An Anti-Noise Bearing’s Fault Diagnosis Method Using Adaptive Deconvolution and Mobile ViT

Aiming at the problems of poor noise immunity and insufficient generalization from traditional deep learning models under small sample conditions and strong noisy background, a method based on adaptive maximum second-order cyclostationarity referenceless deconvolution (ACYCRD) combined with Markov transition field (MTF) and MobileViT for rolling bearing fault diagnosis. First, the impact signal of bearing faults under strong noise background is enhanced by the parameter-adaptive CYCRD algorithm to reduce the influence of strong background noise. Then, MTF is used to transform the preprocessed 1-D bearing vibration signal into a 2-D feature image with temporal correlation. Finally, the MTF image is input into the MobileViT network for training to get the fault diagnosis results. The proposed method is verified by the Southeast University gearbox dataset and the Shenyang University of Technology laboratory rolling bearing dataset under small samples and strong noisy conditions. The results show that the diagnosis model has a higher accuracy when the data are processed by ACYCRD compared to maximum correlated kurtosis deconvolution (MCKD), variational mode decomposition (VMD), and ensemble empirical mode decomposition (EEMD), where the accuracy increased by 1.73, 1.99, and 2.2, respectively. After the modal transformation is realized by the MTF, the accuracy is 2.59%, 3.12%, and 2.72% points higher than that of Gramian angular field (GAF), continuous wavelet transform (CWT), and recurrence plot (RP), respectively. Compared with other deep learning models, the method proposed in this article has higher anti-interference ability and generalization performance under the above conditions.