Recursive PLS soft sensor with moving window for online PX concentration estimation in an industrial isomerization unit

A recursive partial least squares (RPLS) soft sensor with moving window of fixed length is proposed, taking the saturation and integral information of the modeling samples into account. Part of the historical information is recursively retained by the mean and variance updating and the parameters of the model are rolling estimated. The proposed inferential is applied to an industrial isomerization unit for online estimation of the para-xylene (PX) concentration at the outlet of the reactor. The simulation results show that the developed soft sensor has a good performance with the maximum absolute relative error, relative root mean squares error and tracking precision at the level of 2.68%, 0.54% and 0.9543 respectively. The fixed sample length for parameter estimation is detailed discussed and the appropriate length is proved to be 30–50.