The vibration signal of planetary gearbox has the complexity of frequency component and timevarying. To solve this problem, the fault diagnosis method based on deep learning with timefrequency fusion and attention mechanism is proposed. Firstly, the wavelet packet decomposition is used to transform the original vibration signal into two dimensions of frequency band and time, which are adopted as input data. Then, the convolutional neural network is applied to fuse the frequency band characteristics of the data. The bidirectional gated recurrent unit is employed to fuse the timing features. The attention structure is adopted to weight and merge the features of different time point adaptively and dynamically. Finally, the classifier is used to identify the endtoend fault diagnosis of the planetary gearbox. Experimental results show that this method has higher accuracy than the existing deep learning fault diagnosis model. It can accurately diagnose various health states of planetary gearbox.