A Gradient-Penalized Conditional TimeGAN Combined with Multi-Scale Importance-Aware Network for Fault Diagnosis Under...

Area of Science:

• Industrial Engineering
• Machine Learning
• Data Science

Background:

• Industrial fault diagnosis relies on accurate data, but class-imbalanced datasets hinder diagnostic performance.
• Existing methods like TimeGAN struggle with multi-category fault data imbalance due to training instability.

Purpose of the Study:

• To propose a novel method, CTGAN-MSIN, for addressing class imbalance in industrial fault diagnosis.
• To enhance data augmentation and classification accuracy for multi-category fault data.

Main Methods:

• Developed a gradient-penalized Conditional Time-Series Generative Adversarial Network (CTGAN) for controlled generation of high-quality, imbalanced fault samples.
• Constructed a Multi-scale Importance-aware Network (MSIN) featuring Multi-scale Depthwise Separable Residual (MDSR) for feature extraction and Scale Enhanced Local Attention (SELA) for feature selection.
• Validated the CTGAN-MSIN method on the HUST bearing and axial piston pump datasets.

Main Results:

• CTGAN-MSIN achieved high diagnostic accuracies of 98.75% and 96.50% on the HUST bearing and axial piston pump datasets, respectively, at a 15:1 data imbalance ratio.
• The proposed method demonstrated superior performance compared to existing methods across various imbalance ratios.
• The CTGAN component effectively alleviated data imbalance through controllable generation of fault samples.

Conclusions:

• The CTGAN-MSIN method offers a robust solution for industrial fault diagnosis with imbalanced datasets.
• The combination of advanced data augmentation and a specialized classification network significantly improves diagnostic accuracy.
• This approach provides a promising direction for real-world industrial monitoring and maintenance.