Augmented switching linear dynamic latent variable model based for multimode process fault diagnosis

Abstract

Extracting dynamic, uncertain, and multimode features of process data is important and challenging for fault diagnosis. However, the traditional fault diagnosis methods mainly focus on a single feature extraction, and there are few fault diagnosis methods that can extract multimode features. In this paper, a new fault detection and classification method based on augmented switching linear dynamic latent variable (ASLDLV) model are proposed, which can solve the problem of multimode feature extraction, and mine dynamic, uncertain and multimode features simultaneously in process modeling for fault detection and classification. First, a linear dynamic latent variable (LDLV) model is introduced to extract the dynamic and uncertain of process samples. Furthermore, an augmented dynamic switching transition matrix is embedded in LDLV model to describe multiple modes characteristics. In this way, an augmented dynamic switching transition matrix can be used to mine deeper dynamic correlation between multiple modes so that the ASLDLV model is achieved. The proposed ASLDLV model not only can consider the dynamics, uncertain and multimode characteristics of data at the same time, but also it is better to represent dynamic relationship of multiple models so that a more accurate switching model is determined. Finally, the ASLDLV model is applied to Tennessee Eastman process for fault detection and classification, compared with some exiting methods. The simulation results illustrate the effectiveness and superiority of ASLDLV model.