粒子群算法的近红外光谱定性分析预处理及特征提取参数优化方法研究  

Research of Parameter Optimization of Preprocessing and Feature Extraction for NIRS Qualitative Analysis Based on PSO Method

作　　者：李浩光 于云华[1,2] 逄燕 沈学锋[1,2] LI Hao-guang;YU Yun-hua;PANG Yan;SHEN Xue-feng(College of Mechanical and Control Engineering,Shandong Institute of Petrochemical and Chemical Technology,Dongying 257061,China;New Energy College,China University of Petroleum(East China),Dongying 257061,China)

机构地区：[1]山东石油化工学院机械与控制工程学院,山东东营257061 [2]中国石油大学(华东)新能源学院,山东东营257061

出　　处：《光谱学与光谱分析》2021年第9期2742-2747,共6页Spectroscopy and Spectral Analysis

基　　金：国家重大仪器设备开发专项(2014YQ470377);山东省教育厅科技计划项目(J18KA329);东营市科技发展基金项目(DJ2020032)资助。

摘　　要：在近红外光谱定性分析时,为取得良好识别效果,预处理及特征提取是不可或缺环节。预处理主要是为消除各种干扰因素对光谱数据影响,常用预处理方法有平滑、一阶导、归一化等;而特征提取方法能剔除数据中的无关信息,保留有效信息,常用特征提取方法有偏最小二乘、主成分分析、线性判别分析等。不同预处理及特征提取方法具有不同特点,构建定性分析模型时,单一使用某种预处理或特征提取方法往往难以取得理想效果,常需将多种预处理及多种特征提取方法组合使用以提升模型性能。在各预处理及特征提取环节中往往有可变参数如特征提取维数等需要设定,这些可变参数对模型性能有重要影响,因此采用多个预处理及多个特征提取方法就存在多参数需要确定的问题。研究中常采用试凑法求各待定参数最优值,欲求得多个待定参数中某一个参数最优值,首先需据经验固定其他参数值,然后将某一个待优化参数代入近红外定性分析模型进行试凑,以求得模型最优识别率所对应参数值,并将其作为最优值。利用试凑法逐个求得多个待优化参数后,再将参数组合设置到定性分析模型中,最后进行定性鉴别,但试凑法求得的参数组合难以保证为全局最优解。除试凑法外,还可通过多重循环嵌套方法来获取近定性分析模型预处理与特征提取环节最优参数组合,但是该方法需消耗大量计算机内存与计算时间,而且效率低。为此,提出一种基于粒子群算法的近红外光谱定性分析模型预处理与特征提取参数优化方法,可快速获得预处理与特征提取环节的最优参数组合,并保证代入最优参数组合的定性分析模型具有最优识别性能,采用粒子群算法对平滑系数、一阶导系数、偏最小二乘特征提取维数等参数进行寻优,并将该方法与多重循环嵌套方法进行对比实验,实验结果证�In the qualitative analysis of near-infrared spectroscopy,preprocessing and feature extraction are indispensable to achieve good recognition results.Preprocessing is mainly to eliminate the influence of various interference factors on the spectral data.The common preprocessing methods include smoothing,first-order derivatives,normalization,etc.,while the feature extraction methods can eliminate the irrelevant information in the data and retain the effective information.The common feature extraction methods include partial least squares,principal component analysis,linear discriminant analysis,etc.Different preprocessing and feature extraction methods have different characteristics.When building a qualitative analysis model,it is often difficult to achieve ideal results by using a single preprocessing or feature extraction method.It is often necessary to use a combination of multiple preprocessing and feature extraction methods to improve the model’s performance.Variable parameters such as feature extraction dimension need to be set in each preprocessing and feature extraction process.These variable parameters have an important impact on the performance of the model.Therefore,multiple parameters need to be determined in multiple preprocessing and multiple feature extraction methods.In practice,the trial and error method is often used to find the optimal value of each parameter.In order to get the optimal value of one of the parameters,it is necessary to fix the other parameter values according to experience.Then a parameter to be optimized is substituted into the NIR qualitative analysis model for trial and error to get the corresponding parameter value of the optimal recognition rate of the model,and take it as the optimal value.After several parameters to be optimized are obtained one by one by trial and error method,the combination of parameters is set into the qualitative analysis model,and finally,qualitative identification is carried out.However,the combination of parameters obtained by the trial and error

关 键 词：粒子群算法 特征提取 参数寻优 定性分析 

分 类 号：O657.3[理学—分析化学]