Bearing remaining useful life prediction with convolutional long short-term memory fusion networks. (August 2022)

• Record Type: Journal Article
• Authors: Wan, Shaoke; Li, Xiaohu; Zhang, Yanfei; Liu, Shijie; Hong, Jun; Wang, Dongfeng
• Is Part Of: Reliability engineering & system safety. Volume 224(2022)
• Publication Date: 2022-08
• Subjects: Prognostics and health management -- Remaining useful life prediction -- Feature fusion -- Convolutional long short-term memory (CLSTM) network; 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.2022.108528
• Languages: English
• ISSNs: 0951-8320

Abstract

..Highlights: A novel bearing's RUL prediction approach with multi-sensor data is proposed. Multi-branch networks are designed to extract the features of each sensor's data separately. Central network fusing the information between multi-branch networks is designed. Information transfer layer is specially designed to fuse and enhance features of multi-sensor. Abstract: Deep learning methods have improved the performance of RUL prediction, and multi-sensor data has also been found can significantly improve the fault diagnosis's accuracy. Hence, it is also highly motivated to integrate the deeply learned features from multi-sensor data for RUL prediction. In this paper, a novel deep learning framework with multi-branch networks, which is called convolutional long short-term memory fusion networks (CLSTMF), is proposed for RUL prediction with multi-sensor data. In each branch networks, shallow features of single sensor's data are extracted by convolutional layer of convolutional neural network (CNN), and then convolutional long short-term memory (CLSTM) network is employed to capture deep temporal features from these shallow features. Meanwhile, a novel information transfer layer (ITL) is developed to fuse the multi-sensor data's features captured with CLSTM in different branch networks. Experiments are also performed on two real run-to-failure datasets and the results indicates that the proposed approach performs well with respect to higher accuracy.