Abstract: To address the scarcity of labeled samples and complex coupling between wind fields and flight attitudes in flight attitude prediction under complex wind fields, this paper proposes an unsupervised regression prediction method based on the Auxiliary Regression Generative Adversarial Network (ARWGAN). In this method, the generator fuses simulated pitch angles and Gaussian noise to generate wind field samples with consistent physical meaning; the discriminator adopts a dual-branch structure of authenticity classification and pitch angle regression, and combines adversarial loss with Huber regression loss for joint optimization, enabling flight attitude prediction without training labels. Experiments show that the proposed method achieves prediction performance (MSE=34.28, MAE=4.87) under unsupervised conditions, which is close to that of supervised models and significantly better than random benchmark methods. Additionally, the generated samples have high visual similarity to simulated samples, indicating that the method can effectively capture the nonlinear correlation between wind fields and flight attitudes, providing a new technical approach for unsupervised regression tasks with scarce labels and imbalanced distributions.

Key words: Complex wind field, Flight attitude prediction, Generative adversarial networks, Auxiliary regression, Unsupervised regression