Abstract
We develop a multi-fidelity soft-sensing framework to reconcile online (low-fidelity) industrial measurements with sparse (high-fidelity) laboratory samples from an alkylation unit in a refinery. A first-principles model is used to generate an additional low-fidelity dataset and train a surrogate that predicts the output variable. We investigate whether incorporating sparse high-fidelity laboratory data with the low-fidelity data improves prediction accuracy. A multi-fidelity predictor forms a corrected output by learning the residual between the high-fidelity observations and the low-fidelity surrogate using Gaussian process regression. The simplest model structure performs best, reducing test prediction error (computed against laboratory samples) by 31.1% relative to the currently deployed industrial analyzer and outperforming a standard high-fidelity-only model trained on laboratory data. Overall, the simplified surrogate model captures the main industrial trends well enough to serve as a reliable low-fidelity input for the multi-fidelity soft sensor.