ReScConv-xLSTM: An improved xLSTM model with spatiotemporal feature extraction capability for remaining useful life prediction of Aero-engine
Mingxing Huang, Lanying Yang, Gang Jiang, Xingan Hao, Hong Lu, Hang Luo, Peng Wang, Jinyang Li
Abstract
ABSTRACT Remaining Useful Life (RUL) prediction is crucial for Prognostics and Health Management (PHM) of aircraft engines. Although deep learning models based on LSTM and Transformer have achieved significant results in this field, these models typically only extract temporal features, neglecting spatial features, and struggle with parallel computation, leading to a bottleneck in RUL prediction performance. To address these issues, this paper proposes an improved ReScConv-xLSTM model that integrates the characteristics of xLSTM, ScConv, and residual structures. Firstly, the model converts one-dimensional signals from multiple sensors into multi-channel two-dimensional wavelet time-frequency images through time window processing, RobustScaler normalization, and the Continuous Wavelet Transform (CWT) method, enhancing the spatiotemporal features of the data and achieving data augmentation and noise reduction. Subsequently, these spatiotemporal features are input into the ReScConv-xLSTM model for training to learn the spatiotemporal dependencies of the data, thereby obtaining accurate RUL predictions. Experiments on the C-MAPSS and N-CMAPSS datasets of aircraft turbofan engines demonstrate that the model can accurately predict RUL values, exhibiting strong accuracy, generalization capability, and practical application potential, with the average RMSE reduced to 2.07 and the average Score value reduced to 36.42.