Applying Optimization and Artificial Intelligence to NASA’s Communications Networks: Cognitive Real-Time Link Layer Adaptations to Rapid Constellation Planning

Since NASA began operations over 60 years ago, it has been a world leader in all of its space operations: human spaceflight efforts; its robotic exploration of the solar system and beyond; and its Earth science and monitoring. In order to send back to Earth the awe-inspiring videos, images, and data collected by NASA from its portfolio of missions, the space agency relies on its global networks of ground stations and spacecraft. Spacecraft, whether they are as close as a few hundred kilometers above the Earth's surface or as distant as interstellar space, downlink their science data through one of three networks: the Near Earth Network, the Space Network, or the Deep Space Network. The most significant bottleneck to collecting more science data is downlinking the data back to Earth. There are many opportunities for improvement that could ultimately lead to more science return and lower user burden. These improvements can occur on a number of layers in a network and/or at different periods in a project life cycle. Leveraging the tools of software-defined radio and artificial intelligence, the work conducted under this fellowship aims to make improvements on multiple network layers and implementation time scales with four main goals. This provides a diversity of short- and long-term solutions to NASA's needs. The first goal is to improve real-time communications at the link layer while operating within existing standards. This type of improvement is most realizable in the short term as it is designed to fit within legacy software and hardware. The second goal is to develop new scheduling methods that can accommodate a new, large influx of users while working with the current scheduling methods to minimize disruption to legacy users. The third goal is to explore a new framework for spacecraft scheduling designed to allow users to request network resources on-demand. This paradigm shift would dictate new requirements for future spacecraft to take advantage of this new framework. The fourth goal is to develop new methods for orbit planning and optimization to provide designers more insight and the ability to optimally design new networks quicker than conventional orbit propagation methods. To improve point-to-point communications, a proof-of-concept multi-objective reinforcement learning cognitive engine using deep neural networks for the use as a radio-resource-allocation controller was implemented and tested on orbit. In our experiments, the cognitive engine optimized for DC power consumption, transmit power efficiency, target bandwidth, spectral efficiency, throughput, and frame error rate, given the relative importance of these metrics for the mission with the ability to change modulation, coding, transmit power, and excess filter bandwidth. The 20 flight experiments using software-defined radios onboard the International Space Station mark one of the first published experiments of a space communications cognitive engine. Additionally, a generalized and parameterized interacting multiple model (IMM) filter based on a Kalman filter bank was implemented on software-defined radio technology applied to satellite communications applications. This filter can be used to detect the channel state (e.g., clear sky, rain fade, ionospheric scintillation, etc.) in real-time, as well as provide an improved received power estimation. These received power predictions and channel state classifications can be relayed to an adaptive or cognitive system to make more intelligent decisions. For a short-term solution that allows the Deep Space Network scheduling process to manage large influxes of users, block scheduling algorithms were developed and packaged into a software product. Blocking involves finding mutual view periods where groups of spacecraft are within a small angle of each other with respect to a ground station antenna, aggregating their individual requirements, and attempting to schedule these requirements into the mutual view periods. These blocked times are then scheduled into tracks and negotiated through the regular Deep Space Network process. The final tracks are then split apart into individual users based on maximizing the minimum user satisfaction. Blocking has the potential for reducing schedule overhead and provides a more cost-effective method for low-cost missions. It is planned to be used for the SmallSats on NASA’s Artemis 1 mission and other future missions. For a long-term solution to scheduling spacecraft with event-driven science, a beacon-tone--based demand access scheduling system was investigated through quantitative simulations. To illustrate a proof-of-concept demand access system, a simulation campaign (using a custom-built network simulator) for a near-Earth asteroid autonomous SmallSat explorer fleet mission concept was run using the network simulator. When compared with the current static scheduling system, the results showed that the demand-access approach is able to decrease the mean data latency by about 50% for the same number of schedule reservations. It can also decrease the total number of tracks reserved by 25% while still providing a lower mean data latency and equal data collection drop rates compared to the static scheduling approach. The demand access system also provides more frequent (yet more basic) health status reports than the current system. Finally, targeted for the early mission and network architecture planning stages, analytical approximations for common coverage statistics have been derived to be used in speeding up the orbit selection optimization process. Specifically, approximations for maximum gap time, mean gap time, mean pass time, and mean passes per day for circular orbits were derived using a unifying geometric framework. To the best of the author's knowledge, this dissertation presents the first closed-form solution for maximum gap time for non-repeating orbits. Error analysis showed that these approximations will work well for medium- to high-fidelity optimization analysis. All four of these solutions help to improve either the planning, development, or operations of NASA's networks.