Classification of Imbalanced Data by Oversampling in Kernel Space of Support Vector Machines

Area of Science:

• Machine Learning
• Industrial Engineering
• Data Science

Background:

• Industrial machine fault diagnosis data is often imbalanced, posing challenges for traditional classifiers.
• Imbalanced datasets lead to biased models, hindering accurate fault stage diagnosis.
• Existing methods like synthetic minority oversampling technique (SMOTE) have limitations with nonlinear problems.

Purpose of the Study:

• To develop an improved oversampling technique for imbalanced datasets in industrial fault diagnosis.
• To address the limitations of traditional SMOTE in nonlinear feature spaces.
• To create a hierarchical framework for multiclass imbalanced problems with progressive class orders.

Main Methods:

• Proposing weighted kernel-based SMOTE (WK-SMOTE) for oversampling in the feature space of Support Vector Machine (SVM) classifiers.
• Integrating WK-SMOTE with a cost-sensitive SVM formulation.
• Developing a hierarchical framework for multiclass imbalanced problems with ordered classes.

Main Results:

• WK-SMOTE combined with cost-sensitive SVM significantly improves performance on benchmark imbalanced datasets.
• The proposed methods outperform baseline approaches in fault diagnosis tasks.
• Validation on a real-world industrial problem demonstrated effectiveness in detecting high-voltage equipment insulation deterioration.

Conclusions:

• WK-SMOTE offers a robust solution for imbalanced data challenges in industrial fault diagnosis.
• The hierarchical framework effectively handles multiclass imbalanced problems with progressive class structures.
• The developed techniques enhance the reliability and accuracy of industrial equipment monitoring systems.