The continued increase in the number of satellites in low Earth orbit has led to a growing threat of collisions between space objects. On-orbit servicing and active debris removal missions can alleviate this threat by extending the lifetime of active satellites and deorbiting inactive ones, but this requires advanced guidance and control algorithms for the rendezvous phase. Recently, various control policies based on machine learning have been proposed to leverage the advantages of neural networks. One notable technique that has shown much potential in asteroid and planetary landing scenarios is reinforcement meta-learning. This technique consists of training recurrent neural networks in uncertain scenarios in order to develop highly robust control policies that can adapt to unknown conditions in real-time. The goal of this thesis was to apply the meta-learning technique to a rendezvous scenario. Thus, throughout this project a recurrent neural network was trained via reinforcement meta-learning to generate a control policy that can perform the final approach maneuver of a chaser spacecraft towards a rotating target. A feedforward network was also trained for comparison. The learning algorithm used to train the policy is the Proximal Policy Optimization algorithm, which is a modern actor-critic method that has shown good performance in several continuous control settings. A virtual environment was developed in Python to simulate the rendezvous scenario and collect data to train the policy. Before beginning with the training process, the hyperparameters of the model were tuned to ensure a smooth and efficient learning process. The three components that required tuning were the learning algorithm, the architecture of the neural networks, and the reward function. Each of these components was tuned in turn, primarily through trial and error. This process required the learning algorithm to be executed multiple times, using a different combination of hyperparameters on each iteration. By repeating this process over a large search space, suitable hyperparameters were found for the learning algorithm and the neural networks. The hyperparameters were chosen to maximize the amount of reward achieved by the policy while maintaining a reasonable training runtime. The reward function was split into several components to guide the policy towards its objective, thereby improving the speed at which the policy learns. Each of the components of the reward function represented some partial goal that the controller had to accomplish. Tuning the relative weight of these components was a challenging process since it often leads to trade-offs between different policy behaviors. Once the tuning process was completed, a sensitivity study was performed to ensure that the model can be used for different kinds of rendezvous trajectories. The sensitivity analysis was performed by training the model on different scenarios, including different orbit altitudes, different distances from the target, different target sizes, and different target rotation speeds. The results of this study showed that the model could be applied to most of these scenarios without the need for any major changes. After completing the tuning and the sensitivity analysis, the recurrent and the feedforward policies were each trained on a partially observable environment, and their performance was evaluated using a Monte Carlo simulation for a total of one thousand trajectories. The results showed that the recurrent policy was able to learn how to infer hidden information from the environment, which led it to have a much better performance than the feedforward policy. However, the recurrent policy was not without its limitations, since it could not always generate collision-free trajectories, especially when the target rotated at a faster rate. Overall, this thesis showed that reinforcement meta-learning can be a valuable tool for executing complex rendezvous maneuvers, which may be useful now that active debris removal missions are becoming a reality. Furthermore, this thesis also presented a description of how the model was designed and tuned, so that other machine learning practitioners may apply the technique to different scenarios.

Satellite tracking is used to predict a satellite’s orbit to communicate and perform other key functions. It is commonly performed using large ground-based radars with an accuracy of 1 to 10 km or specialized onboard satellite systems. Ground optical systems are a proliferating technology for satellite tracking, identification, and communications. Such optical systems have a narrow field of view, and to ensure smooth and quick satellite acquisition, the satellite must be tracked with higher accuracy (in the order of 100 m). Hence, to allow for optical ground systems to function smoothly they require satellite tracking systems providing sufficient accuracy which is not common for small satellites. This work analyses two radio-based methods to provide sufficiently accurate satellite tracking for the next TU Delft satellite to be acquired optically. These methods utilize the innate satellite radio communications subsystem of the satellite without modifications. To optimize the analysis work, the model complexity was gradually built up. The first method analysed is phase interferometry, which based on a first-order model, turned out to be non-viable. The second method investigated is time difference of arrival. Increasingly more sophisticated models were developed to understand its performance and limitations. For example, for Delfi-PQ, by utilizing 25 ground stations in a square configuration with 1000 km baseline and a noise level of 236 ns (71 m), the target satellite in a 500 km circular orbit can be acquired, with at least the required accuracy, for 18% of the area where the satellite can be received by a minimum of 4 ground stations. It is expected that this performance can be improved through future work. Specifically, by utilizing multiple measurements taken at different times, the satellite orbital parameters can be estimated directly. We conclude time difference of arrival to be a promising method for further research, development, and eventual implementation by TU Delft. The approach and structure of an even more sophisticated model is detailed for future work.

