残差卷积自编码网络无监督迁移轴承故障诊断

深度学习类轴承故障智能诊断研究中，一般会假设训练数据与测试数据同分布且典型故障样本充足，而实际工况复杂多变，难以获得大量标签数据。将残差学习引入卷积自编码，并结合迁移学习，提出了基于残差卷积自编码无监督域自适应迁移的故障诊断方法。堆叠一维卷积自编码进行特征提取，通过残差学习避免过拟合，提高学习效率；融合多层多核概率分布适配来约束网络学习域不变特征；实现了基于无监督域自适应迁移学习的故障诊断，并获得了较高准确率的识别结果。采用凯斯西储大学轴承数据集进行验证，结果证明了所提出方法的有效性，此外还对主要参数及其影响进行了探讨并给出了对比结果。

Unsupervised Transfer Learning with Residual Convolutional Autoencoder Networks for Bearing Fault Diagnosis

Intelligent fault diagnosis was widely studied based on deep learning for rolling bearings, and most researches assumed that the training and test data possessed identical distributions and there were sufficient samples of typical faults. However, considering complicated and variable operating conditions, it was difficult to obtain the real fault samples in large quantities in practical applications. Introducing residual learning into convolutional auto-encoder and combining with transfer learning, the paper proposed an unsupervised domain adaptive transfer learning method based on residual convolutional auto-encoders. Firstly, multiple one-dimensional convolutional auto-encoders were stacked to extract fault features, which adopted residual learning to avoid over-fitting and improved learning efficiency; secondly, integrated multi-layer multi-core probability distribution adaptation to constrain the domain learning of invariant features. Finally, the unsupervised and domain adaptation transfer learning were realized and good diagnosis results were obtained. The effectiveness of the proposed method was verified on bearing datasets from the bearing datasets of Case Western Reserve University. And the effects of the main parameters were discussed and comparisons were also presented.