Feasibility of Machine Learning Methods for Predictive Alerting of the Energy State for Aircraft

Abstract

This paper discusses the feasibility of using machine learning methods, including deep model architectures, for the prediction of near future hazardous energy states (i.e., stall, overspeed, high and fast, low and slow, unstable approaches). Aircraft state prediction and specifically energy state prediction is an important step in providing the flight crew with visual and aural cues to improve their Aircraft State Awareness (ASA). Lack of ASA has been identified as one of the leading contributing factors in commercial aviation accidents, thus improving ASA has the potential to enhance aviation safety. In previous research, various Predictive Alerting of Energy (PAE) methods for flight crew information management and decision support (IMDS) were developed and tested with data from a NASA flight simulator study (AIME-l) in which eleven commercial airline crews (22 pilots) completed more than 230 flights. The previously tested aircraft state prediction methods included predictor stages of: (i) sequential stochastic filters, (ii) batch estimators and (iii) fast-time 3DOF model simulations. Successful predictions of stall, overspeed and high-fast/low-slow conditions were generated with these methods for time horizons ranging up to 300s. The paper discusses the use of machine learning techniques for energy state prediction, and considers its fundamental safety implications and algorithmic limitations, such as a lack of off-nominal training data, while also examining performance characteristics and providing insight into the underlying structure of the algorithms used.