A Novel Dual-Domain Adversarial Method for Vibration Signal Denoising in Bearing Fault Diagnosis

Rolling bearings are critical components in rotating machinery, and their malfunction can result in significant losses. Therefore, accurate fault diagnosis is essential for maintaining the performance and reliability of such machinery. However, this process is often hindered by the substantial noise in industrial environments, which can contaminate bearing vibration data and obscure critical features. To address the challenge of recovering vibration signals contaminated by noise, we propose a deep learning (DL) framework named dual-domain adversarial residual encoder-decoder networks (DAREDNs). This framework utilizes a cascade structure composed of adversarial residual encoder-decoder networks (AREDNs) for denoising. The first AREDN processes noisy signals in the time domain to generate initial denoised signals, while the second AREDN refines the signals further in the frequency domain. Additionally, we introduce dual-domain reconstruction and adversarial regularization strategies during training to enhance the feature extraction capabilities of the network. To evaluate its performance, we design targeted experimental strategies and compare DAREDN with three other DL-based denoising models using both public and private bearing fault datasets. Experimental results demonstrate that DAREDN effectively reconstructs vibration signals under high noise conditions, significantly improving fault diagnosis accuracy. These findings highlight the potential of DAREDN as a robust solution for noise-robust fault diagnosis in industrial applications.