With the widespread application of deep learning in fault diagnosis of mechanical equipment, its "black-box" nature has limited its deployment in safety-critical fields. Especially in transfer learning, where models need to adapt to different working conditions or equipment, the unobservability of their decision-making processes has exacerbated user distrust. Focusing on the interpretability of both the transfer process and the decision-making process, this paper proposes an interpretability analysis framework for deep learning-based bearing transfer fault diagnosis tasks, integrating the physical mechanisms of bearing faults. By analyzing the correlation between the transferable features of the model and the final diagnostic decisions, the internal logic of the transfer learning model in cross-condition diagnosis is revealed, thereby enhancing the transparency of the diagnostic process and the credibility of the results.