Multipoint Optimal Minimum Entropy Deconvolution Adjusted (MOMEDA) is an advanced deconvolution method, which can effectively inhibit the interference of background noise and distinguish the fault period by calculating the multipoint kurtosis values. However, multipoint kurtosis (MKurt) could lead to misjudgment since it is sensitive to spurious noise spikes. Considering that L-kurtosis has good robustness with noise, this paper proposes a multipoint envelope L-kurtosis (MELkurt) method for establishing the temporal features. Then, an enhanced image representation method of vibration signals is proposed by employing the Gramian Angular Difference Field (GADF) method to convert the MELkurt series into images. Furthermore, to effectively learn and extract the features of GADF images, this paper develops a deep learning method named Conditional Super Token Transformer (CSTT) by incorporating the Super Token Transformer block, Super Token Mixer module, and Conditional Positional Encoding mechanism into Vision Transformer appropriately. Transfer learning is introduced to enhance the diagnostic accuracy and generalization capability of the designed CSTT. Consequently, a novel bearing fault diagnosis framework is established based on the presented enhanced image representation and CSTT. The proposed method is compared with Vision Transformer and some CNN-based models to verify the recognition effect by two experimental datasets. The results show that MELkurt significantly improves the fault feature enhancement ability with superior noise robustness to kurtosis, and the proposed CSTT achieves the highest diagnostic accuracy and stability.
Keywords: Vision Transformer; fault diagnosis; fault visualization; multipoint envelope L-kurtosis; rolling bearing.