Various reports from entities such as the Federal Aviation Administration (FAA) and the National Transportation Safety Board (NTSB) have shown a recent increase in the number of incidents involving helicopters. The versatility of rotorcraft operations makes the establishment of safety metrics challenging. Yet, flight data monitoring (FDM) programs enable the implementation of data-based models and analyses that can contribute to improving the safety of helicopter operations. Traditionally FDM programs have featured exceedance-based data analyses by defining safety thresholds. However, recent advances in data science, and more particularly in deep learning techniques, have paved the way for a more reliable definition of safety thresholds via the use of outlier detection algorithms. This paper focuses on the implementation of an anomaly detection model for the takeoff phase which represents a large portion of incidents in rotorcraft operations. After generating training data and augmenting the dataset, the takeoff segment is extracted from each flight data record. Then, the type of takeoff performed is identified through a classification algorithm, and finally, a recurrent neural network composed of long short term memory cells is implemented to detect anomalies or outliers in the input takeoff data.
Gabriel Achour, Ricardo F. Silva, Alexia P. Payan, Charles C. Johnson, and Dimitri N. Mavris, “Rotorcraft takeoff analysis and classification to detect outlier operations that could present a safety risk,” AIAA SciTech 2023 Forum, AIAA 2023-3578, 12-16 June 2023, San Diego, CA https://doi.org/10.2514/6.2023-3578