Process fault detection using Stan

Effective and reliable monitoring and diagnostics of process control installations are of utmost importance, as they are an important part of the world’s economy. Algorithms for fault detection and isolation allow extension of system lifetime, reduction in operation interruption, and can lead to significant savings.

The fundamental difficulty in their development is that process plants have a considerable level of complexity, are nonlinear, and influenced by stochastic disturbances and parameter variations. Therefore, approaches based on first principles models are difficult or even impossible to use on a wider scale. While modeling of systems’ healthy operation is possible, the faulty states are unique. Not to mention, that data for such cases is rather scarce.

In this talk, we want to tackle fault detection, by functional analysis of time series of process measurements. Our intention is to consider the problem as recognition of outliers, which are uncharacteristic for normal work. We investigate a spline representation of measurements and attempt to create a Bayesian model of a healthy state. Applying it and a data depth statistic, we analyze the feasibility of this approach with both simulated and experimental data. We conduct experiments using the laboratory water tank system, which has a possibility of simulating faulty behavior.

Presenter biography:
Jerzy Baranowski

My name is Jerzy Baranowski and I am an associate professor of automatic control at AGH UST in Kraków, Poland. I did my PhD on state estimation in dynamical systems and my DSc on numerical methods for fractional calculus. For last few years my work focuses on process diagnostics and Bayesian methods. I head the Laboratory of Computer Science in Control and Management (, which my co-authors belong to. I am a fresh addition to Bayesian community, but I hope to be a productive member.