Performance assessment of multivariate control systems



Journal Title

Journal ISSN

Volume Title



Modern process plants and refineries contain hundreds to thousands of loops under automatic controls. Poorly performing control loops cause off-spec product, waste energy, and cause maintenance and safety issues. Much of the process control engineer's time is spent finding the root cause of poor performance, meanwhile many underperforming loops are left alone for sake of insufficient knowledge and resources. The field of control loop performance assessment and diagnosis provides tools for control and plant engineers to identify poorly performing loops, diagnose root causes, and find corrective actions. Much as process controls automate much of the regulation and set point tracking functions within the process plant, these new tools aim to automate, as much as possible, plant monitoring and root cause diagnosis tasks.

Two of the main problems for poor plant performance are bad controller tuning and stiction in control valves. Concerning the former problem, most process controls in industrial environments use a proportional-integral (PI) algorithm. There exists a large body of literature on PI controller diagnosis and retuning, and one aim of this document is to classify and compare these methods to allow better understanding of their potential uses. Also herein, the use of the Hurst exponent as a tuning diagnosis measure is proposed, as well as retuning algorithms that can incorporate a wide variety of existing tuning diagnosis measures in order to make retuning decisions. With respect to valve stiction, new theoretical results on the efficacy of Hammerstein model based valve stiction detection are presented. In addition, the Hammerstein model based stiction detection method is extended to the important case of interacting systems.

One key component of control loop performance assessment is benchmarking performance. When creating control loop performance benchmarking tools for multivariate control loops, there exists a large number of possible objectives to pursue. Early attempts focused on creating mathematical tools that could provide the theoretical minimum sum of variances of output variables. However, this objective is not always aligned with the goal of plant personnel, which is to increase economic benefit of the plant operation. Therefore, this thesis pursues economic benchmarking of multivariate control systems, which involves trading off variances of different process variables in order to achieve an optimal operating point. Earlier work in this area relied on non-convex programming with approximate solution methods. The current work instead performs a piecewise linear approximation of the non-convex constraints, allowing for a linear programming solution of the problem.

This document proposes tools that allow for an overall control loop assessment and diagnosis framework to be implemented. In the single-input single output case, this framework allows for completely data-based performance assessment and root cause diagnosis to be performed. For multivariate systems, the tools require more system information than in the single-loop case, but with this cost comes the additional benefit of performance data directly related to the operation's economic objectives. In each case, the performance assessment and diagnosis use only non-invasive methods, so plant operation is not disrupted to perform the analysis. Adoption of these tools will allow for more efficient and better performing plant operations.



Process control