An Adaptive Online Learning Model for Flight Data Cluster Analysis

Safety is a top priority for civil aviation. To help airlines further improve safety, various clustering-based methods were developed to better understand their current flight operations and detect unknown risks from onboard flight data. However, existing methods can only be carried on historical data in batches, resulting in its inability to update and adjust as new data come in. New onboard flight data related to anomaly detection are generated at airlines every day. The addition of new data will inevitably cause changes in the clustering results. Yet it would be computational costly to run clustering on all data as they accumulate. Therefore, anomaly detection methods that allow real-time update of cluster models as new data come in are more practical for airlines. This paper presents a reinforcement learning method to identify common patterns in flight data via cluster analysis and update its clusters as new data come in. This method is based on Gaussian Mixture Model (GMM) and uses online (recursive) expectation-maximization (EM) algorithm to update clustering results over time. An initial result of clusters can be obtained by performing GMM-based clustering on historical flight data. Then, as new data come in, the parameters of GMM are updated via an online EM algorithm. By recording the GMM parameters, the method can also track changes in clusters over time. We demonstrated the proposed method using Flight Data Recorder (FDR) data from real operations of an airline. The evolution of clusters was observed as new batches of flight data are fed into the proposed method.