From synthetic data to actionable insights: fault detection in industry

Reliable fault detection is key requirement in industrial manufacturing, where even minor defects can result in substantial financial losses and safety risks. However, the reliable deployment of machine learning (ML) models in such settings is constrained by two key challenges: severe class imbalance and limited actionable explainability. Class imbalance occurs as defective instances are rare, causing models to favor the majority class, while insufficient explainability limits experts from interpreting or acting on model predictions. To address these issues, this paper proposes a VAE–SHAP framework that integrates generative data augmentation with explainable defect mitigation. The VAE-guided module generates realistic minority-class samples through a discriminator-guided reward mechanism to address class imbalance, while the SHAP-guided module transforms feature attributions into actionable, feature-level adjustments to reduce predicted defect risk. We conduct empirical evaluations on a real-world dataset, illustrating that our proposed framework generates realistic, distributionally aligned synthetic samples, achieving a significantly lower FID score of $$3.18 \times 10^{6}$$ compared to state of the art oversampling methods