A novel transfer learning fault diagnosis method of rolling bearing based on ViT-DANN under variable conditions

Abstract

In the actual engineering system, the damage of rolling bearings can cause significant accidents, resulting in significant losses. There are still some unresolved issues even though the development of rolling bearing fault diagnostic algorithms has been aided by the use of both conventional and novel deep learning techniques. Aiming at the problems of inconsistent data distribution and inefficient feature extraction of rolling bearings under variable working conditions, a vision Transformer-domain adversarial neural network model (ViT-DANN) was proposed. First, the short-time Fourier transform (STFT) is utilized to derive the two-dimensional time–frequency features from the bearing vibration signal. Second, to extract common features from the source and target domains, the vision Transformer is utilized, which strengthens the model’s ability to extract features for long-distances. Then, by the adversarial learning of ViT-DANN model, an association between the source and target domains is created, implicitly bringing these two domains closer together and enabling the proposed method to exhibit better domain adaptability. Finally, to demonstrate the proposed method’s effectiveness and generalizability, multiple open data sets are applied to imitate the transfer learning scenario. As the experimental results show, the proposed model outperforms previous models in terms of resilience and cross-domain diagnostic capacity.