Machine Learning Based Condition Monitoring for SiC MOSFETs in Hydrokinetic Turbine Systems

This work demonstrates a machine learning (ML) based condition monitoring system for silicon carbide MOSFETs in a hydrokinetic turbine (HKT) energy conversion system. In this application, the power electronics are underwater, and their maintenance is challenging and expensive. At the device level, MOSFET on-state resistance ( $R_{dson}$ ) can be monitored to track MOSFET degradation. Conventionally, the variation in $R_{dson}$ with temperature is compensated for by explicit measurement or estimation of junction temperature $T_{J}$ , which can be difficult to implement. Instead, in the proposed system, $R_{dson}$ load and temperature dependencies are accounted for via a ML model of the system, which first predicts the $R_{dson}$ of a healthy MOSFET given the system operating conditions, and then this prediction of healthy $R_{dson}$ is compared to the actual $R_{dson}$ measurement, with the difference tracking the change in $R_{dson}$ due to degradation. This ML based method is particularly advantageous for the HKT system, since the dynamics of the electrical and thermal systems as well as their variation with water speed or temperature do not need to be modeled. The proposed condition monitoring (CM) systems using this ML approach are demonstrated by simulation and experimental testing.