A novel framework for automated feed phase identification

Bioprocesses are the principal driver for innovation in the pharmaceutical industry as well as for the sustainable production of bio-based chemicals and polymers. However, monitoring and controlling bioprocesses is challenging due to the complex interplay between biology and process technology. A key step towards reliable control- and monitoring systems involves automatic identification of physiological- and technological process phases from the available process data. Along these lines we here present a novel classification framework for feed phase identification which is i) accurate, ii) robust, iii) time independent and iv) efficient when coping with shifted time profiles. Our method breaks down the multiclass learning task into the binary sub-problems of identifying the transitions between adjacent process phases imposing process specific constraints on the model (i.e. unidirectionality). More precisely, we employ a soft controller element that switches between binary multivariate classifiers upon prediction of a phase transition and robustify the design by introducing a lag parameter in order to counteract misclassifications (Figure). We demonstrate the superiority of our framework over classical machine learning (ML) approaches on a real world dataset from a 4-phases bioprocess comprising 16 batches from 4 different reactors (4 batches each) where it achieves close to 100% accuracy, significantly outperforming current state-of-the-art ML techniques.