Research on Forecasting Aeroengine Vibration Signals Based on the MAE Model

For understanding the vibration states of aero-engines deeply and acquiring warning signs in real-time, a T-step forecast method of aero-engine vibration is proposed based on a powerful and advanced deep learning model, the MAE (Masked Autoencoders) model. Unlike previous common algorithms, the MAE model performs pre-training tasks by reconstructing information to achieve the purpose of analyzing and learning the latent information among parameters. Through the use of this technique, downstream tasks will be able to fine-tune and better suit the objective function, increasing the forecast’s accuracy. Compared with simulation data and open-source datasets, this paper uses the real flight data recorded by aircraft data acquisition systems. This implies that the real vibration states of aero-engines can be learned and built. And from this, fairly practical conclusions can be reached. The result shows that it is feasible to forecast the vibration signal of aero-engines. This means that not only is that the first time series forecasting application of Transformer models that has a pre-training mechanism with masked code, but it also takes the lead in exploring the feasibility of vibration signal forecast in the aero-engine area. In addition, the possibility of forecasting in different types of aero-engines is also tested. Finally, to make our theories more reasonable and convincing, experiments on different aero-engine states involving the transition state and the steady state are carried out.