The ML-airport-arrival-runway software is developed to provide a reference implementation to serve as a research example how to train and register Machine Learning (ML) models intended for predicting arrival runway assignments. The software is designed to point to databases which are not provided as part of the software release and thus this software is only intended to serve as an example of best practices. The software is built in python and leverages open-source libraries kedro, scikitlearn, MLFlow, and others. The software provides examples how to build three distinct pipelines for data query and save, data engineering, and data science. These pipelines enable scalable, repeatable, and maintainable development of ML models.
The ML-airport-arrival-runway model is one of several ML models tied together by the Airport Surface Model Orchestrator shown below.
The Airport Surface Model Orchestrator runs at fixed intervals and is driven from real-time Fused FAA System Wide Information Management (SWIM) data feeds and also pulls in additional D-ATIS airport configuration data and weather data. The data input manager within the Orchestrator prepares data for each ML service and data is exchanged between the Orchestrator and the ML services via API interfaces.
The input data sources for the individual ML prediction services are shown in the below diagram which also illustrates dependencies between the individual ML services.
The ML Airport Surface Model forms the building blocks of a cloud based predictive engine that alerts flight operators to pre-departure Trajectory Option Set (TOS) reroute opportunities within the terminal airspace. The ML Airport Surface Model was designed to be a scalable replacement for the capabilities provided by NASA's Surface Trajectory Based Operations (STBO) subsystem, which is a component of the fielded ATD2 Phase 3 System in the North Texas Metroplex. The STBO subsystem relies heavily upon detailed adaptation, which defines the physical constraints and encodes Subject Matter Expert knowledge within decision trees, and creates a costly bottleneck to scaling the pre-departure TOS digital reroute capability across the National Airspace System.
Arrival Runway Prediction is part of a suite of softwares designed to model the airport surface:
- ML-airport Airport Configuration Model
- ML-airport Arrival Runway Model
- ML-airport Departure Runway Model
- ML-airport Taxi-In Model
- ML-airport Taxi-Out Model
- ML-airport Estimated-On-Time Model
- ML-airport Data Services
In addition to the steps below, take a look at the Kedro documentation and the MLflow documentation for more help getting started.
Create new environment from conda.yml file
conda env create -f conda.yml
Then activate the environment.
conda activate arr-rwy-env
Credentials
credentials.yml defines username, host, port to access a database which we assume has FUSER data.
A Kedro credentials.yml file needs to be created in conf/local/.
An example of such a file is in credentials-template.yml.
Per-airport Globals
In this project, each model is trained for a particular airport.
The airport and potentially airport-specific items are specified in conf/base/globals.yml or conf/base/<airport_icao>.globals.yml, where <airport_icao> is the ICAO identifier for an airport (e.g., KCLT).
Which particular globals.yml is used is specified at run time as a parameter specified in the kedro run command line interfaces (described below) or the airport specific file can be copied into globals.yml.
This approach is based on Kedro's templated configurations capability.
Implementing it required us to overwrite the _create_config_loader method in src/arr_rwy/run.py to use a TemplatedConfigLoader in a way similar to what is described in the Kedro documentation, but extended to allow for the global_config_yml command line interface parameter.
If using Kedro's Jupyter notebook or IPython integrations, the overall globals.yml is always used (i.e., there is not currently any way to specify which airport-specific <airport_icao>.globals.yml to use when initiating the Jupyter notebook session).
The overall globals.yml can be updated during a Jupyter or IPython session (e.g., to specify a different airport) if, after changing the file, the %reload_kedro command is executed in the Jupyter notebook or IPython console session.
This will update the Kedro context variable to take into account the updated globals.yml.
See the "Working with Kedro from notebooks or IPython" section below and the relevant Kedro documentation for additional information about Kedro's Jupyter and IPython capabilities.
Parameters
The conf/base/parameters.yml file contains any parameters that control various aspects of the pipelines that engineer data and train and test models.
The parameters file specifies, for instance, the type of machine learning model that should be used and what hyperparameters will control its training.
You may wish to update some of these parameters.
Items in parameters.yml that are surrounded by ${ and } will be imported from the globals.yml.
Kedro's configuration templating documentation provides additional information about templated configurations via globals.yml.
Data Catalog
Any data sets used in this Kedro project must be declared in the conf/base/catalog.yml "data catalog."
There should not be any reason for you to update this data catalog, but it contains items surrounded by ${ and } that will be imported from the globals.yml specified for a particular run.
Kedro's configuration templating documentation provides additional information.
Kedro Nodes & Pipelines
Kedro pipelines specify a directed acyclic graph of Kedro nodes to be run, with various data sets declared in the catalog.yml or parameters specified in parameters.yml serving as inputs.
Other than parameters or data sets declared in the data catalog, model inputs can be names of objects that are output from other nodes, such as intermediate data sets or trained models.
The overall project pipelines are defined in src/arr_rwy/pipeline.py.
These are defined by combining partial pipelines defined in other places throughout the project.
For example, some data engineering pipelines are defined in src/arr_rwy/pipelines/data_engineering/pipeline.py.
Nodes are simply wrappers around functions, which in turn can be defined or imported from anywhere, but are often and conventionally defined in various nodes.py files in various directories throughout the project.
MLflow server
MLflow can be configured to track runs locally or remotely on a server.
MLflow API
Throughout the code are some MLflow API calls. These do things like log run parameters and metrics to the MLflow server or log run artifacts (e.g., pickled trained models) to the MLflow artifact store.
Most of the data required to run train models in this project can be acquired by running "data query and save" (DQS) pipelines.
