Performance Monitoring and Diagnosis of a Binary Distillation Column

The life-time performance of chemical processes is limited due to changes in the plant dynamics and disturbance characteristics over time. Such systems often make use of model-based controllers. When a dynamic change arises over time, a difference occurs between the dynamic models contained in the controller and the true system dynamics. The difference in dynamics could deteriorate the performance of a model-based control system. Monitoring the performance on-line is therefore of importance. Detection of a dynamic plant or disturbance change occurs by a classical performance monitoring method, which estimates the variance of the controlled outputs. A change is detected at the moment that a maximum performance bound is violated. An important step is to distinguish between control-relevant plant changes and variations in disturbance characteristics due to different solutions strategies. With an existing performance diagnosis method which makes use of closed-loop prediction error identification the true plant dynamics are identified. Then it is verified whether a performance drop is caused by a change in control-relevant plant dynamics by making use of hypothesis testing. A set is considered that contains all plant dynamics which achieve a satisfactory performance and it is verified whether the identified model is located in or outside the set to make a decision. With an alternative second decision rule, a heuristic method is used where models are constructed around the estimated model by making use of a normal distribution. It is verified which percentage of these models are located outside the set to make a decision. A third decision rule is used which is a combination of the first and second decision rules and the confidence is compared between all considered decision rules. In a simulation case study of a binary distillation column, the performance monitoring method detects a performance drop satisfactory without creating many false alarms. Furthermore, it is shown that with the heuristic method a significant increase in confidence is achieved compared to the first decision rule and only a minor difference with respect to the third decision rule.Delft Center for Systems and ControlMechanical, Maritime and Materials Engineerin