In-Excess Attenuation Detection Using Satellite Link Channel Measurements at Ka- and Q-Bands With Deep-Learning Architectures

In this article, a robust process to detect the existence of in-excess attenuation values in the next future time-steps based on measured data from previous time-steps is provided for satellite channel measurements campaign at frequencies above 10 GHz. The straightforward process of anomaly detection is applied, where "normal" data, i.e., attenuation values lower than a threshold, are used to learn their predictive error probability. Afterward, a sequence is detected as "normal" or "abnormal" based on the Mahalanobis distance of its errors from the learned distribution. deep-learning architectures such as encoder-decoder and traditional paradigms are investigated and tested for different regions (Greece and U.K.) and frequency bands (Q and Ka downlinks). The method's performance is very promising and the rainfall rate information seems valuable and adequate for attenuation's detection. Finally, under a "plug-and-play" logic, the trained detectors in Greece exhibit high performance also in U.K. data showing the robustness of the proposed detection ability.