LSTM-Based Incremental Learning for Multi-FOV Star Sensors' Structural Alignment Calibration

A multi-field-of-view (FOV) star sensor is an innovative attitude measurement device with superior star acquisition capabilities. Because of the excessive maneuvering of sensors and the influence of the space thermal environment, the structural alignment errors of multi-FOV star sensors are easy to change in real time during the orbit operation, which affects the overall attitude measurement fusion accuracy. To address the significant variations in the structural alignment error caused by the thermal radiation exposure (TRE) of the structure in real time, a long short-term memory (LSTM)-based incremental learning framework based on multiple learner models is designed. Each learner is mainly composed of 1-D convolutional neural network (1DCNN) and bi-directional long short-term memory (Bi-LSTM). To improve the generalization ability of the model, the simulation data with noise are generated by interpolation to form an offline training dataset. The LSTM-based incremental learning framework learns knowledge from the real-time measured data stream. The updated dataset composed of the base dataset and incremental data is used to train the latest learner model. Then, the models in the buffer pool can be updated to adapt to the changes in the current data stream. The simulation experiment and ground experiment verify the advantages of the proposed algorithm compared with the traditional algorithms in the error estimation of structural alignment calibration.