A data-driven approach for flutter boundary prediction is proposed, which adopts a deep learning model to extract flutter features from measured signals, thereby enabling the analysis of aeroelastic system stability. Following a modeling strategy that leverages structural acceleration response signals, a dataset is constructed from experimental data acquired from wind tunnel tests, and a multidimensional feature system and a set of comparative models are subsequently established for performance evaluation. Comparative analytical results reveal that the integration of power spectral density (PSD) with the Bayesian optimization Transformer–long short-term memory hybrid model can markedly enhance the efficacy of feature extraction and the accuracy of flutter boundary prediction. Experimental validation demonstrates that the proposed method achieves a mean prediction error of 3.73% for unseen working conditions of the wind tunnel model and 4.14% for data obtained from a single flight test sortie. Furthermore, a prediction error below 10% is achievable at approximately 70% of the critical flutter speed, which could contribute to enhanced early warning performance in flutter tests.
Wang et al. (Sat,) studied this question.