A Time-Varying Probabilistic Latent Variable Model for Process Monitoring with Missing Variables

In process monitoring, multivariate statistical analysis models assist engineers in timely fault detection within industrial production systems. However, actual industrial data is often incomplete, autocorrelated and prone to noise contamination, diminishing the effectiveness of traditional databased models. To address the challenges of noisy incomplete time series, this paper proposes a process monitoring method based on a time-varying probabilistic latent variable (TPLV) model. In this model, observed samples are represented as latent variables governed by a first-order Markov chain, with each measured variable independently generated from its associated latent variable. Moreover, the model parameters are estimated using the expectation maximization (EM) algorithm, enabling the trained latent variables to capture both dynamic and static correlations within the data. The performance of the DPLV model is validated on the Tennessee Eastman (TE) dataset, where results demonstrate its superior monitoring capabilities compared to similar approaches.