Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (12): 289-298.DOI: 10.3778/j.issn.1002-8331.2011-0070

• Engineering and Applications • Previous Articles     Next Articles

Deep Manifold Transfer Learning Method for Fault Diagnosis of Rotating Machinery Under Different Working Conditions

QIU Yingyu, ZHANG Ke, YANG Xinyi   

  1. School of Information Engineering, Xuchang University, Xuchang, Henan 461000, China
  • Online:2022-06-15 Published:2022-06-15

面向旋转机械故障诊断的深度流形迁移学习

邱颖豫,张柯,杨欣毅   

  1. 许昌学院 信息工程学院,河南 许昌 461000

Abstract: Deep learning has gradually become the mainstream fault diagnosis method of rotating machinery because of its strong feature extraction ability. However, the deep models lack domain adaptation capability, and the performance degrades greatly when the working conditions change. Because transfer learning can learn domain-invariant features to some extent, it provides a new way to solve the problem of fault diagnosis under different conditions. However, most of the existing deep transfer learning methods only align the mean centers of different domain distributions, and do not consider the manifold structure of feature distributions. Thus their adaptive performance is still insufficient to deal with complex mechanical fault signals from different working conditions. To solve this problem, a deep manifold transfer learning method is proposed. Based on stacked auto encoder framework, the source domain and target domain samples are simultaneously used in the unsupervised pre-training to fully explore the features of data. For fine-tuning, a manifold transfer framework is proposed, which adapts the distribution mismatch and maintains the consistency of feature distribution structure among domains simultaneously. The new method has been compared with the existing transfer learning method in rotating machinery fault diagnosis cases. Experimental results show that the new method is superior to the existing transfer learning method and can significantly improve the fault diagnosis accuracy under different working conditions. In addition, by the effectiveness analysis, it is further proved that the unsupervised training using target domain data and fine tuning by manifold regularization are effective to improve the accuracy of cross-condition fault diagnosis tasks.

Key words: deep neural network, transfer learning, manifold learning, rotating machinery, fault diagnosis

摘要: 深度学习因强大的特征提取能力已逐渐成为旋转机械故障诊断的主要方法。但深层模型缺乏领域适应能力,工况变化时性能衰退严重。迁移学习为解决变工况诊断问题提供新的途径。然而现有深度迁移学习方法大多仅对齐不同领域分布的均值中心,未考虑特征分布的流形结构,其适配性能仍难以应对不同工况复杂的机械故障信号。针对该问题,提出一种深度流形迁移学习方法,以堆叠自编码器为框架,在无监督预训练阶段同时利用源域和目标域样本训练,充分挖掘数据本质特征;针对模型微调,提出流行迁移框架,在适配分布差异同时还保持领域间特征分布结构的一致性。将新方法与现有迁移学习方法在旋转机械故障诊断案例进行充分的比较实验,结果表明,新方法优于现有方法,能显著提高变工况故障诊断精度。通过有效性分析在机理上进一步证明了融合目标域数据的无监督预训练策略和流形迁移微调策略对提高变工况故障诊断的有效性。

关键词: 深层神经网络, 迁移学习, 流形学习, 旋转机械, 故障诊断