Temporal Precursor Discovery Using Long Short-Term Memory with Feature Attention

Abstract

The continuous growth of demand on commercial airlines has made it crucial to guarantee the safety of airspace operations. Although adverse events are rare, once they happen, they can cause unpredictable risky factors and degrade airspace efficiency. Thus, studying historical air traffic data to discover precursors, features, or events that contribute to the occurrence of the adverse event in the future is important and has gained interest in recent years. In this paper, a novel and real-time applicable temporal precursor discovery (TPD) framework based on the long short-term memory neural network and the feature attention mechanism is proposed. The feature attention mechanism enables the framework to pay attention to certain features at a certain time, and the attention score is defined as the temporal precursor. The temporal precursor reflects the rationale behind the neural network’s prediction at each time step, providing a data-driven explanation of how the adverse event occurs. The proposed TPD framework was tested with real air traffic data and weather data recorded at Incheon International Airport in South Korea in 2019.