Improving bearing fault diagnosis method based on the fusion of time-frequency diagram and a novel vision transformer

Bearings are indispensable components in mechanical equipment, it is crucial to realize accurate and reliable fault diagnosis of bearings. Traditional bearing fault diagnosis methods suffer from insufficient feature extraction and poor robustness. This paper presents an improved bearing fault diagnosis method based on the fusion of time-frequency diagram and a novel vision transformer. On the one hand, the method adopts continuous wavelet transform to map the time-domain feature relationship of vibration onto the time-frequency domain. On the other hand, the method designs a novel vision transformer for bearing fault diagnosis model which can effectively improve the fault diagnosis performance and reduce the computational complexity on the basis of retaining the advantage of local feature extraction and dealing with long-range feature dependencies. In this paper, a new multi-head attention module called SRWA is designed to be utilized on the novel vision transformer model. Experiments are conducted to assess and analyze the performance of the proposed models using the bearing datasets: Case Western Reserve University data set and Harbin Institute of Technology inter-shaft bearing fault diagnosis data set. The experimental results demonstrate that the classification performance of the novel model put forward in this paper surpasses the state-of-the-art bearing fault diagnosis models on different datasets, even under variable operating conditions and noise conditions.