Towards a machine learning operations (MLOps) soft sensor for real-time predictions in industrial-scale fed-batch fermentation

Abstract

Real-time predictions in fermentation processes are crucial because they enable continuous monitoring and control of bioprocessing. However, the availability of online measurements is limited by the availability and feasibility of sensing technology. Soft sensors - or software sensors that convert available measurements into measurements of interest (product yield, quality, etc.) - have the potential to improve efficiency and product quality. Machine learning (ML) based soft sensors have gained increased popularity over the years since they can incorporate variables that are measured in real-time, and exploit the intricate patterns embedded in such voluminous datasets. However, ML-based soft sensor requires more than just a classical ML learner with an unseen test set to evaluate the quality prediction of the model. When a ML model is deployed in production, its performance can deteriorate rapidly leading to an unanticipated decline in the quality of the output and predictions. Here a proof concept of Machine Learning Operations (MLOps) to automate the end-to-end soft sensor lifecycle in industrial scale fed-batch fermentation, from development and deployment to maintenance and monitoring is proposed. Using the industrial-scale penicillin fermentation (IndPenSim) dataset that includes 100 fermentation batches, to build a soft sensor based on Long Short Term Memory (LSTM) for penicillin concentration prediction. The batches containing deviations in the processes (91–100) were used to assess concept drift of the LSTM soft sensor. The evaluation of concept drift is evidenced by the soft sensor performance falling below the set threshold based on the Population Stability Index (PSI), which automatically triggers an alert to run the retraining pipeline.