Online Factor Analysis Model With Kullback–Leibler Constraint for Satellite Target Recognition

Satellite target data are usually obtained over a long period in streaming mode rather than at one time. In satellite target recognition, most of existing methods are in offline mode. With the arrival of new data, the offline learning methods have to retrain a new model with both historical and new data. However, the repetitive retraining not only leads to a huge burden for storage but also results in poor real-time performance, due to the storage and repeated calculation of all historical data. To address this problem, this paper presents a novel satellite target recognition method based on online factor analysis model with Kullback-Leibler constraint (OnFA-KL model). On the one hand, the parameters of OnFA-KL model can be updated by only using new data without revisiting the past data; on the other hand, a Kullback-Leibler (KL) divergence term is introduced to preserve the previous learned knowledge of the past data, thus alleviating the catastrophic forgetting. In addition, owing to the streaming mode of the satellite target data, streaming variational Bayesian (SVB) algorithm is employed to infer the parameters of our OnFA-KL model. In the simulation experiments, the data of five satellite targets we use are generated by using real orbit parameters and computer aided design (CAD) models. Numerical experiments show that compared with the offline learning method, the proposed method achieves comparable recognition performance and improves the learning efficiency.