Adaptive Semi-Supervised Random Forest for Transformer Fault Diagnosis in the case of limited labeled data

Transformer fault diagnosis based on dissolved gas analysis (DGA) is often constrained by two practical conditions: limited fault labels and pronounced class imbalance. To address these constraints, we propose an adaptive semi-supervised random forest framework (Adaptive-SSRF) that incorporates unlabeled DGA observations while controlling the risk of error propagation from pseudo-labeling. The framework integrates three mechanisms. First, a Confidence-Anchored, Performance-Scaled (CAPS) weighting scheme assigns pseudo-label weights as a function of prediction confidence and validation performance, moderating the influence of medium-confidence cases in early iterations. Second, an out-of-bag (OOB) individual review procedure updates a seniority score for each adopted pseudo-label and down-weights observations that repeatedly conflict with the model’s internal consensus. Third, a trust-stratified augmentation strategy generates synthetic samples preferentially around high-trust regions to alleviate imbalance without mechanically equalizing class counts. Using real-world DGA datasets, Adaptive-SSRF improves performance relative to supervised and self-training baselines across multiple label regimes. At a 10% labeled ratio, it achieves 79.76% average accuracy and exhibits reduced variance across repeated splits. Ablation analyses indicate that each mechanism contributes substantively to gains under low-label conditions. The proposed framework offers a practical approach to learning reliable diagnostic models when labels are scarce and the empirical class distribution is skewed.