Shock Wave Signal Compression Sensing Method Base On Deep Learning

The shock wave generated by muzzle will cause serious harm to the people around and the acquisition instrument,so the accurate measurement of the shock wave signal plays an important role in the development and improvement of weapons.As a transient signal,the percentage of effective information of shock wave signal is relatively small in the whole acquisition period,and its frequency component is complex and contains high frequency components.When collecting shock wave signal based on traditional Nyquist principle,high sampling rate must be maintained,which will produce a lot of redundant data and seriously consume network transmission bandwidth and memory space.In order to break through the limit of Nyquist principle,compressed sensing theory was then introduced to process the shock wave signal.However,the signal reconstruction effects are influenced by the following conditions: Firstly,suitable sparse matrix must be selected when adopting compression sensing;Secondly,sparse matrix and the observation matrix must satisfy the principle of irrelevance;Thirdly,in the process of reconstruction of multiple measurement vectors,signal must meet the requirements of joint sparse apriority.To meet the above conditions,deep learning theory is introduced into the compressed sensing technology,and a deep learningbased compressed sensing method for shock wave signals is designed in this paper.The main research contents of this paper are as follows:1)Aiming at solving the problem of improper selection of sparse matrix in compressed sensing technology,which leads to large signal reconstruction error,this paper proposes an algorithm combining deep convolutional generating network and compressed sensing.The algorithm takes fixed random signals as network inputs,optimizes network parameters through the designed loss function,and finally outputs reconstructed signals.In addition,this algorithm is a kind of lazy data-learning method,which does not need a large amount of data to train the network model.Instead,each signal is learned separately,so as to realize end-toend signal recovery.This algorithm can avoid the designing of sparse matrix,and the reduction of reconstruction error is verified through simulation.Experimental results of measured shock wave signals from 15 psi and 5psi range sensors show that the proposed algorithm has a better reconstruction effect than the traditional compressed sensing technology,and its reconstruction error is merely about 0.5 times of that of DFT-OMP algorithm and DCT-OMP algorithm when the number of measurements is 2400.2)In order to meet the requirement that the distributed compressed sensing reconstruction algorithm requires the signals of different channels to meet the joint sparse prior conditions when processing multiple measurement vectors,this paper proposes an algorithm combining long and short time memory network and compressed sensing.By giving the initial residuals of all channels,the algorithm calculates the conditional probability of each non-zero value in each vector to capture the previously unknown dependencies,and then uses the learned dependency structure in the reconstruction algorithm.It’s essential to note that the long short term memory network is used to estimate the conditional probability of non-zero values of the vector,and the network is a data-driven model specialized in dealing with time series problems.It trains the parameters of the network model by minimizing the cross-entropy cost function.In order to find the value of the non-zero item of the vector,the least square method is needed.Finally,the experimental results of measured shock wave signals from 15 psi and 5psi range sensors show that the reconstruction performance of the proposed algorithm is better than that of the traditional SOMP,BCS and MT-BCS algorithms,and can improve the reconstruction performance of distributed compressed sensing technology to a certain extent.