Neural Network-based Magnetometer Calibration and Attitude Determination for Magnetic-device-based Small Satellites

Recently, small satellites have garnered increasing interest due to their advantages such as cost efficiency, rapid development cycles, and frequent launch opportunities. Magnetometers, characterized by their high reliability, low weight, low cost, and high energy efficiency, are favorable instruments for these satellites, playing a critical role in attitude determination and control. However, the accuracy of magnetometer measurements can be significantly affected by interference from other satellite instruments and the space environment, leading to reduced precision in attitude determination. Traditionally, calibration of magnetometers has been achieved using data from high-precision instruments like star trackers. This study focuses solely on magnetometers and sun sensors, common sensors in small satellites, and employs a neural network approach for magnetometer calibration to achieve high-accuracy satellite attitude determination without the need for costly instruments like star trackers, further reducing the development costs of small satellites. Firstly, this paper introduces a tiered filtering method for attitude determination using sun sensors and magnetometers: initially, the satellite uses magnetometer data for a first-level filtering to obtain rough attitude information. This preliminary attitude information, along with the covariance matrix, is then input into a second-level filtering algorithm that relies solely on sun sensor data to determine the satellite's attitude information. This filtering approach allows for flexible addition or removal of filtering layers from different instruments, improving stability while reducing computational load. For the magnetometer calibration, the attitude information filtered through the sun sensor serves as the reference for the neural network. Satellite state information, such as temperature and magnetic torque strength, is input into the neural network, which is trained to calibrate magnetometer measurement data. Unlike traditional correction methods, the neural network does not require a predefined calibration model and can incorporate multiple factors affecting the magnetometer by adding inputs. Since the improvement in attitude determination accuracy using sun sensors alone is limited, a single calibration may not entirely eliminate magnetometer biases. Therefore, it is necessary to repeat the calibration process based on the results of the previous correction, gradually enhancing the measurement accuracy of the magnetometer. Numerical simulations validate the feasibility of this approach. The results indicate that this method can effectively reduce measurement errors such as biases in magnetometer readings. Additionally, using tiered filtering and the corrected magnetometer data, the accuracy of satellite attitude determination can reach 0.5 degrees.