Class-Conditioned Diffusion Augmentation and Explainable In-Context Ensemble Learning for Imbalanced Maintenance Action Classification in Aircraft Systems

Class imbalance in aircraft predictive maintenance data causes classifiers to favor frequent actions and miss rare but operationally important outcomes. This study proposes a hybrid framework that combines a Class-Conditional Tabular Denoising Diffusion Probabilistic Model (TabDDPM-CC) with a k-Center Ensemble Tabular Prior-Data Fitted Network (TabPFN-CE). It was applied to a 647-record aircraft oil pump maintenance dataset with four classes and an imbalance ratio of 9.0. TabDDPM-CC was compared on fidelity, privacy and utility with ten augmentation methods, among them state-of-the-art diffusion models. TabPFN-CE combined k-center greedy context selection with posterior averaging and was compared with eleven classifiers. Ablations, paired statistical tests and noise and shift experiments examined each design choice. The proposed pipeline raised balanced accuracy from 0.454 to 0.841 and G-mean from 0.439 to 0.840. After Holm–Bonferroni correction, its advantage was significant against most comparators but not against SMOTE, three diffusion baselines or a single k-center context. Among the generator components, class conditioning had the largest effect. SHapley Additive exPlanations (SHAP) showed that the predictions rely on operationally plausible variables. On the public AI4I 2020 dataset, the framework reached a balanced accuracy of 0.901. Class-conditioned augmentation with context-aware classification can therefore support interpretable maintenance decisions on small imbalanced datasets.