A correlation-aware and deployable architecture for reliable unsupervised satellite anomaly detection

Satellite anomaly detection plays a critical role in ensuring the reliability and safety of space missions. Nevertheless, most existing methods rely extensively on labelled data and face difficulties in effectively modelling the complex dependencies among multivariate telemetry signals, thereby limiting their applicability in practical satellite operations. This study proposes Co-GAN, a correlation-aware generative adversarial network developed for unsupervised satellite anomaly detection. The method captures both temporal patterns and inter-variable correlations within high-dimensional telemetry data by combining generative modelling with dynamic correlation analysis. In contrast to conventional threshold-based or supervised approaches, Co-GAN operates in unlabeled environments and adapts automatically to variations in operational conditions. Experimental results using real-world satellite telemetry data indicate that the proposed approach substantially improves anomaly detection precision while reducing false alarm rates, thereby demonstrating its effectiveness and practical significance in aerospace applications.