Abstract: To solve the problem that it was difficult to obtain a large number of high-quality rolling bearing fault data in the actual industrial environment， and the generalization performance of the intelligent diagnosis model was poor， a fault diagnosis method was proposed based on simulation data driven and domain adaptation. Firstly， a physical model was established to obtain rich simulation data， which simulated different failure forms of bearings according to bearing parameters and corresponding operating conditions. Secondly， the transfer learning method was used to solve the problem of inconsistent data feature distributions between simulation and actual fault data. The residual channel attention mechanism network was used to extract the transfer fault features of different domains， and the adaptive operation of different domains in the network training processes was carried out through the condition maximum mean discrepancy metric criterion， which taken into account the conditional distribution discrepancies between different domains. Finally， different transfer model tests were carried out on the bearing data sets damaged by man-made damage and accelerated life test. The results show that the method proposed may obtain better recognition accuracy when the target domain contains a small number of labels. 

Key words: fault diagnosis, rolling bearing, simulation data, domain adaptation

摘要： 针对实际工业环境中较难获取大量的高质量滚动轴承故障数据，智能诊断模型泛化性能差的问题，提出了一种基于仿真数据驱动和领域自适应的故障诊断方法。首先，建立仿真数据驱动故障诊断所需的物理模型，并根据轴承的型号及相应工况模拟不同故障形式，获得丰富的仿真数据；其次，采用迁移学习的方法解决仿真与实际故障数据存在数据特征分布不一致的问题，通过残差通道注意力机制网络提取不同领域的迁移故障特征，采用考虑了源域与目标域数据特征的条件分布差异的条件最大均值差异度量准则进行网络训练过程中不同领域的自适应操作；最后，在人为损坏和加速寿命实验损坏的轴承数据集上进行了不同迁移模型的实验验证，结果表明所提方法能在目标域小样本监督情况下获得较高的识别精度。

关键词: 故障诊断, 滚动轴承, 仿真数据, 领域自适应