To ensure the safety of a spacecraft, operators collect thousands of telemetry signals and monitor them for anomalies, which is both expensive and time-consuming. Space agencies have been researching Machine Learning (ML)-based Time Series Anomaly Detection (TSAD) methods to improve automation, but this is hindered by a lack of high-quality benchmarks. This thesis explores the application of ML-based TSAD on instrument telemetry data for the XMM-Newton space telescope. The data was explored to find several unique challenges, such as recurring eclipses and a high volume of missing data. A methodology was then developed to pre-process the raw data into an ML-compatible format, detect anomalies using a semi-supervised forecasting approach and post-process the detections into a benchmark-suitable format. The method yielded over 40 detections, which were partially validated in discussion with instrument engineers. A refined methodology, incorporating areas for improvement, is presented to proceed with an eventual XMM-Newton anomaly benchmark.

The design of an optimized retroreflector array for satellite laser ranging (SLR) is critical for enhancing satellite tracking accuracy and providing a passive space segment backup for conventional attitude determination systems. This thesis documents the research aimed at developing an effective retroreflector array for the new TU Delft platforms: Delfi-Twin and the Da Vinci satellite. The goal is to enable SLR measurements via both the International Laser Ranging Service (ILRS) network and the laser communication terminal at Delft University. Following a structured approach, the most suitable array configurations were identified through a Probability of Visibility (PoV) simulation. Consequently, a link budget analysis was performed, coupling the selected arrays with the considered ground stations to define the required retroreflector size to enable SLR measurements. Improvements to the Delft laser communication terminal have also been proposed to enhance its SLR capabilities. Future work will focus on qualifying the acquired retroreflectors and finalising the holder structure for satellite integration.

With the projected expansion of the small satellite market and the rise of the lunar economy, the development of cost efficient and reliable navigation in cislunar space—defined as the region of space between Earth and Moon including the region around the surface of the Moon—has become increasingly important. However, the increase in the number of satellites strain existing communication networks in terms of availability, which exerts pressure on orbit determination quality and the financial budgets. One way to release the pressure from the existing ground station communications is through the concept of Autonomous Orbit Determination (AOD) by using inter-satellite radio two-way range between at least two satellites. The three-body problem of cislunar space-the most common region for small satellite deep space missions-makes it an ideal environment to perform AOD due to strong third body perturbations. Previous literature also has not clearly linked AOD with mission and spacecraft design related parameters, and is primarily focused on the orbit estimation aspect rather than including navigation as well.

The analysis of this work is based on a case study in which an L2 Lagrange point orbiter (LPO), called LUMIO, and an elliptical lunar orbiter (ELO), called LPF (based on the SSTL Lunar Pathfinder), perform AOD based on two-way inter-satellite ranging with a Gaussian noise level of 2.98m1σ, observed at a 300 s interval. This thesis aims to research the strategic adjustment of the timing of inter-satellite tracking sessions while optimizing for 1 year of station keeping cost for LUMIO, defined in ∆V.  Since the accuracy of the maneuvers relies on the magnitude of estimation errors from the OD process, which in turn depends on the state observability and thus relative satellite geometry during a tracking arc, solving for lowest ∆V is a complex optimization problem. Three different timing strategy categories were set up in which tracking is performed with varying levels of complexity: tracking based on constant tracking duration and interval, tracking around the perilune or apolune of the LPF satellite, and tracking based on the solution of a heuristic optimization routine that adjusts each tracking arc individually.

The overarching conclusion of this work is that improvements can be made compared to this value in each of the three strategy categories but with varying levels of ∆V. Initial observation windows were defined as a set of tracking arcs of1.0 day with a 3.0-day interval between each arc, where the predicted annual cost equated to 0.613±0.0066 1σ m/s, serving as the baseline value. The best of the constant tracking category showed an an annual ∆V of 0.375±0.0020 m/s, but this comes at the cost of large relative tracking time with a length-interval combination of 0.5-0.5 days. The best of orbit-based solutions reduce cost to an annual mean of 0.5 m/s and have relatively short tracking arcs, which makes it possible to spend more time on collecting scientific data. The Particle Swarm Optimization (PSO) algorithm shows that a reduction to 0.280±0.0134 1σ m/s can be made. All in all, from a mission design perspective, it means that it pays to adapt to a more complex tracking scheme, but employing a constant-type tracking arc timing scheme can also already yield ∆V improvements.  Advances in reducing power and fuel budgets can extend mission duration by lowering fuel consumption. Optimized tracking reduces total tracking time, requiring less power for signal transmission and freeing more power for payloads or other subsystems. This allows more time for scientific observations, increasing the mission's output.

