Data-knowledge-driven distributed monitoring for large-scale processes based on digraph

Abstract

Large-scale processes play a pivotal role in modern industry. For process safety concerns, this work proposes a novel digraph-based data-knowledge-driven method (DG-DKD) for large-scale process monitoring. Different from traditional purely data-driven methods, DG-DKD combines process data and knowledge to improve the capability of fault detection and diagnosis. Firstly, the large-scale process is converted into a digraph based on process knowledge and then decomposed into multiple physical meaningful subblocks through a digraph partition method, which eliminates the spurious correlation and improves interpretability. On this basis, the spatial information between subblocks is characterized by a one-hop digraph diffusion method and the temporal information within subblock is captured by canonical variate analysis. With concurrent analysis of temporal and spatial information, the accurate detection of process faults can be guaranteed. Subsequently, a two-stage distributed fault diagnosis method is developed where process data and digraph are leveraged for contribution analysis and causality analysis respectively, which can reduce the smearing effect and identify the fault root cause and propagation path. Finally, the effectiveness of the proposed method is illustrated through the Tennessee Eastman benchmark process and a reactor separator process.