基于1D-CNN提取Cl2紫外吸收谱特征的浓度反演方法研究  

Research of Chlorine Concentration Inversion Method Based on 1D-CNN Using Ultraviolet Spectral

作　　者：贾彤华 程光旭[1] 杨嘉聪 陈昇 王海容[3] 胡海军[1] JIA Tong-hua;CHENG Guang-xu;YANG Jia-cong;CHEN Sheng;WANG Hai-rong;HU Hai-jun(Department of Process Equipment and Control Engineering,School of Chemical and Technology,Xi'an Jiaotong University,Xi'an710049,China;China Special Equipment Inspection&Research Institute,National Market Supervision Technology,Beijing 100029,China;State Key Laboratory of Machinery Manufacturing System,School of Mechanical Engineering,Xi'an Jiaotong University,Xi'an 710049,China)

机构地区：[1]西安交通大学化学工程与技术学院化机系,陕西西安710049 [2]国家市场监管技术创新中心中国特种设备检测研究院,北京100029 [3]西安交通大学机械工程学院机械制造系统国家重点实验室,陕西西安710049

出　　处：《光谱学与光谱分析》2024年第11期3109-3119,共11页Spectroscopy and Spectral Analysis

基　　金：国家重点研发计划子课题(2022YFC3004504)资助。

摘　　要：开放环境下氯气泄漏的准确检测一直是氯碱生产企业亟待解决的难题,差分吸收光谱技术(DOAS)可以实现大气中的污染气体的痕量远距离测量,而氯气的紫外吸收光谱呈现“慢变化”的特征,无法用差分的方法分离吸收特征与噪声信号。提出了一种基于一维卷积神经网络(1D-CNN)的氯气浓度反演算法来充分利用光谱信息,通过逐层提取氯气的吸收特征,解决了传统算法容易受噪声干扰导致反演精度下降的问题。与常用的最小二乘法(LS)、多层感知机(MLP)、支持向量机(SVR)和k近邻(KNN)方法相比,该算法的反演结果相比实测数据的准确度最高(R2=0.996,RMSE=4.40,MAE=2.64,SMAPE=8.51%)。由于系统中不可避免的随机噪声会对检测产生干扰,对比了S-G滤波、傅里叶变换、奇异值分解和小波变换分解算法的预处理效果。结果表明,S-G滤波和小波分解算法可以在去除噪声的同时保留氯气的吸收特征信息,进一步提高氯气浓度反演模型的性能。所提出的浓度反演算法为实现开放环境下氯气泄漏的远距离定量检测提供了新的可行方法。The accurate detection of chlorine leakage in an open environment has been an urgent problem for chlor-alkali manufacturers.Differential optical absorption spectroscopy(DOAS)can realize long-distance measurements of trace polluting gases in the atmosphere.Due to the flat characteristic of the UV absorption spectrum of chlorine,it is impossible to differentiate the absorption characteristics from the noise signal by normal methods.A new algorithm based on a one-dimensional convolutional neural network(1D-CNN)is proposed to solve the problem of poor accuracy caused by noise interference,which can fully use spectral information and extract chlorine absorption characteristics layer by layer.Compared with commonly used models such as least squares(LS),multilayer perceptron(MLP),support vector machine(SVR),and k-nearest neighbor(KNN),the inversion result of this algorithm has the highest accuracy(R~2=0.996,RMSE=4.40,MAE=2.64,SMAPE=8.51%).Due to the inevitable random noise in-the system,the preprocessing effects of the S-G filter,Fourier transform,singular value decomposition,and wavelet transform decomposition algorithms are compared.The results show that S-G filtering and wavelet decomposition algorithms can retain the characteristic information of chlorine while removing noise and further improving the model's performance.The concentration inversion model based on 1D-CNN provides a new feasible method for long-distance quantitative detection of chlorine leakage in the open environment.

关 键 词：氯气泄漏 远程检测 差分吸收光谱技术 深度学习 一维卷积神经网络 

分 类 号：O433.4[机械工程—光学工程]