Enhancing Multi-operating Modes Knowledge Embedding and Data Augmentation Method for Final Boiling Point Prediction in Hydrogenation Distillation Unit

Abstract

In process industries, critical quality indicators (CQIs) like diesel Final Boiling Point (FBP) are challenging to monitor online due to equipment limitations, while valuable human expertise remains unformalized for operational guidance. We propose a knowledge-embedded framework integrating three innovations: 1) A modified UMAP optimization injecting human knowledge constraints on data affiliation to align visualization with domain patterns; 2) Production mode identification through labeled operational records, enabling intuitive decision support; 3) Elastic net-based pseudo-mechanistic modeling combined with mode-specific data augmentation for real-time CQI prediction. Experiment was validated the on a hydrogenation distillation unit (HDU). The method proposed in this paper demonstrates excellent performance compared to conventional models (LSTM,SVR,GRU,PSO_MLP) and state-of-the-art approaches (JMSDL), particularly excelling in small dataset scenarios and fluctuating operating modes. The framework systematically solidifies human knowledge into adaptive multi-mode predictions while maintaining low computational complexity for online deployment. The joint optimization of human knowledge and the use of multi-mode approaches enable the application and solidification of human knowledge, embedding it to help improve the prediction of indicators under different conditions.