Multiple in-space experiments with deep learning have shown promising results in applying deep neural networks for automation of satellite fault detection tasks. However, the deployment of such neural networks on small satellites with low-power onboard computers is hindered by the highly limited computational resources of these devices. In this thesis, an effort has been made to create a lightweight neural network solution for satellite temperature monitoring, to explore whether such networks can improve the quality of fault detection even when deployed on low-power devices. The designed neural network solution tested on the MSP432P401R microcontroller uses only 29.5 kB of RAM and 65.2 kB of flash memory and can detect small-sized deviations in the satellite temperature sensor readings long before they grow into anomalies that exceed the safe operating range.

The first mission proposals to visit the Alpha Centauri system use photon-sail acceleration as a mode of propulsion to reach this stellar system closest to our own Solar System. To prepare for a future mission, the photon-sail dynamics in the system is investigated. Planar Lyapunov orbits around the colinear classical Lagrange points are designed to explore the Alpha Centauri system. This has been done before in other systems like the Earth-moon and Sun-Earth systems, but not yet in an elliptical binary star system. Starting with an initial guess in the circular restricted three-body problem without photon-sail acceleration, a Multiple Shooting Differential Correction (MSDC) algorithm changes the trajectory to a periodic orbit. A continuation method increases the eccentricity to match e = 0.5208, which is the eccentricity of the inner binary system of Alpha Centauri. The lightness number of the photon sail is increased to add photon-sail acceleration to the model up to a defined maximum of ?ZNe = 2. A set of five constant steering laws is chosen to investigate its effect. Next to that, the moment at which the periodic orbit starts in terms of the true anomaly is varied as well. This results in a set of 40 families of periodic orbits with increasing lightness numbers. Depending on the orientation, the augmented Lyapunov orbit either shrinks into smaller orbits or expands into larger orbits when increasing the lightness number. If the orbit shrinks, it can either converge into an artificial equilibrium point or the photon-radiation pressure on the sail can become minimal. In that case, the Lyapunov orbit becomes (almost) independent on the lightness number and reaches ?ZNe = 2. If the orbit expands, the maximum velocity will eventually go to infinity. At this vertical asymptote, the maximum lightness number is found. The initial true anomaly of Alpha Centauri 0 has a great effect on the Lyapunov orbits around L2 and L3 in the classical ER3BP. For 0 = 0, the orbit either converges to an AEP or the maximum velocity goes to infinity. For 0 = , a few orientations can reach ?ZNe = 2. To further explore Alpha Centauri, an adaptive differential evolution algorithm is used to design trajectories between the Lyapunov orbits. The performance of the algorithm is expressed as the Euclidean difference between the states at the end of the departure leg and the beginning of the arrival leg. Three different lightness number of ? = 0.1, 0.5 and 2 are used for these trajectories. With a lightness number of 0.1, the dimensionless Euclidean error is in the range of 1E-1 to 1E-3, depending on the Lyapunov orbits. With this lightness number, the stars are also used as a gravity assist. For larger lightness numbers, the Euclidean error becomes negligible in the range 1E-7. With a lightness number of 2, the time of flight during the trajectory is significantly lower. In future research, this can be further decreased using an MSDC algorithm.

The use of small satellites, enabled by the standardization of the CubeSat specifications and miniaturization in electronics, has seen a rapid increase in the past decades. The low-cost and short development time of these satellites has made them an attractive option for both commercial and academic applications, making space exploration more accessible. However, these small satellites are prone to failures, leading to lost scientific potential. Mitigation of these failures forms the motivation for this thesis. Recent advances in neural networks have shown promise in the field of anomaly detection. The black-box nature of such models, however, makes it challenging to understand the reasoning behind their predictions.

Constraining the data-driven models with known physics can not only help us understand the reasoning behind their predictions, but also ensuring the model is consistent with the real-world behavior of the system. The work presented in this Master's thesis aims to demonstrate the advantages of such first-principles neural networks over purely data-driven models in thermal behavior modeling of small satellites. Baseline performance of data-driven Long Short-Term Memory (LSTM) networks is established using FUNCube-1 telemetry data, quantifying the temperature prediction accuracy of the models under ideal conditions. The limitations of these models, especially with sparse data, are then investigated, to highlight the need for more robust models.

