Physics-Informed Simulation-to-Hardware Alignement for DC-Link Capacitor Condition Monitoring

Reliable condition monitoring of DC-link capacitors is essential for the safety and longevity of traction inverters, yet long-term experimental degradation data are scarce. This work presents a physics-informed, simulation-to-hardware alignment framework for non-invasive capacitor health assessment using low-frequency (100 Hz) measurements from standard shutdown events. A high-fidelity electrothermal inverter model generates synthetic discharge profiles across a broad range of operating conditions and degradation levels. To bridge the synthetic–experimental gap, we introduce a degradation-aware latent space structuring mechanism combining domain alignment and class ordering, optimized via an adaptive multi-task loss balancing strategy. Experiments on a three-phase traction inverter demonstrate robust degradation tracking with 98.9% classification accuracy, full recall for the most degraded state, and low computational cost suitable for embedded deployment. The method enables interpretable and deployment-ready capacitor monitoring without additional sensors or control-layer modifications.