关键词: 航空运输, 扇区复杂度评估, 条件生成对抗网络, 复杂度指标, 增广数据集, 样本多样性

Abstract: Sector complexity is an important reference for controller workload and dynamic airspace configuration, which needs to be accurately evaluated in advance. To handle the difficulty of the small sample size in supervised complexity data sets, a sector complexity evaluation framework based on Conditional Generative Adversarial Networks (CGAN) was proposed. Firstly, three types of complexity factors, i.e., traffic flow, aircraft performance, and potential conflicts, were constructed, and subjective complexity levels were combined to obtain calibration samples. Then, the CGAN was used to design a labeled sample generation algorithm to get an augmented data set. Finally, Logistic Regression (LR), Support Vector Machines (SVM), and Random Forest (RF) algorithms were used to build complexity evaluation models. Taking the sector in the central-south region as an example, the effectiveness of the generated samples was verified from qualitative and quantitative perspectives. The evaluation results of each model were compared under multiple training set configurations. The results show that CGAN gradually converges to stability after 200 iterations. The relative error between the generated sample and the real sample for most factors in the mean is less than 5%, and the relative error in the standard deviation is more than 5%. Under the multi-class evaluation indicators, the augmented data set improves the overall evaluation accuracy of the three models by 11.77%, 11.04%, and 8.34%, respectively. The proposed evaluation framework can improve sample diversity under limited data conditions, and it is an effective method for the sector complexity evaluation.

Key words: air transportation, sector complexity evaluation, Conditional Generative Adversarial Networks, complexity index, augmented data set, samples diversity