Flight Phase Identification from Trajectory with LSTM

Flight phase estimator for trajectory data based on K-means clustering and LSTM. The following flight phases are identified based on the ICAO ADREP standard.

phase used for classification |ICAO primary phase| ICAO sub-phase | ---|---|---| taxi | taxi | all take-off | take-off | take-off run initial climb | take-off | initial climb climb | en route | climb to cruise cruise | en route | cruise & change of cruise level descent | en route | descent approach | approach | all landing | landing | level off-touchdown & landing roll

Training data inlcludes trajectory (Altitude, Speed, Rate of Climb) and flight phases found with the flight phase finder tool on X-plane simulator data. After training only trajectory data is required to estimate flight phase in order to transfer the model to ADS-B data.

Requirements

The list of requirements can be found in the requirements.txt file.

Main requirements: - Pytorch - Numpy - Matplotlib

Using pretrained model on custom ADS-B data

1) Preprocess ADS-B data

bash python src/ADSB_preprocessing.py --folder custom_files_folder

If no storage folder is given a new folder is created with the same name as the original folder and '_preprocessed' appended The results of the preprocessing are the files themselves, the images that compare before and after and the reports on the quality.

In order to obtain reports on the quality of the analysed flights an overview file has to be provided.

For more options and personalisation see bash python src/ADSB_preprocessing.py --help

2) Run evaluation on preprocessed flights

bash python src/evaluation.py --custom_data_path custom_files_folder_preprocessed/csvs

This stores the images of the labeled flights together with the CSV files that include the labels in the results folder.

Training a model

The data used to train this model can be found under the following link.

If one wishes to train a new model either on their own FDR data or on the provided data a dataset can be created using the following steps: - Use the find_flight_phases(pandas_df) function in src/poffunctions/flightphasefindercore.py to label each flight.\ (For X-plane data python src/pof_functions/flight_phase_finder_xplane.py takes raw x-Plane txt log files from the data/xplaneraw folder and separates, labels and stores them.) - Store trajectory data seperate from its labels respectively in data/preprocessed/trajectoriestrain and data/preprocessed/labelstrain - ```python src/poffunctions/create_dataset.py``` takes the preprocessed files and generates a training and test dataset for training.

The src/parameter_girdsearch.py module allows to train different models with different hyperparameters in parallel each in its own terminal. Please consult:

bash python src/parameter_gridsearch.py --help

for details.

Contributors

Emy Arts

Alexander Kamtsiuris

Reference

This code is part of the publication "Trajectory based Flight Phase Identification with Machine Learning for Digital Twins".

If you use this repository for your research please reference:

Arts, E.; Kamtsiuris, A.; et al. (2022): Trajectory based Flight Phase Identification with Machine Learning for Digital Twins. Deutsche Gesellschaft für Luft- und Raumfahrt - Lilienthal-Oberth e.V.. https://doi.org/10.25967/550191

@article{Arts.2021, author = {Arts, E. and Kamtsiuris, A. and Meyer, H. and Raddatz, F. and Peters, A. and Wermter, S.}, date = {2021}, title = {Trajectory based Flight Phase Identification with Machine Learning for Digital Twins}, publisher = {{Deutsche Gesellschaft f{\"u}r Luft- und Raumfahrt - Lilienthal-Oberth e.V}}, doi = {10.25967/550191} }