A bidirectional recursive gated dual attention unit based RUL prediction approach. (April 2023)

• Record Type: Journal Article
• Authors: Yang, Lei; Liao, Yuhe; Duan, Rongkai; Kang, Tao; Xue, Jiutao
• Is Part Of: Engineering applications of artificial intelligence. Volume 120(2023)
• Publication Date: 2023-04
• Subjects: Bidirectional recursive gated dual attention unit -- Attention mechanism -- Gated recurrent unit -- Remaining useful life -- Health stage division; Engineering -- Data processing -- Periodicals; Artificial intelligence -- Periodicals; Expert systems (Computer science) -- Periodicals; Ingénierie -- Informatique -- Périodiques; Intelligence artificielle -- Périodiques; Systèmes experts (Informatique) -- Périodiques; Artificial intelligence; Engineering -- Data processing; Expert systems (Computer science); Periodicals; 620.00285
• DOI: 10.1016/j.engappai.2023.105885
• Languages: English
• ISSNs: 0952-1976

Abstract

Abstract: With the increasing requirements for the reliability and safety of high-end equipment, the predictive maintenance of high-end equipment has been indispensable. The remaining useful life (RUL) prediction is a key part of predictive maintenance. Most existing studies regard the process of health degradation as a single-stage process, which demands the RUL prediction model to have high adaptability and generalization performance. However, current RUL prediction models cannot meet this requirement and predict accurately therefore cannot be achieved as expected. In view of that, this paper proposes a stage division method and constructs a powerful RUL prediction model to solve this problem. Firstly, a novel continuous gradient recognition algorithm is developed to identify the degradation initial time, and then the whole health degradation process is divided into two stages, called the normal operation stage and the accelerated degradation stage. Secondly, a bidirectional recursive gated dual attention unit is proposed to predict the RUL during the accelerated degradation stage. It introduces two attention gates into the classical gated recurrent unit and constructs a bidirectional structure to fully learn the forward and backward degradation law of time series as well as the initial hidden state of the forward network is corrected by the final hidden state of the backward network. Two real bearing datasets are analyzed to verify the effectiveness and capability of the proposed method. Finally, the comparative analysis is implemented and the results further show that the proposed method has better performance both in prediction accuracy and robustness.