Accurate prognostics and health management of rotating machinery is crucial for the long-term stable operation of industrial equipment. However, developing a fault diagnosis model that simultaneously possesses the ability to effectively capture comprehensive data information and strong inductive bias capabilities has always been a challenge due to their opposing nature. To address this issue, this study proposes a comprehensive domain transformer based on spatial reduction attention and dynamic convolution for fault diagnosis in rotating machinery. First, the proposed model employs complete ensemble empirical mode decomposition with adaptive noise to decompose the input data into IMFs components of different scales, which are then independently fed into a multi-branch model along with the original data to integrate information from various components. Next, leveraging a spatial reduction attention mechanism and dynamic adaptive deep convolution, a dynamic feed-forward parallel network is constructed to enhance the model's ability to extract both global information and local details. Finally, a multi-scale feed-forward deep convolutional model is established to fully capture multi-scale information, thereby improving the model's inductive bias capabilities across different datasets. The proposed model is tested on datasets from both laboratory equipment and real industrial equipment. Experimental results demonstrate the robust performance of this method in fault diagnosis of rotating machinery, outperforming several state-of-the-art fault diagnosis methods.