Deep Neural Network-Based 4-Quadrant Analog Sun Sensor Calibration

Abstract

Many error sources influence the calibration experiment of 4-quadrant sun sensors, making the calibration of sun sensors cumbersome and its accuracy difficult to improve. Any continuous function on a bounded closed set can be approximated by a deep neural network. This paper uses the deep neural network model to approximate the error model. The data-driven training network is adopted to continuously modify the model parameters to fit the error compensation model and ensure that the accuracy reaches the target requirements after calibration. Considering that the deep neural network model needs a considerable amount of data, the neural network model training is divided in 2 stages. In the preliminary stage, cubic surface fitting is used to generate a dataset, which is small in size and controllable. After the completion of the initial training, the experimental data are used to fine-tune the model to achieve error compensation. The accuracy can be improved from 1deg (1-sigma) to 0.1 deg (1-sigma) after the incident angle of the sun sensor is corrected. The error compensation model eliminates the loss of accuracy caused by the distortion of light spots at the edge of the field angle and provides a favorable condition for the expansion of the field angle of the 4-quadrant analog sun sensor.