Using Machine Learning in the Development of ARAIM


SST provides engineering and operations research support to the Federal Aviation Administration’s (FAA’s) National Satellite Navigation Division at the William J. Hughes Technical Center (WJHTC). The WJHTC is the nation’s premier federal aviation laboratory for advancing the United States National Airspace System (NAS) and sustaining its continued safe and efficient operations.

Since the mid-1990s, the FAA has used the Receiver Autonomous Integrity Monitoring System (RAIM) to monitor the nation’s Global Positioning System (GPS). In 2004, a joint agreement was signed between the United States and the European Union to incorporate services of the Galileo Satellite system with GPS to offer improved safety-of-life services. SST, in coordination with the FAA, developed the Advanced Receiver Autonomous Integrity Monitoring System (ARAIM). ARAIM is a robust airborne monitoring system that improves the signal-in-space accuracy and availability of an ionosphere-free solution for positioning, navigation, and timing (PNT) using the dual-frequency and multi-constellation satellite navigation systems of GPS and Galileo.

SST supported the FAA by incorporating research from the universities into a robust solution using Global Ground Station Receiver Systems’ observation of satellite messages. Sequential observations are used to compute and monitor the message parameters evolution, compensating for the problem of satellite failure, constellation failure, the nominal bias, and sigma of user range errors. The parameters are monitored and evaluated, then used to generate integrity messages. The integrity messages are then transmitted to warn pilots of satellite faults so these satellites can be excluded when developing user positioning solutions.

This is accomplished by using both supervised and unsupervised machine learning. The first component of the solution is the generation of an integrity message, which is developed by assessing message data originating from multiple ground stations observing signals from the satellites. The ground station processes and packages the observation messages into a time-series set of messages with varying degrees of data quality. Unsupervised learning methods are employed to initially clean the observation message data using binary clustering algorithms. A consensus or majority rule algorithm is then applied to a representative sample of messages to determine the most accurate message. During the computation of the integrity message, a gradient descent algorithm, which is a supervised learning algorithm, is employed. Its objective function uses a variable value to determine when the message data has reached an approximately optimal state. Supervised learning is used again during assessments with known good results from the National Geospatial Agency, where the precise data is used to verify the results of the orbit-clock dynamic model.  

The second component of the solution is the development of an independent orbit-clock dynamic model to eliminate dependency on the external Air Navigation Service Provider. To accomplish this task, we used data from multiple source files to compute the precise position, velocity, orientation, and trajectory of each satellite. Several algorithms are required to filter and optimize the results.  In short, the message data from observing ground stations and ephemeris data from a satellite are used in the solution. Because the data quality varies, the system uses redundant signals to feed an unsupervised learning algorithm. This is done by using a form of the gradient descent algorithm to iteratively filter and solve an objective function that updates the parameters until the solution meets a threshold value or reaches a maximum number of iterations.

ARAIM enhances safety-of-life operations by incorporating multiple satellite constellations and dual frequencies, providing deeper threat analysis, and augmenting integrity monitoring systems update capabilities.  The system is envisioned to enable aircraft navigation that completely relies on Global Navigation Satellite Systems from takeoff through the final approach to landing.

Using Machine Learning in the Development of ARAIM