On-line modeling and monitoring for multi-operation batch processes with infinite data types

Abstract

For complex industrial processes with frequent operating characteristics, process data types will be infinite due to the randomness and uncertainty of operation. Additionally, the process data follow serious non-Gaussian distribution. In this paper, an efficient q-nearest-neighbor standardization principal component analysis (q-NNS PCA) based on-line modeling method is proposed to handle complex data distributions and incursive frequent operation shifts. Due to the limitation of initial modeling data, the modeling data structure needs to be continuously replenished with accumulation of new normal batches. The on-line modeling method is proposed to avoid the complexity of model updating and establishment of massive offline models as well as the difficulty of the multiple models selection. For each test sample online, the q-NNS method can search its modeling data coming from the same operation to establish monitoring model, which settles non-Gaussian distribution problem. The proposed method is illustrated with a 120t ladle furnace (LF) steelmaking process. The comparison of monitoring results demonstrates that the proposed method is superior to multiple PCA and MKPCA methods and can achieve accurate and prompt detection of various types of faults in multi-operation processes.