Anomaly Detection and Cause Analysis During Landing Approach Using Recurrent Neural Network

Abstract

The stabilized approach is important to avoid aircraft accidents during landing. Although there are many possible factors that can lead to an unstable approach, actions can be taken to mitigate the risk if the reason for unstable approaches is identified properly. However, because unstable approaches are rarely observed, it is difficult to identify the relevant reasons by analyzing unstable approaches only. Instead, the author proposes to identify untypical flights, which are not necessarily categorized into unstable approaches, by using neural network. Using the proposed method, more flights are available to identify the reason of unstable approaches. The stability index is estimated by neural network considering the current flight conditions such as wind and initial deviation. Therefore, untypical flights can be detected by comparing the actual stability index and the estimated stability index. The long-term interaction between the stability index and flight conditions is modeled by using gated recurrent unit. As a result of modeling, some of flights are categorized into untypical flights, and the false localizer beam is observed for most of these flights, which could be a potential hazard for unstable approaches. The proposed approach has a big potential to identify such potential hazards based on historical flight data.