Data-driven Bayesian approach for control loop diagnosis

Many methods and algorithms have been proposed for control performance monitoring and process monitoring. However, there are few methods available for synthesis of different monitoring algorithms to form a control loop diagnostic system. Determination of the underlying reason of poor control performance is challenging. In this paper, we investigate a novel data-driven Bayesian approach for control loop diagnosis. The new approach can synthesize information from different monitoring techniques to give an appropriate inference even if the performance of each individual monitor may be low. Some other merits of the new approach include, for example, probabilistic inferences which can be easily used by optimal decision making system, robustness to missing data, and ability to incorporate a priori knowledge. Simulation of Bayesian diagnostic system for a binary distillation column is presented. Data missing handling feature using causality structure and marginalization is discussed. Performance of the Bayesian diagnostic system is examined under different operating modes to demonstrate the information synthesizing ability of the proposed approach.