Remaining Useful Life Prediction Based on Multi-Scale Temporal Dynamic Graph Attention Network

Remaining useful life (RUL) prediction is pivotal for ensuring the safe operation and efficient maintenance of aero-engines. To address the challenges of capturing dynamic degradation patterns in high-dimensional, nonlinear, and multi-condition data, the multiscale temporal graph attention network (MST-DGAT) is designed in this paper. The model employs K-means clustering to categorize operational states into early, stable, and degradation phases. The MST-DGAT then uses phase-specific temporal convolutional networks (TCNs) with tailored kernel sizes to capture long-term trends and short-term anomalies. These phase-specific TCN outputs are integrated via cascaded fusion to form multi-scale temporal features. The TCN outputs are used as initial graph node representations. To model the correlations between these nodes, a dynamic adjacency matrix is then constructed based on Pearson correlation coefficients. This graph model integrates a graph attention mechanism with a gated fusion strategy to enable it to adaptively combine temporal, spatial, and statistical-learned dimensions. This process generates high-dimensional representations, which are subsequently mapped to RUL predictions. Experiments on the C-MAPSS dataset demonstrate that MST-DGAT achieves desirable performance.