Research On Moving Target Recognition With Seismic Signals

It is of great scientific and practical significance to carry out the research on intelligent border monitoring.As terminal equipment in modern digital environments,seismic sensing systems can be used to identify moving targets.Seismic sensing systems have the characteristics of strong concealment and non-line-of-sight detection,so it is promising to apply this technique in boundary monitoring.Although this field has been developed for two decades,related works encounter new challenges in term of datasets,target detection algorithms and target classification algorithms.This dissertation studies the current challenging problems in this field.The main contributions are as follows:(1)Study on the mechanism of seismic signals induced by ground moving target and low-altitude flying target.The equivalent mechanical models of various types of targets are established.The seismic signals generated by the frictions between ground moving targets and the earth surface,and the acoustic-seismic coupling between flying targets and the earth surface are discussed.The moving target recognition principle based on Rayleigh wave is determined,and the propagation mechanism of Rayleigh wave in layered geological medium is analyzed.(2)Construction of independent and abundant moving target datasets.To solve the problem of insufficient seismic datasets,seismic signals generated by six kinds of common moving targets are acquired in two seasons.The collated dataset is named JL dataset.Data analysis shows that all types of data accord with the objective facts and the motion law.In addition,the JL dataset possesses good SNRs.The SNR spans are wide,and the SNR ranges are all over 15 d B.The collected signals provide necessary data foundation for the research of target detection algorithms and target classification algorithms.(3)Study on target detection algorithm based on capacity dimensions and linear support vector machine.Most of the traditional target detection algorithms rely on signal energy,which results in high false positive rate and false negative rate when detecting weak signals.To further improve the detection rate,a new target detection algorithm using capacity dimensions joint support vector machine(FD-SVM)is proposed in this dissertation.Capacity dimensions can quantitatively describe the chaotic and nonlinear behavior of seismic signals,and can complete the extraction of seismic features at the same time.The linear support vector machine is used to identify effective signal component and interference noise component so as to accurately detect moving targets.Two case studies show that the proposed FD-SVM algorithm achieves promising accuracy rate,recall rate and F1 score.Comparative experiments demonstrate that the FD-SVM algorithm is better than benchmark algorithms in detecting weak seismic signals.(4)Study on target classification algorithm based on compressed seismic measurements and deep neural networks.Deep learning has become the most popular method to classify seismic signals induced by moving targets.However,the previously established deep neural networks are now faced with two challenges,i.e.,redundant feature extraction and weak ability to resist noise.To solve the two problems simultaneously,a novel target classification algorithm(CO-SDL)using compressed measurements is proposed in this thesis.This algorithm can take advantage of the sparsity of seismic signals in the time domain.Specifically,the measurement matrix in the CO-SDL model linearly projects original seismic signals onto a compressed domain to obtain compressed measurements.Compression observation greatly reduces data dimensions,but can retain most of the valuable information and suppress the noise energy.Next,the deep neural networks in the CO-SDL model can comprehensively extract nonlinear features from the compressed measurements,and then infer the mapping relations between compressed measurements and target categories.Two case studies prove that the proposed CO-SDL algorithm can efficiently learn the deep relations of seismic signals.It obtains a high classification accuracy only by using the low dimensional compressed measurements.Comparative experiments demonstrate that the presented CO-SDL algorithm is 10 times faster than the state-of-the-art algorithms with comparable accuracy.Furthermore,the CO-SDL algorithm shows the best anti-noise ability.