Is control theory better than AI for improving plant performance?

Understand the strengths and weaknesses of artificial intelligence (AI) and machine learning (ML) versus control theory, particularly model predictive control (MPC) for improving process and manufacturing applications and operations.

By John F. Carrier October 7, 2022
Courtesy: John F. Carrier, MIT; Control Engineering/CFE Media and Technology Virtual Training Week 2022

 

Learning Objectives

  • Review some strengths and weaknesses of AI, ML.
  • Understand when control theory, particularly model predictive control, can be more effective than AI and ML.
  • Link to on online course from Control Engineering and CFE Media and Technology with more on related topics.

Artificial intelligence (AI) and machine learning (ML) have been successfully applied to improve plant performance over the past decade. Even so, the uptake of AI techniques for improving existing operating facilities has significantly lagged expectations, with many projects stuck in what McKinsey and Co. refers to as “pilot purgatory.”

AI and ML’s greatest strengths (model free learning based on massive data sets) can be a weakness when applying to real time-varying systems. Indeed, AI’s “data-centric” paradigm, while suitable for tasks such as image classification, outlier detection and consumer behavior, is inferior to classical model-based control and system identification approaches for operating physical asset-based operations.

Model predictive control (MPC) framework often exceeds AI and ML performance for the control and adaptation of asset-based facilities. Specific AI techniques may be selectively used within this framework for performance improvement. Understanding observability and controllability helps with understanding critical control functions and where AI/ML methods perform poorly in real-world applications. Some success stories have been miscategorized as AI, when they are applications of traditional control techniques.

AI, ML, MPC properly applied

These points are covered in a Virtual Training Week session on Oct. 25, with the following learning objectives:

  • Understand the strengths and limitations of using AI and ML for process control and risk management
  • Review the fundamental role of the model in the design, implementation, maintenance, and adaptation of an effective control system, and uncover why an AI/ML approach on its own is insufficient
  • Identify the weaknesses of your organization’s AI approach, and how to address these weaknesses through the use of the MPC framework.
Interactions among the process, model and data elements of a control system are critical to creating transformational change, explained John F. Carrier, senior lecturer at the MIT Sloan School of Management in the System Dynamics Group. Courtesy: John F. Carrier, MIT; Control Engineering/CFE Media and Technology Virtual Training Week 2022

Interactions among the process, model and data elements of a control system are critical to creating transformational change, explained John F. Carrier, senior lecturer at the MIT Sloan School of Management in the System Dynamics Group. Courtesy: John F. Carrier, MIT; Control Engineering/CFE Media and Technology Virtual Training Week 2022

Control frameworks for safe, reliable operations

It’s important to go over this information, because AI marketing has taken over the industrial transformation mind space. AI and ML have limits when applied to real, noisy, non-stationary processes. MPC provides a superior framework for designing, operating and adapting real processes over the more limited AI/ML approach. Finally, the MPC framework complex problem solving, providing decision guidance to managers as well as control engineers.

While AI and ML approaches can be successfully applied in many industrial applications, AI and ML focus on the model/data set, and ignore the real process. AI and ML also are relatively weak with respect to modeling dynamic response and “brittle” with respect to noise and time-varying systems.

Control frameworks provide a much more comprehensive framework for safe and reliable operations of facilities. In particular, MPC provides a framework to guide all of plant operations, including operations and maintenance personnel. Keep in mind that simplifying the system makes it easier to model, increases productivity, and reduces operational risk.

John F. Carrier, Doctor of Science (ScD) in control systems from MIT, is a senior lecturer at the MIT Sloan School of Management in the System Dynamics Group. Edited by Mark T. Hoske, content manager, Control Engineering, CFE Media, mhoske@cfemedia.com.

KEYWORDS: Control theory, artificial intelligence, machine learning

LEARNING OBJECTIVES

CONSIDER THIS

Are you using AI/ML and control theory for the right applications?


Author Bio: John F. Carrier, Doctor of Science (ScD) in control systems from MIT, is a senior lecturer at the MIT Sloan School of Management in the System Dynamics Group.