Vibration signals of rolling bearings exhibit the characteristics of a mixture of transient impacts and non-stationarity. Although traditional continuous wavelet transform  is able to provide time-frequency localization analysis, it has insufficient time-frequency focus on transient impacts, which easily leads to aliasing of non-stationary signal components and makes it difficult to accurately extract fault features. To address this issue, this paper proposes a fault diagnosis method that integrates adaptive wavelet transform  and improved MobileNetV3. To resolve the problem of insufficient frequency coverage caused by the fixed scale of traditional CWT, an adaptive wavelet scale optimization strategy is adopted to dynamically calculate the optimal scale range of the Morlet wavelet, thereby achieving accurate coverage of the characteristic frequencies of bearings. Meanwhile, through harmonic enhancement and multimodal feature fusion, the  fault features are improved. Aiming at the problems of high computational cost and significant time consumption in the training of traditional convolutional neural networks , an improved MobileNetV3 model is constructed: SE  attention modules are embedded in specific layers to enhance the ability of fault feature extraction; stochastic depth technology is introduced to improve the model's generalization performance; and Bayesian search is combined to optimize hyperparameters. Experimental results based on the dataset show that this method is capable to effectively process complex vibration signals, and realize model lightweighting while ensuring diagnostic accuracy, providing a new solution for the effective and efficient diagnosis for rolling bearings in engineering application scenarios.

rolling bearings, continuous wavelet transform, MobileNetV3, Squeeze-and-Excitation, Bayesian search