Spatio-temporal hypergraph-driven evolutionary Graph-Mamba method for remaining useful life prediction

Abstract

The effective fusion of multi-sensor information is crucial for predicting the remaining useful life of aero-engines. However, due to the complexity of variable operating conditions, sensor signals exhibit time-varying and nonlinear characteristics, making degradation information ambiguous. This poses challenges in constructing predictive models that can accurately extract degradation trends and effectively integrate the spatio-temporal characteristics of signals with prior knowledge. Therefore, this paper proposes a remaining useful life prediction method based on evolutionary Graph-Mamba. First, the mapping relationship between operating conditions and sensor signals in the healthy stage is learned through the Kolmogorov–Arnold Networks, and the residual between the output value of the network and the original signal is characterized as degradation information. Meanwhile, the energy transfer paths within the aircraft engine are embedded as knowledge to construct a hypergraph, thereby creating a spatio-temporal hypergraph to achieve information fusion. Second, we design a gating mechanism to simulate the crossover operation, fusing information from the previous generation to enhance the diversity of embeddings generated by Graph-Mamba, thereby leading to superior graph representations. Simultaneously, we add Gaussian white noise to simulate mutation operations, improving the robustness of the prediction model. Finally, the prediction model was validated on NASA’s N-CMAPSS dataset and further verified for its effectiveness using the C-MAPSS dataset. Experimental results demonstrate that this method has excellent predictive performance.