Abstract
<title>Abstract</title> <p>Predictive maintenance (PdM) in Industry 4.0 relies on data-driven models that learn the signatures of impending failures from condition-monitoring data. In real operating conditions, however, fault events are scarce and unevenly distributed across failure modes, producing severe class imbalance that biases learning toward the healthy class and degrades the detection of exactly the rare, high-cost failures that PdM is meant to anticipate. This paper investigates conditional generative data augmentation as a remedy, and in particular evaluates a class-conditional denoising diffusion probabilistic model (DDPM) for synthesizing minority-class samples. Using the publicly available AI4I 2020 Predictive Maintenance dataset, we cast the problem as a five-class, single-label fault-diagnosis task with four rare failure modes and benchmark the proposed diffusion augmentation against no augmentation, random oversampling, SMOTE, a conditional GAN (cGAN) and a conditional VAE (cVAE). Each augmentation strategy is assessed both intrinsically, through distributional fidelity metrics (maximum mean discrepancy and a classifier two-sample test), and extrinsically, through downstream diagnosis performance measured with two independent classifiers (a multilayer perceptron and a random forest) over eight random seeds. The diffusion-based augmentation yields the highest rare-failure recall and the highest balanced accuracy with both classifiers, improving minority-class recall from 0.61 to 0.76 (multilayer perceptron) and from 0.41 to 0.76 (random forest), with the improvement over every baseline being consistent across all eight runs (Wilcoxon signed-rank test, p = 0.008). A transparent precision recall trade-off is observed and analyzed. The results indicate that diffusion models are a competitive and reproducible tool for mitigating class imbalance in predictive maintenance, especially when the operational priority is to maximize the detection of rare failures.</p>