An hybrid domain adaptation diagnostic network guided by curriculum pseudo labels for electro-mechanical actuator. (December 2022)

• Record Type: Journal Article
• Authors: Wang, Jianyu; Zeng, Zhiguo; Zhang, Heng; Barros, Anne; Miao, Qiang
• Is Part Of: Reliability engineering & system safety. Volume 228(2022)
• Publication Date: 2022-12
• Subjects: EMA fault diagnosis -- Hybrid domain adaptation -- Inter-domain clustering -- Intra-class domain clustering -- Curriculum pseudo labeling; Reliability (Engineering) -- Periodicals; System safety -- Periodicals; Industrial safety -- Periodicals; Fiabilité -- Périodiques; Sécurité des systèmes -- Périodiques; Sécurité du travail -- Périodiques; 620.00452
• DOI: 10.1016/j.ress.2022.108770
• Languages: English
• ISSNs: 0951-8320

Abstract

Highlight: We firstly carry out an unsupervised domain adaptation comparison experiment to address the domain difference challenge in EMA fault diagnosis. The introduction of intra-class domain clustering is beneficial to improve domain adaptation ability compared with only considering inter-domain clustering. The participation of target domain samples in triplet loss can further improve the domain adaptation accuracy compared with only considering source domain samples. The CPL can dynamically adjust adaptive thresholds for different classes and improve the domain adaptation accuracy compared with the fixed threshold. Abstract: Electro-mechanical actuator (EMA) usually operates in complex working conditions. When developing data-driven fault diagnosis models for EMA, training and testing data might come from different working conditions, reducing the generalization ability of traditional data-driven models. To address the challenge of domain difference between training and testing data, we propose a hybrid domain adaptation network, whose loss functions comprise of adversarial loss, triplet loss and cross-entropy loss. Adversarial loss and triplet loss can enhance the inter-domain and intra-class domain clustering, respectively. A softmax classifier with cross-entropy loss is used to predict pseudo labels for unlabeled target domain training samples. Compared to traditional transfer learning models that only reduces the global inter-domain difference between two domains, the strength of our model is that both the intra and inter-class domain difference are reduced. Curriculum pseudo labeling (CPL) is further applied to dynamically adjust thresholds for different classes during training phases. Compared to the fixed threshold in previous efforts, CPL can take into account the difference in pseudo label prediction and improve the performance of the developed model. The experiment results show that, compared to several transfer learning models, the developed model can achieve better classification accuracy in target domain.