(342d) A Testbed for Studying the Interactions between Model Predictive Control (MPC) and Human Operators
AIChE Annual Meeting
Friday, November 20, 2020 - 8:00am to 9:00am
As a cyber-physical system, chemical processes are usually operated under closed-loop control while being supervised by specialized human operators. Complex, large industrial-scale processes have 100s of low-level single loop controllers, which would be supplemented by advanced multi-variate control schemes, such as Model Predictive Control (MPC) controllers that seek to drive the process towards further higher-order objectives while recognizing various constraints. The algorithm in the MPC controller is complex in nature, and the control actions triggered can be non-intuitive to the operator even when it is correct. In case of failures and disturbances, the control actions can be completely inappropriate necessitating the operator to override them. Further, the process can be occasionally operated in new regimes that were not accounted for during the design of the MPC. Control room operators are ultimately responsible for safe and smooth operation of the process. Hence, if control actions triggered by any controller (low-level or advanced) seem suspect, the operator can turn it off at any moment. Hence, the issue of the operatorâs trust on the controller is paramount. This paper seeks to develop a systematic method to understand the operatorâs confidence in MPC through a human factors-based approach.
While advanced control techniques have been widely deployed in various process industries, it has been known for over two decades that there are still outstanding challenges in engendering a smooth interaction with the operator. The challenges in operating an MPC based controller have been described by Forbes et al. . They describe various situations where the operators preferred to switch off the MPC. The common reasons mentioned were the MPCâs move was different than the operatorsâ or the operators did not understand the control moves that were conducted by the MPC and therefore, returned to manual process control which they knew from their experience. Forbes et al also noted that operators used to single loop control did not always find it easy to understand MPCs actions, especially in large systems. This operatorsâ inability to understand the actions taken by MPC is particularly aggravated in processes having high degrees of mass and heat integration. One approach to enhance the trust between MPC and operators that has been proposed in literature is to provide well-designed operator interfaces containing carefully chosen information about the actions taken by the controller. Guerlain et al.  point out that human factors considerations are often overlooked by MPC experts who design and build advanced control systems. They suggest visual representations and HMI design principles to better support the interaction between the human operator and the automation. However, there are few studies that seek to systematically understand and quantify the challenges faced by operators while using MPC. We seek to address this gap.
In this work, we propose a testbed that enables systematic studies of the specialized interactions between the control room operator and process control systems involving MPC controllers. The proposed testbed enables human factors studies of advanced controller similar to that reported by Adhitya et al.  for alarm systems. The test environment consists of a simulated chemical process, instrumented with various sensors, comprising multiple low-level PID-type single loop control loops as well as an MPC controller.
The process is motivated by Pourkargar et al.  and consists of two continuous stirred tank reactor (CSTRs) and a liquid-vapor separator with a recycle arrangement. It involves two first order exothermic reactions in series (A -> B -> C). B is the desired product; at high temperatures, an undesired byproduct C is formed. The process has 9 measurements available for monitoring or control and 9 control valves. Various single-loop control schemes have been deployed. Specifically, the level in the reactors and separators are controlled using P controllers that manipulate the outlet flowrates. Also, PI controllers are used to control temperature by using cooling and heating load of reactors and separator respectively as manipulating variables. A MPC-PI cascade arrangement is used to control the overall objective of the process where MPC acts as a supervisory controller (master controller) and PI loops are the regulatory slave controllers (secondary controller). MPC is used for controlling composition of desired product by manipulating temperature of reactors/separator. The temperature calculated by MPC is used as a set point by the PI controller. The MPC has been designed using the MPC toolbox of MATLAB. Next, various disturbances and failures have been implemented in the process and its automation layer.
To explicitly enable human-automation interaction, a graphical user interface for controlling the process that resembles a typical distributed control system (DCS) interface has been provided. Additionally, an MPC visualization screen, displaying information about the actions taken by the MPC in tabular form is also provided. These interfaces are instrumented with various mechanisms to observe the operatorâs interactions in order to enable human factors studies â specifically all keystrokes and mouse movements and clicks are logged continuously. Further, the testbed has been designed so that it can be seamlessly operated with various cognitive engineering sensors such as eye trackers, electroencephalogram (EEG) etc to measure the operatorâs cognitive processes.
The testbed can thus be used to conduct controlled human factors experiments to study the interactions between operators and the advanced control system under various normal and abnormal conditions. For instance, the modus operandi utilized by different operators with different levels of skills and expertise can be studied. Such studies can used enhance identify shortfalls in operatorâs understanding of the control system and customized training can be provided; alternatively the graphical interfaces for interaction between the operator and the MPC can be modified or improved to suit operatorsâ cognitive processes. In this paper, we will describe the testbed and report observations on the patterns of interaction between the automation system and different human subjects.
- Forbes M.G., Patwardhan R.S., Hamadah H. and Gopaluni R.B. (2015). Model predictive control in industry: Challenges and opportunities, IFAC-PapersOnLine, 48(8), pp. 531â538.
- Guerlain, P. Bullemer, G A. Jamleson, âThe MPC elucidator: A case study in the design for human-automation interaction,â IEEE, February 2002.
- Adhitya, S.F. Cheng, Z. Lee and R. Srinivasan, âQuantifying the effectiveness of an alarm management system through human factors studies,â Computers and chemical engineering, Elsevier, April 2014.
- B. Pourkargar, A. Almansoori and P. Daoutidis, âImpact of Decomposition on Distributed Model Predictive Control: A Process Network Case Study,â Ind. Eng. Chem. Res., 56, 9606-9616, July 2017.