Ensemble transfer learning assisted soft sensor for distributed output inference in chemical processes

Abstract

Chemical processes with distributed outputs are characterized by various operating conditions, and the scarcity of labeled data poses challenges to the prediction of product quality. An ensemble transfer Gaussian process regression (ETGPR) model is developed for prediction of different quantities of distributed outputs. First, for each test instances from target domain, just-in-time learning is adopted to select distance-based similar instances from source domain in related operating conditions. Mutual information helps create various local models by building diverse input variable sets. Subsequently, Bayesian inference is used to produce the posterior probabilities relative to the test instance, then set as the weights of local prediction. The instance transfer is thus completed via distance-based similar instance selection from source domain for local model construction, and the model performance is improved by the ensemble weighting strategy, concerning the target domain, under diverse operating conditions. Therefore, by utilizing and transferring information from source domain, unsupervised transfer can be implemented with available unlabeled target data. The superiority of ETGPR model is confirmed in the case of modeling the polymerization process with distributed outputs.