Machine Learning Based Constellation Optimisation with S²VSE

Kurzfassung

This presentation shows how machine learning techniques like genetic algorithms can be used to adapt an augmentation constellation for GNSS or a LEO PNT system to a specific scenario or use case by finding and optimal or near optimal constellation design. The optimisation is based on existing capabilities of S²VSE including the possibility to rapidly create and analyse single- and multi-layer constellations and create realistic user environments with digital surface models from multiple data sources for the performance evaluation. Besides single-objective optimisation the work also explores the possible impact of multi-objective optimisation on the proposed solutions. Topics such as multilayer systems and mega constellations have gained prominence in recent years, along with envisioned systems for positioning, navigation, and timing (PNT) in low Earth orbit (LEO). As a result, it is increasingly important to acquire the capability to analyse and assess both existing and future constellations in terms of their designs and capabilities. To satisfy this demand the System and Service Volume Simulation Environment (S²VSE) for global navigation satellite systems (GNSS) has been extended over the last years from the analysis of existing constellations, to include the capability to rapidly create new single- and multi-layer constellations either as standalone or as argumentation to existing constellations. The generation is based on predefined constellation blueprints to specify the architecture of each layer or constellation with just a few parameters. Currently this includes commonly used constellation schemes such as Walker Delta (e.g. used for GNSS), Walker Star (commonly used in newly proposed LEO constellations) and Streets of Coverage (used in communication constellations e.g. Iridium), but also less frequently used frameworks like Flower (more suited for regional or local services) and Draim (allowing global coverage with a minimal number of satellites). These two established capabilities of creation and analysis of constellations are the foundation for the work presented here, the automatic optimisation of a constellation design to suit either a global, regional or local scenario or use case. As the well-established GNSS constellations positioned in medium earth orbit (MEO) have stabilised in their design over time, and improvements are more focused on onboard technologies, augmentation constellations or emerging standalone LEO PNT systems more suitable for optimisation. These constellations are more diverse and adaptive in their design and therefore constantly evolving. For regional augmentation systems the goal is to enhance GNSS performance in specific regions, for this traditionally few satellites in geostationary earth orbit (GEO) and geo synchronous orbit (GSO) are used (e.g. the Japanese Quasi Zenith Satellite System - QZSS). Newer proposed systems are more frequently situated in LEO aiming for a global augmentation or even operation as an independent new global PNT system, these systems are planned by governments as well as from commercial companies and can either be dedicated to PNT or host a secondary payload to support PNT. There are certain advantages of constellations in LEO, for example more satellites can be used due to less expensive components and reduced launch costs. However, these additional satellites are also necessary, the number of satellites needed for a comparable global performance in regards to classic GNSS is around ten times higher. Based on the increased number of satellites and the higher degrees of freedom in the architecture augmentative constellations and new standalone LEO PNT systems are well suited for optimisation. For the optimisation genetic algorithms (GAs) are used, they are a machine learning technique inspired by the process of natural selection. They simulate the process of natural evolution, using a population of candidate solutions that evolve over generations. Each solution, represented as a "chromosome," is evaluated for its fitness based on how well it solves the presented problem. The best solutions are selected, recombined (crossover), and mutated to create new solutions. This process continues iteratively until an optimal or near-optimal solution is found. GAs are commonly used in complex optimisation problems, where traditional methods may be inefficient. In the beginning three algorithms where chosen, the non-dominated sorting genetic algorithm 2 (NSGA-II), its improved version NSGA-III, and the strength pareto evolutionary algorithm 2 (SPEA-II). These algorithms differ, for example, in the usage of an archive and the method of determining pareto-dominance, but are frequently used and provide many configuration and extension possibilities. There are some parameters that provide the possibility for the optimisation of a constellation, but while focusing on single objective optimisation the first parameter used for navigation constellations is the dilution of precision (DOP). The quality of a positioning solutions is in direct correlation with the DOP value as it takes the user aspect with the receiver position as well as the system design based on the satellite geometry into considerations. The DOP is based on the relations of the locations of the satellites in line of sight (LOS) to the receiver on ground. Within these lines of sight, a reversed squared pyramid can be constructed between four satellites and the receiver. Depending on the volume of the resulting pyramid the value changes, a larger volume leading to a smaller DOP and thereby to a better positioning solution. The scenarios covered in the analysis combine both sides of S²VSE, the constellations on the system side is evaluated with the support of the digital surface models created to represent the user environment on the service volume side. Most satellite-based services including GNSS have an excellent performance in open field conditions but experience reduced user performance in obstructed areas. The premise for most augmentation systems is the performance of GNSS in deep urban environment and therefore in urban canyons. The Digital Surface Module integrated in S²VSE can tap into multiple databases to create a realistic representation of the user environment, this includes natural elevations, forest heights and densities as well as street layouts and building heights. Based on these created environments the new constellations are evaluated in different user scenarios ranging from open field conditions and rural areas over suburbs into deep urban canyons. The next step is to move on from single-objective to multi-objective optimisation, the chosen algorithms allow to compare and choose solutions that balance the objectives differently, and thereby enable the selection of the most appropriate for a specific optimisation scenario. Other parameters that will be used besides the DOP are the depth of coverage (DOC), satellite visibility and re visit time. These parameters are important for constellations which establish connections either with a user on ground or a station for the operation of the system. Naturally another parameter for optimisation is the overall reduction of the number of satellites in the constellation without having a significant impact on the overall capabilities. These later parameters are applicable not only for PNT constellations but also for other applications relying on satellites such as communication and earth observation, here the availability and the continuity of the provided services can be improved.