Dynamic process monitoring based on parallel latent regressive models

Abstract

Abstract To comprehensively characterize the underlying time-serial behaviors in a dataset obtained from normal operating conditions, a novel modeling algorithm with the goal of constructing parallel latent regressive models (PLRMs) is proposed for dynamic process monitoring. Instead of exploiting the time-serial variation in a given dataset through covariance or correlation, a directly derived LRM is considered to understand the time-serial behavior inherited from the extracted latent variable. More importantly, the direct derivation of latent regressive relationships is not restricted to just estimating the current from the past. In contrast, a more comprehensive regressive modeling strategy based on multiple LRMs in parallel is considered, with respect to a straightforward argument that a latent variable can be estimated by its time-serial neighbors, including the past and future, within consecutive sampling time steps. Consequently, more comprehensive dynamic behavior can be uncovered from the given dataset. Next, salient performance achieved by the proposed PLRMs-based dynamic process monitoring approach can be expected, as demonstrated through comparisons with counterparts.