A Highly Imbalanced Assembly Vibration Prediction of Aero-engine Using Feature Selection and Data Augmentation

Purpose: In the process of aero-engine assembly, the vibration value of the whole aero-engine has an important impact on flight safety. To limit the assembly vibration to a reasonable range, the data-driven method is used to establish a direct mapping model between the assembly parameters and the vibration. The data collected from the assembly process are small samples with high dimensions and high imbalance, which is difficult for the data-driven model to predict. Method: This paper proposes a prediction method via MCA + 1DWGANGP, which can significantly improve the sample quality through feature selection and data augmentation. First, a novel MCA (Multiple Correlation Analysis) is developed, integrating different coefficients for redundancy analysis and sensitivity analysis, which is used to select key features. Second, a novel 1DWGANGP (1-Dimensional Wasserstein GAN with Gradient Penalty) is developed, which is used for data augmentation in small samples. Then, various classifiers are used to predict the interval range of vibration, which can represent the assembly quality. Results and Conclusions: The public dataset and assembly dataset are used to verify the reasonableness and superiority of the proposed model. The experimental results show that the MCA + 1DWGANGP can achieve accurate vibration prediction via feature selection and data augmentation. The accuracy of vibration prediction improves by [41–44%], and there is a significant enhancement in precision, recall, and F1 score.