Soft sensor development for nonlinear and time-varying processes based on supervised ensemble learning with improved process state partition

The nonlinearities and time-varying characteristics are two major causes of low performance of soft sensors in process systems. Motivated of solving the two problems, this paper proposes an adaptive soft sensing method under the ensemble learning framework. An improved process state partition scheme is proposed to construct independent local models, which not only inherits the merits of the original process state partition method but also possesses the function of detecting and deleting redundant models. These prepared local models are weighted by a supervised weighting mechanism and then combined via the Bayesian inference to predict the y-value of the query sample. Because the weighting mechanism can fully exploit the historical data set and quantify each local model's generalization ability for the query sample, it is potential to compute the combination weights more accurately. Simulations are conducted on two benchmark data sets from two real-life chemical processes. Extensive performance evaluations of the proposed soft sensor are conducted, and results show its effectiveness. (c) 2015 Curtin University of Technology and John Wiley & Sons, Ltd.