Anomaly Detection in Initial Climb Segments for Helicopter Operations

Helicopters are used in a variety of operations and recent studies show that the number of accidents associated with helicopters is stagnating, if not increasing. Flight data monitoring (FDM) is a useful tool to review the data retrospectively for risk mitigation. Exceedance analyses are typically used in FDM for anomaly detection. However, they typically rely on pre-defined thresholds which might vary depending on the type of operations or vehicles considered. With recent advancements in data mining techniques, many efforts have been put into anomaly detection in the commercial fixed-wing aviation and this provides a new perspective beyond traditional methods. In this research, a sequential approach is proposed to detect anomalies in initial climb segments for helicopter operations. The stepwise methodology contains three elements: trajectory pattern mining, time series length analysis, and shape analysis for identifying different levels of anomalies. To ensure the effectiveness of the methods selected, synthetic and simulated data are used for testing before applying candidate methods to the actual initial climb segments. A specific group of initial climb segments is used to demonstrate the validity of the methods chosen in this study. Our tests show that functional principal component analysis and a convolutional variational autoencoder along with DBSCAN are capable of identifying shape anomalies in flight parameters. Although the detected anomalies might not directly be associated with hazardous events, it is useful to assist helicopter operators in discovering patterns not conforming to the norms.