During the last decades, large deployable structures are starting to be seen as a plausible configuration to multiple space missions, such as solar sailing, LEO (Low Earth Orbit) deorbiting missions or solar power generation. Technological advances in key areas, such as thin film solar cells or new deployment methods, as well as the miniaturization of satellites and their components, have considerably increased the usefulness of this design. The research presented in this document aims to contribute to these technological developments, helping to unfold the whole potential of this structural solution. This research comprises the MSc thesis of the author, in partial fulfillment of the MSc in Aerospace Engineering by the Technical University of Delft. The research was conducted during an internship at the department of Guidance Navigation and Control (GNC) Systems, Institute of Space Systems, German Aerospace Center (DLR). The research focuses in the study of the behavior of large thin flexible structures (LTFSs) in space from the perspective of the attitude determination and control system (ADCS). This work was performed in close relation to the DLR’s mission GoSolAR (Gossamer Solar Array). The contributions of the research can be classified in two areas, related to the modeling of the structure and to the design of the ADCS. In relation to the first one, a methodology to integrate the flexibility of a structure into its equations of motion is explained from a theoretical perspective and implemented over a flexible spacecraft. This methodology is based in Lagrange’s equations and can be applied to diverse structures, allowing modeling accurately the dynamics of the system while maintaining the model comprehensive and of relatively low order. The implementation of this approach leads to equations of motion containing both rigid and flexible motion, enabling to conduct an evaluation of the impact of the flexibility of a structure on its dynamics in a space environment. The second major area in which this thesis aims to contribute to the body of knowledge is related to the design and evaluation of attitude controllers targeting flexible satellites. Different controller’s designs are proposed, based in: 1) a linear-quadratic regulator (LQR) approach, 2) robust control theory, particularly in the minimization of HInf and H2 norms, and 3) a control approach based in analytical dynamics, known as the Udwadia–Kalaba approach. The performance of these designs was evaluated not only in relation to control instructions related to the rigid motion of the satellite, e.g. angular rate, but also to the capability of each controller of damping the system, reducing the vibrations that appear due to its flexibility. The particularities that need to be considered when implementing each controller are also studied, in order to give a clear idea of when would each controller be an adequate control solution. The result is a set of different control designs, including for each of these designs: 1) the methodology to derive the controller, 2) the performance of the controller and 3) the particularities and limitations of its implementation.

This thesis work proposes an approach to integrate design and reliability assessment, based on a single modeling environment by deployment of Systems Modeling Language (SysML). The main purpose of the developed approach is to help enhance reliability assessment in student-based CubeSat projects. Considering the limited project resources, that CubeSat university design teams typically have to cope with, an attempt is made to create a methodology, that allows adoption of single modeling environment for both system design and reliability analysis, thereby focusing on the main principles and advantages offered by the Model-Based Systems Engineering (MBSE) practice. To develop the considered methodology and to demonstrate its practical application the following steps were taken: A SysML design model is built, consisting of both functional and physical architectures, based on a hypothetical simplified spacecraft subsystem. Thereby an assessment is made to what detail a system needs to be modeled and which characteristics it should at least comprise to provide a sufficient basis for a risk model, as a choice of a certain risk assessment methodology may influence the extent to which a system must be modeled, while certain system design model characteristics influence the selection of the appropriate risk assessment methods to be implemented. To determine how risk analysis can be performed in SysML for a given hypothetical design problem, various risk assessment methods were considered for the sake of risk assessment methodology development, based on specific criteria. After it became clear how the implementation of both qualitative and quantitative parts of risk assessment had to be realized in SysML, the actual risk analysis was initiated; the design model was extended by a reliability model. The evaluation on the obtained results was finally performed, and general observations from integrating risk analysis in MBSE were presented, based on the comparison with traditional methods, considering the major risk assessment aspects.

