as trust screening, random fluorescence intensity was added to the Ex450/Em530 fluorescence pair (related to flavins) in some samples, creating a defective dataset. The dataset was evaluated with the trained AE and the ResNet model to compare reconstruction errors and bioprocess concentrations. The AE was able to identify the samples with added errors, and, as expected, the defective samples also presented higher predictive errors in general. The higher the AE's reconstruction RMSE, the less the new sample should be trusted to avoid misleading predictions.