Soft sensor for nonlinear processes based on ensemble partial least squares with adaptive localization

This paper proposes a high-accuracy soft sensing approach for nonlinear processes based on ensemble learning. To partition the process state into local model regions, an adaptive localization scheme is developed, through which the correlation among process variables can be modeled and the pre-setting of sub-model number is not required. These prepared local models are weighted by a proposed supervised weighting mechanism and then combined via the Bayesian inference to predict the y-value of the query sample. Such weighting mechanism can fully exploit the historical data set and quantify each local model´s generalization ability for the query sample, thus it is potential to compute the combination weights more accurately. In addition, extensive performance evaluation of the proposed soft senor is conducted over a real-life industrial debutanizer column process. The effectiveness of the proposed soft sensor is demonstrated through comparison results in contrast with several other soft sensor modeling methods.