Study Of Identification Of Illicit Drugs In Complex Matrix Using Raman Spectroscopy Based On Machine Learning

Drug abuse not only harms the lives and health of drug addicts,but also induces illegal crimes,bring risks and hidden dangers to public safety.So rapid drug screening and detection analysis are of great significance for effectively combating drug crimes and maintaining social stability.Nowadays,the detection methods of drugs are mainly GC-MS or LC-MS in the laboratory,colloidal gold immunochromatography and chemical method in rapid screening.However,there are still various problems such as complicated operation,poor accuracy,or narrow application range.At the same time,the current detection methods mainly focus on the pre-processing of samples for complex matrix and human inspection materials,and lack of detection data mining and intelligent identification.In response to the above problems,we combine surface-enhanced Raman spectroscopy(SERS)and machine learning,uses machine learning to analyze SERS data in different complex matrix and perform feature extraction,establishes a suitable recognition model,and realizes accurate identification of SERS data in the different system,and develop a simple,fast,intelligent and accurate drug identification method.The research is mainly focus on as follow:(1)A recognition platform based on the Deep Architecture Search Network(DASN)with a handheld Raman spectrometer was developed.We explored the feasibility of Raman spectroscopy for drug identification,studied the accuracy of DASN classification model for drug identification,and compared it with KNN,RF and SVM.The results show that the accuracy identification of drugs is ACCc=100%,ACCv=100%,and ACCp=100%by DASN,which is higher than other machine learning algorithms.The feature map shows that DASN has a strong feature extraction ability,and its 1.0 AUC indicates that it has excellent robust.The combination of handheld Raman spectrometer,data transmission system,and cloud server synchronization realizes rapid and accurate analysis of the drug scene.This platform provides new ideas for solving drug identification at the scene of drug manufacturing and trafficking.(2)A recognition method based on support vector machine(SVM)and dynamic enhanced Raman was established to realize rapid determination of drugs in urine.By preparing a highly uniform self-assembled gold nanorod film as a substrate,SERS spectra with high repeatability and high stability are obtained.Using PCA to preprocess the spectral data,combined with SVM to successfully classify and identify different types of spectral data:50,2.5 and 0.1mg/Kg in urine.The average full-spectrum classification accuracy rates of MAMP and MDMA respectively They were 96.1%and 95.9%,95.3%and 95.3%,94.2%and 94.0%.The identification analysis method provides a new way for the detection of drugs in human samples.(3)A recognition method based on the combination of multi-strategy optimized deep network and SERS was established to realize the detection and recognition of drugs in the hair.The SERS substrate with high reproducibility and high sensitivity was prepared,and the detection limit of common drugs in hair was studied.The recognition accuracy of deep learning methods such as Lenet-5,VGG-16,Alexnet,Resnet50,Inception and Inception-Resnet under different conditions of original spectrum,data augmentation and wavelength selection are explored.The results showed that the detection limits of methamphetamine,ketamine and morphine,which are common in hair,reached 0.05,0.1,and 0.1ng/mg.For all recognition models,under the condition of variable selection,Inception-Resnet achieved the highest recognition accuracy,with ACCT,ACCV and ACCP increasing to 100%,99.01%and 99.25%.This method provides an effective means for long-term traceability in drug testing.(4)A rapid detection and screening method for hair samples based characteristic peaks and FWHM is proposed.By comparing and analyzing the SERS spectra of true and false positive,blanks samples and drug standards,a similar data analysis method is proposed.When the content of methamphetamine is high,using the peak position difference and the idea of decision tree,the discrimination accuracy can reach 98.98%;When the content of methamphetamine is low,using the feature that the half-peak width(13.3987±2.1589)of the SERS spectrum of false positive samples at 1004 cm-1 is larger than that of true positive samples(6.8814±0.7194),the discrimination accuracy can reach 98.76%.This method provides an effective solution for hair drug detection in the actual system.