Random Forest-Based Early Fault Detection and Classification in DC-DC Converters Using Multi-Sensor Electrical and Thermal Features

Abstract

Intelligent condition monitoring of Direct Current (DC) to DC converters requires fault detection that is accurate, reproducible and early enough to act before conventional protection trips. This paper presents a Random Forest (RF) framework that fuses output-voltage, inductor-current and heatsink-temperature measurements of a 48 V/10 A synchronous buck converter (96 V input, 20 kHz switching, cascaded PI control). A fully specified averaged-model simulation with sensor models (10 kHz sampling) and per-run component tolerances generates 200 labelled runs (40 per class, 60 s each; 59,000 windows of 0.2 s) covering normal operation, overcurrent (13.8–15.2 A), overvoltage (55.5–59.0 V), overtemperature caused by cooling degradation, and a combined severe fault. Ten features are extracted per window, including ripple factors, ripple Total Harmonic Distortion (THD), temperature slope and a thermal-model residual. A five-class RF (128 trees, maximum depth 14) tuned by Bayesian optimisation under run-wise cross-validation achieves 97.9% accuracy and 98.1% macro-F1 on 60 independent test runs (16,843 windows), and 97.7 ± 1.4% accuracy in repeated group 5-fold cross-validation. A tuned support vector machine attains higher in-distribution accuracy (99.5%), whereas RF gives the highest mean accuracy (98.0%) under six unseen operating conditions. Fusing electrical and thermal features raises accuracy from 70.2% (electrical only) and 75.4% (thermal only) to 97.9%. With a three-window alarm rule, all 48 test faults are flagged before the conventional trip threshold is crossed, with mean lead times of 3.9 s (overcurrent), 7.5 s (overvoltage), 26.5 s (overtemperature) and 2.5 s (severe fault). A classifier-weighted Fault Severity Index (FSI) rises from 0.04 in normal operation to 2.62 under severe fault, providing a quantitative severity ranking.