Comparative Analysis of Regression Models for Prediction : An analysis in the context of Satcom-On-The-Move Antenna Performance Based on IMU Placement

Sammanfattning

Satcom-On-The-Move (SOTM) platforms are an emerging technology that enables communication while in motion by stabilizing an antenna towards a target location. Designing such systems, which must withstand various disturbances and operate in challenging and critical conditions, requires careful design choices. One of the most essential components for handling movement in these systems is the inertial measurement unit (IMU), making its optimal placement crucial for performance. Simulating these systems is a common approach to inform design decisions and determine the ideal IMU location. In this work, a dynamic model of a SOTM system was developed and implemented in Simulink, including a test setup with the system mounted on a vehicle. The model was used to evaluate system performance based on different IMU placements. However, conducting numerous simulations to identify optimal positions can be time-consuming. To expedite this process, machine learning algorithms were employed to estimate system performance at untested locations. These models also aid in design decisions when placement constraints are imposed or when different components and configurations are considered. Both Random Forest Regression (RFR) and Gaussian Process Regression (GPR) were applied to estimate system performance, and their comparative effectiveness was analyzed to determine the most suitable approach for this specific case. The results indicate that RFR has a clear advantage in terms of predictive performance based on IMU position of the simulated SOTM system. The conclusion drawn from this, based on the relative strengths of the models, is that the generated data is more suitable to the robust nature of RFR, which can adapt to more complex and irregular data, whereas GPR is more suited to data of a more continuous nature. While the results are promising, they are based on simulated data under specific assumptions, and future work should explore validation against real-world systems and broader model configurations to ensure generalizability and practical relevance.