Finite State Machine enriched Deep Learning

In pharmaceutical manufacturing, process control is based, among others, on estimating the duration of some (bio)chemical reactions. These processes are highly complex and challenging to model analytically. In addition, the lack of data makes it difficult to learn predictive models since the reaction progress may not be directly measurable with sensors but requires costly analysis of samples. In this project, we address this problem through a Deep Learning approach. We propose a novel method where the design models of the process are used to enrich the sensor data with additional labels to improve the estimation.