Spectrum Reconstruction via Deep Convolutional Neural Networks for Satellite Communication Systems

Abstract

Satellite based spectrum sensing is studied for a system consisting of multiple satellites and a gateway (GW), where these satellites perform spectrum sensing and send mass spectrum-sensing data to the GW. To address the challenges of mass spectrum-sensing data and limited transmission capacity of the links from spectrum-sensing satellites to the GW, we propose a method called joint anomalous data repairing and deep convolutional neural network based spectrum reconstruction (ADRD-SR), which can reconstruct the original spectrum-sensing data from the incomplete data. Specifically, the GW preprocesses the incomplete data using the anomalous data repairing algorithm. A deep convolutional neural network is constructed and well trained, then it is activated to reconstruct the preprocessed spectrum data. Additionally, to sustain good reconstruction performance by tracing the dynamical spectrum-sensing data, we design a real-time evaluation oriented spectrum reconstruction framework, through seeking the events when the mean absolute error (MAE) becomes larger than a predefined threshold. Furthermore, the ADRD-SR method can reduce the MAE by more than 68% over the conventional reconstruction methods. Moreover, the reconstructed spectrum data can be used to assist spectrum sensing, and the corresponding probability of correct detection is only degraded by 5% even when 75% of the data is discarded.