基于多维度并联卷积神经网络与质谱数据的芬太尼分类模型研究  

Research on fentanyl classification model based on multi-dimensional parallel convolutional neural network and mass spectrometry data

作　　者：唐龙 薛凌云[1] 徐平[1] 刘亦安 TANG Long;XUE Lingyun;XU Ping;LIU Yian(College of Automation,Hangzhou Dianzi University,Hangzhou 310028,China)

机构地区：[1]杭州电子科技大学自动化学院,浙江杭州310018

出　　处：《杭州电子科技大学学报（自然科学版）》2024年第1期23-30,共8页Journal of Hangzhou Dianzi University：Natural Sciences

基　　金：浙江省自然科学基金资助项目(2022C03G2072815)和(TGY23H180036)。

摘　　要：近几年,由于芬太尼及其类似物质很容易合成,所需要的化学物品和设备等也并不难找,导致全球发生了无数滥用芬太尼及其类似物质过量死亡的案例。为避免该新型毒品流通,提高芬太尼及其类似物质的识别率,设计了一种基于多维度并联卷积神经网络与质谱数据的芬太尼分类模型,并且利用SeLU激活函来防止梯度消失。该模型利用Focal loss作为损失函数来提高对相似品种的识别准确率,将查询质谱和5条经典参考谱拼接成5张二维图像,最终将原始查询一维谱图和其拼接得到的5张二维图像输入到并联卷积神经网络对查询质谱进行分类。实验结果表明,该方法公开数据集中识别率达99.73%的识别率,验证了该方法的有效性。In recent years,fentanyl and its analogues are easily synthesized,and the required chemicals and equipment,etc.,are not difficult to find,leading to numerous cases of overdose deaths from the abuse of fentanyl and its analogues worldwide.To avoid the circulation of this new drug and to improve the recognition rate of fentanyl and its analogs,a fentanyl classification model based on a multidimensional parallel convolutional neural network with mass spectrometry data is designed,and a SeLU activation function is used to prevent gradient disappearance.The model uses the Fcal loss function as a loss function to improve the recognition accuracy of similar species,splices the query mass spectrum and five classical reference spectra into five two-dimensional images,and finally inputs the original query one-dimensional mass spectrum and its spliced five two-dimensional images into the parallel convolutional neural network to classify the query mass spectrum.The Experimental results show that the recognition rate of 99.73%in the public dataset of the method verifies the corresponding effectiveness of the method.

关 键 词：芬太尼 质谱 卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]