Just-in-time latent autoregressive residual generation for dynamic process monitoring

Abstract

With a goal of timely and adaptively exploiting the inconsistency inherited in the monitored samples of current interest, a novel dynamic process monitoring method based on just-in-time latent autoregressive residual generation (JITLAR²G) model is proposed. Different from the mainstream dynamic modeling and monitoring methods which usually train a signature generating mechanism and then repeatedly apply it for online monitored samples, the proposed JITLAR²G-based approach provides a JITLAR²G model for the online monitored samples after data augmentation, so that the corresponding inconsistency within the given consecutive samples could be timely and adaptively uncovered. Instead of expressing the time-serial relationship that generally accepted by the normal samples in the given dataset, solving the objective function designed for JITLAR²G in a just-in-time manner can adaptively and correspondingly seek but only one projecting vector as well as coefficient vector to generate residual, which points to the potential inconsistency inherited in the monitored samples, for the sole purpose of fault detection. As demonstrated through comparisons, the proposed JITLAR²G model can consistently guarantee its effectiveness, in terms of reducing both false alarm rate and missed alarm rate, for dynamic process monitoring, the salient performance achieved by the proposed JITLAR²G-based method in contrast to the counterparts can be always confirmed.