Abstract:
Automated non-invasive diagnosis and localization of the root cause of oscillations in process plants is a widely held industry goal sought in order to stabilize product qualities and reduce equipment problems and energy costs. As stiction in control valves is one of the leading causes of oscillations in plant variables, detection and localization of valve stiction is a major part of any method for root cause diagnosis. Previous contributions have introduced a number of techniques that seek to achieve this objective, but most of the approaches presented are only applicable for the case of single-input single-output (SISO) processes. The current work seeks to extend one widely used approach, Hammerstein-model-based stiction detection, to the case of interacting plants. Because of the difficult nature of the problem, we first consider the case where a nominal linear plant model is known for the interacting system in question. With these approximate linear dynamics, we introduce an approach to identify in which loop the valve stiction is originating from, under the assumption that only a single valve in the interacting system is afflicted by stiction. The method efficacy is explored using simulation studies. The feasibility of extending the method to the case of unknown plant model via multivariate time-series identification of the linear plant model is then briefly discussed.