Research On Feature Extraction And Classification Method Of Terahertz Time Domain Signal Of Liquid Dangerous Goods

With the development of social economy,we will face many new security threats.Among them,liquid safety cannot be ignored.As for the liquid detection technology,the liquid detection technology based on terahertz has gradually been paid attention by researchers.The researchers have conducted in-depth theoretical research on the detection technology of specific substances and proved the practical feasibility of the terahertz liquid detection technology.However,the domestic research on terahertz-based liquid identification and classification technology is still in its infancy.Based on the terahertz-based liquid detection technology,this paper studies the terahertz-based liquid feature extraction and classification method,mainly doing the following work:(1)In terms of preprocessing of terahertz time-domain signal data,a data processing method is proposed,which can smooth the terahertz time-domain signal data.This method aims at the noise caused by the coincidence jitter and sampling jitter of the delay line in the terahertz time-domain spectral system,and proposes to use the empirical mode decomposition-R / S analysis method to smooth the terahertz spectral signal.Experimental results show that,compared with common algorithms such as Gaussian smoothing algorithm,EMD-R / S analysis algorithm can effectively smooth the terahertz time-domain spectral signal,and can effectively restore the characteristics of the terahertz time-domain signal.(2)In terms of terahertz time domain signal feature extraction and classification of liquid dangerous goods,this article explores common feature extraction methods and classification methods.In this paper,we first try to extract the terahertz time-domain signal characteristics of liquid dangerous goods using multi-dimensional scaling.According to the principle of multi-dimensional scaling method,this paper explores two factors that may affect the effect of multi-dimensional scaling method feature extraction: distance measurement method and extracted feature dimension,and draws corresponding conclusions.Aiming at the problem that the performance of the classification algorithm affects the classification accuracy rate,this paper uses the features extracted by the multi-dimensional scaling method to classify using a variety of classification algorithms.The experimental results prove that the support vector machine classification effect is better than other classification algorithms.Next,this paper compares the multi-dimensional scaling method with the four methods of Laplace transform,local linear embedding,principal component analysis,and t SNE.The experimental results show that the multidimensional scaling method is slightly superior to the feature extraction of liquid dangerous goods.Finally,in view of the limitation that the multi-dimensional scaling method cannot extract the distinguishable features of multiple types of liquid dangerous goods,this paper proposes an improved method to successfully classify multiple types of liquid dangerous goods.(3)In addition to using traditional pattern recognition ideas for research,this paper also tries to use popular deep learning techniques to extract and classify the terahertz time-domain signals of liquid dangerous goods.This paper first analyzes the internal structure of the restricted Boltzmann machine and studies the calculation process of the restricted Boltzmann machine.Next,this paper analyzes the principle of deep belief network,and proposes a terahertz time-domain signal classification method for liquid dangerous goods based on deep belief network.Finally,this paper studies the influence of the network parameter setting of the deep belief network on the classification of terahertz time-domain signals of liquid dangerous goods.The experimental results show that,with the selection of appropriate network parameters,the deep belief network can achieve terahertz time-domain signal classification of various types of liquid dangerous goods,and the classification accuracy rate can reach 95%.This paper explores the preprocessing,feature extraction and classification of the terahertz time domain signal of liquid dangerous goods,which improves the accuracy of the classification of the terahertz time domain signal of liquid dangerous goods.It has practical significance and good commercial prospects in security inspection applications.However,in the future,a large number of additional experiments are needed to prove the method proposed in this paper practical and feasible.