Physics-Guided Soft Sensing under Sparse Laboratory Measurements and Temporal Regime Shifts in Industrial Monoammonium Phosphate Production
Year: 2026
Type: preprint
Abstract: Industrial plants collect high-frequency telemetry while critical quality variables may be measured only every few hours. We study soft sensing for an industrial monoammonium phosphate process using 260,642 one-minute records and 1,325 paired laboratory observations. The grey-box model reconstructs latent acid properties from conductivity and combines them with material-balance calculations to estimate molar ratio and downstream density. Validation is entirely chronological: nine pre-gap rolling tests, a short-term regime-transfer test, a strict external post-gap test on 438 future laboratory pairs, and ten post-gap rolling comparisons. To prevent uncontrolled extrapolation, the conductivity input to the fitted latent-property mappings is limited to the range observed in the corresponding training data. Ridge and LSBoost remain more accurate than the physics-guided model, particularly in the independent long-horizon evaluation. The results show that physical structure improves interpretability but does not by itself ensure temporal robustness, motivating regime-aware monitoring, recalibration, and fallback logic for industrial virtual telemetry.
Source: ChemRxiv
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7220649278