Adaptive Multi-Parameter Closed-Loop PWM Control Using Machine Learning

Abstract

Pulse Width Modulation (PWM) is a popular control method used in embedded and power electronics systems to control voltage, speed, and power in applications like motor drives, converters, lighting systems, and Internet of Things (IoT) devices. The traditional PWM controllers tend to be implemented by fixed or rule-based control schemes and usually have one operating parameter which restricts their flexibility and operation in dynamically changing environmental and load conditions. Adaptive PWM strategies have received more and more attention with increasing demand of intelligent and flexible control systems built on embedded platforms. In this regard, the concept of machine learning (ML) can be effectively utilized as a data-driven solution to the problem, whereby it learns the connection between various system parameters and the duty cycle that is needed. In this paper, a multi-parameter closed-loop adaptive PWM control approach based on machine learning is proposed, in which temperature, speed, and load would be collectively counted in predicting the duty cycle. A comparative analysis of lightweight machine learning models is conducted to determine the most appropriate one to be used in embedded implementation regarding the accuracy of prediction and calculation efficiency. Moreover, a feedback loop is used to help build a more stable system and counterbalance prediction errors and external disturbances. The framework proposed is more adaptable, robust and more accurate in control and has low computational complexity, which makes it suitable to resource-constrained embedded and IoT-based control applications.