Rolling Bearing Fault Diagnosis Using Improved Deep Residual Shrinkage Networks
To improve feature learning ability and accurately diagnose the faults of rolling bearings under a strong background noise environment, we present a new shrinkage function named leaky thresholding to replace the soft thresholding in the deep residual shrinkage networks (DRSNs). In this work, we discover that such improved deep residual shrinkage networks (IDRSNs) can be realized by using a group searching method to optimize the slope value of leaky thresholding, and leaky thresholding in the IDRSNs can more effectively eliminate the noise of signal features. We highlight that our techniques can significantly improve the performance on various fundamental tasks. Experimental results show that IDRSNs achieve better fault diagnosis results on noised vibration signals compared with DRSNs. Moreover, we also provide a normalized processing to further improve the fault diagnosing accuracy of rolling bearing under a strong background noise environment.