Abstract: In response to the issues of early fault signals in rolling bearings exhibiting strong nonlinearity and non-stationary characteristics and being easily submerged by strong background noise resulting in low diagnostic accuracy, a fault diagnosis method based on a combined denoising algorithm and SE-GoogLeNet v2 is proposed. Firstly, a refined denoising strategy integrating CEEMDAN and wavelet packet multi-thresholding was developed to achieve high-fidelity reconstruction of weak transient impact features. Secondly, the denoised one-dimensional sequence was mapped into a two-dimensional time-frequency spectrum via continuous wavelet transform and input into the SE-GoogLeNet v2 network. Experiments show that the average diagnostic accuracy of this method reached 98.98% with a standard deviation of only 0.15%, significantly outperforming traditional models such as ANN and SVM; moreover, it maintained strong robustness even under severe conditions with a substantial imbalance between normal and abnormal samples, providing a highly reliable intelligent solution for mechanical condition monitoring in complex noisy environments.