<p>Rolling bearings are widely used in various types of machinery and equipment, and effective fault diagnosis is crucial to ensuring personnel safety and production efficiency. Due to the non-stationary characteristics of vibration signals, adaptive analysis and feature extraction that can reduce the influence of subjective experience factors have also been the focus of research. In this paper, a dual adaptive bearing fault diagnosis scheme is proposed. In the proposed method, the classical adaptive decomposition method is first employed to decompose the bearing signal into multiple components. Then, the multi-branch joint attention network (MBJANet) is developed to abstract and fuse the features of each component separately. Different from the traditional channel and spatial attention, the attention in the proposed MBJANet is based on the Pearson coefficient between the components obtained by signal decomposition and the original data, and the attention mechanism is extended to the branch level of the deep architecture to achieve autonomous adjustment of the importance of different component features. Besides, the use of adaptive decomposition methods fully considers the multi-component characteristics of bearing signals. Compared to the unified processing of mixed vibration signal directly using deep models, the separate processing of component signal in the proposed dual adaptive strategy is more aligned with the characteristics of bearing vibration signals, which basically does not introduce additional subjective empirical factors. The effectiveness of the proposed method is verified through two public rolling bearing datasets.</p>