Multivariate Condition Monitoring of SMPS Output Aluminum Electrolytic Capacitors via Frequency-Aware Transfer Learning

This paper presents an AI-driven transfer learning framework for multivariate condition monitoring (CM) of output aluminum electrolytic capacitors (AECs) in switched mode power supplies (SMPS). Existing methods are limited to single-parameter estimation in narrow frequency bands. In contrast, our approach simultaneously predicts four degradation indicators-equivalent series resistance (RS), impedance (Z), dissipation factor (D), and capacitance (CS)-across the midfrequency diagnostic band (10-100 kHz), where aging effects are most pronounced. The key innovation is cross-domain adaptation: trained on 2200 µF capacitors, the model generalizes to 470/1000 µF types with < 5% calibration data, addressing a critical industrial challenge. Experimental validation using precision LCR measurements demonstrates that the Extra Trees model achieves a < 5% MAPE for RS/Z predictions and sub-millisecond inference times, enabling real-time deployment. While CS estimation requires log-scaling for stability, our model provides integrated visualizations (residual maps, radar plots) that offer interpretable diagnostics, empowering users with a deeper understanding of the system's health. A comparative analysis confirms a statistically robust performance (p = 0.1865), with Extra Trees outperforming neural networks in terms of accuracy and stability. This work bridges lab-grade monitoring and scalable AI-enhanced maintenance with direct applications in smart grids and electrified transportation, where predictive capacitor health management can significantly reduce downtime and lifecycle costs.