Transformer model is being gradually studied and applied in bearing fault diagnosis tasks, which can overcome the feature extraction defects caused by long-term dependencies in convolution neural network (CNN) and recurrent neural network (RNN). To optimize the structure of existing transformer-like methods and improve the diagnostic accuracy, we proposed a novel method based on the multiscale time-frequency sparse transformer (MTFST) in this paper. First, a novel tokenizer based on shot-time Fourier transform (STFT) is designed, which processes the 1D format raw signals into 2D format discrete time-frequency sequences in the embedding space. Second, a sparse self-attention mechanism is designed to eliminate the feature mapping defect in naive self-attention mechanism. Then, the novel encoder-decoder structure is presented, the multiple encoders are employed to extract the hidden feature of different time-frequency sequences obtained by STFT with different window widths, and the decoder is used to remap the deep information and connect to the classifier for discriminating fault types. The proposed method is tested in the XJTU-SY bearing dataset and self-made experiment rig dataset, and the following work is conducted. The influences of hyperparameters on diagnosis accuracy and number of parameters are analysed in detail. The weights of the attention mechanism (AM) are visualized and analysed to study the interpretability, which explains the partly working pattern of the network. In the comparison test with other existing CNN, RNN, and transformer models, the diagnosis accuracy of different methods is statistically analysed, feature vectors are presented via the t-distributed stochastic neighbor embedding (t-SNE) method, and the proposed MTFST obtains the best accuracy and feature distribution form. The results demonstrate the effectiveness and superiority of the proposed method in bearing fault diagnosis.

Dongba characters are ancient ideographic scripts with abstract expressions that differ greatly from modern Chinese characters; directly applying existing methods cannot achieve the font style transfer of Dongba characters. This paper proposes an Attention-based Font style transfer Generative Adversarial Network (AFGAN) method. Based on the characteristics of Dongba character images, two core modules are set up in the proposed AFGAN, namely void constraint and font stroke constraint. In addition, in order to enhance the feature learning ability of the network and improve the style transfer effect, the Convolutional Block Attention Module (CBAM) mechanism is added in the down-sampling stage to help the network better adapt to input font images with different styles. The quantitative and qualitative analyses of the generated font and the real font were conducted by consulting with professional artists based on the newly built small seal script, slender gold script, and Dongba character dataset, and the styles of the small seal script and slender gold script were transferred to Dongba characters. The results indicate that the proposed AFGAN method has advantages in evaluation indexes and visual quality compared to existing networks. At the same time, this method can effectively learn the style features of small seal script and slender gold script, and transfer them to Dongba characters, indicating the effectiveness of this method.