Modeling Tail-Specific Performance Using Historical Flight Data and Machine-Learning Techniques

Abstract

Aircraft performance has always been a focus of attention in aviation. The work of aircraft designers, certifying agencies, aircraft operators, and air traffic controllers relies on aircraft performance models. Current aircraft performance models are based on performance data of brand-new aircraft, independent of airline configuration and customizations. Nonetheless, over time aircraft suffer structure, engine and aerodynamic deterioration, as well as maintenance actions. These factors, which vary with tail number, make aircraft performance deviate from the theoretical and create the need for aircraft performance monitoring, and ultimately for aircraft performance tailoring. This thesis proposes two novel approaches to develop up-to-date, tail-specific performance models. These approaches are based on the use of historical flight data, namely Quick Access Recorder (QAR) data, and machine-learning techniques. First, a methodology is designed to calibrate Base of Aircraft DAta (BADA), a widely consolidated physics-based performance model. As a result, more accurate performance models are generated, maintaining the same applicability over the entire flight envelope and during all phases of flight as BADA nominal models. The second approach is purely data-driven, and in contrast to the first approach, describes aircraft performance without modeling the underlying physics. The resulting models proved to be more accurate tan BADA nominal models. Additionally, they provided insights on the parameters that have a significant impact on aircraft performance. This research project is the first to provide a comparison between physics-based and non-physics-based performance tailoring approaches. Despite the differences in accuracy achieved with both methods, each one has its own advantages and disadvantages. What approach to follow is determined first by the application, and secondly by the demanded level of accuracy.