Predicting remaining useful life of a machine based on embedded attention parallel networks. (1st June 2023)

• Record Type: Journal Article
• Authors: Zhang, Xiong; Guo, Yunfei; Shangguan, Hong; Li, Ranran; Wu, Xiaojia; Wang, Anhong
• Is Part Of: Mechanical systems and signal processing. Volume 192(2023)
• Publication Date: 2023-06-01
• Subjects: Attention mechanism -- Gate recurrent unit (GRU) -- Long short-term memory (LSTM) -- Remaining useful life (RUL) prediction; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2023.110221
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: A new LSTM-based improved structure ResLSTMa is proposed. A new embedded attention structure (EA) is proposed for feature extraction. Parallel Networks Based on Basic and Advanced Feature Extraction. Abstract: Remaining useful life (RUL) prediction of mechanical equipment is the core task of the mechanical system of prediction and health management (PHM). To find data that is highly correlated with equipment degradation in enormous data, the mutual use of long short-term memory (LSTM) network and attention mechanism achieves good performance. However, the different combination of the two will greatly affect the efficiency of data feature extraction, and some disadvantages of the traditional sequential LSTM structure and the basic attention mechanism also limit the performance of the network. In this paper, we propose an embedded attention-based parallel network (EAPN) framework for RUL prediction on machine equipment. A new LSTM-based structure, ResLSTMa, is designed for feature extraction of data, and the Query branch of the attention mechanism is improved. Moreover, the ResLSTMa structure is embedded in the Query branch to improve the attention of the relevant features of the attention network, which is combined with the gate recurrent unit (GRU) basic feature extraction network of the parallel branch to achieve multi-level feature extraction and fusion to improve the performance of RUL prediction. The proposed EAPN is evaluated on the aircraft engine dataset and compared with some state-of-the-art prediction methods, and the experimental results demonstrate that our algorithm outperforms these state-of-the-art methods.