Abstract:In order to meet the challenge of early weak failure of rolling bearings, a new method is proposed. In this method, PCA (principal component analysis) is used to screen the features of the vibration signal to reduce the data dimension, effectively simplify the structure of the vibration data, and enhance the expressiveness of the features. Then, the CEEMDAN (Fully Adaptive Noise Set Empirical Mode Decomposition) algorithm is used to decompose the weak fault vibration signals disturbed by background noise, which enhances the ability to identify weak features by introducing adaptive noise on the basis of Empirical Mode Decomposition (EMD), effectively separates the trend and noise data, and significantly improves the accuracy of fault diagnosis. Finally, the Transformer model is introduced to further optimize the extraction and representation of features, and realize the efficient processing of long sequence data for the extraction and characterization of weak fault features. This comprehensive method has multiple advantages such as dimensionality reduction, noise suppression and long sequence processing, and is expected to achieve remarkable results in rolling bearing fault detection.