The Mars Ascent Vehicle will bring samples back from the surface of Mars to its low orbit in 2031. Its preliminary design has been performed by NASA, defining it as a two-stage solid propellant rocket. However, a research gap exists in finding what the optimum solid rocket motors geometries are to bring the Martian samples to a target orbit between 300km and 375km above Mars. To this extent, the ascent and propellant burn have been modelled, allowing for variation in the geometry of the motors, but also in the launch and separation angles, and in thrust vectoring control. This model has been incorporated into an optimisation, from which a minimum lift-off mass of 343kg was found, 57kg less than required. In the process, 18 thousand different solutions have been found that satisfy all requirements, would the priority shift from the mass of the vehicle to for instance the burn time, or the final altitude.

Spacecraft orbit and attitude dynamics have been classically seen as two separated subjects, since the effect of attitude in orbit dynamics was deemed too small to be considered, or simply, modeled as noise in the estimation process for most satellites. In the eighties, a study by [1] showed that for a very large spacecraft, approximately the size of the International Space Station, the effect of the attitude in the orbit dynamics should be considered.@en

As satellite systems become more and more complex and interact with each other in space, close proximity operation becomes an important aspect of many satellite missions. Simultaneously, systems are becoming increasingly autonomous, for example in rendezvous and docking operations. This poses harsh requirements for the guidance, navigation and control systems on-board of satellites, and especially on satellite navigation filterswhich estimate the state of the satellite and of other systems and objects that it is interactingwith, often based on information input from visual sensors. The commonly used Extended Kalman Filter (EKF) performs well but is not necessarily ideally suited for these emerging challenges, which is why the viability of potential filter alternatives in close-proximity satellite operations was studied. The project was conducted in cooperation with DLR Oberpfaffenhofen in Germany, where currently an EKF is used in close-proximity satellite operations. This filter serves as a baseline for performance testing conducted throughout the project. Potential filter alternatives were identified based on an extensive background study and the mission needs for close-proximity satellite operation. Furthermore, the purpose of the project is to identify and document concrete mismatches between the different testing methods used to serve as a reference in future projects. From the background study two filters, the Extended Kalman Filter with intermediate smoothing step (EKFS) and the Unscented Kalman Filter (UKF) were selected based on a qualitative performance trade-off that focussed on the expected performance under the demands of closeproximity operation in space. The filters were judged based on the available documentation. For the selected filters, as well as for the EKF currently used by DLR, performance criteria were formulated, with a focus on the accuracy of satellite state parameter estimation and filter convergence speed. To collect test data, two approaches were taken: a newly developed simulation test assessing the theoretical performance of the filters in different test scenarios; and a hardware-in-the-loop test in the EPOS 2.0 facility in Oberpfaffenhofen where approaches using two physical satellite models can be performed. The latter test is used to identify problems in the filter performance that have not been found using the simulation test and to validate the filters for more representative real-world performance. An analysis of the test results from the simulation performance test have shown that the EKFS and the UKF can outperform the EKF in the convergence speed and the estimation of some, but not all satellite state parameters. However, it was also identified that the UKF using its current implementation struggles to assess the attitude of the satellite state accurately. Apart from the attitude estimation from the UKF the filters were considered verified and were implemented in the hardware test facility. The hardware test could not confirm the previously seen performance consistently and both filters showed state estimation divergence at closeproximity of the satellites. Thus, their performance could not be validated and they cannot yet under the current implementation be called viable filter alternatives to the EKF. This is due to the fact that sudden state estimation divergence is potentially catastrophic, especially at close distances of the satellites. In addition, differences in the quality of the measurements between the tests of the different filters highlighted potential problems in the comparability of test results. It was concluded that the UKF is the more promising alternative filter for the future since it showed better performance in the hardware test prior to divergence than the EKF and outperformed the EKFS in all hardware tests. In addition it converged the fastest of the three filters. Further studies need to be performed to correct implementation problems, however. Possible approaches for validating the suggested approaches are presented. Different test approaches should also be taken to make the hardware tests more representative and the simulation test should be updated to be more reflective of real world conditions. Several observations on the challenges in moving from simulation to hardware testingwere identified and are presented. Primarily, challenges were found to arise from the faulty selection of test cases for comparability, the neglect of certain inputs observed in hardware testing with previously unpredicted effects in the simulation test and the continuous change of inputs in the real world which were modelled constant in the simulation test.

