Abstract: In response to the issue of low accuracy in handling complex work conditions in rolling bearing fault diagnosis, a multi-task learning model and corresponding hierarchical training mode is proposed. This model combines multi-stage, multi-task-united training to achieve a hierarchical information sharing mode. It further improves the model's generalization and fault recognition accuracy beyond the ordinary multi-task learning mode, enabling the model to perform excellently on both complex and simple datasets. Meanwhile, by incorporating the bottleneck layer structure of one-dimensional ResNet, the depth of the network is ensured while avoiding issues such as gradient explosion and vanishing, fully extracting relevant features of the dataset. Lastly, experiments are designed using the Paderborn University bearing fault dataset as the test dataset. The results demonstrate that under varying degrees of operational complexity, compared to the OMoE (One-gate Mixture-of-Experts)-ResNet model without multi-task learning, accuracy improvement ranges from 5.45 to 9.30 percentage points. Compared to Ensemble Empirical Mode Decomposition Hilbert spectral transform (EEMD-Hilbert), MMoE (Multi-gate Mixture-of-Experts), and Multi-Scale multi-Task Attention Convolutional Neural Network (MSTACNN), accuracy improvement ranges from 5.47 to 18.26 percentage points.