Deep Learning Based Compressive Sensing for UWB Signal Reconstruction

Compressive sensing (CS) can greatly reduce the number of sampling points of signals, and therefore it is widely adopted in ultrawideband (UWB) sensor systems. However, how to reconstruct the sensing signal from the compressed signal accurately is still an open problem because original signals do not always satisfy the sparse hypothesis that is required in CS. Typically, an appropriate CS reconstruction algorithm should be designed for a particular scenario, such as signal encoding, optical imaging, and soil dynamic monitoring, etc. Unfortunately, soil data is susceptible to climatic factors, which leads to unsatisfactory performance of traditional reconstruction algorithms. To improve the accuracy of CS reconstruction for volatile signals as UWB soil echoes, we propose a novel deep learning (DL) based CS algorithm, named select-first-decide-later CS (SFDLCS) for UWB sensor signal reconstruction. In this algorithm, a search network is designed to perform the nonlinear mapping from compressed residuals to nonzero elements in sensor signal, and a decision network is designed to characterize the distribution of UWB signals. These two networks form a "select first, decide later" structure, which greatly improves the accuracy of signal reconstruction by utilizing the correlation of nonzero elements of the sensor signal. The effectiveness of this SFDLCS is demonstrated based on measured UWB soil data acquired by a P440 UWB sensor. Compared with traditional reconstruction algorithms, the proposed algorithm achieves both lower reconstruction error and stronger robustness in the noisy environment.