A Transformer–PINN Fusion Model for Soft Sensing of Product Purity in Reactive Distillation Columns

Abstract

Reactive distillation is an integrated reaction-separation process that has been intensified; however, due to non-linear coupling of the products, measuring product purity is slow, infrequent, and easily affected by operational fluctuations. A soft sensor integrating a temporal Transformer and a physics-informed neural network is developed for continuous estimation of top-product purity in this study. Multiple-rate process signal is alignment through quality-aware resampling, masked self-attention encoding, and residual constraints derived from component balance, phase equilibrium and monotonic operating relation correction. A pilot-scale methyl-acetate column data set with steady-state operation, feed disturbances, reflux ramps, catalyst ageing and sensor degradation was used for evaluation. Among Partial Least Squares, Long Short-Term Memory, Temporal Convolution, and a data-only Transformer, the proposed model achieved a test root-mean-square error of 0.0028 mole fraction and a mean absolute error of 0.0019, and an R-squared value of 0.987. Under an unseen feed-composition shift, the error increased by only 18.6% for it and 52.4% for the data-only Transformer. The model found the purity excursion 11.7 minutes before the delayed laboratory analyzer and maintained 96.1% interval coverage. Based on the ablation results, balance residuals, uncertainty-weighted fusion and missing-signal masking all contribute to improved robustness. Thus, the fusion can offer a high-accuracy, physically consistent and deployable estimator for supervisory control and quality supervision in reactive distillation.