Abstract: Due to the problem of small amount of data and uneven distribution of bearing fault data, transfer learning is introduced into the field of fault diagnosis. At the same time, because the distribution of bearing fault data is very different from the distribution of the source data set, the direct use of transfer learning methods will produce negative transfer effects. That is, because the distribution difference between the source data set and the target data set is too large, it is impossible to learn the knowledge of the source data set. A new diagnosis method for improving transfer learning is proposed：the two-step transfer learning method, using DCGAN to make the auxiliary for the data set. It first performs transfer learning on the auxiliary data set, and then puts the network on the target data set for transfer learning training again. According to the comparison experiment with ordinary transfer learning and without transfer learning, the new method has faster speed and higher accuracy compared with the current methods.

Key words: transfer learning, bearing fault diagnosis, deep convolutional generative adversarial network

摘要： 由于轴承故障数据存在数据量少和分布不均衡的问题，将迁移学习引入故障诊断领域，同时由于轴承故障数据的分布与源数据集分布差异巨大，直接采用迁移学习的方法会产生负迁移效应，即由于源数据集与目标数据集间分布差异过大而导致无法学习到源数据集的知识，提出一种对迁移学习进行改进的诊断新方法：即两步迁移学习法，使用DCGAN来制作辅助数据集，在辅助数据集上进行迁移学习，再将网络放在目标数据集上再次进行迁移学习训练，根据与普通迁移学习和不使用迁移学习的对比实验，新方法相较于目前已有的方法具有更快的速度与更高的准确率。

关键词: 迁移学习, 轴承故障诊断, 深度卷积对抗生成网络