Dynamic ensemble fault diagnosis framework with adaptive hierarchical sampling strategy for industrial imbalanced and overlapping data

The coexistence of class imbalance and class overlap significantly challenges fault diagnosis in modern industrial processes. Class imbalance, characterized by the scarcity of fault data, and class overlap, arising from similarities between normal and fault data as well as correlations among fault types, are intertwined issues that jointly degrade fault diagnosis performance. To address these coupled issues, this paper proposes a dynamic ensemble fault diagnosis framework with adaptive hierarchical sampling strategy (DEAHS). The framework employs a boosting ensemble structure, effectively mitigating class imbalance through dynamic majority class undersampling and reducing class overlap by focusing on minority classes in high-overlap regions. In the outer layer, a Markov decision process guides the adaptive undersampling of majority class, achieving relatively balanced subsets. In the inner layer, a membership entropy-based method identifies overlap regions, and a weighted oversampling strategy improves minority classes’ representation in these regions. The proposed framework is validated on the Tennessee Eastman process and a real-world polyester esterification process, where its performance is evaluated using four metrics commonly employed for imbalanced datasets. The results demonstrate that the proposed method achieves superior performance across a majority of metrics, highlighting its effectiveness in handling imbalanced and overlapping industrial fault data.