A poisson flow-based data augmentation and lightweight diagnosis framework for imbalanced rolling bearing faults

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Accurate diagnosis of rolling bearing faults is critical for safe rotating machinery operation.
• Real-world fault datasets often exhibit severe class imbalance, impeding deep learning model performance.
• Existing methods struggle with imbalanced data, necessitating novel approaches for reliable fault detection.

Purpose of the Study:

• To develop a novel diagnostic framework, PFRNet, capable of accurately diagnosing rolling bearing faults.
• To address the challenge of class imbalance in fault datasets using advanced generative modeling.
• To improve the robustness and generalization of fault diagnosis systems in industrial settings.

Main Methods:

• A Poisson Flow-based generative model is integrated with a lightweight residual network (PFRNet).
• Raw vibration signals are converted to time-frequency representations using Continuous Wavelet Transform (CWT) for feature extraction.
• The Poisson generative mechanism synthesizes realistic minority-class samples by learning data distributions, mitigating class imbalance.

Main Results:

• PFRNet demonstrated superior diagnostic accuracy, robustness, and generalization compared to state-of-the-art methods on the CWRU benchmark.
• Quantitative evaluations confirmed that synthesized samples closely resemble real data in quality and diversity.
• The framework effectively mitigates the impact of class imbalance on diagnostic performance.

Conclusions:

• PFRNet offers a promising solution for reliable rolling bearing fault diagnosis under imbalanced industrial conditions.
• The integration of Poisson Flow generative models enhances the ability to handle imbalanced datasets.
• The proposed method contributes to safer and more efficient operation of rotating machinery.