Replacing PI-type current controllers using reinforcement learning

Abstract

The usage of various learning-based solutions across fields has increased in the recent years. This umbrella term includes any method where a system learns patterns or behavior from data, such as supervised, unsupervised, and reinforcement learning. Such rapidly rising interest is driven by the recent expansion of computational power and capacities, novel algorithms, and available training data. Reinforcement learning approaches, in particular, have become prominent across various domains for their ability to handle uncertainty in their environment and adapt to high-dimensional spaces through continuous interactions, which represent a trial-and-error approach. This utility is applied in fields such as robotics, behavioral science, and control engineering, among many others. In the latter field specifically, these methods replace the existing controllers for parameter tuning, current control, etc.

This thesis contributes to the research of learning-based solutions of current control. Throughout this work, a literature review is conducted on the replacement of existing proportional-integral (PI) controller with a reinforcement learning approach. Particularly, an axial active magnetic bearing scenario is used as the example for the sake of demonstrating the flexibility of this application. Comparison of both approaches takes place and architectures are evaluated. We then propose a process of creating an architecture of the system to replacement of a traditional PI current controller with a reinforcement learning one.