Study On Hybrid Model-driven And Data-driven Based Compressed Sensing Methods And Its Application

With the rapid development of society,more and more people suffer from chronic diseases due the busy and unhealthy lifestyle.The contradiction between the trend of development and limited medical resources has become an urgent global public health issues.In recent years,the rapid development of sensors and Internet of Things technology provides technical support for the implementation of remote medical monitoring systems.The system presents major advantages such as constant patient observation,a reduction in healthcare costs and increased patient mobility,provides a promising solution for the monitoring of chronic diseases.However,constant monitoring requires sensor nodes to send large amounts of biomedical data through wireless transmission.Due to the high energy consumption of wireless link,this process needs to consume a lot of energy,which greatly shortens the lifetime of sensor nodes.Compressed sensing technology can achieve extremely low complexity of compression coding through the design of measurement matrix and the energy consumption of sensor nodes in the process of data transmission is greatly reduced.It has become a research hot spot in this field.But there still are some shortcomings in existing research: First,the most approach of design of the measurement matrix is to construct a measurement matrix randomly,or construct the measurement matrix from the perspective of general performance index optimization,lacking an adaptive construction method for specific types of data.Second,existing methods of reconstruction rely on a prior signal information which is designed by hand.But a prior signal information that designed by hand is difficult to learn the deep prior signal information,which makes it difficult to further improve the performance of reconstruction.Third,concerned on the quality of reconstruction of signals more than complexity of computation of reconstruction of signals.With more and more applications need to deployed on mobile terminals with limited resources,the shortcomings of this status are becoming more and more prominent.This paper studies the above problems and the main contents of research and contributions include the following:(1)For the construction of adaptive measurement matrix,this paper propose a partialconnected autoencoder(Pc AE).Pc AE integrates the advantages of both model-driven and data-driven methods to joint optimize a sparse binary matrix and a noniterative recovery solver,so that low complexity of encoding and high compression ratio can be obtain simultaneously.Experiments on neural spikes in wireless neural recording applications demonstrate that the proposed Pc AE outperforms several state-of-the-art CS-based methods and deep learning-based methods significantly.(2)With the powerful learning capabilities of deep learning,this paper proposes a class of methods based on anti-convolution ECG signal reconstruction,which try to learn the deep prior information from ECG data and insert the deep network that equivalent the deep prior information into the ECG reconstruction method.In this way can we obtain efficient capability of reconstruction and low complexity of reconstruction.Experiments on PTB ECG signals demonstrate that this paper proposed method of reconstruction outperforms several state-of-the-art methods.Especially,the proposed algorithm obtain good performance at low sampling rate than performance of other methods at high sampling rate.Moreover,the proposed method with extreme low complexity compared other deep learning-based methods.