Machine Learning for Industrial Processes: Prediction, Monitoring, and Adaptive Control

• Author / Creator

  Nian, Rui

• Machine learning (ML) has shown great potential to create tremendous value and growth to all sectors around the world, enhancing productivity, health, and longevity of humanity. ML differentiates itself from all previous methods through its adaptive and self-learning capabilities. In recent years, the energy sector experienced significant setbacks due to collapsing commodity prices and increasing pressure from environmental groups. As such, the sector is now actively seeking new innovative strategies to improve their bottom line. One such avenue is to centralize and leverage historical data for process optimization and enhanced business decisions, a concept known as Industry 4.0. This thesis aims to explore and demonstrate the capabilities of ML in process modelling, monitoring, and control.

  Central to Industry 4.0 is the ability and necessity to create value for stakeholders. As technology continues to progress and data continues to accumulate, it becomes increasingly difficult for engineers to fully understand and optimize modern processes. By leveraging ML, whose performance is highly correlated with the amount of training data, highly multi-variate relationships within modern processes can be identified. Through their discovery, multi-variate optimizations can be leveraged to further enhance process performance and push the bounds operating efficiency. In Chapter 2, a comprehensive process for identifying the multi-variate relationships and optimization step was shown and applied onto an industrial pipeline.

  Just as important to process optimizations is the topic of process safety. Currently, the strongest line of defense against process upsets is proactive risk management, where the hazards are eliminated or isolated before they escalate. If this fails, industrial alarms will warn plant operators of the potential dangers. Unfortunately, many industrial alarm systems are poorly designed, resulting in thousands of flooding alarms during process upsets. Here, ML was first used to construct proactive anomaly prediction tools for passive risk monitoring. To tackle alarm floods, a ML-based alarm management system was introduced to mitigate redundant alarms and prioritize safety critical alarms.

  Lastly, process control and optimal control are perhaps the most important subjects in the modern process industry for safety and operation excellence. Traditional optimal control used methods such as model predictive control (MPC) where a model of the process is identified and leveraged to perform multi-variate optimization. Such methods were widely demonstrated on small systems; however, their application in large, multi-variate systems are still limited due to computation constraints. Furthermore, the identification of such processes may not be feasible. In Chapter 5, a ML-based optimal control algorithm, known as reinforcement learning (RL), was leveraged instead to perform optimal control. The two main advantages of RL are its unreliance on a process model and cheap online computation cost, making it a convincing method for processes with un-identifable and/or fast dynamics.

  Each application presented was then applied onto an engineering system to validate its effectiveness and feasibility in a physical process. Pipelines, distillation towers, and wastewater treatment plants were selected as the engineering systems due to their importance to society, making them prime targets for optimization. By leveraging ML and RL, the pipelines and wastewater treatment plants undergone significant cost savings while still meeting strict government regulations. Moreover, safety and reliability were greatly enhanced on the distillation tower through a RL fault-tolerant control system. To explore the current progress of RL, this thesis was concluded with a literature review of its current applications in the process industry.

• Subjects / Keywords

  • reinforcement learning
  • data analytics
  • optimization
  • anomaly detection
  • machine learning
  • optimal control
  • control systems
  • process monitoring

• Graduation date

  Spring 2020

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-az2y-sy21

• License

  Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.