Learning-based Compressive Sensing for UWB Receiver

Ultra-wideband (UWB) communication is an emerging technology for high data rate information transfer in a medium range wireless communication network. It has different applications such as UWB radars, wireless sensor networks, and medical imaging. The Federal Communication Commission (FCC) requires UWB signals to have a very short width and a very low power. Such a low power signal demands Analogue to Digital Converter (ADC)s with high sampling rates which are hard to realize. The research community has put forth a number of approaches to address this issue using Compressive Sensing (CS) with random or semi-random measurement matrices. However, these approaches are computationally demanding when a higher accuracy is desired. In this research, we propose a data-driven approach for extracting rich signal segments. Such segments are identified using autoencoders which have been trained on training examples that are stochastically analogous to those of our interest. The learning-based approximation of the measurement matrix enables us to achieve a high accuracy by eliminating the need for sampling signal segments which are not quite effective in the reconstruction phase. Empirical results show our approach outperforms state-of-the-art solutions by yielding a superior Bit Error Rate (BER) especially in environments with low Signal to Noise Ratio (SNR).