A two-stage remaining useful life prediction method based on adaptive feature metric and graph spatiotemporal attention rule learning

Rolling bearing remaining useful life (RUL) prediction is crucial for ensuring reliability and developing maintenance strategies. However, current researches on data-driven RUL prediction methods faces limitations. Firstly, they often overlooks the differences in degradation properties across different stages. Secondly, recurrent neural networks (RNN) and long short-term memory networks (LSTM), commonly used for modeling temporal characteristics, fail to adequately capture the correlation of spatiotemporal features and struggle with processing long-term sequences. To address these issues, this paper proposes a two-stage RUL prediction method for rolling bearings. In the first stage, a feature space metric-based degradation point identification method is introduced. By employing the deep feature metric method, an adaptive decision threshold is established to determine the degradation stage. In the second stage, multi-domain sensitive features are extracted and a spatiotemporal graph is constructed. Subsequently, a features and spatiotemporal attention graph neural network (FSTAGNN) is developed. This network incorporates gated channel transformation (GCT) and graph self-attention aggregation network (GSAAN) modules to focus on feature sensitivity and spatiotemporal dependencies, respectively. Finally, the RUL value is evaluated online based on the input data. The proposed method is validated using two bearing datasets, and experimental results demonstrate its superiority over existing graph neural network methods.