Design of an online learning scheme for quality prediction in refining and chemical units

Abstract The refining and chemical production process is highly complex, with various uncertainties leading to fluctuations in product quality or even non‐compliance with standards. This makes real‐time monitoring of product quality essential. In recent years, with the advancement of digitalization and intelligent technologies, many quality prediction methods based on mechanistic or data‐driven modelling have emerged. However, for refining and chemical processes characterized by high dimensionality, strong coupling, and multiple working conditions, these methods may suffer from model mismatches and other limitations. Therefore, this study introduces an online learning approach for quality prediction in refining and chemical units, aiming to extract feature data across multiple working conditions from historical records without relying on process models. The method begins by efficiently extracting steady‐state data from process variables using filtering and fluctuation detection techniques. K‐means clustering and support vector data description (SVDD) are then applied to extract features at typical process points, while an innovative incremental principal component analysis (IPCA) method is employed to determine the optimal number of clusters for precise data classification. Additionally, a novel solution addresses matching issues between data sampled at different intervals, integrating process variables, working conditions, and quality indicators based on expert knowledge. A case study of a catalytic cracking distillation unit demonstrates the feasibility of the approach, highlighting its significant implications for quality prediction and online soft sensors, as well as its strong potential for future application across multiple units.