TF-GAN: Satellite Anomaly Detection via Generative Adversarial Networks and Time–Frequency Spectrum
Jian Jiao, Gang Li, Jianwen Wang, Zhichun Zhao, Jun Li, Hongmeng Chen · IEEE Transactions on Aerospace and Electronic Systems · 2024
Anomaly detection for a noncooperative satellite is a fundamental and significant step in space situation awareness. To effectively detect the abnormal state of an on-orbit satellite, this article proposes a novel generative adversarial network with a time–frequency (TF) spectrum (TF-GAN). First, the TF spectrum of radar echoes is acquired, and the artificially selected features are extracted from the TF spectrum of the normal state of the on-orbit satellite. Second, the proposed model is trained to learn the TF spectrum latent representation and the pre-extracted features. Then, the trained model can distinguish abnormal samples from the normal ones and output the evaluated anomaly score. Finally, the proposed TF-GAN is compared with several GAN-based methods. Experiments show that the accuracy and F1-score of the TF-GAN are superior to those of the comparison methods, and the anomaly scores of abnormal and normal classes are more separable.