Research On Compressed Sensing Tenique In Distributed Systems

With the widely promotion and application of mobile Internet and Internet of Things, distributed system has broad potential for growth, and becomes an essential technique for future wireless network. In distributed system, communication among a large number of nodes gives rise to massive data traffic, which calls for efficient data compression technique.In traditional data processing methods, the way of first sampling and then compression wastes many processing resources, and causes high cost of hardware implementation and power consumption. Compressed sensing breaks the traditional mode of signal processing, and directly acquires compressed representation of the signal, which omits the sampling of large amount information of no useful. By utilizing compressed sensing technique into the signal processing of distributed system, one can reduce the cost of each node, and reduce the communication cost among nodes.This dissertation investigates the compressed sensing technique in distributed systems, and the main contents and contributions are listed as follows.Firstly, we study the compressed sampling techniques, which acquires compressed digital signals from analog signals directly. In current literature, the number of samples generated by analog-to-information converter (AIC) is limited by the deployed integrator branches, while segmented AIC suffers from high complexity.Considering these problems, this paper proposes two improved segmented AIC with lower complexity, that is, partial segmented AIC and Toeplitz-like segmented AIC. The former divides all branches into groups,and each group only implements integration in partial time period, which generates equivalent measurement matrix (EMM) that consists of multiple block diagonal matrices. In the latter, the integration waveforms are multiplexed among different branches, which generate EMM that consists of multiple Toeplitz block matrices. We prove the restricted isometry property of these two EMMs to verify the effectiveness of the proposed sampling techniques, and derives the upper bound of mean squared error of partial segemented AIC. Simulation results verify that these two sampling methods both enjoy excellent sampling performance.During the generation of samples, these two structures both adopt permutation operations within some branches, thus they are suitable for the compressed sampling at each node in distributed systems.Secondly, to reduce the transmitted data amount, we propose a compressed spectrum-sensing scheme based on one-bit quantization for cognitive radio network with multiple cognitive nodes. Each cognitive node acquires the same primary information, and then adopts one-bit quantization to obtain the sign information of compressed samples as the measurements. After receiving the sign information from different nodes,the fusion center employs two joint reconstruction algorithms to recovery the primary signal. The first one adopts a single norm to construct cost function, and the second one adopts different norms to deal with signals with different receiving qualities, and then construct cost function.Simulation results show that the proposed algorithms recover the spectrum of primary user effectively, and the second algorithm outperforms the first one.Thirdly, we propose a cooperative transmission scheme based on one-bit compressed sensing for a distributed system with multiple nodes,where the signal at each node consists of a common component and an innovation component. The proposed scheme can reduce the transmitted data amount that is required. Each node derives the measurements as the sign information of the compressed samples by using on-bit quantization.Based on the received sign information from different nodes, four joint recovery algorithms are designed to recover the signal of each node.These algorithms use different norms to construct reconstruction cost functions, and solve the optimization problems in an iterative manner. In each iteration, the former two algorithms directly reconstruct common component and innovation component, while the latter two algorithms adopt indirectly reconstruction, where the measurements of each component are recovered firstly, and then the corresponding component.Simulation results show that the proposed scheme with the four reconstruction algorithms outperforms both the independent reconstruction scheme and the scheme based on multi-bit quantization. In addition, the latter two algorithms outperform the former two.