iSTformer: A Spectral-Temporal Dual-Channel Framework With Adaptive Graph Convolutional Network for Remaining Useful Life Prediction

In recent years, data-driven approaches have achieved significant success in the fields of prognostic health management (PHM) and remaining useful life (RUL) prediction, due to their minimal requirement for prior physical knowledge. However, most existing results frequently neglect three critical issues: (1) the spatial features of spectral information; (2) the integration of time-domain features and spectral-domain information from multi-sensor data; (3) the way of embedding in Transformer-based models. Addressing these challenges could substantially improve the performance of the model and provide a novel perspective for RUL prediction. By leveraging continuous wavelet transform (CWT), we extract the spectral features from multi-sensor data, and the adaptive graph convolutional network (AGCN) is utilized to capture the spatial feature of multi-sensor data and its spectral information. Upon integration with the input data, the gated recurrent unit (GRU) is utilized for the dimensionality reduction of the features. Additionally, designed modules called ST-Block are stacked using residual connections, and the overall framework is constructed with a Transformer backbone that incorporates inverted embedding. The experimental results show that the proposed prediction framework outperforms the baseline models in almost all cases.