Kenny Andersson, Swedish Defence Research Agency; Tim Kinnunen, Swedish Defence Research Agency; Semeli Papadogiannakis, Swedish Defence Research Agency; Rolf Ragnarsson, Swedish Defence Research Agency
Keywords: Space situational awareness, Sensor planning, Optimization, Radar, Sensor resource allocation
Abstract:
As the population of active satellites and orbital debris continues to grow, the ability to track, catalogue, monitor, and characterise these objects has become critical for maintaining space situational awareness. This, in turn, enables essential functions such as collision avoidance, anomaly detection, and operational planning. Achieving this relies on accurate and timely observations, often using ground-based radar systems. Efficiently allocating radar resources to track a large number of satellites is a non-trivial task. This is not only due to limited observation windows and diverse orbital trajectories but also due to the necessity to balance competing priorities. The sensors used for satellite tracking can be dual-purpose, meaning that their primary functions may have to take precedence over satellite observations. Furthermore, misalignment or obstructions, coupled with the decision to forfeit measurement opportunities in favour of other critical sensor tasks, inherently leads to some measurement attempts being unsuccessful.
Sensor planning, also referred to as sensor scheduling, encompasses a range of techniques aimed at optimising the utilisation of one or more sensor resources to maximise the information obtained. Based on information theory, sensor data and the desired information are modelled as stochastic variables, allowing the sensor planning problem to be formulated as an optimisation problem that minimises the entropy of the target variables. In this work, we solve this optimisation problem using a genetic algorithm to determine the optimal schedule for measuring a set of satellites, seeking to minimise and maintain the uncertainties below a specified threshold while preferably allocating sensor resources to the satellites with higher priorities.
Our proposed method recognises that enforcing strict uncertainty requirements can be counterproductive. Allocating sensor resources to marginally reduce the uncertainty of an already well-known satellite orbit may come at the expense of opportunities to reduce the uncertainty for others. Therefore, we introduce an alternative approach that replaces hard constraints with soft constraints, allowing our algorithm to optimise the overall performance across the satellite population at the cost of a potentially slight increase in uncertainty. Additionally, unlike approaches that relax selection variables, our method maintains them as strictly binary (either a measurement is taken or it is not), preventing fractional selection variables which could lead to ambiguity.
To evaluate our proposed method, three different planning strategies were implemented and compared. The first, and least dynamic strategy is hierarchical, which strictly prioritises the observation of satellites with higher priorities. The second strategy is conservative, and is designed to minimise the utilisation of sensor resources. The third, and most dynamic strategy tested is redistributive, balancing resource allocation across satellites with varying priority levels. It achieves this by reallocating measurement opportunities initially assigned to one satellite to another when the former’s uncertainty reaches a certain threshold.
Our results show that using more dynamic observation planning strategies decreases the overall uncertainties in the states of all satellites. In an ideal case, the most dynamic strategy would involve updating the schedule after every single measurement. However, such an approach may be computationally infeasible. Instead, a balance must be found where the schedule is updated frequently enough to adapt to changing conditions while avoiding excessive rescheduling. Achieving this balance requires solving the scheduling problem quickly, which we suggest can be accomplished using a genetic algorithm. While genetic algorithms do not guarantee convergence to the global optimum, a well-tuned algorithm improves the likelihood of converging to it within the defined search space while enabling responsive tracking operations through reduced computation time. In principle, the ability to iterate and rapidly update the schedule increases the likelihood of correcting mistakes that may arise from non-optimal solutions, which ultimately reduces the necessity for a guaranteed optimal solution. The potential delay in searching for the perfect solution can outweigh the benefits of accuracy, making it more beneficial to generate a near-optimal solution quicker.
This work discusses the importance of sensor planning to maintain a robust space situational awareness. Our scenario specifically uses a generic radar sensor to observe satellites in low Earth orbits with high inclinations, but the same concepts can likewise be applied to other sensors and orbital regimes. The core aspects of this paper will be focused on the formulation of the optimisation problem to be applicable for a genetic algorithm, which is based on previous work done by colleagues at the Swedish Defence Research Agency. The use of evolutionary algorithms, specifically variants of the genetic algorithm, to solve scheduling and tasking problems has a long history and is well documented. This formulation of the optimisation problem inherently incorporates the cost of performing measurements and the penalties for failing to satisfy desired constraints, as well as resolving measurement conflicts among satellites with differing priorities.
Date of Conference: September 16-19, 2025
Track: Machine Learning for SDA Applications