Adaptive dynamic predictive monitoring scheme based on DLV models

In this paper, we propose an adaptive method for dynamic predictive monitoring of industrial processes based on dynamic latent variable (DLV) models. DLV models extract dynamic latent variables with descending predictabilities and provide explicit modeling of the dynamics. By exploiting these two characteristics, the proposed method provides predictions conditional on fault-free data, such that potential faults lie in the prediction errors only. When a fault has been detected, the prediction horizon will increase in real time to avoid using faulty data for predictions. However, the prediction errors will grow as the prediction horizon increases. Based on the descending order of the predictability built in the DLVs, a DLV's prediction will be adaptively turned off when its prediction error variance is greater than the variance of the DLV. The DLV's prediction will be resumed when the fault data period is over. In general, the most predictive DLVs will survive the longest prediction horizon. In the limiting case when all DLV predictions are turned off, the monitoring scheme uses the data mean for prediction, which is equivalent to a static monitoring scheme. Case studies are provided to illustrate the effectiveness of the proposed method. Copyright (C) 2021 The Authors.