A Transformer Model-Based Approach to Bearing Fault Diagnosis

Bearings are an important component in rotating machinery and theirfailure can lead to serious injuries and economic losses, therefore the diagnosis ofbearing faults and the guarantee of their smooth operation are essential steps inmaintaining the safe and stable operation of modern machinery and equipment.Traditional bearing fault diagnosis methods focus on manually designing complexnoise reduction, filtering, and feature extraction processes, however, theseprocesses are too cumbersome and lack intelligence, making it increasingly difficultto rely on manual diagnosis with large amounts of data.With the developmentof information technology, convolutional neural networks have been proposed forbearing fault detection and identification. However, these convolutional modelshave the disadvantage of having difficulty handling fault-time information, leadingto a lack of classification accuracy. So this paper proposes a transformer-basedfault diagnosis method, using the short-time Fourier transform to convert the onedimensionalfault signal into a two-dimensional image, and then input the twodimensionalimage into the transformer model for classification. Experimentalresults show that the fault classification can reach an accuracy of 98.45%.