Fault Diagnosis of Rolling Bearings Based on Adaptive Denoising Residual Network

To address the vulnerability of rolling bearings to noise interference in industrial settings, along with the problems of weak noise resilience and inadequate generalization in conventional residual network frameworks, this study introduces an adaptive denoising residual network (AD-ResNet) for diagnosing rolling bearing faults. Initially, the sensors collect the bearing vibration signals, which are then converted into two-dimensional grayscale images through the application of a continuous wavelet transform. Then, a spatial adaptive denoising network (SADNet) architecture is incorporated to comprehensively extract multi-scale information from noisy images. By exploiting the improved pyramid squeeze attention (IPSA) module, which excels in extracting representative features from channel attention vectors, this unit substitutes the standard convolutional layers present in typical residual networks. Ultimately, this model was validated through experiments using publicly available bearing datasets from CWRU and HUST. The findings suggest that with −6 dB Gaussian white noise, the average accuracy of recognition achieves a rate of 90.96%. In scenarios of fluctuating speeds accompanied by strong noise, the recognition accuracy can reach 89.54%, while the training time per cycle averages merely 3.65 s. When compared to other widely utilized fault diagnosis techniques, the approach described in this paper exhibits enhanced noise resistance and better generalization capabilities.