Feature extraction from vibration signals is a critical component in the diagnosis of machine faults. However, traditional wavelet transform relies on empirical and spectral characteristics to determine scale parameters, making it difficult to ensure optimal transformation results. To address this issue and further extract information from bearing vibration signals, improving the accuracy of fault type diagnosis, this paper proposes a novel rolling bearing fault diagnosis method called DWCCA (Dynamic Wavelet-Convolution Cross Attention). This methodology incorporates an autoencoder for data denoising, subsequently utilizes wavelet transform with trainable scale parameters alongside convolutional neural networks for feature extraction, and achieves a comprehensive integration of multi-scale features through a cross-attention module. This process bolsters the reliability of feature representation and the model's adaptability, rendering it more effective in the analysis of non-stationary signals. Experimental results show that the DWCCA model performs excellently in fault diagnosis, achieving a high classification accuracy that significantly outperforms other deep learning models. Consequently, the DWCCA proposed in this paper provides an efficient and reliable technical means for rolling bearing fault diagnosis.