PDRNet: A Novel Physical Feature-Driven Residual Network for Motor Vibration Signal Denoising.

Area of Science:

• Mechanical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Bearing signal denoising is crucial for industrial machinery maintenance.
• Conventional methods struggle to balance noise suppression and feature preservation.
• Need for advanced techniques to handle complex industrial signal variability.

Purpose of the Study:

• To develop a novel denoising framework for bearing signals.
• To incorporate physical feature priors using manifold-based simulated data.
• To enhance signal denoising and bearing fault diagnosis.

Main Methods:

• A novel denoising framework integrating a regression branch with convolutional and residual neural networks.
• Exploiting intrinsic signal structure via low-dimensional manifold-based priors.
• Embedding manifold-informed priors to enhance denoising and retain physical features.

Main Results:

• The proposed approach outperforms traditional methods in signal denoising.
• Demonstrated superior performance in bearing fault diagnosis.
• Manifold-derived priors improved the model's ability to capture inherent signal features.

Conclusions:

• The novel framework offers a robust denoising paradigm for industrial machinery.
• The method effectively preserves critical physical features in bearing signals.
• This approach advances condition monitoring and predictive maintenance capabilities.