Soft Sensors are mathematical models used to predict the behavior of real systems. They are usefully applied to estimate hard-to-measure quantities in the process industry. Many Soft Sensors are designed by using data-driven approaches and exploiting historical databases. Machine learning is widely used for this aim. Here, the potentialities of deep learning in solving some challenges raising in industrial applications are introduced. More specifically, the paper focuses on three specific aspects: labelled data scarcity, computational complexity reduction, and unsupervised feature exploitation. The state of the art of Soft Sensors based on deep learning is described. Then, the focus is on Soft Sensors based on Deep Belief Networks, as a research field that the authors have been investigating since years. The improvements offered by Deep Belief Networks, over more conventional data-driven approaches, in designing Soft Sensors for real-world applications will be shown. Soft Sensors for specific cases study are described.
CITATION STYLE
Graziani, S., & Xibilia, M. G. (2020). Deep Learning for Soft Sensor Design. In Studies in Computational Intelligence (Vol. 867, pp. 31–59). Springer Verlag. https://doi.org/10.1007/978-3-030-31764-5_2
Mendeley helps you to discover research relevant for your work.