Intelligent Fault Diagnosis for Bearing Dataset Using Adversarial Transfer Learning based on Stacked Auto-Encoder

Jipu Li, Ruyi Huang, Weihua Li
2020 Procedia Manufacturing  
Intelligent fault diagnosis of bearing dataset based on transfer learning has made more attention because transfer learning has powerful transferability. However, in previous studies, a major assumption accepted by default is that the source and target domain have the same number of fault categories. Unfortunately, this assumption is invalid in real industrial work conditions because the target domain samples are unlabeled, which leads to the number of fault categories are unknown. In this
more » ... the existing transfer learning methods fail to achieve well transfer results. Therefore, inspired by the idea of open set domain adaptation, an adversarial transfer learning based on stacked auto-encoder method is proposed to address new fault emerging problem for the target domain. In the proposed method, a stacked auto-encoder (SAE) network is adopted to effectively extract transferable features. Then, adversarial learning and the gradient reverse layer are utilized to achieve model training and parameters backpropagation. Finally, the known fault types are recognized in the known fault classifier and the new fault type is identified in the new fault classifier. The proposed method is verified by a famous bearing dataset. The experiment results show that the proposed method not only able to effectively obtain the known fault categories transfer, but also separate the new fault type from the target domain samples. Abstract Intelligent fault diagnosis of bearing dataset based on transfer learning has made more attention because transfer learning has powerful transferability. However, in previous studies, a major assumption accepted by default is that the source and target domain have the same number of fault categories. Unfortunately, this assumption is invalid in real industrial work conditions because the target domain samples are unlabeled, which leads to the number of fault categories are unknown. In this case, the existing transfer learning methods fail to achieve well transfer results. Therefore, inspired by the idea of open set domain adaptation, an adversarial transfer learning based on stacked auto-encoder method is proposed to address new fault emerging problem for the target domain. In the proposed method, a stacked auto-encoder (SAE) network is adopted to effectively extract transferable features. Then, adversarial learning and the gradient reverse layer are utilized to achieve model training and parameters backpropagation. Finally, the known fault types are recognized in the known fault classifier and the new fault type is identified in the new fault classifier. The proposed method is verified by a famous bearing dataset. The experiment results show that the proposed method not only able to effectively obtain the known fault categories transfer, but also separate the new fault type from the target domain samples. Abstract Intelligent fault diagnosis of bearing dataset based on transfer learning has made more attention because transfer learning has powerful transferability. However, in previous studies, a major assumption accepted by default is that the source and target domain have the same number of fault categories. Unfortunately, this assumption is invalid in real industrial work conditions because the target domain samples are unlabeled, which leads to the number of fault categories are unknown. In this case, the existing transfer learning methods fail to achieve well transfer results. Therefore, inspired by the idea of open set domain adaptation, an adversarial transfer learning based on stacked auto-encoder method is proposed to address new fault emerging problem for the target domain. In the proposed method, a stacked auto-encoder (SAE) network is adopted to effectively extract transferable features. Then, adversarial learning and the gradient reverse layer are utilized to achieve model training and parameters backpropagation. Finally, the known fault types are recognized in the known fault classifier and the new fault type is identified in the new fault classifier. The proposed method is verified by a famous bearing dataset. The experiment results show that the proposed method not only able to effectively obtain the known fault categories transfer, but also separate the new fault type from the target domain samples.
doi:10.1016/j.promfg.2020.06.014 fatcat:6tkpxeecyjeqlggidpptxyxar4