Image-Driven Hybrid Model for Sun Sensor Attitude Determination

Practical Applications Sun sensors face two major challenges during attitude determination: inaccurate localization of irregular spots and inadequate compensation for nonlinear and unknown errors. To solve these problems, this paper proposes a hybrid model that combines neural networks with physical models. Specifically, a neural network is designed to compensate for errors in spot localization. The optimized coordinates of the spot center are then input into the physical model to compute the sun sensor's attitude angles. The parameters of both models are optimized together during joint training. Experimental results show that the hybrid model not only significantly improves the measurement accuracy of the sun sensor but also maintains stable performance across all viewing angles. Its single-sample inference time fully meets the real-time needs of spacecraft attitude control. Additionally, as a compact hybrid model, it can be easily integrated into small sun sensors used in space missions such as microsatellites. This study provides a reliable method to enhance the measurement accuracy of sun sensors, offering robust assurance for the efficient execution of space missions.