Abstract The performance degradation of aircraft engines is inevitable and exhibits high correlation and strong coupling. Predicting the remaining useful life (RUL) can effectively ensure flight safety and improve operational economy. Traditional RUL prediction methods typically rely solely on single-source data streams, neglecting the correlations among sensor data, and struggle to integrate temporal evolution characteristics. This paper proposes a physics-data fusion prediction method for aircraft engine RUL based on a static-dynamic hybrid graph neural network, which effectively captures the inherent co-evolution characteristics in multi-sensor data streams. Firstly, a temporal variational autoencoder (TVAE) framework is introduced to process time-series sensor data, extracting key feature representations of the raw data through a probabilistic encoding mechanism. Subsequently, a static graph structure is constructed based on physical prior knowledge, while dynamic temporal features extracted by the TVAE are used to compute feature similarities among sensor nodes, thereby establishing a dynamic graph structure. Furthermore, a hybrid graph neural network is built by fusing the physics-informed static graph structure and the data-driven dynamic graph structure, incorporating a multi-head graph attention mechanism to achieve adaptive learning of complex sensor correlation patterns. The proposed method significantly enhances the model's ability to effectively process temporal evolution characteristics of raw multi-source sensor data streams.Finally, various comparative experiments are designed, showing that the prediction accuracy is improved by 37.85% and 37.75% compared to the baseline methods, respectively. The results validate that the proposed method can significantly enhance the prediction performance of complex engineering systems characterized by interdependencies among multi-source sensors.
Zhang et al. (Fri,) studied this question.