Applying artificial neural networks for multidimensional anomaly detection based on flight data monitoring during final approaches

Abstract

Flight Data Monitoring (FDM) programmes have become a key part of every major airline’s safety management system. They are primarily based on learning from unwanted deviations in flight parameters encountered during normal flight operations. Owing to its unique nature, anomaly detection of FDM presents distinct problem complexities from the majority of analytical and learning tasks. This methodology, while useful, concentrates only on a small part of the operation, leaving most of the data unprocessed, and does not allow for analysing events that had the potential to go wrong but were recovered in time by the crews. This research focused on analysing an FDM dataset of 1332 approaches between January 2018 and July 2022 at Tenerife South Airport (Spain), where there is a known phenomenon of increasing headwinds during the final approach. The flights were clustered using self-organising maps (SOM) by patterns of increasing headwinds, and the clusters were assessed in terms of clustering performance. The clusters were well differentiated. A further comparison between the results from the airline showed that 88 flights were affected by wind shifts, while 27 flights were picked up by the airline. The results demonstrate that SOMs are a meaningful tool for clustering flight data and can complement the current FDM analysis methodology. Combining both methodologies could shift FDM data analysis to look beyond exceedances into what went well, thus shifting the FDM paradigm towards a more safety-II-based method.