Aviation bearings are critical components in aircraft engines, and their fault diagnosis is of utmost importance. However, obtaining a sufficient number of labeled fault samples for these bearings is challenging due to the difficulty of collecting data across various operating conditions. This creates a cross-domain and few-shot learning scenario. To address this challenge, this paper proposes a fault diagnosis method for aviation bearings based on mutual learning and multi-scale knowledge transfer. First, a mutual learning strategy is proposed, where dual student models collaborate and teach each other during the training process. Next, multi-scale knowledge transfer is performed from decision boundaries, structured features, and state identification across the network’s hidden layers, feature layers, and decision layers, respectively, to mitigate the impact of cross-domain and limited training samples. Finally, several experiments from multiple perspectives are carried out to validate the effectiveness and superiority of the proposed method.