Spatio-temporal degradation modeling and remaining useful life prediction under multiple operating conditions based on attention mechanism and deep learning

Abstract

Accurately predicting the Remaining Useful Life (RUL) is useful to avoid unexpected significant failure of engineering system and reduce maintenance costs effectively. Meanwhile, the diverse operating conditions and high-dimensional feature variables from diverse sensors spatially located in the system are two main obstacles for building an accurate and stable RUL prediction model. Considering that these two obstacles have not been fully considered in the state-of-art work, a novel RUL prediction method based on the improved Transformer model is proposed in this work, which resorts to the attention mechanism and deep learning considering spatio-temporal characteristics and multiple operating conditions. First, the difference of the original sequence data caused by operating conditions is eliminated by clustering and standardization in the data preprocessing process. The future condition information is also considered in RUL prediction calculation. Meanwhile, spatio-temporal feature is extracted by self-attention to realize the information fusion of multi-dimensional sensors and long-term series, in which position encoding is designed to retain the sequence information in the original sequence data. Finally, experimental results on the C-MAPSS and N-CMAPSS datasets show that the proposed method achieves a better performance compared with other existing methods.