基于改进的粒子群优化-反向传播神经网络的CO_(2)红外吸收光谱定量分析  

Quantitative Analysis of CO_(2) Infrared Absorption Spectrum Based on Improved Particle Swarm Optimization-Back Propagation Neural Network

作　　者：吴旭阳 管港云 刘志伟 朱冰洁 耿子迅 郑传涛[1] 严国锋 张宇[1] 王一丁[1] Wu Xuyang;Guan Gangyun;Liu Zhiwei;Zhu Bingjie;Geng Zixun;Zheng Chuantao;Yan Guofeng;Zhang Yu;Wang Yiding(State Key Laboratory of Integrated Optoelectronics,College of Electronic Science and Engineering,Jilin University,Changchun 130012,Jilin,China;Research Center for Optical Fiber Sensing,Zhejiang Laboratory,Hangzhou 311100,Zhejiang,China)

机构地区：[1]吉林大学电子科学与工程学院集成光电子学国家重点联合实验室,吉林长春130012 [2]之江实验室光纤传感中心,浙江杭州311100

出　　处：《光学学报》2024年第11期305-314,共10页Acta Optica Sinica

基　　金：国家自然科学基金(62175087,62235016,62205301,61960206004,62105118);吉林省科技发展计划(20230201054GX);浙江省自然科学基金(LTGY24F050001);长春市重点研发项目(21ZGN24)。

摘　　要：在吸收光谱气体传感领域,实测光谱存在信噪比低和由光谱失真带来的线性度低的问题,使得传统的线性分析方法难以实现高准确度的气体体积分数反演。为此,本文提出了一种基于进化策略、参数调整策略双重改进的粒子群优化(IPSO)算法,并结合误差反向传播神经网络(BPNN),建立了网络初始连接权值和阈值优化的反向传播(BP)神经网络(IPSO-BPNN)气体体积分数反演模型。基于光频梳直接吸收光谱技术测量CO_(2)红外吸收光谱,构建了由训练集、验证集和测试集构成的多体积分数光谱数据集,用于IPSO-BPNN模型的气体体积分数反演性能测试。利用IPSO-BPNN模型对14种体积分数的CO_(2)气体进行了反演,结果表明,与粒子群优化算法优化的BP神经网络(PSO-BPNN)、BPNN、极限学习机(ELM)、支持向量机(SVM)、最大吸光度提取(MAE)法五种气体体积分数的反演方法相比,IPSO-BPNN模型的均方误差最小(1.95×10^(-6)),相对误差绝对值的平均值最低(0.0112),决定系数最大(0.9997)。上述结果验证了IPSO-BPNN模型优异的鲁棒性以及在高准确度分子吸收光谱分析中重要的应用潜力。Objective In absorption spectroscopy for gas sensing,there are problems with low signal-to-noise ratio and low linearity caused by spectral distortion in measured spectra,which makes it difficult for traditional linear analysis methods to achieve high-precision gas concentration inversion.Artificial back propagation neural networks(BPNNs)are suitable for solving nonlinear problems.However,in the optimization problem of multiple local extrema,the final convergence value of the gradient descent algorithm which is usually employed as the training algorithm of artificial BPNNs is related to the initial value.Thus,traditional BPNNs may converge to the local optimal value due to the random initial connection weights and thresholds between neural nodes.Traditional particle swarm optimization(PSO)algorithms are prone to converge to local optima,thereby reducing the optimization effect.Therefore,we adopt an improved particle swarm optimization(IPSO)algorithm to optimize the initial connection weights and thresholds of the artificial BPNN and build an IPSO-BPNN gas concentration inversion model which has been proven to be high-precision and robust.Methods To enhance the local and global search capabilities of PSO algorithms,we conduct improvements on traditional PSO algorithms in terms of evolutionary strategy and parameter settings.Meanwhile,mutation operations are introduced into the PSO algorithm to increase the diversity of particles and enable them to jump out of local optima.In each iteration,each particle has a certain probability of mutation,and the position and velocity of the mutated particles will be randomly initialized again.To better balance the local and global search capabilities of PSO algorithms,we carry out dynamic adjustments to inertia weights,individual learning factors,group learning factors,and maximum speed.Then the IPSO algorithm is constructed.Additionally,we optimize the initial connection weights and thresholds of the BPNN using the IPSO algorithm to enhance the prediction accuracy of the BPNN.

关 键 词：光谱技术 红外气体检测 气体体积分数反演 粒子群优化算法 反向传播神经网络 

分 类 号：TN21[电子电信—物理电子学]