A dynamic spatial-temporal graph transformer with multi-frequency attention for remaining useful life prediction

作者：Yu Xia, Hui Liu · 发表于：Measurement Science and Technology · 年份：2025 · DOI：10.1088/1361-6501/ae228a · 被引用次数：2 · 研究领域：Machine Fault Diagnosis Techniques、Reliability and Maintenance Optimization、Advanced Data and IoT Technologies

Abstract The remaining useful life (RUL) prediction is essential for cost-effective production and reliable predictive maintenance in intelligent manufacturing. Existing deep learning-based approaches often struggle to capture complex degradation patterns across temporal, spatial, and frequency domains. To address this limitation, a dynamic spatial-temporal graph transformer with multi-frequency attention (DSTGT-MFA) is proposed in this paper for RUL prediction. The proposed DSTGT-MFA model consists of three key components: a multi-scale gated convolutional neural network for extracting hierarchical local features, a graph convolution transformer for modeling long-term spatial-temporal dependencies with dynamic and static adjacency matrices, and a multi-frequency spatial-temporal attention mechanism to enhance temporal and spatial attention in the frequency domain. This integrated architecture enables the model to comprehensively capture degradation trends and fuse multi-domain features. Extensive experiments conducted on the commercial modular aero-propulsion system simulation (CMAPSS) and new CMAPSS datasets demonstrate that the DSTGT-MFA model achieves superior prediction accuracy compared to twelve baseline methods.