Continuous Alert Level Detection In Flight Data Time Series Using Machine Learning

Abstract

Situational Awareness and Alerting Systems in rotorcraft and other aircraft are typically designed with discrete triggering thresholds. While these rule-based systems fulfill strict certification requirements, they lack the ability to gauge proximity to critical thresholds in a continuous manner. As a result, they often pose a trade-off between excessive false alarms and delayed warnings. In this paper, a data-driven methodology is proposed to create a continuous surrogate model of such alerting systems using machine learning, specifically time-series regression models. The approach includes data preprocessing, a Multilayer Perceptron model architecture trained to predict a postprocessed continuous metric that reflects the proximity of flight data to an alert threshold. The methodology is illustrated on real flight data, showing how it can build a new kind of distance of alarm occurrence. A discussion concludes with challenges posed by low sampling rates, lessons learned, and directions for future research, including the possibility of leveraging previous predictions to enhance model performance and identifying patterns of approaches to alert triggering.