Abstract:

Remaining useful life (RUL) prediction is crucial for maintaining the reliability and safety of industrial equipment, but the existing RUL prediction methods still face many challenges in processing the high-dimensional sensor data and capturing the temporal degradation patterns. To address the above issues, this paper proposes a RUL prediction method based on bidirectional long short term memory network-variational auto encoder (BLSTMN-VAE) under the constraint of degradation trend smoothing. This method is used for data preprocessing, including data noise reduction, sliding window segmentation, and label correction. Then, a BLSTMN-based VAE type feature extractor is designed to effectively extract the nonlinear relationships and long-distance dependencies in time series data. Finally, a degradation trend smoothing constraint module based on manifold learning is proposed to enhance the robustness and generalization ability of the proposed model through the assumption of local invariance. The proposed RUL prediction method is verified using the aero-engine dataset. The results show that the proposed RUL prediction method outperforms various existing RUL prediction methods, and has lower prediction errors and higher stability.

Key words: remaining life prediction, bidirectional long short-term memory network, variational autoencoder, smoothness constraint, manifold learning