Latent space transformation based on principal component analysis for adaptive fault detection

Principal component analysis (PCA) has been effectively applied in fault detection and in diagnosis of industrial processes to deal with a large number of variables with high correlations. However, normal changes often occur in real process, which always result in false alarms for a fixed-model approach. The authors' research is focused on the traits of normal process changes, which are classified into three scenarios, including process drifting, enlarging and bias, and then three latent space transformation-based PCA algorithms are proposed to obtain an adaptive model described by a new set of coordinates for adaptive fault detection. Finally, the proposed algorithms are applied to imperial smelting furnace.