Electronic nose signals-based deep learning models to realize high-precision monitoring of simultaneous saccharification and fermentation of cassava

Abstract

This study innovatively proposes a high-precision monitoring method for key parameters in the process of ethanol production from simultaneous saccharification and fermentation (SSF) by electronic nose technology combined with recurrent neural network (RNN). A PEN3 electronic nose system was employed to acquire the odor information of the fermented samples, and four deep learning algorithms based on the RNN architecture were employed to design reasonable network structures to realize the deep learning of the electronic nose signal features and model calibration. The results obtained showed that each deep learning model based on the RNN architecture has good generalization performance for the determination of cassava SSF process parameters. Among them, the bidirectional long short-term memory network (BiLSTM) model has the best monitoring effect on ethanol content, with root mean square error of prediction (RMSEP) of 3.7 mg·mL−1 and coefficient of predictive determination (RP2) of 0.98 and the relative percent deviation (RPD) of 8.1. The bidirectional gated recurrent unit (BiGRU) model had the best monitoring effect on glucose content, and its RMSEP, RP2 and RPD were 2.9 mg·mL−1, 0.99 and 9.1, respectively. The overall results reveal that deep learning algorithms have promising application prospects in the feature mining and model calibration of electronic nose signals, which provides an effective analysis tool for in-situ monitoring of electronic nose technology in modern industrial fermentation processes.