Modeling of aircraft fuel consumption using machine learning algorithms

With the aid of recording systems such as the flight data recorder, information from aircraft sensors can be transmitted in-situ to airlines or maintenance providers during the flight in form of reports, or stored for subsequent analysis in the postprocessing. This allows the determination of current aircraft performance with regard to fuel consumption or emissions. At present, data in a highly aggregated form (predominantly averages) are used for statistical estimations or physical models. The metrics are calculated on a rolling basis as performance indicators of the aircraft and are then compared with book values from the manuals or with data from the performance monitoring system. However, this procedure represents only a situational, aggregated point evaluation of stable flight conditions. The data aggregation is based on strict validity limits for parameter variations. Furthermore, trigger conditions of the recording logic determine the number and quality of the transmitted reports, such, that only a few data points are available for performance analyses. To improve realistic performance analyses, the time series of all aircraft sensors over the entire flight mission (full-flight data) can be used. In contrast to physical models, this study presents data-based approaches using machine-learning tools from the field of artificial intelligence, to develop fuel flow models based on full-flight data. The proposed methods result in detailed statements for the diagnosis of aircraft fuel consumption. The paper deals with the model development and results of different analyses based on a variety of an airline’s operational flight data records. This study describes the learning methods and shows results for two different data-based models, including neural networks and decision trees. Finally, future applications for the models and an outlook on the authors’ activities will be provided