Scalable active subspace low-rank graph representation for continuous system online security evaluation with input corruption

Abstract

Continuous online system operation evaluation aims to monitor the occurrence of abnormal conditions. However, most of the currently popular monitoring methods are developed from the perspective of historical data integrity. In industrial processes, missing or corrupted entries generally exist due to improper recording or sensor drop-out, which brings difficulties to online monitoring evaluation. In this article, a scalable active subspace low-rank graph representation method is presented to deal with missing data in the offline phase of fault monitoring. The method realizes the reconstruction of a clean matrix by the linear combination of low-rank representation and scalable orthogonal matrix. Meanwhile, the low-rank representation and sparse error matrix with l2,1-norm are arranged in the loss function to improve the optimal feature selection and robustness of the model. Apart from that, to preserve the consistency of the adjacent structure between the reconstruction space and the original corrupted space, a Laplacian manifold regularization is designed to constrain the sparse error. Finally, we establish an optimal graph discriminant model of recovery data for online safety monitoring of continuous systems. Confirmatory simulations on benchmark TE process and actual multi-phase flow process illustrate that the proposed approach is superior to the state-of-the-art methods, fully verifying the robustness and detectability performance in the presence of specific corruption.