First-principles models, based on a physics-informed curve-fit and simplified thermal network models, are then developed to constrain the data-driven model predictions. The first-principles models are shown to be more robust to sparse data, with the predictions on data not seen during training being more consistent with the real-world thermal behavior of the satellite. Methods to relate the first-principles model parameters to the physical properties of the satellite are also proposed and explored, to help extract the evolution of the thermal behavior of the satellite over time.

Small satellite systems such as CubeSats and PocketQubes have strict requirements in terms of size, weight, and power available onboard. In light of these constraints, small satellite systems typically omit the inclusion of a ranging system due to its power and specialized hardware requirements. However, new research in satellite ranging invented the telemetry ranging class of techniques which do not place any such requirements on the satellite. The thesis explores the possibility of implementing a telemetry ranging technique on the Delfi-PQ satellite, which is already in orbit using only software modifications. The working and performance of the technique in determining the position of the satellite were explored using its engineering model and have positive implications for small satellite navigation and science.

As space technology continues to advance, the landscape of telecommunication is witnessing a transition from traditional radio frequency to laser communication technology, driven by its enhanced efficiency. Notably, laser communication offers the potential for achieving millimeter-level ranging accuracy, thanks to its significantly higher data rates. This thesis presents the design of a system utilizing a field-programmable gate array (FPGA) and commercial off-the-shelf (COTS) multi-gigabit optical transceivers. The system tests laser communication ranging through time-of-flight measurements, leveraging the inherent structure of optical data. Through extensive testing, the thesis aims to characterize the system's maximum achievable accuracy, identifying constraints in the avenues of improvement. The use of COTS components ensures scalability and facilitates seamless integration into future test setups.

This thesis report has explored the complex field of satellite formation flying, focusing on Guidance, Navigation and Control (GNC) strategies for maintaining formations of multiple satellites. The research has been guided by a series of research questions aimed at improving the understanding and implementation of GNC systems for satellite formations. The report begins by providing an overview of the background and context of satellite formation flying, highlighting its significance in modern space missions. It emphasizes the need for precise control systems to maintain formations and addresses the challenges posed by various disturbances and constraints. To address these challenges, the report presents a detailed methodology for developing an end-to-end GNC system. This methodology involves the use of nonlinear dynamics models, including considerations for Earth’s oblateness and drag effects, to describe the motion of satellites within the formation. Additionally, the report explores the use of both absolute and relative dynamics models to enable the control of large satellite formations. Throughout the report, the performance of the GNC system is analyzed through various simulations and experiments. Different thruster models, including variable and fixed thrusters, are evaluated, shedding light on their effectiveness in maintaining formations. The analysis also considers the impact of energymatching conditions on reducing the frequency of maneuvers. The results of the simulations demonstrate the challenges and limitations of the Sliding Mode Controller (SMC) based control system, particularly in the context of real-world thruster implementations. Tracking errors and drift in formations are observed, necessitating further research into mitigating these issues. In the final chapters, the report explores potential areas for future research and improvements. These include investigating alternative thruster models, optimizing control algorithms, and developing strategies for reducing drift in formations. In conclusion, this thesis report provides valuable insights into the complexities of satellite formation flying, and the challenges faced by GNC systems. It offers a comprehensive methodology for developing
and analyzing these systems and highlights areas for future research. Ultimately, the report contributes to the ongoing advancements in satellite formation flying, paving the way for more precise and efficient space-based operations.

CubeSats have risen in popularity since its first launch in the year 2003. The low mass, lower launch cost, lower development cost, possibility for piggyback with larger satellites and lower development time involved compared to larger satellites have opened them to be commercialised by private companies. In contrast to being used for educational, research and technology demonstration purposes in their early years, they are now used for varied applications such as communications, Earth observation, military surveillance, in-orbit manufacturing, asteroid exploration, Internet-of-things and interplanetary exploration missions. Their lucrative features make them favourable over their larger counterparts. As a result, they are predicted to launch in higher numbers in the coming years, considering their preference for missions involving constellations or distributed space systems. This trend of increasing demand for CubeSats and their application in advanced missions requires more electrical power for their operation. The increased electrical power demand can be solved using a Solar Array Drive Mechanism. The SADM allows relative rotary motion of the solar arrays with respect to the satellite structure so that the solar panels are always perpendicularly positioned to the Sun independent of the payload pointing requirements. They can produce up to 185% more power than the panels just deployed in the case of a 3U CubeSat. Only a handful of six such SADM products were found in the commercial market, and four of them had a very similar design that drove two solar arrays and could be used in limited panel mounting configurations. A need in the commercial market for a SADM system that is modular and scalable was identified. Space mechanisms such as the SADM were found to be one of the major causes of mission failure after communications and unknown causes. It was found that tribological elements were the prominent root cause of such space mechanism failure. A research gap was identified to find the root causes of tribological failure in space mechanisms and design a SADM system that minimises failure caused due to tribological elements. This thesis has succeeded in designing a SADM that is scalable to multiple sizes of CubeSats (3U to 12U), applicable to more than three panel mounting configurations that were possible with the existing SADM and minimising failure chances due to tribological elements. The current SADM has minimised the chances of failure due to common tribological elements such as roller bearings and sliprings by eliminating the cause of failure. This includes eliminating rolling elements, liquid lubricants and metals in the case of bearings and a novel power and data transfer mechanism alternative to sliprings called "Flex-wrap" has been designed in this project. The current SADM is the smallest in the market in dimensions (70x50x6.9 mm) and applies to 5 different solar array mounting configurations.

