Subjective Air Traffic Complexity Analysis Based on Weak Supervised Learning

Controller subjective evaluation is one of the most important ways to assess air traffic complexity. However, the inconsistency of human experts has a negative impact on the inference of complexity analysis models. To solve this problem, this paper proposes to construct a weakly supervised air traffic complexity dataset using highly reliable traffic situation similarity as labelling information. On this basis, a distance metric learning model is trained to generate a distance metric matrix that satisfies the similarity relationship. Finally, the K-means algorithm is combined to realise preferred complexity situation level classification and evolution analysis. Taking the actual operating data of a mid-southern area sector of China as an example, the effectiveness of the proposed method is verified. Experimental results show that the aircraft density, aircraft ground speed variance, heading disorder, convergence speed and horizontal conflict have a greater impact on the complexity situation. Compared with the K-means algorithm based on Euclidean distance, metric learning improves the optimal silhouette coefficient and Davidson-Boldin index by 31.80% and 12.97%, respectively. In addition, it is confirmed that the situation evolution is driven by one or two key influencing factors.