Imbalanced fault diagnosis based on sample-weighted counterfactual

In fault detection tasks, the scarcity of fault samples often leads models to learn primarily from normal samples, resulting in biased predictions. To address class imbalance, this study introduces a data augmentation framework. It combines sample weighting based on stable learning with counterfactual generation. First, sample weighting is applied to enhance the model’s ability to capture true feature-outcome causal relationships. Then, a clustering algorithm selects high-weight samples with substantial distribution differences. Based on these representative normal samples, a multi-objective counterfactual generation method synthesizes fault samples under physical constraints, while weighted feature importance is used to identify key diagnostic features. The proposed approach effectively alleviates data imbalance in complex industrial fault diagnosis, improving both the accuracy and interpretability of fault detection. Experiments conducted on the Tennessee Eastman Process and the Multi-phase Flow Process show that our method significantly improves fault diagnosis performance under imbalanced data conditions.