Health indicators for remaining useful life prediction of complex systems based on long short-term memory network and improved particle filter

• Published in: Reliability engineering & system safety 2024-01, Vol.241, p.109666, Article 109666
• Main Authors: Zhang, Yadong, Zhang, Chao, Wang, Shaoping, Dui, Hongyan, Chen, Rentong
• Format: Article
• Language: English
• Subjects: Health indicator; Long short-term memory network; Particle filter; Performance degradation; Prognostics
• ISSN: 0951-8320; 1879-0836
• DOI: 10.1016/j.ress.2023.109666

•The ranking of features is guided by transforming sensor features into distance vectors.•Constructs HIs by learning implicit relationships and differences between features.•Overcome particle filter deficiencies by automatically generating proposal distributions.•The dynamic estimation is combined with the HI to improve the prediction accuracy.•The proposed framework is able to quantify the uncertainty in the RUL prediction. In recent years, the development of sensing technology has enabled engineers to collect large amounts of data for condition monitoring and life prediction of complex systems. Although some research has explored the health indicators (HIs) of degraded systems, Conventional methods mostly define and assume initial conditions, which may lead to inconsistencies with the actual degradation. In this paper, on the basis of long-short-term memory (LSTM) network, a HI construction method is proposed, which is integrated with improved particle filter to predict the remaining useful life (RUL) of complex systems. Firstly, considering that the traditional LSTM-based HI construction ignores the different contributions of different signals, we propose to combine LSTM and Euclidean distance (ED-LSTM) to select degenerate signals so as to construct the system's HI. Afterward, a Bayesian neural network (BNN) is introduced and embedded into the particle filter (PF) framework to replace the traditional prior distribution and overcome the defects of particle filter. Finally, the proposed integrated methodology is used to predict the RUL of a complex system before failure, and experiments are carried out on a turbofan engine dataset to verify its effectiveness. Experimental results show that the proposed framework outperforms other state-of-the-art methods.