The optimization of interplanetary, low-­thrust trajectories is a computationally expensive aspect of preliminary mission design. To reduce the computational burden associated with it, surrogate models can be used as cheap approximations of the original fitness function. Training the surrogate models in a fully online manner can be done to remove the need of having previously generated datasets, which is another source of computational cost. The Sims­-Flanagan transcription is used to model an Earth-Mars transfer which is optimized through different optimization routines. The development of a C++ library with machine learning tooling was initiated, containing implementations for Generalized Regression Neural Networks (GRNNs) and Radial Basis Function Networks (RBFNs) that are used in global and local surrogates, respectively, having their hyperparameters tuned through cross-­validation. A surrogate model was constructed using Differential Evolution (DE) operators and an uncertainty-based infill criterion for the global search phase, and approximation of the derivative of the original fitness function which is provided to SNOPT (Sparse Nonlinear Optimizer), in the local search phase. An ablation study was performed to assess how each of the components of the surrogate model contribute to the results. It was verified that neither the derivative information nor the local search as a whole led to better results. The surrogate model was also outperformed by the standard optimization strategy found in literature, Monotonic Basin Hopping (MBH). Two new surrogate models incorporating ideas of this strategy were created, with one of them outperforming every other model that was tested. Despite not having performed a full study of the computational effort due to the simulations having been run in a server with a variable load, the new models present better results for similar amounts of fitness function evaluations. A Wilcoxon rank-­sum test was performed to assess whether the results have statistical significance, leading to the conclusion that a surrogate model can be used to improve the optimization of low-­thrust trajectories modeled with the Sims­-Flanagan transcription when inserted in a monotonic basin hopping optimization scheme.

Autonomous orbit determination could help reducing the growing load on ground-based deep space networks. For small satellites, orbit determination based solely on inter-satellite measurements could be a strong alternative for conventional ground-based navigation support. The possibilities of this autonomous orbit determination method are investigated, using LUMIO and Lunar Pathfinder mission aspects. The feasibility strongly depends on the estimated two-way ranging error, so the dependency between the radio parameters and performance in autonomous orbit determination is explored. To estimate the two-way error a link budget analysis is performed in which the bit rate and antenna gain are free parameters. In an iterative process, the radio parameters are obtained in which autonomous orbit determination can be performed in compliance with LUMIO’s mission requirements.

Aerosol is defined as the suspension of solid or liquid in the atmosphere. While some aerosols do not pose any serious threat to humankind, others have devastating effects. Thus, it is important to understand the type and distribution of these aerosols in the atmosphere. While traditional satellite data pose delays due to the huge amount of data that it has to process through the traditional pipeline, machine learning is quickly proving to be a likely winning candidate in providing accurate and efficient models. The advances in machine learning and cloud computing combined with the terabytes of data from the earth observation satellites open up avenues for creating newer and variant data products of better accuracy in the domain of aerosol classification. There is a recognized need for distinguishing and characterizing different kinds of aerosols in the 5.6 billion dollar air quality market. This research focuses on the investigation and designing of machine learning models for the aerosol retrieval process. To meet this end we implement supervised learning on satellite data to achieve aerosol classification to distinguish the different types of aerosols. The two satellites whose data will be analyzed are POLDER-3 and OMI. In this study, the three supervised learning algorithms SVM, RF and KNN were implemented to classify aerosol types for the year 2006 on POLDER-3 and OMI satellite data. We used results from previous studies on POLDER-3 as eight input aerosol class labels along with selected microphysical parameters for supervised learning and could achieve a very high reproducibility of the aerosol classes with a reduction in training time. Similarly, we used three aerosol label classes as input along with selected microphysical parameters to generate a high reproducibility. The results showed that SVM performed best on POLDER-3 data while RF was the best performing algorithm on OMI data. Using SVM on POLDER-3 dataset with hyperparameter tuning we reached an overall accuracy of 99%, precision of 99%, recall of 98%, and f1-score of 99% on POLDER-3 dataset for the eight classes of aerosols. Using RF on OMI dataset with hyperparameter tuning we reached an overall accuracy of 99%, precision of 99%, recall of 99%, and f1-score of 99% on OMI dataset for the three classes of aerosols. It was concluded that while machine learning still has a long way to go, it shows promising results in the field of satellite data processing for aerosol classification.

