A DEEP GENERATIVE MODEL FOR REMAINING USEFUL LIFE PREDICTION OF AERO-ENGINES

Predictive maintenance is helpful for building smart, sustainable, and resilient mechanical systems. Remaining useful life (RUL) prediction is the core procedure for estimating the operation time of systems before repair or replacement. Recent, deep learning-based RUL prediction methods have achieved good performance in some cases. However, the assets with same type have different RUL which results in the sample imbalance problem. In real operating condition, samples with healthy status are the majority of training samples. Few researches discussed the problem decreasing accuracy of prediction models. In this work, a deep generative model based on bidirectional long short-term memory network (Bi-LSTM) and graph neural network is proposed for RUL prediction of aero-engines considering the sample imbalance problem. The Bi-LSTM is firstly used to generate samples with minority labels by regression and the exponential weighted moving average technique (EWMA) was used to smooth the synthetic samples. Next, the graph convolution layers are used to fusing multi-sensor signals and a deep encoder-decoder structure is proposed to adaptively extract latent features. Finally, the latent features are input the predictor for RUL prediction. C-MAPSS aircraft engine simulator data is used to test the deep generative model in this work. The experiment results show that the proposed method improving prediction accuracy by shoveling the sample imbalance problem. © 2024 Computers and Industrial Engineering. All rights reserved.