Dynamic Probabilistic Latent Variable Model with Exogenous Variables for Dynamic Anomaly Detection

A novel dynamic probabilistic latent variable model with exogenous variables (DPLVMX) is proposed in this article to capture system dynamics with the existence of random noises. The dynamic auto-regressive relations between current and past latent variables are extracted in a Markov state-space form in the proposed model. Furthermore, to strengthen the utilization of valuable information in the collected data, a composite loading index is designed to select some interested variables as the exogenous variables, which is explicitly incorporated into the model relations of DPLVMX. An improved DPLVM based monitoring scheme is also designed, where a new dynamic monitoring index is proposed to detect dynamic anomalies. The Tennessee Eastman process is used to illustrate the superiority of the proposed algorithm.