A soft-sensability index for pre-modeling predictability assessment and a mechanism-constrained latent-variable model for grade-structured soft sensors

초록

Data-driven soft sensors for grade-structured and campaign-structured chemical processes are conventionally validated by cross-validation, yet in these plants the procedure reports optimistic accuracy because grade identity is correlated with both the process features and the laboratory labels. A model that only reproduces the mean of each grade can then appear highly predictive. We develop a two-part chemometric methodology. The first part is the soft-sensability index, a pre-modeling statistic that separates the grade-corrected signal from a grade-confounding null. The null is generated by permuting labels within grade-by-campaign blocks, and both the time-ordered score and the null are estimated with time-ordered and moving-block resampling so that serial correlation does not distort the test. The second part is a mechanism-constrained latent-variable model that fixes a Mayo-Lewis copolymerization manifold as a physics anchor and fits a machine-learning residual on features deflated against that anchor. On an industrial styrene-acrylonitrile solution-polymerization soft sensor, naive validation rejects the no-signal null on every synthetic replicate, whereas the proposed test stays at its nominal level of 0.04. The index also shows that a color property with a low time-ordered coefficient of determination of 0.47 contributes the largest grade-corrected gain of any property, a reversal that conventional metrics obscure. The physics anchor recovers apparent reactivity ratios that absorb the steady-state offset between feed and reactor composition, and lowers unseen-grade extrapolation error from 2.3 to 0.9 wt% relative to a purely data-driven model. With four held-out grades this is not distinguishable from the physics anchor alone. © 2026 Elsevier B.V.

키워드