With the maturization of the cubesat market, the requirements for finer control over satellite attitude drives the development of higher performance Attitude Determination and Control Systems. (ADCS)

The ADCS uses reaction wheels as an actuation method, which in turn use brushless DC (BLDC) motors to rotate a heavy rotor wheel. BLDC motors use an external controller to drive armature currents, rotating the stator field, which in turn rotates the rotor.

The purpose of this MSc thesis is to improve rotor state estimation, improving reaction wheel performance by allowing for the implementation of more advanced control algorithms. This state estimation is improved through the development of a capacitive sensing encoder. This encoder was developed, a prototype built and tested. The encoder has an accuracy of 0.101 degrees, which is a 300x improvement over the currently used method of state estimation.

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.

Stereo matching, the process of inferring depth maps from stereo images, is one of the most heavily investigated topics in computer vision. It is part of the first module of navigation systems of planetary rovers, e.g., NASA’s Mars Exploration Rover (MER) missions, NASA’s Mars Science Laboratory (MSL) mission, and ESA’s ExoMars mission. Many stereo matching algorithms, traditional and deep learning based, are available and their performance is typically evaluated on indoor objects or outdoor city scenes. In this thesis, our goal is to improve insight into the performance of stereo matching algorithms for the Martian surface on a single-board computer. First, we search for and compare stereo matching algorithms ranked on stereo vision datasets to obtain a manageable set of algorithms and verify their implementations. Second, we derive performance metrics from the requirements and validate our set of verified algorithms on the Katwijk Beach Planetary Rover dataset. Through experimental evaluation, we gain insight into the performance of four stereo matching algorithms.

Low downlink data rates limit the scientific and commercial return from small satellites. Now with constellations of small satellites becoming popular, the amount of data generated in low Earth orbit will be greater than ever. A novel data downlink architecture for small satellites is investigated that uses non-geostationary orbit (mega-)constellations to relay data to ground. This data-relay architecture provides more contact time with the satellite allowing for higher a throughput per orbit to be achieved. A data-relay orbital simulator is developed that can find the contact opportunities for a satellite in low Earth orbit with non-geostationary orbit (mega-)constellations. A design optimization is performed using the results of this simulation for a data-relay downlink on two concept small satellite missions: an Earth Observation mission and an Internet-of-Things mission. This optimization shows that with the novel downlink architecture higher throughputs and lower latencies can be achieved while using the current state-of-the-art in small satellite communication technologies.

The Lunar Zebro is a small six-legged robot. It has the potential to be used in a swarm carrying out objectives like exploring planetary surfaces. A new step towards autonomous navigation is made with the newly developed Obstacle Processing ALgorithm (OPAL) using primarily open-source libraries. This study showed that the initial iteration of OPAL could detect rocks and determine their absolute position to the rover’s low-positioned cameras using a stereo vision system. Obstacles and their relative distances are detected using the disparity map—the amount of shift of pixels in the stereo image pair. When translating the disparity to V-disparity, a histogram of the disparity per row, the ground and the obstacle could be isolated. It took six steps to realise this thesis goal. After setting the requirements, a test model, called Bars, was developed and tested at a location containing a Mars-like environment (Decos). This test model uses, along with Lunar Zebro’s hardware, mostly Commercial Off-The-Shelf products. With the footage, all the different components of OPAL were integrated into one algorithm. Hereafter, a pipeline on a server was created, and multiple test cases were run to establish results. The predetermined requirements of the algorithm were validated using measuring tape measurements and ground-truth bounding boxes tracked by a CSRT-algorithm. Together, Bars and the initial iteration of OPAL prove feasibility and expose opportunities and challenges, which could be a starting point for optimisations or other approaches.