Congestion Recognition of the Air Traffic Control Sector Based on Deep Active Learning

Abstract

The air traffic control sector (ATCS) is the basic unit of the airspace system. If we can identify the congestion of an ATCS, it will help provide decision support for planning and daily operations. However, current methods mainly characterize congestion from the static structure and the dynamic operational features, resulting in poor generalization and operability. To this end, we propose a deep learning method from the perspective of complex networks. It takes aircraft as nodes to construct an aircraft network and utilizes the complexity indices to characterize it. So, the problem of identifying congestion becomes the complexity of the aircraft network. Inspired by active learning methods, we construct a deep active learning (DAL) model for congestion recognition. It adopts an iterative semi-supervised approach to reduce the number of labeled samples while ensuring recognition performance. To make full use of a large number of unlabeled samples, the sparse autoencoder is employed to characterize all labeled samples and unlabeled samples. The hidden layer of the deep neural network is constructed by stacking. In the process of active learning iteration, minimum confidence, marginal sampling, and information entropy are introduced as measures to select samples from the unlabeled sample set with significantly different features from the labeled sample set. The model is applied to three representative sectors in China’s airspace as cases. Results suggest that DAL can reduce the labeled sample set’s redundancy and achieve the desired performance with the smallest number of samples. Additionally, DAL is superior to the existing mainstream methods in the four objective evaluation indices.