These run some queries per data set declarations in conf/base/catalog.yml (with some aspects of these declarations imported from the specified conf/base/*globals.yml), accessing databases per credentials in conf/local/credentials.yml, and save resulting CSVs locally in the data/ folder (typically in the 01_raw/ subfolder).
The CSV data set naming convention puts the relevant airport ICAO code as a prefix (e.g., if running with a KJFK.globals.yml for KJFK, then the data sets will be named things like data/01_raw/KJFK.MFS_data_set.csv).
Other data engineering or full pipelines start from those CSVs.
These commands do not need to be rerun over and over for the same airport and time period; once they are run once the data will be available in the CSVs in the data/ folder.
Since the data query and save pipelines run queries that retrieve and then save to disk potentially large amounts of data, they can take tens of minutes to hours to run, depending in particular on the duration of time between start_time and end_time in globals.yml.
To track pipeline run progress during these potentially long runs, Kedro informational logging statements will be printed to the screen.
To get the data needed to train, run the command below from the ML-airport-arrival-runway/ directory.
kedro run --pipeline dqs --params global_config_yml:<airport_icao>.globals.yml
The <airport_icao> must be replaced by the ICAO identifier. A few other data query pipelines are available, and are listed in the file src/arr_rwy/pipeline.py.
After running the queries to get the required data sets, the data can be cleaned with the engineering pipeline.
To perform data engineering, run the command below.
kedro run --pipeline de --params global_config_yml:<airport_icao>.globals.yml
After running the queries to get the required data sets, models can be trained and tested.
To train and test an arrival runway model, run the command below.
kedro run --pipeline ds --params global_config_yml:<airport_icao>.globals.yml
In order to use Jupyter notebooks or IPython in your Kedro project, you need to install Jupyter from within the project's conda environment:
conda install jupyter
For using Jupyter Lab, you need to install it:
conda install jupyterlab
After installing Jupyter and/or Jupyter Lab, you can start a local notebook server:
kedro jupyter notebook
You can also start Jupyter Lab:
kedro jupyter lab
And if you want to run an IPython session:
kedro ipython
Staring Jupyter or IPython this way executes a startup script in .ipython/profile_default/startup/00-kedro-init.py.
This creates a Kedro context variable in scope; the context can be used to access the data catalog, parameters, execute pipelines or portions thereof, and more.
See the relevant Kedro documentation for details.
In each of these cases, the session uses the global settings (e.g., airport ICAO) in conf/base/globals.yml to populate various items in parameters.yml and catalog.yml.
If you wish to adjust those global settings after starting up a Jupyter notebook or an IPython session, simply change the contents of globals.yml and run the %reload_kedro line magic command.
The Kedro viz tool can be used to visualize Kedro pipelines. To use it, first install it in the project environment via pip.
pip install kedro-viz
Then, from the top-level ML-airport-arrival-runway/ directory, visualize a pipeline by providing its name after the --pipeline command line interface command.
For example, the command below will launch a browser with an interactive visualization of the data science pipeline.
kedro viz --pipeline ds
The ML-airport-arrival-runway code is released under the NASA Open Source Agreement Version 1.3 license
Copyright © 2021 United States Government as represented by the Administrator of the National Aeronautics and Space Administration. All Rights Reserved.
No Warranty: THE SUBJECT SOFTWARE IS PROVIDED "AS IS" WITHOUT ANY WARRANTY OF ANY KIND, EITHER EXPRESSED, IMPLIED, OR STATUTORY, INCLUDING, BUT NOT LIMITED TO, ANY WARRANTY THAT THE SUBJECT SOFTWARE WILL CONFORM TO SPECIFICATIONS, ANY IMPLIED WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, OR FREEDOM FROM INFRINGEMENT, ANY WARRANTY THAT THE SUBJECT SOFTWARE WILL BE ERROR FREE, OR ANY WARRANTY THAT DOCUMENTATION, IF PROVIDED, WILL CONFORM TO THE SUBJECT SOFTWARE. THIS AGREEMENT DOES NOT, IN ANY MANNER, CONSTITUTE AN ENDORSEMENT BY GOVERNMENT AGENCY OR ANY PRIOR RECIPIENT OF ANY RESULTS, RESULTING DESIGNS, HARDWARE, SOFTWARE PRODUCTS OR ANY OTHER APPLICATIONS RESULTING FROM USE OF THE SUBJECT SOFTWARE. FURTHER, GOVERNMENT AGENCY DISCLAIMS ALL WARRANTIES AND LIABILITIES REGARDING THIRD-PARTY SOFTWARE, IF PRESENT IN THE ORIGINAL SOFTWARE, AND DISTRIBUTES IT "AS IS."
Waiver and Indemnity: RECIPIENT AGREES TO WAIVE ANY AND ALL CLAIMS AGAINST THE UNITED STATES GOVERNMENT, ITS CONTRACTORS AND SUBCONTRACTORS, AS WELL AS ANY PRIOR RECIPIENT. IF RECIPIENT'S USE OF THE SUBJECT SOFTWARE RESULTS IN ANY LIABILITIES, DEMANDS, DAMAGES, EXPENSES OR LOSSES ARISING FROM SUCH USE, INCLUDING ANY DAMAGES FROM PRODUCTS BASED ON, OR RESULTING FROM, RECIPIENT'S USE OF THE SUBJECT SOFTWARE, RECIPIENT SHALL INDEMNIFY AND HOLD HARMLESS THE UNITED STATES GOVERNMENT, ITS CONTRACTORS AND SUBCONTRACTORS, AS WELL AS ANY PRIOR RECIPIENT, TO THE EXTENT PERMITTED BY LAW. RECIPIENT'S SOLE REMEDY FOR ANY SUCH MATTER SHALL BE THE IMMEDIATE, UNILATERAL TERMINATION OF THIS AGREEMENT.