The ongoing digital transformation in the process industry creates a solid data basis for the comprehensive analysis of process data. Combined with machine learning methods and hybrid approaches that complement first-principles modelling with data-driven models make it possible to develop soft sensors that can recognize and follow complex relationships in plant behaviour.
The subproject "Feature extraction from process data" includes selected use cases from industry to investigate how task-relevant information can be extracted from image data and used for online process analysis of multiphase processes such as distillation, extraction, sublimation, and aeration.
Another focus of the subproject is on process prediction based on historical data, supervised and unsupervised phase detection and changepoint analysis in multidimensional time series. Here, the potential lies in automated analysis and optimisation of batch processes despite a high variance.
A hybrid modelling approach allows a reduction of the required process data by describing already known physical-chemical relationships with existing first-principles models. The subproject studies how information from a process simulation software can be combined with real data from a production plant using surrogate modelling and transfer learning approaches.
To address the challenge of extremely complex labelling of process data, tools for interactive labelling together with Active Learning are developed in the subproject and evaluated using industrial examples.
Contact: Prof. Leon Urbas, TU Dresden
P2O-Lab and ZIH of the TU Dresden