The goal was to provide a solution to the thermal problem of high-performing CubeSats. These CubeSat generate much heat due to a combination of increased power generation capabilities and miniaturisation. A deployable radiator is designed to solve this problem using HiPeR technology developed by Airbus Defence and Space Netherlands. This design is analysed, a prototype is produced and tested. The design is proposed after iteratively sizing the shape, dimensions, location of the interface and number of layers. The design is a T-shaped laminate of 12 sheets of PG, and a top and bottom protective sheet of Kapton adhered with 3M966. This shape allows for easy folding over the outside of a CubeSat to minimise volume while being compliant with the P-POD requirements. Some parts of the PG sheets are not adhered, allowing for local bending with a small bending radius. A tape spring hinge provides support and the actuation for deployment. The radiator allows for an efficient heat flow path between the payloads and space. This reduces the temperatures of the payloads and allows for CubeSats to continuously generate up to 30W of heat.

The launch vehicle industry is undergoing a major shift as new commercial organizations are staking claim to large market shares with bold technologies. To stay competitive in this new era, launch vehicle companies must develop new technologies to reduce launch costs while serving the growing demand for space accessibility. A possible solution is to use additive manufacturing in combination with topology optimization to relieve manufacturing and assembly burdens in addition to reducing launch mass. Topology optimization is a mathematical method that optimizes the distribution of material in a design domain for a given combination of loads, boundary conditions and constraints with the objective of maximizing mechanical performance. Now an increasingly popular design tool in the aerospace and automotive industries, topology optimization is used to redesign structures to reduce mass while increasing stiffness. To effectively apply this design tool to launcher structures, the optimization process must take into account multiple load cases to represent the large number of load cases such structures are subjected to, throughout the various flight phases. In this thesis, a design methodology was established for topology optimization of launcher structures under multiple load cases by redesigning a launcher structure produced by Airbus Defence and Space Netherlands. The thesis used two analytical models; a simple cantilever beam model to set baseline expectations of design methodologies which were then verified on the launcher structure demonstrator model. To achieve the objective, a suitable multiple load case objective function was identified for the design problem at hand by testing different objective functions on both analytical models. In the process, efficient ways to accommodate the large number of load cases in the objective were also studied through which promising methods to reduce the number of load cases were identified. The final optimized design was compared to the original launcher structure which showed that the optimized design exhibited organic design features, higher stiffness with the same mass. This thesis provided insight into multiple load case topology optimization which is expected to accelerate its adoption in the launch vehicle industry and pave a new path for more efficient space travel.

Laser communication provides numerous benefits over typical Radio Frequency communication, such as lower power, not occupying regulated frequency bands, possibilities for much higher data rates and resistance to jamming. Combined with satellite constellations, Laser Inter-satellite Links (LISL) can enable global connectivity. However, satellites move at fast relative velocities, while optical beam divergence angles are in micro-radian levels. Thus, low-latency and precise position data between linking satellites is crucial. This thesis investigates the LISL conditions in a combined LEO/MEO constellation and the applicability of on-board GNSS-based Orbit Determination (OD) and Orbit Prediction (OP). Novel methods, such as Preprocessing Extended and Single-propagation Unscented Kalman Filters are tested and compared to typical GNSS-OD methods. Analyzing Pointing Uncertainty contributions in various link cases, results indicated that fully-kinematic methods could support LISL for 100-s periods, whereas reduced-dynamic OD-OP methods performed more consistently.

