基于迁移学习的新精神活性物质拉曼光谱识别研究

Research on Raman spectral identification of new psychoactive substances based on transfer learning

随着精神活性物质(NPS)对社会的潜在危害越来越严重,提高新型毒品的快速检测、识别能力是当前打击毒品犯罪的重中之重。本研究采集了196份新精神活性物质缴获样本拉曼光谱,结合连续小波变换和卷积神经网络迁移学习模型,对五种检测仪器获得的拉曼光谱数据进行分析和识别。对VGG16、InceptionV3和ResNet50三种深度学习模型分类效能的比较结果显示,单仪器新精神活性物质识别能力为80.2%~100%不等,多仪器综合识别能力为88.6%。通过合理的光谱预处理方法,能够将不同仪器的NPS光谱检测数据,统一格式并进行批量的分析和识别,精准提高了拉曼光谱的数据提取和利用效率。

As the potential harm of psychoactive substances(NPS)to society is becoming more and more serious,improving the rapid detection and identification of new drugs is a top priority in the current fight against drug crimes.In this study,196 Raman spectra of seized samples of new psychoactive substances were collected,and the Raman spectral data obtained from five detection instruments were analyzed and identified by combining continuous wavelet transform and convolutional neural network migration learning models.The comparison of the classification efficacy of three deep learning models,VGG16,InceptionV3 and ResNet50,showed that the single-instrument new psychoactive substance recognition ability ranged from 80.2%to 100%,and the combined multi-instrument recognition ability was 88.6%.Through reasonable spectral pre-processing methods,the NPS spectral detection data from different instru-ments can be unified in format and analyzed and identified in batch,which precisely improves the efficiency of data ex-traction and utilization of Raman spectroscopy.