Transfer-learning-based bearing fault diagnosis between different machines: A multi-level adaptation network based on layered decoding and attention mechanism. (15th November 2022)

• Record Type: Journal Article
• Authors: Wan, Shaoke; Liu, Jinyu; Li, Xiaohu; Zhang, Yanfei; Yan, Ke; Hong, Jun
• Is Part Of: Measurement. Volume 203(2022)
• Publication Date: 2022-11-15
• Subjects: Bearing fault diagnosis -- Transfer learning -- Domain adaptation -- Layered decoding -- Attention mechanism; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2022.111996
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: A novel transfer-learning-based bearing fault diagnosis model using unlabeled data is proposed. Shared and private feature extraction module is designed for source and target domains. Layered decoding operation is specially adopted in the shared feature extraction module. Multi-level adaptation with attention mechanism is specially proposed during distribution adaptation. Abstract: It has been a challenge to use the learned knowledge from collected labeled data of one machine to achieve the intelligent fault diagnosis of other machines. In this paper, a novel multi-level domain adaptation network based on layered decoding and attention mechanism (LDAM-MAN) is proposed for the transfer bearing fault diagnosis across machines using unlabeled data of practical machine. The architecture consists of shared and private feature extraction module, and layered decoding operation is adopted in the shared feature extraction module. Multi-level domain adaptation is developed to align the domain distribution. Attention mechanism is introduced to distribution adaptation to guarantee the features from source and target domains belong to same fault type. Six tasks of transfer fault diagnosis are designed using three different bearing datasets to validate the performance of proposed method, and the comparative experiment results show that the proposed method can achieve higher diagnosis accuracy and better transferability.