Spatial reconstruction-based dynamic latent variable model for fault detection in industrial processes

Abstract How to construct a model that comprehensively describes the signatures of underlying variation for a data set measured from normal operation, is always the most important task in implementing data-driven fault detection for an industrial process. Given that spatial and time-serial relations are two dominant metrics for characterizing underlying normal variation, a novel statistical analytic algorithm called a spatial reconstruction-based dynamic latent variable (SRDLV) is proposed, with the sole goal of enhancing fault detectability. Instead of designing an objective that simultaneously exploits spatial and time-serial relations in a tradeoff manner, the proposed SRDLV requires that we first reconstruct each sample point from its spatial neighborhood, and then seek a projection vector as well as a coefficient vector to uncover the sequential dynamics inherited in the reconstruction residual. From this standpoint, the proposed SRDLV thus provides a different scheme for exploring underlying normal variation from both spatial and time-serial aspects. A novel optimization problem is derived and an iterative algorithm is designed to realize the two main concerns of SRDLV. Finally, the effectiveness of the SRDLV-based fault detection approach is evaluated via the Tennessee Eastman process and an industrial-scale multiphase flow facility. The superiority of the proposed method is validated through comparisons with other comparative approaches.