At R2 Inc., we monitor in real-time tens of thousands of sensors in industrial plants, looking for anomalous behaviour. Industrial processes are usually very well understood and predictable, making them ideal candidates for classical Bayesian analysis. Off-the-shelf modelling approaches like Stan offer great flexibility in formulating a model, but the Markov Chain Monte Carlo algorithms they use are slow on large datasets, which hinders model iteration. Kalman filters can provide much faster, exact posterior probabilities, by assuming that the uncertainties are Gaussian and that the hidden state evolution is linear.

We have built a probabilistic programming language around the switching Kalman filter algorithm. It was implemented as a Julia macro, and supports Unitful time series with missing data. I will present its design, implementation and some of its applications.