Dynamic graph embedding for fault detection

Abstract

Using sequence information can improve performances in fault detection for serial (temporal) correlated process data. Classical methods firstly construct extended vectors through concatenating current process data and a certain number of previous process data, and then take dimension reduction methods. However, the simple extension of process data may distort the correlation between variables and largely increase the dimensionality. This paper proposes a novel algorithm, called Dynamic Graph Embedding (DGE), for fault detection. DGE adopts augmented matrices instead of extended vectors to encode sequence information. Furthermore, DGE incorporates both time information and neighborhood information to form similarities of different process data. And then DGE is designed to obtain embedding matrices with Markov chain analysis of the similarities. Extensive experimental results on the Tennessee Eastman (TE) benchmark process show the superiority of DGE in terms of missed detection rate (MDR) and false alarm rate (FAR).