Adaptive remaining useful life prediction method for aircraft engines under variable operational conditions based on multi-channel transformer

Abstract Aiming at the problem of degradation trend being covered under time-varying operating conditions and insufficient generalization ability of prediction models, which leads to reduced predictive performance, a domain adaptive method (MC-TraMMDNet) combining improved multi-channel transformer and Maximum Mean Discrepancy (MMD) is proposed for end-to-end prediction of remaining useful life (RUL) of aircraft engines. Firstly, the monitoring data of variable operating conditions are clustered and normalized at the same lifespan scale based on the fault state data. Secondly, corresponding transformer encoders are established for each operating condition, and degradation information is comprehensively extracted for each operating condition through a multi-head self-attention mechanism and multi-channel convolutional network. In addition, to enhance the generalization of the model, the maximum mean difference heterogeneous domain adaptive layer is used to extract domain invariant features. Next, a multi-layer perceptron model is utilized for forecasting the remaining useful lifespan of aircraft engines in diverse operational environments. Finally, validation is carried out on the publicly available dataset C-MAPSS. In comparison to other methods, the results demonstrate that this method excels in both RMSE and Score. The evaluation indicators have decreased by about 9% and 35% respectively.