| Laser-induced breakdown spectroscopy(LIBS)is a multi-element analysis technique based on the study of plasma emission lines generated by the interaction of high-power laser pulses with sample media.Each spectrum in plasma radiation corresponds to a unique transition in an atom,ion or molecule that can be used as a "fingerprint" to identify the structure of the sample.LIBS detection technology provides a fast and versatile method for organic explosive identification under long-range conditions that are difficult to achieve using conventional analytical techniques.However,based on the current research status at home and abroad,most of the cases are directed to the identification of organic explosives on aluminum(Al)pedestals.When organic explosives are deposited on organic susceptors,the spectral similarity between samples increases and the recognition performance is unstable.Therefore,identifying organic explosives deposited on organic susceptors remains a particularly challenging and promising subject.In the classification of LIBS-based explosives,high energy materials are usually organic compounds that require multivariate analysis.At present,when using LIBS for explosives identification,the partial least squares discriminant method(PLS-DA)is often used for spectral data analysis and processing.For applications that identify organic explosives from trace samples similar to organic explosive structures on pedestals of similar nature,the linear additivity of PLS-DA will result in limitations in recognition performance.BP neural network(BPNNC)has unique self-organization,self-learning adjustment and natural nonlinear ability,which can reduce the influence of matrix effect on recognition,thus effectively improving the accuracy of recognition.At the same time,in order to overcome the defects of BP neural network classifier(BPNNC)requiring large data volume and black box structure,scatter plot analysis is introduced to describe the relationship between characteristic line and sample category,and to determine the optimal spectral line feature to reduce The model input features are identified to establish a recognition model under small sample conditions.This thesis first builds an experimental platform for the long-distance LIBS detection system,and obtains a certain amount of sample spectral data through LIBS experiments,including three organic explosives: RDX,TNT,DNT,and potentially confusing samples with common atomic composition of explosives: penicillin,acyclovir,butter,hand cream,seven samples are located on two organic substrates.On this basis,the denoising of spectral data and the identification of specific compounds are studied.The discrete wavelet transform method is used to deduct the continuous background noise,and the wavelet threshold denoising method is used to remove the general noise,so as to determine the best approximation of the clean LIBS signal.Finally,experiments verify that PLS-DA and scatter plot analysis combined with BPNNC's ability to identify organic explosives deposited on organic susceptors.Based on the seven sample spectra,PLS-DA modeling and scatter plot analysis were combined with BPNNC training,and the optimal models of the two methods were determined by cross-validation.The two identification models were validated based on a certain amount of “unknown” spectral data to evaluate the ability of the two methods to identify organic explosives deposited on organic susceptors.The experiment found that in the PLS-DA recognition model,the accuracy of specific compounds at 10 meters was 77.68%,and the false positive rate was 22.95%.In the BPNNC recognition model,organic explosives with the same conditions were identified and accurate.The rate was 88.39% and the false positive rate was 9.38%.It was finally determined that the scatter plot analysis combined with BPNNC in the test for the identification of organic explosives from trace samples with similar organic explosive structures on similar substrates has better results. |
PDF Full Text Request