Mitigation strategies which eliminate existing space debris, such as with Active Space Debris Removal (ASDR) missions, are now more important than ever regarding the ever-growing space debris population problem. One of the considered ASDR approaches uses a net as a capturing strategy. The benefits of such strategy are to allow for a large capturing distance, high compatibility for different space debris sizes and reduced accuracy requirements. A key requirement of any ASDR missions is that during capture, no new space debris is to be generated during the process. However, when simulating net capturing in the literature, the potential to break of vulnerable structures, like antennas or solar panels is often neglected. Such elements may show an enhanced risk of failure, especially if these are already damaged, potentially contributing to even more space debris. A discrete Multi-Spring-Damper net model was used to simulate the 20 m/s-frontal impact of a 30 m x 30 m net onto an ESA Envisat mock-up. The Envisat was modelled as a two rigid-body system with a Single-Degree-of-Freedom hinge connection. A sequential modelling strategy was implemented, which de-coupled all the necessary dynamic and structural models. More than two large sub-structures (the Ka-band antenna dish and solar array) were found to have a high likelihood of breaking, leading to the recommendation of several design mitigation strategies using two types of sensitivity analysis. With secondary space debris being generated, net capturing is found to be riskier than originally assumed throughout the literature.

Programmers usually write test cases to test onboard software. However, this procedure is time-consuming and needs sufficient prior knowledge. As a result, small satellite developers may not be able to test the software thoroughly. A promising direction to solve this problem is reinforcement learning (RL) based testing. It searches testing commands to maximise the return, which represents the testing goal. Testers need not specify prior knowledge besides the reward function and hyperparameters. Reinforcement learning has matured in software testing scenarios, such as GUI testing. However, migration from such scenarios to onboard software testing is still challenging because of different environments. This work is the first research to apply reinforcement learning in real onboard software testing and one of few studies that perform RL-based testing on embedded software without a GUI. In this work, the RL agent observes current code coverage and the interaction history, selects a pre-defined command, or organises a command from pre-defined parameters to maximise cumulative reward. The reward function can be code coverage (coverage testing) or estimated CPU load (stress testing). Three RL algorithms, including the tabular Q-Learning, Double Duelling Deep Q Network (D3QN), and Proximal Policy Optimization (PPO), are compared with a random testing baseline and a genetic algorithm baseline in the experiments. This study also performs regression testing with a trained RL agent, i.e., to test a version of onboard software that it has never seen before. To do that, the agent processes graph input with code coverage information. The graph is extracted from the onboard software source code via static code analysis. The work tries two graph neural network architectures (GGNN and GAT) with several graph pooling mechanisms to process the graph input. Apart from the test command generation algorithms, some middleware is also implemented, including a command/response parser, a state identification module, a branch coverage collection tool, and a tool to extract the graph representation and node features. During onboard software testing, the onboard computer (OBC) or the electrical group support equipment (EGSE) can be the master of the bus. The command generation algorithms can run on a lab PC or a cloud server. The research reveals the advantages and drawbacks of using reinforcement learning to test onboard software. On the one hand, RL-based testing performs well in non-deterministic environments (e.g., stress testing) and regression testing. On the other hand, more straightforward methods like random testing and the genetic algorithm are more useful in deterministic environments. This document also introduces relative background knowledge. It leaves many recommendations for future work, such as improving sampling efficiency, generalization, and learning a model for fault detection in satellite operation.

Certain Formation Flying missions rely on their orbital control thrusters to maintain the formation, making a Fault Detection and Isolation (FDI) system that is capable of detecting thruster faults very valuable. In addition, communication between satellites in a formation can be expensive. In this thesis, a distributed FDI approach was developed making use of Recurrent Neural Networks (RNNs), utilizing the already available relative position and velocity data as input. To test the approach, a numerical simulation of a Virtual-Rigid-Body formation in low Earth orbit was developed in MATLAB. The RNNs were trained on noisy data from the simulation, utilizing the tensorflow library. The resulting system was analyzed and compared against a centralized Kalman-filter based approach, in detection capability, isolation capability and robustness. The developed approach showed overall worse performance, but offers reduced communication needs compared to the comparison method.