Highly Imbalanced Fault Diagnosis of Rolling Bearings Based on Variational Mode Gaussian Distortion and Deep Residual Shrinkage Networks

In the realm of data-driven intelligent diagnosis for rolling bearings, a prevalent challenge arises from the limited number of fault samples present in the training set in comparison to the healthy samples. This imbalance contributes to a high rate of misdiagnosis in intelligent diagnosis models. In order to address this issue, a novel fault diagnosis approach is developed that employs variational mode Gaussian distortion (VMGD) and deep residual shrinkage networks (DRSNs). Initially, the faulty training samples are decomposed into intrinsic mode functions (IMFs) using variational mode decomposition (VMD). Subsequently, one of the IMFs is selected at random for distortion, and the distortion coefficients are generated according to a Gaussian distribution. The distorted IMF is then combined with the other IMFs to synthesize augmented fault samples, ensuring that the augmented samples possess mean values and standard deviations (STDs) consistent with the original samples. Finally, DRSNs are trained using the augmented training samples and employed to classify the test samples. Through a series of experiments, the proposed method is demonstrated to be effective and robust against imbalanced datasets.