改进型PSO-SVM算法对井下多组分气体定量分析的研究

引用本文

段小丽, 王明泉. 改进型PSO-SVM算法对井下多组分气体定量分析的研究[J]. 光谱学与光谱分析, 2019,39(9): 2883-2888.
DUAN Xiao-li, WANG Ming-quan. Quantitative Analysis of Multi-Component Gases in Underground by Improved PSO-SVM Algorithm[J]. Spectroscopy and Spectral Analysis, 2019,39(9): 2883-2888.
Doi:10.3964/j.issn.1000-0593(2019)09-2883-06 复制到剪切板

Permissions

《光谱学与光谱分析》期刊社所有

改进型PSO-SVM算法对井下多组分气体定量分析的研究

段小丽^1,², 王明泉^1,^*

1. 中北大学信息与通信工程学院, 山西太原 030051

2. 晋中学院机械学院, 山西晋中 030600

*通讯联系人 e-mail: wangmq@nuc.edu.cn

作者简介: 段小丽, 女, 1982年生, 中北大学信息与通信工程学院博士研究生 e-mail: duanxiaolinuc@sina.com

收稿日期: 2018-08-22 接受日期: 2018-12-27

基金: 国家重大科学仪器设备开发专项(2013YQ240803), 山西省科技攻关项目(20140321010-02)资助

摘要

对于多组分混合气体定量分析而言, 基于特征光谱的定量分析技术具有不可比拟的优势, 而定量检测效率与精度取决于其采用的光谱数据处理算法的优劣。优化光谱分析算法参数与改进光谱数据处理方式是提高定量分析速度与精度的重要手段。针对井下多组分气体定量分析建模过程中支持向量机(SVM)参数难以确定, 并且随组分数增多而呈指数增长的光谱数据运算量的问题, 提出了一种改进型粒子群优化-支持向量机(PSO-SVM)算法。该算法主要针对多组分气体混合光谱数据量大, 光谱特征信息存在交叠的问题进行研究。通过粒子变异约束PSO算法的收敛路径, 再通过粒子信息共享提高模型优化效率, 最后利用设置动态不敏感区提高模型精度。设计了一种井下多组分气体快速定量检测系统。该系统由CPU控制信号调制模块驱动红外光源, 信号光经过滤尘除湿后的气室照射在探测器上。在压力与温度传感器补偿的基础上, 由信号处理模块将探测得到的光信号量化传入CPU, 最终, 结合改进型PSO-SVM算法实现各组分气体浓度的定量分析。在完成井下实际样气采集、预处理的基础上, 对浓度范围0~10.0%的CH₄和浓度范围0~1.0%的C₂H₆, C₃H₈, SO₂和CO₂共5种组分的混合气体进行了测试, 获得了800组红外光谱数据, 其中训练集400组, 验证集400组。采用SVM建立了多组分气体的定量分析模型, 利用改进型PSO对SVM中的参数进行了优化, 并将获得的最优参数重建了定量分析模型。对采集的红外光谱数据分别由本算法与传统BP网络算法进行各组分气体浓度反演, 实验结果显示, 由于变异粒子对其产生的约束, 使最优值收敛范围变小, 从而提高了收敛速度, 该算法建模时间仅为传统方法的1/10; 由于通过气体光谱特性给出不敏感区, 使特征光谱计算时交叉敏感效率降低, 从而提高了模型预测的准确度, 平均误差约为传统方法的1/5。由此可见, 该算法在全局优化及快速收敛方面得到了显著提升, 改进型PSO结合SVM用于井下多组分气体定量分析是可行的。改进型PSO-SVM算法对于多组分气体混合红外光谱数据的分离具有很好的适用性, 其有一定的实际应用价值。

关键词: 多组分气体定量分析; 粒子群算法; 支持向量机; 红外光谱

中图分类号:O657.3 文献标志码:A

Quantitative Analysis of Multi-Component Gases in Underground by Improved PSO-SVM Algorithm

