Intelligent control of GaN power converters using reinforcement learning and MEMS sensors in smart manufacturing

Abstract

Gallium nitride (GaN) power converters that are run under dynamic industrial loads have significant problems of preserving their energy efficiency and thermal stability because their switching between dynamic, rapidly changing operational atmospheres cannot be fully managed by traditional fixedparameter controllers. In this work, a new reinforcement learning (RL)-based adaptive control architecture based on the use of microelectromechanical systems (MEMS) multi-modal sensor feedback is presented, allowing to optimize the work of GaN converters in real-time and autonomously. The suggested Deep Q-Network agent is able to actively learn vibration, acoustic, temperature and current signals as measured by MEMS and dynamically change the switching frequencies and gate timing parameters using a tailored reward scheme that balances efficiency maximization, thermal control and transient stability. The validation of hybrid simulation and hardware-in-the-loop experimental protocols show significant 14 % energy efficiency and 22 % of transient response time reduction as compared with proportional-integral-derivative control benchmark. The self-optimizing nature of the framework relates to critical weaknesses in traditional methods of control, and it defines a radical re-thinking of autonomous power electronics control in smart manufacturing systems that need resilience and flexibility in unpredictable operating environments. The paper represents progress in the integration of artificial intelligence and wide-bandgap semiconductor technology and cyber-physical sensing in next-generation industrial automation.