Comparison of Algorithms for Anomaly Detection in Flight Recorder Data of Airline Operations

Abstract

In order to improve the current high level of safety in air carrier operations, there is increasing emphasis on developing proactive safety management systems to identify and mitigate risk areas before they manifest in aircraft accidents or incidents. One way to conduct proactive safety management for airline operations is to utilize the operational data archived in modern Flight Data Recorders (FDRs) equipped on aircraft. Recently, efforts have been made to develop algorithms to detect anomalies in sensor data from a complex engineered system in a dynamic operating environment. 1 These algorithms take a data-driven approach to build a model for detecting anomalies directly from data collected during system operation, rather than building it based on domain knowledge, standard operating procedure or human expertise. The knowledge discovery processes, in general, face the challenge of validating new discoveries from real-world data as in many cases there exists no “ground truth” that can confirm the presence of these anomalies. In this study, we compared two data-driven anomaly detection algorithms and contrasted the two datadriven methods with the traditional flight data analysis method Exceedance Detection. The two datadriven anomaly detection algorithms are Cluster-based Anomaly Detection (ClusterAD) and Multiple Kernel Anomaly Detection (MKAD). Both algorithms were developed to detect anomalous flights in recorded flight data. The Exceedance Detection method is a Flight Operational Quality Assurance (FOQA) analysis tool widely used in the airline industry. It detects exceedance events when certain flight parameters exceed pre-specified thresholds. Only known safety concerns are examined by this method. All three methods were independently tested on the same set of flight data from an airline’s normal operations. The entire dataset contains recorded flight data of a narrow-body aircraft. The data are from short to medium range flights of a commercial passenger jet airline. The comparison results of the three