Soft sensor for quality estimation of product in cyclohexanone unit

Abstract

Accurate measurement of quality variables are importance for the complete quality monitoring and control tasks in chemical process. Unfortunately, few analysers or hardware sensors for measuring the key variables are available and difficult to maintain. Moreover, these have also limitations such as very high cost and the large time delays. Therefore, it is necessary to utilize soft sensors to estimate the quality variables using other directly measurable secondary variables. In this work, three estimating approaches by using empirical models being multilayer feedforward (MLFF) artificial neural networks (ANNs), partial least squares (PLS) regression, and neural network partial least squares (NNPLS) are exploited to build soft sensors able to estimate the top product of the distillation column in the cyclohexanone unit using available temperature measurements in the column. However, buildings of soft sensor models using only real plant data have the hard limitation since they cannot implement in wide range estimation because of the plant data having smooth responses or small data variations. In order to handle this problem, this work used two data sources which were real plant data and wide range simulated data for constructing the soft sensors. These two data sources were mixed together to calculate parameters of the soft sensor model based on PLS model. For MLEF and NNPLS model, the wide range simulated data were used to pre-train or pre-calibrate parameters of soft sensor models and used the real plant data for find-tuning the parameters again. In case study, the results proved that all of the soft sensors showed satisfactory estimating performances and soft sensor based on MLFF method gave better estimating performances than both of the PLS methods.