Industrial Imbalanced Fault Diagnosis Method Based on Borderline SMOTE Integrated with NPE and CatBoost

The data collected in modern process industry have imbalanced, high-dimensional, and non-linear features, which bring great challenges to chemical process fault diagnosis. Facing these features of data, we present a new fault diagnosis method based on Borderline Synthetic Minority Over-Sampling Technique (BorSMOTE) integrated with Neighborhood Preserving Embedding (NPE) and CatBoost named BSNC. In the proposed BSNC, BorSMOTE is an improved oversampling method based on SMOTE, which improves the class distribution of samples by using only minority class sample on the boundary to synthesize some new samples; NPE is used for dimensionality reduction (DR) to extract critical features associated with faults; finally, CatBoost is used as a classifier to identify the fault types. In order to verify the feasibility of the proposed BSNC methodology, the Tennessee Eastman process (TE) with different types of fault data is chosen for simulation experiment validation. The simulation results show that the BSNC methodology in this paper has considerable performance compared with the data in the imbalanced state and the related DR methodologies.