Traditional time difference (TD) model may deteriorate as time increases. In order to enhance reliability and prediction accuracy of TD model

an adaptive soft sensor was proposed on the basis of a selective ensemble of local time difference Gaussian process regression (LTDGPR) models and consideration of process time-delay characteristics. First

data for modelling was reconstructed by time delay and dynamic information extracted from database. Then

an adaptive localization step was used to statistically classify the reconstructed time-difference dataset and to establish an LTDGPR model set. For new input samples

prediction of dynamic drift for primary variables at certain time lapse was achieved through selective ensemble of LTDGPR models which had strong generalization capability. Finally

spontaneous online prediction of primary variable was achieved on the basis of TD model theory. Simulation results of a real debutanizer process indicated the effectiveness and accuracy of the proposed soft sensor.