Abstract: To address the challenges of strong non-stationarity and insufficient feature fusion in gear fault diagnosis, this study proposes a novel method based on continuous wavelet transform (CWT) and a hybrid convolutional neural network-vision transformer (CNN-ViT) architecture. First, one-dimensional vibration signals are converted into time-frequency representation via CWT to extract their non-stationary characteristics. A dual-branch framework is then constructed, wherein a CNN branch captures local features and a vision transformer (ViT) branch models long-range contextual dependencies. Furthermore, a cross-attention fusion module is designed to enable deep interaction among the ViT, CNN, and statistical features. A gating mechanism is introduced to adaptively weight the contributions from these different feature modalities. Experimental results demonstrate that the proposed method outperforms single-modality models in gear fault classification, achieving higher recognition accuracy and superior generalization capability. The framework effectively enhances diagnostic robustness under complex operating conditions.