Soft sensors development for industrial reactive distillation processes under small training datasets



Under real industrial conditions, there are usually missing values in the data. This is due to measurement errors, sensor failures, missing values in real-time databases, irregular measurement intervals, and data that are not covering the total operation range of the plant. Due to the lack of data, the soft sensor (SS) model is of poor quality. The functioning of this model is also unsatisfactory in the new operating points of the plant in the case of a small training sample. We propose the use of a calibrated rigorous (first- principles) process model with acceptable limits of parametric uncertainty to extend the training dataset, which allows us to take into account the physicochemical characteristics of the process. It is shown that the extension of the training sample based on the rigorous model makes it possible to obtain a nonlinear SS of higher accuracy.