kNN-RVM lazy learning approach for soft-sensing modeling of fed-batch processes

Fed-batch processes are inherently difficult to model owing to non-steady-state operation, small-sample condition, instinct time-variation and batch-to-batch variation caused by drifting. Furthermore, when the process switches to different operation phrases, global learning modeling methods would suffer poor performance due to the negative impact of overdue training samples. In this paper, a k nearest neighbor relevance vector machine (kNN-RVM) based lazy learning method is proposed to model the fed-batch processes to soft-sense the corresponding production indices. A recursive algorithm is developed to effectively obtain the kernel matrices used by previous kNN step and following modeling process. Simulative soft-sensors of penicillin production process and rubber mixing process are implemented to valid the proposed method. Comparative results indict that proposed method has better precision and much lower computational complexity than relevance vector machine (RVM) on soft-sensing modeling of fed-batch processes.