Visualization of process data with dynamic Bayesian networks

We describe a novel visualization algorithm for high-dimensional timeseries data. The underlying model is a switching linear dynamical system, a particular variant of a dynamic Bayesian network. An important difference with most existing visualization techniques is the possibility to incorporate time dependencies between data points. Exact inference in switching linear dynamical systems is intractable, but new techniques for approximate inference enable fast computation of accurate posteriors. The model can be learned with a standard EM-algorithm. We illustrate our method on a real-world data set with sensor readings from a paper machine.