Distributed plant-wide monitoring via modularity-optimal NMF decomposition based on graph embedding

Abstract

A new distributed process monitoring framework is proposed in this paper that aims to effectively decompose industrial process variables for process monitoring. Existing variable decomposition methods focus on direct correlations between variables, such as adjacency matrix, while neglecting the potential correlations. Graph embedding captures the potential correlations between variables precisely through distances in the embedding space. Consequently, the information of variables themselves used to construct graph embedding deserves greater attention. We propose a novel variable decomposition method based on Graph Auto-Encoder(GAE)-Nonnegative Matrix Factorization (NMF). GAE with attention mechanism is employed to derive graph embedding by taking into account the information of the variables themselves. Then, NMF with optimal modularity is run in graph embedding, and the process variables are divided into sub-blocks for distributed process monitoring. To verify the performance of the method, Canonical Correlation Analysis(CCA), as a distributed process monitoring, obtains the final monitoring results. The Tennessee-Eastman process(TEP) is used to demonstrate the performance of distributed process monitoring.