Robust DNN-Based Recovery of Wideband Spectrum Signals

Compressive sensing has been proved as an effective approach for the wireless communications. However, it is challenging to efficiently recover the wideband spectrum signals with low signal-to-noise ratio (SNR). To solve this problem, we exploit a deep neural network-based robust alternating direction method of multipliers (R-ADMM) network. This neural network can suppress the noise and optimize the learnable parameters and operations of the signal reconstruction to speed up the convergence. Furthermore, we adopt the numerical differential-based gradient computation method to enhance the robustness of the network training. Numerical results show that the proposed R-ADMM achieves markedly improved noise robustness in low SNRs and reduced reconstruction time with different experiments on the real-world and simulated signals.