DUAN Xiao-li^1,², WANG Ming-quan^1,^*

1. School of Information and Communication Engineering, North University of China, Taiyuan 030051, China

2. College of Mechanical Engineering, Jinzhong University, Jinzhong 030600, China

*Corresponding author

Abstract

For quantitative analysis of multi-component gas mixtures, there are incomparable advantages for the quantitative analysis technology of characteristic spectrum. However, the efficiency and accuracy of quantitative detection depends on the capabilities of the spectral data processing algorithms. Optimizing the parameters of spectral analysis algorithms and improving the processing of spectral data are important means to improve the speed and accuracy of quantitative analysis. According to the problem in selecting parameter of support vector machine(SVM) when detecting quantitatively the concentration of multi-component gas in underground mine, an improved Particle Swarm Optimization-Support Vector Machine (PSO-SVM) algorithm was proposed. The algorithm is mainly used to research the problem that the multi-component gas mixed spectrum data is large and the spectral feature information overlaps. The algorithm constrains the convergence path of the PSO algorithm through particle variation, and it improves model optimization efficiency through particle information sharing, and it uses the setting of dynamic insensitive areas to improve model accuracy. A rapid quantitative detection system for multi-component gas was designed in underground mine. The infrared light source was drived by signal modulation module controlled by PC, and the signal light was irradiated on the detector by the air chamber with dust and steam filter. On the basis of the pressure and temperature sensor compensation, the detected optical signals were transmitted to the CPU by signal processing module. Finally, quantitative analysis of the gas concentrations for the various components was achieved by the improved PSO-SVM algorithm. On the basis of the actual sample gas collection and pretreatment in the underground, five kinds of gas components of CH₄, C₂H₆, C₃H₈, SO₂ and CO₂ were tested. The concentration range of CH₄ was 0~10%, and the concentration range of other gases was 0~10%. Infrared spectral data of these five gases were collected with Fourier infrared spectrometer. 800 groups of these gases were divided into 400 groups for calibration set and 400 groups for validation set. The quantitative analysis model of multi-component gas was established by SVM. The parameters of SVM were optimized by improved PSO, and the quantitative parameters were reconstructed by the obtained optimal parameters. The infrared spectral data collected by the algorithm and the traditional BP network algorithm were used to invert the gas concentration of each component. The experimental results show that the convergence range of the optimal value is reduced due to the constraint of the mutated particle, which improves the convergence speed. The modeling time of the algorithm is only 1/10 of that of the traditional method; Since the insensitive area is given by the spectral characteristics of the gas, the cross-sensitivity effect of the characteristic spectrum is reduced, which improves the prediction accuracy of the model. It improves the accuracy of model predictions, with an average error of about 1/5 of traditional methods. It is feasible to use improved PSO combined with SVM for quantitative analysis of multi-component gas in underground. The improved PSO-SVM algorithm has good applicability for the separation of multi-component gas mixed infrared spectral data, and it has certain practical application value.

Keyword: Multicomponent gas quantitative analysis; Particle swarm optimization(PSO); Support vector machine (SVM); Infrared spectroscopy

文章图片

引言

煤炭是我国能源的主要组成部分, 而国内矿井类型复杂, 以地下矿、高浓度瓦斯矿居多, 也是我国矿难多发的主要因素之一^[1]。故研究能够快速定量井下各类有毒、有害气体对保障井下人员生命安全和煤炭工业发展具有重要意义。

多组分混合气体的定量浓度检测主要方法有电化学传感器法、气相色谱法以及光学检测法^{[2, 3]}。电化学传感器法若拟实现多组分探测, 需要采用电解质, 在井下应用非常不便, 而仅能测试瓦斯的单一传感器虽在矿井中曾被大量应用, 但其由于存在探测器“ 中毒” 问题, 每月至少需要更换两次, 操作复杂, 同时还不适用于高瓦斯矿; 气相色谱法通过色谱分离实现, 该过程时间长, 系统结构复杂, 不适合工作人员在井下日常工作使用; 光学检测法利用各气体特征光谱经算法分离实现定量, 其结构紧凑、物理测量无辅材, 精度较低。

为了提高光学检测法的精度, 通常采用人工神经网络^{[4, 5]}(artificial neural network, ANN)、支持向量机^{[6, 7, 8, 9]}(support vector machine, SVM)、蚁群算法等。刘全金^[10]等利用RBF-ANN实现了传感器阵列采集的方式, 大幅降低了运算数据量, 经实验验证了其方法的可行性; 王智文^[11]等曾对多组分气体建模, 利用主成分筛选特征标志, 从而优化ANN网络, 实现了检测时间从4 000~8 s的提高; 曲健^[12]等通过优化SVM参数, 对CO₂进行定量分析, 其分类收敛精度提升了1个数量级。

本文在改进了粒子群优化(particle swarm optimization, PSO)^{[13, 14, 15]}算法的基础上, 完成了支持向量机参数的确定, 实现了对多组分气体混合光谱数据的有效降维。对瓦斯浓度在1.0%~10.0%范围内多组分混合气体进行了测试, 并与传统人工神经网络的测试结果进行了比较。

1 井下多组分气体检测系统

多组分气体快速检测系统如图1所示, CPU控制信号调制模块发出控制信号, 控制信号使驱动电路产生需要的调制波形。红外光源受驱动电路控制产生相应的初始光信号, 经过检测气室后由探测器采集出射光信号。气室待测气体首先由入气口进入过滤池, 其采用滤网及分子筛完成对煤尘、粉尘以及水蒸气的滤除, 再进入气室, 待气体稳定后测试完毕再由出气口排出。出射光信号经信号处理电路, 处理电路根据温度传感器和压强传感器的参数值完成对出射光信号的数据处理, 并通过A/D转换电路传输给CPU。最终, CPU实现对多组分气体的混合光谱数据的分离与定量反演, 研究中所设计的改进型PSO-SVM算法将对光谱反演过程进行优化, 利用各待测气体光谱特征构造数据联系, 从而通过PSO实现SVM参数的确定, 提高光谱数据中特征数据分类的准确度与收敛速度。

	Figure Option View Download New Window
	图1 多组分气体快速检测系统Fig.1 Multi-component gas rapid detection system

