Neural Network Assisted Adaptive Covariance Tuning for Quaternion Based CubeSat EKF
- Paper ID
108154
- author
- company
Turkish-German University
- country
Türkiye
- year
2026
- abstract
The performance of Attitude Determination and Control Systems (ADCS) in CubeSat platforms is significantly affected by time-varying sensor noise characteristics and environmental disturbances. Conventional Extended Kalman Filters (EKF) employ fixed process and measurement noise covariance matrices ($Q$ and $R$), which may lead to suboptimal estimation accuracy under dynamic orbital conditions and sensor degradation. This work presents a hybrid estimation framework that integrates a neural network-based adaptive covariance tuning mechanism into a quaternion-based EKF architecture. The proposed method aims to dynamically adjust $Q$ and $R$ matrices in real time by analyzing sequences and residual statistics generated during filter operation. Unlike conventional fixed-covariance approaches, the adaptive strategy seeks to capture nonlinear and time-dependent noise behaviors that are difficult to model analytically. The system is implemented in simulation environment, incorporating realistic gyroscope and magnetometer sensor models under Low Earth Orbit (LEO) conditions. A feedforward neural network is trained using simulated sensor data to predict optimal covariance scaling factors. The evaluation framework includes Monte Carlo simulations and Root Mean Square Error (RMSE) comparison against a baseline fixed-covariance EKF. The study aims to assess the feasibility and robustness of neural-network-assisted covariance adaptation for resource-constrained small satellite platforms. The proposed framework is expected to improve estimation stability under varying noise profiles while maintaining computational feasibility for resource-constrained CubeSat platforms.