Axlebox bearings are the key components of rail transit train bogies, which are crucial to the operation of the train. However, the conditions of transit are relatively complex, coupled with the fact that the optimization of the deep neural network of the existing fault diagnosis methods requires a large number of manual modifications and experience of experts, which makes it difficult to effectively diagnose the faults of axlebox bearings. To address the above issues, this paper proposes an improved deep reinforcement learning strategy based on Dueling Double Deep Q-Network (D3QN) and Prioritized Experience replay (PER) for the fault diagnosis of axlebox bearings. Firstly, D3QN combines Double DQN and Dueling DQN to solve the problem of overestimation of the Q-value of Deep Q Network (DQN) and improve the stability of training. Secondly, by introducing the prioritized experience replay and the elevated reward mechanism, the efficiency of experience utilization and the speed of model convergence is improved. Finally, the proposed method is validated by using subway train transmission system fault simulation datasets of Beijing Jiaotong University (BJTU). The results show that compared with the original DQN method, PER-DR3QN has significant improvement in training efficiency, accuracy and stability.