PaEDNet: A Robust Denoising and Classification Framework for Vibration-Based Fault Diagnosis with Measurement Noise

Area of Science:

• Mechanical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Fault diagnosis in rolling bearings is challenging due to coupled fault structures and noise in vibration signals.
• Existing methods often struggle with severe noise and complex signal patterns, limiting diagnostic accuracy.
• Effective noise reduction and feature extraction are crucial for reliable fault identification.

Purpose of the Study:

• To propose a robust fault diagnosis framework, the Phase-space adaptive Expert Denoising Network (PaEDNet), for rolling bearing vibration signals.
• To enhance the capability of extracting fault-related structures from noisy signals through novel representation and restoration techniques.
• To achieve accurate and robust fault classification even under severe noise conditions.

Main Methods:

• Phase-space reconstruction to create a 2D similarity representation from 1D vibration signals.
• A CoPaMoE-augmented adaptive denoising module for structural restoration in the 2D representation domain.
• DenseNet for fault classification, forming an integrated end-to-end diagnostic pipeline.
• Stage-wise training for the integrated framework.

Main Results:

• PaEDNet consistently outperformed comparative models on CWRU and PU datasets across various signal-to-noise ratios (SNRs).
• Achieved high classification accuracies (93.98% and 90.45%) under a low SNR of -6 dB.
• Demonstrated superior robustness in low-SNR scenarios compared to existing methods.
• Ablation studies confirmed the effectiveness of structured representation and adaptive expert restoration.

Conclusions:

• The proposed PaEDNet framework offers a novel and effective approach for rolling bearing fault diagnosis in complex noisy environments.
• The integration of phase-space reconstruction and adaptive expert denoising significantly improves diagnostic performance and robustness.
• PaEDNet provides a promising new modeling perspective for vibration signal analysis in challenging industrial conditions.