A semi-supervised learning method for air traffic complexity evaluation

Air traffic complexity is an essential indicator to evaluate the air traffic situation within one given airspace, with respect to both the workload of traffic control and the degree of operational safety. In practice, it is helpful for airspace reconfiguration and air traffic flow management. Hence, identifying an accurate method for air traffic complexity evaluation is important. Considering there are many factors that may influence air traffic complexity in complicated ways, machine learning-based complexity evaluation models based on labeled samples have been proposed recently. However, the high cost of sample labeling work usually results in insufficient labeled samples for training. Compared with labeled samples, unlabeled samples can be easily obtained by automatically processing flight track data. Hence, in this paper, we propose a semi-supervised learning model for air traffic complexity evaluation. In the training process of this model, the unlabeled samples are iteratively labeled by both active labeling and automatic labeling strategies. The active labeling strategy means the model will actively query the labels of a few samples that have high learning values from experts; and the automatic labeling strategy refers to automatically labeling the samples which can be credibly labeled, while having certain informative degree. The experimental results using the real operational data of Chengdu Flight Information Region in China show that our model can effectively utilize the complexity evaluation information contained in unlabeled samples, and achieve better performance than the traditional supervised evaluator when sacrificing the same amount of labeled samples.