Physics-Guided Domain Alignment for Condition Monitoring of Three-Phase Inverters

Simulation-generated synthetic data and hardware-based experimental data are both crucial for predictive maintenance in power electronics. However, their integration in condition monitoring is challenged by simulation-experimental gaps due to modeling assumptions and unmodeled dynamics. In this work, we propose a physics-guided alignment strategy to bridge this gap by leveraging a contrastive learning strategy. Based on large volume of synthetic data, we study the impact of introducing limited degraded experimental samples balancing alignment effectiveness and classification accuracy. Our results demonstrate that controlled inclusion of degraded experimental data improves domain alignment and enhances predictive performance with minimal reliance on real experimental data.