Domain knowledge-enhanced spatio-temporal adaptive fusion network for aeroengine remaining useful life prediction

The integration of graph convolutional networks (GCNs) with time-series models has gained popularity in predicting aeroengine remaining useful life (RUL), as it allows for the capture of both spatial and temporal dependencies from multidimensional time-series (MST) monitoring data. However, existing methods encounter several limitations: (1) the physical essential correlations among sensor parameters are often overlooked, which hinders the precise extraction of actual spatial dependencies from MST data; (2) the mainstream sequential network architectures often lead to mutual interference between spatial and temporal dependencies, consequently causing degradation information loss. To address these challenges, a novel Domain Knowledge-Enhanced Spatio-Temporal Adaptive Fusion Network (DKESTAFN) is developed, which integrates aeroengine domain knowledge with parallel spatiotemporal feature extraction to significantly improve RUL prediction accuracy. First, MST monitoring data are transformed into both grid-structured and graph-structured formats. Meanwhile, the physical essential correlations among sensor parameters are revealed through an analysis of aeroengine working principles. These correlations are then leveraged to construct the graph-structured data, thereby more accurately representing spatial dependencies and enhancing spatial feature expression. Second, a novel dual-stream spatio-temporal feature extraction module is designed to simultaneously extract representative degradation features from both temporal and spatial streams via parallelly connected temporal and spatial branches, which effectively mitigates the mutual influence between spatial and temporal dependencies. Furthermore, a gated mechanism is designed to adaptively fuses the extracted spatial and temporal features, yielding more comprehensive and representative degradation features, ultimately boosting RUL prediction accuracy. Extensive comparative experiments demonstrate that DKESTAFN substantially outperforms ten state-of-the-art prognostic methods in prediction accuracy, thereby validating its effectiveness, robustness, and practical potential for aeroengine RUL prediction.