Safety is key to civil aviation. To further improve its already respectable safety records, the airline industry is transitioning towards a proactive approach which anticipates and mitigates risks before incidents occur. This approach requires continuous monitoring and analysis of flight operations; however, modern aircraft systems have become increasingly complex to a degree that traditional analytical methods have reached their limits - the current methods in use can only detect 'hazardous' behaviors on a pre-defined list; they will miss important risks that are unlisted or unknown. This paper presents a novel approach to apply data mining in flight data analysis allowing airline safety experts to identify latent risks from daily operations without specifying what to look for in advance. In this approach, we apply a Gaussian Mixture Model (GMM) based clustering to digital flight data in order to detect flights with unusual data patterns. These flights may indicate an increased level of risks under the assumption that normal flights share common patterns, while anomalies do not. Safety experts can then review these flights in detail to identify risks, if any. Compared with other data-driven methods to monitor flight operations, this approach, referred to as ClusterAD-DataSample, can (1) better establish the norm by automatically recognizing multiple typical patterns of flight operations, and (2) pinpoint which part of a detected flight is abnormal. Evaluation of ClusterAD-DataSample was performed on two sets of A320 flight data of real-world airline operations; results showed that ClusterAD-DataSample was able to detect abnormal flights with elevated risks, which make it a promising tool for airline operators to identify early signs of safety degradation even if the criteria are unknown a priori.@en

A method has been developed to support Flight Operations Quality Assurance (FOQA) by identifying anomalous flights based on onboard-recorded flight data using cluster analysis techniques. Unlike current techniques, the method does not require pre-defined thresholds of particular parameters, but detects data patterns which differ from the majority of flights by considering all the available flight parameters. The method converts time series data from multiple flight parameters into a high dimensional data vector. Each vector captures all the available information for a single flight. Cluster analysis of the vectors is performed to identify nominal flights which are associated with large clusters and anomalous flights that do not belong to a specific cluster. The method was applied to a representative Digital Flight Fata Recorder (DFDR) dataset from an international airline. Detailed analysis was performed on takeoff and approach for 365 B777 flights. Abnormal flights were detected using the cluster technique which was able to identify anomalous behaviors including: high and low energy states, unusual pitch excursions, abnormal flap settings, high wind conditions. In addition, data clusters representing nominal conditions were also detected. Three distinct takeoff clusters were identified in the B777 data: one represented a majority of the takeoff cases, one correlated with a specific high altitude airport, one correlated with reduced power takeoffs. This initial evaluation indicates that cluster analysis is a promising approach for the identification of anomalous flights from onboard-recorded flight data.@en

The airline industry is moving toward proactive risk management, which aims to identify and mitigate risks before accidents occur. However, existing methods for such efforts are limited. They rely on predefined criteria to identify risks, leaving emergent issues undetected. This paper presents a new method, cluster-based anomaly detection to detect abnormal flights, which can support domain experts in detecting anomalies and associated risks from routine airline operations. The new method, enabled by data from the flight data recorder, applies clustering techniques to detect abnormal flights of unique data patterns. Compared with existing methods, the new method no longer requires predefined criteria or domain knowledge. Tests were conducted using two sets of operational data consisting of 365 B777 flights and 25,519 A320 flights. The performance of cluster-based anomaly detection to detect abnormal flights was compared with those of multiple kernel anomaly detection, which is another data-driven anomaly detection algorithm in recent years, as well as with exceedance detection, which is the current method employed by the airline industry. Results showed that both cluster-based anomaly detection to detect abnormal flights and multiple kernel anomaly detection were able to identify operationally significant anomalies, surpassing the capability of exceedance detection. Cluster-based anomaly detection to detect abnormal flights performed better with continuous parameters, whereas multiple kernel anomaly detection was more sensitive toward discrete parameters.@en