The fault diagnosis of the planetary gearbox often relies on strong professional knowledge, and the universality of the diagnosis model is poor. Based on deep reinforcement learning, an intelligent fault diagnosis method of the planetary gearbox using the deep wide convolution Q network is proposed. Firstly, fault diagnosis of the planetary gearbox is resolved into a sequential decision problem, which is described by the classification Markov decision process. The fault diagnosis simulation environment is established. Secondly, a deep wide convolutional neural network is designed as an action-value network in the deep Q network model to enhance the perception ability of the environmental state. Finally, the model learns the best diagnostic policy autonomously by interacting with the environment and according to the reward of the environment. In this way, the state identification of the planetary gearbox can be achieved. Experiment and case results show that this method can effectively and accurately realize the intelligent diagnosis of the planetary gearbox under multiple working conditions. The diagnostic accuracy is more than 99% , which enhances the generalization and universality of the diagnosis model.