A comparative study of unsupervised machine learning methods for anomaly detection in flight data : case studies from real-world flight operations

Abstract

This paper provides a comparative study of unsupervised machine learning (ML) methods for anomaly detection in flight data monitoring (FDM). The study applies various unsupervised ML techniques to real-world flight data and compares the results to the current state-of-the-art flight data analysis techniques applied in industry. The results are validated by the industrial experts. The study finds that a hybrid Local Outlier Factor (LOF) approach provides significant advantages compared to the current state of the art and other ML techniques because it requires less hyperparameter tuning, reduces the number of false positives, provides an ability to establish trends amongst the entire fleet and has an ability to investigate anomalies at each timestep within every flight. Finally, the study provides an in-depth review for some of the cases highlighted by the hybrid LOF and discusses the particular cases providing insights from an academic and flight safety/operational point of view. The analysis conducted by the human expert regarding the outcomes produced by an ML technique is predominantly absent in scholarly research, thereby offering extra value. The study presents a compelling argument for transitioning from the current approach, based on analyzing occurrences through the exceedances of a threshold value, towards an ML-based method which provides a proactive nature of data analysis. The study shows that there is an untapped opportunity to process flight data and achieve valuable information for enhancing air transport safety and improved aviation operations.