Graph embedding dictionary pair learning for robust process monitoring

Process data collected from industrial processes is frequently corrupted with outliers, which arouses a challenge to monitor industrial processes. In this context, robustness to corruption is critical for an efficient process monitoring approach. In this paper, we propose a novel method termed as graph embedding dictionary pair learning (GEDPL) for robust process monitoring. The proposed method jointly learns a synthetic dictionary and an analytical dictionary for data representation, which reduces the computation burden during the training phase over the monitoring method based on traditional dictionary learning with the l1 or l0-norm sparsity constraint. To alleviate the negative effect of corruption, an l2,1-norm constraint is forced on the projective analytical dictionary to enhance robustness of the dictionary. Moreover, a graph embedding regularizer is constructed to preserve local geometric information and global structures in data. In this way, the faithful reduced-dimensional representations can be effectively captured, improving the subsequent process monitoring capability. Besides, a reconstruction-based contribution plot based on the GEDPL is developed to identify fault variables. Finally, two case studies, including the Tennessee Eastman benchmark processes and a real-world industrial application, validate the efficacy of the proposed process monitoring scheme.