Abstract

Probabilistic Graphical Models (PGMs) are a growing data-driven AI approach towards unraveling patterns as a complex network representation useful in many applications such as Social Networks, Media Analytics, and Smart Cities. In which the biggest challenge in monitoring multiple sensors is related to the high dimensionality that a process can easily obtain. Additionally, Statistical Process Control (SPC) techniques have difficulty in simultaneously detecting changes in quality characteristics when extrapolated from the normality (or Gaussian) assumption. For that, understanding multivariate time series (TS) sensors is a must but also a challenge, which may also be even more complex if the sensors' locations behave differently rather than as a homogeneous system; that is, under the existence of spatiotemporal relationships. This study focused on the SPC multivariate case (dimension d ≥ 2) using a Dynamic Bayesian Network class of models, enabling estimate causal inference, which takes over the data dependence or marginal asymmetry, as well as dynamic correlation TS estimation into account.