A Machine Learning-Based Model for Flight Turbulence Identification Using LiDAR Data

Abstract

By addressing the imbalanced proportions of the data category samples in the velocity structure function of the LiDAR turbulence identification model, we propose a flight turbulence identification model utilizing both a conditional generative adversarial network (CGAN) and extreme gradient boosting (XGBoost). This model can fully learn small- and medium-sized turbulence samples, reduce the false alarm rate, improve robustness, and maintain model stability. Model training involves constructing a balanced dataset by generating samples that conform to the original data distribution via the CGAN. Subsequently, the XGBoost model is iteratively trained on the sample set to obtain the flight turbulence classification level. Experiments show that the turbulence recognition accuracy achieved on the CGAN-generated augmented sample set improves by 15%. Additionally, when incorporating LiDAR-obtained wind field data, the performance of the XGBoost model surpasses that of traditional classification algorithms such as K-nearest neighbours, support vector machines, and random forests by 14%, 8%, and 5%, respectively, affirming the excellence of the model for turbulence classification. Moreover, a comparative analysis conducted on a Zhongchuan Airport flight crew report showed that the model achieved a 78% turbulence identification accuracy, indicating enhanced recognition ability under data-imbalanced conditions. In conclusion, our CGAN/XGBoost model effectively addresses the proportion imbalance issue.