Remaining Useful Life Prediction Based on Cross-Temporal Dynamic Graph Convolutional Network

Taking advantage of deep learning (DL) to extract hidden degradation signals from machinery monitoring data has led to significant advancements in predicting equipment's remaining useful life (RUL). However, existing methods that use similarity and adaptive adjacency matrices to construct graphs fail to reflect sensor relationships accurately. This article presents a cross-temporal dynamic graph convolutional network (CTDGCN) for RUL prediction to address this issue. The CTDGCN combines cross-temporal modeling with dynamic spatio-temporal graph construction, collecting multi-sensor time series signals to create dynamic graph embeddings. By constructing a cross-temporal sensor network, temporal and spatial features are extracted to design a decay graph based on temporal distance. This model utilizes decay and cross-temporal pooling layers to aggregate information and capture complicated spatio-temporal dependencies. Studies conducted on two cases indicate that the CTDGCN model significantly outperforms existing models in RUL prediction tasks.