Density‐Weighted Similarity Label Distribution With Dual‐Boundary SVDD for Multi‐Modal Intermittent Process Soft‐Sensing

ABSTRACT Existing soft sensors for penicillin fermentation often fail to address batch‐to‐batch variability and the complex dynamics of transition modes, leading to inaccurate mode recognition and degraded prediction. To overcome this, we propose WSDPC‐DBSVDD‐RVR, a novel soft‐sensing strategy integrating three core components. First, a density‐weighted similarity label peak clustering (WSDPC) algorithm is developed for robust mode division. It employs adaptive weighting to enhance low‐density region contributions and a distance‐density similarity index for accurate label assignment, coupled with a modal evaluation index (MEI) for identifying normal/abnormal modes. Second, a collaborative modeling framework combines dual‐boundary support vector data description (DBSVDD) and relevance vector regression (RVR). DBSVDD characterizes stable modes (inner hypersphere) and permissible transition fluctuations (outer hypersphere), while RVR provides probabilistic predictions. A sliding window enables dynamic online mode identification. Third, a multi‐strategy improved Kepler Optimization Algorithm (Imp‐KOA) integrates chaotic mapping and opposition‐based learning to achieve online adaptive parameter optimization, preventing local optima. Experimental validation on penicillin fermentation data demonstrates that WSDPC‐DBSVDD‐RVR significantly enhances prediction accuracy (e.g., reduced RMSE), dynamic adaptability, and industrial applicability compared to benchmarks, effectively addressing the limitations of global modeling and density imbalance.