Abstract:
The changes in rotating speed and load of mechanical systems cause the varying working conditions of gearboxes or bearings, so deep learning-based fault diagnosis models ...Show MoreMetadata
Abstract:
The changes in rotating speed and load of mechanical systems cause the varying working conditions of gearboxes or bearings, so deep learning-based fault diagnosis models face inconsistent distribution of training and test sets, resulting in a great reduction in diagnosis accuracy. To solve this problem, various effective transfer learning-based fault diagnosis methods are proposed in the literature; however, the majority of current works still suffer the following drawbacks: First, they usually align the distribution at the same-level layer of the parameter-shared feature extractor, but in fact, the same-level fault features may distribute in different-level layers of the feature extraction network. Second, the model trained under supervised information of the source domain cannot guarantee discriminability for the target domain samples. To address the above issues, a novel unsupervised adaptive cross-layer alignment network with norm constraints (ACLAN-NC) framework is developed in this article, which aligns the same-level fault features among the same-level and different-level model layers. By using those cross-layer pairs, we can align the fault features at a fine-grained level and reduce inappropriate knowledge transfer. The batch nuclear-norm minimization and maximization technique is, moreover, used to restrict the prediction output of the source domain and target domain, ensuring the discriminability and diversity of the model. To demonstrate the effectiveness of ACLAN-NC for fault diagnosis under varying working conditions, extensive experiments are conducted to manifest the superior performance of our proposed ACLAN-NC framework.
Published in: IEEE Transactions on Instrumentation and Measurement ( Volume: 72)