Remaining Useful Life Interval Prediction for Complex System Based on BiGRU Optimized by Log-norm

• Published in: IEEE access Vol. 10; p. 1
• Main Authors: Yan, Xiaojia, Liang, Weige, Xu, Dongxue
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Big Data; Complex systems; Data models; Degradation; Feature extraction; Fusion model; gated recurrent unit; Logic gates; Mechanical systems; prediction intervals; Predictive models; Probability density functions; remaining useful life; Statistical analysis; Statistical methods; Statistical models; Stochastic processes; system prognostics; Turbofan engines; Uncertainty; uncertainty management; Useful life
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2022.3212694

The task of remaining useful life (RUL) uncertainty management is the major challenge in solving the failure of the complex mechanical system. Primary research methods use statistical models or stochastic processes to fit the distribution of historical degradation data. However, it is difficult to accurately capture the degradation information of monitoring big data through statistics in practice. In this paper, the prediction interval (PI) obtained by the proposed feature attention-log-norm bidirectional gated recurrent unit (FA-LBiGRU) model is adopted to quantify the prediction uncertainty of RUL. Initially, the critical feature vectors are extracted from multi-dimensional, nonlinear, and large-scale sensor signals using the feature attention mechanism. Additionally, the BiGRU network is used to model and learn the time-varying characteristics of the attention-weighted features from the forward and backward directions, and the network parameters are trained by the maximum log-likelihood loss function. Ultimately, the probability density function based on the lognormal distribution is calculated to measure the uncertainty of the equipment RUL. The effectiveness of the proposed method is verified through the well-known benchmark data set of the turbofan engines provided by NASA. The experimental results show that the proposed methods can obtain higher point prediction accuracy for the complex system compared with state-of-the-art approaches and high-quality PIs satisfying real-time requirements.