A Study Of Compressed Sensing Reconstruction Method Based On Restricted Boltzmann Machine

Under the era of "Internet of Everything",data has the characteristics of large scale,various types,complex structure and high dimensionality.Traditional signal processing methods based on Quest's sampling theory face many challenges,such as hardware cost,device power,and redundant data.In order to solve the above problems,a new sampling method,compressed sensing,is proposed.Under the assumption of sparsity,compressed sensing can sample the signal far below the Nyquist sampling rate,and the original signal can be accurately reconstructed by a small number of measurements obtained by sampling.After more than ten years of development,compressed sensing has gradually formed a relatively complete theoretical system.Among them,the reconstruction algorithm is the core technology of the compressed sensing theory,and it is also the hotspot of current research.In the research of reconstruction algorithm,how to use signal prior information to improve the performance of compressed sensing reconstruction algorithm is a very important topic.At present,there are some problems to be solved in this direction:the acquisition of prior information is mainly through artificial observation or simple mathematical statistics,and it is difficult to ensure the accuracy of information;some single and shallow prior information is obtained;It is often based on structural information under a specific sparse basis.Therefore,this paper focuses on the fusion of prior information and convex optimization class reconstruction algorithms,mainly to study the following two points:First,how to accurately obtain more prior information,and provide support for subsequent reconstruction algorithms;The fusion of prior information and reconstruction algorithm,in order to make full use of prior information as a principle,establish a reasonable optimization model to achieve the purpose of improving the performance of the reconstruction algorithm.Based on the above research ideas,the RBM-WL1M algorithm proposed in this paper.First,a limited Boltzmann machine is used to model the signal sparse mode distribution to learn the high-order dependencies between the elements of the sparse mode.Secondly,based on this prior distribution,the common support set of the same type of signal and the probability that each element is non-zero value are obtained,and the weighting norm minimizes the weight parameter in the reconstruction algorithm.Finally,the reconstruction of the signal is obtained by solving the weighted norm minimization problem.Among them,for the common support set of the signal and the probability that each element of the signal is non-zero,the greedy algorithm and Gibbs sampling are used to obtain the estimated value.Experiments show that compared with other reconstruction algorithms,such as BSBL algorithm,algorithm and RBM-OMP algorithm,the RBM-WL1M algorithm proposed in this paper has certain advantages in the reconstruction performance and sampling complexity of ECG signals.At low sampling rates,the reconstruction performance of the RBM-WL1M algorithm is also well guaranteed.Therefore,the RBM-WL1M algorithm is suitable for low-sampling or power-limited compression sensing problems.