A fault diagnosis method based on multi-task deep learning is proposed, which classifies fault diagnosis tasks into fault classification and defect severity recognition. The shared layer uses convolution neural network to extract fault characteristic information contained in monitoring vibration signal, and the two subtask modules use gated recurrent unit to classify fault and recognize defect severity respectively. In multi-task deep learning method, the two subtask modules can interact with each other through the shared layers to promote the feature extraction ability of the model and make the whole model have better fault diagnosis performance. Fault diagnosis experiments are carried out on bearing dataset and compared with fault classification single-task model and defect severity identification single-task model to verify the fault diagnosis performance of multi-task deep learning method. The experimental results show that the accuracy of the multi-task deep learning model is 99. 79% for both tasks on the test set. In order to further verify the feature extraction capability of the multi-task deep learning method, different degrees of Gaussian noise were added to the test set for fault diagnosis experiment. Under the condition of strong noise, the accuracy of the multi-task deep learning model was significantly higher than that of the single-task deep learning model. The research results show that the multi-task deep learning model can diagnose fault more accurately and has better denoising than the single-task deep learning model, which has certain practical value.