A mechanism‐consistent hierarchical framework for pulp quality prediction with risk‐aware attribution in industrial refining

Abstract Accurate and timely prediction of pulp quality is essential for stable refining control but is usually challenged due to instrument delays and limited observability. To address these challenges, this paper proposes a mechanism‐consistent hierarchical prediction framework. First, a two‐stage paradigm is established: a data‐driven soft sensor reconstructs the intermediate refining state represented by the Schopper–Riegler degree (SR), a widely used indicator of pulp drainage behaviour. This predicted SR then serves as a critical intermediate indicator to assist the inference of the average fibre length (AFL). In the second stage, a physically interpretable FiLM–SE–gating network (FSG‐Net) is developed. The architecture combines SR‐conditioned feature modulation with a gated decoupling design to explicitly distinguish refining behaviours dominated by external fibrillation and fibre cutting. Additionally, to suppress spurious correlations between average fibre length and control variables, we introduce mechanism‐consistent gradient regularization (MCGR), which aligns the gradient sensitivities of SR and AFL with respect to key control variables. Furthermore, the deterministic framework is extended into a risk‐aware prediction system via residual‐based quantile regression. By coupling predictive uncertainty with internal gating states, the framework enables mechanism‐specific risk attribution, identifying whether operational risks originate from external fibrillation effects or fibre cutting mechanics. Industrial pulp refining case studies demonstrate that the proposed framework achieves high prediction accuracy for AFL ( R 2 = 0.921), while maintaining strong physical consistency, and supports reliable process monitoring under varying operating conditions.