A novel deep dynamic hybrid domain adaptation network with physical information enhancement for cross-machine fault diagnosis
Feiyu Lu, Qingbin Tong, Xuedong Jiang, Ziwei Feng, Jianjun Xu, Jingyi Huo, Zengqiang Ma
Abstract
Intelligent fault diagnosis has important practical significance in the maintenance stage of industrial manufacturing. Numerous intelligent approaches have been raised to address cross-machine fault diagnosis issues. However, these approaches still suffer from two shortcomings: (1) Most methods overlook domain adaptation guided by physical information. (2) Many methods usually focus on the static joint of marginal distribution and conditional distribution, unable to assess the importance of distributions Therefore, we propose a new method for cross-machine fault diagnosis. Firstly, the physical information enhancement (PIE) algorithm can highlight fault frequencies in the envelope spectrum to extract physical prior features, enhancing the sensitivity of model to fault characteristics. Secondly, in the deep dynamic hybrid domain adaptation network (DDHDAN), a dynamic hybrid distribution alignment (DHDA) mechanism is proposed, and the joint maximum mean discrepancy (JMMD), and a dynamic weighting factor based on A−distance is designed to balance relative contributions. Finally, the diagnostic capability is enhanced by minimising the hybrid domain distribution differences between physically relevant features associated with fault frequencies. Under the unsupervised transfer learning framework, 12 cross-machine fault diagnosis tasks are conducted. Experimental results confirm the superiority of the proposed approach.