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

Abstract

Bearing signal denoising is a pivotal task in predictive maintenance and condition monitoring of industrial machinery. However, conventional denoising methods often face difficulties in simultaneously suppressing noise and preserving essential physical features. To address this challenge, we propose a novel denoising framework that incorporates physical feature priors derived from low-dimensional manifold-based simulated data. Specifically, a regression branch—constructed using convolutional and residual neural networks—is integrated into the main denoising model to exploit the intrinsic structure of bearing signals. By embedding manifold-informed priors, the regression branch enhances denoising performance and ensures the retention of critical physical features. Experimental results demonstrate that the proposed approach surpasses traditional methods in both signal denoising and bearing fault diagnosis. Notably, the incorporation of manifold-derived priors improves the model’s capability to capture the underlying physical characteristics of the signals, indicating that the network has learned their inherent features rather than simply minimizing the loss function. Overall, this study introduces a robust denoising paradigm for complex industrial environments where bearing signals exhibit significant variability.