Multilabel Transfer Learning Method With Dynamic Multimetric for Coupling Fault Diagnosis

In industrial practice, as systems become increasingly complex and integrated, the simultaneous failure of multicomponents, namely, coupling faults, has become more prevalent, which can be viewed as multilabel data. In addition, due to the changing industrial tasks, coupling fault diagnosis problems under varying operating conditions can be treated as cross-domain multilabel learning problems, which can be solved by multilabel transfer learning methods. However, existing multilabel transfer learning methods are all preliminary explorations lacking an in-depth exploring the multilevel similarity and complex features of coupling faults. To address this challenging problem, we propose a novel multilabel transfer learning method for coupling fault diagnosis, which achieves dual domain alignment at two levels. At the global feature level, the hypothesis space is reduced by minimizing the maximum mean discrepancy (MMD) at multistages of the network to align the global distribution. Furthermore, we decomposed the overall similarity into a combination of multiple local similarities and innovatively designed a dynamic multimetric structure to capture the diverse similarity characteristics of the data. By integrating the multimetric structure and dynamic mapping selection technique to form mathematical representations of this diverse similarity, this approach constrains the consistency of the local spatial structure of the two domains to achieve local space structure alignment. This method performs high superiority in multiple transfer tasks on the public and laboratory datasets, strongly demonstrating its effectiveness.