2 改进型PSO-SVM算法

2.1 SVM

SVM常被应用于分类与回归问题的解决, 适用于多组分气体光谱数据的分离。该算法具有较高的适用性和鲁棒性, 针对多组分分离的情况是比较合适的。设训练数据(x_i, y_i), i=1, 2, 3, …, n(n表示样本个数), x_i和y_i均属于R, x_i为样本输入, y_i为期望输出, 其拟合函数有

$f (x) = ωx + b$ (1)

式中, ω 是空间超平面, b偏移量。

根据结构风险最小准则, 该问题可被优化为

$\begin{array}{l} \min \frac{1}{2} ‖ ω ‖^{2} + C \overset{l}{\sum_{i = 1}} (ξ_{i} + ξ_{i}^{*}) \\ s.t. \{\begin{array}{l} y_{i} - ω x_{i} - b \leq ε + ξ_{i}, - y_{i} + ω x_{i} + b \leq ε + ξ_{i}^{*} \\ ξ_{i} + ξ_{i}^{*} \geq 0 (i = 1, 2, \dots, l) \end{array} \end{array}$ (2)

其中, C表示惩罚因子, 用于调节对样本误差超限的惩罚程度; ξ 表示松弛因子; ε 表示不灵敏参数, 用于体现系统对误差的容忍程度。

由拉格朗日乘子获得上式的对偶表述有

$\begin{array}{l} \max [- \frac{1}{2} \overset{l}{\sum_{i, j = 1}} (α_{i} - α_{i}^{*}) (α_{j} - α_{j}^{*}) (x_{i} x_{j}) - \\ \overset{l}{\sum_{i = 1}} α_{i} (ε - y_{i}) - \overset{l}{\sum_{i = 1}} α_{i}^{*} (ε + y_{i})] \\ s.t. \{\begin{array}{l} \sum (α_{i} - α_{i}^{*}) = 0 \\ 0 \leq α_{i}, α_{i}^{*} \leq C (i = 1, 2, \dots, l) \end{array} \end{array}$ (3)

其中, α _i和 $α_{i}^{*}$ 可通过二次规划求解。再由KKT条件计算可得输出有

$\{\begin{array}{l} f (x) = \overset{l}{\sum_{i = 1}} (α_{i} - α_{i}^{*}) (x_{i} x) + b \\ b = \frac{1}{N_{s}} \sum_{S} [t_{s} - ε - \overset{l}{\sum_{m \in S}} (α_{m} - α_{m}^{*}) (x_{m} x_{S})] \end{array}$ (4)

2.2 PSO

PSO的核心思想是通过样本个体间的信息交换寻找最优解的, 其公式为

$\{\begin{array}{l} V_{ij}^{k + 1} = σ V_{ij}^{k} + c_{1} r_{1} (P_{ij}^{k} - X_{ij}^{k}) + c_{2} r_{2} (P_{gj}^{k} - X_{ij}^{k}) \\ X_{ij}^{k + 1} = X_{ij}^{k} + V_{ij}^{k + 1} \end{array}$ (5)

其中, i∈ N(N为粒子总数), j表示维度, σ 为惯性权重; D表示维数; k和k+1分别表示当前和下一次迭代数; V_ij表示粒子速度; X_ij表示粒子位置; P_ij和P_gj分别表示个体和群的极值, c₁和c₂表示加速度因子。虽然PSO收敛快, 适用于多组分光谱分离, 但其存在早熟收敛、后期迭代效率低的问题, 故本工作在改进了PSO的基础上与SVM结合实现对多组分气体光谱的反演。

2.3 改进型PSO-SVM

算法改进主要从两个方面实现, 一是针对PSO的改进, 通过拟测气体光谱特性设定限制条件使相应粒子变异, 从而

引导种群达到最优收敛效果; 二是针对SVM的改进, 在支持向量回归计算中设置一个动态变化不敏感区ε (y_i), 即采用ε (y_i)取代原算法中的ε 。在此改进思路的基础上, 式(4)被替换后化简可得

$\{\begin{array}{l} f (x) = \overset{l}{\sum_{i = 1}} (α_{i} - α_{i}^{*}) (x_{i} x) + b \\ b = \frac{1}{N_{s}} \sum_{S} [t_{s} - ε (y_{i}) - \overset{l}{\sum_{m \in S}} (α_{m} - α_{m}^{*}) (x_{m} x_{S})] \end{array}$ (6)

ε (y_i)的引入可抑制模型过学习的效果, 提高模型预测精度和鲁棒性。由此完成优化的具体流程如图2所示。

如图2所示, 改进后的PSO-SVM算法在设置变异粒子和动态不敏感区后再完成分类计算。优化步骤如下: (1) 取光谱测试数据总量1/3作为训练样本, 其他作为测试样本; (2) 初始化C和ξ , 构建SVM模型; (3) 设置粒子群维度为2, 每个维度选取20个粒子, 迭代数设为200, 根据设定范围优化V_ij和X_ij, 并计算适应度函数的均方误差; (4) 设置变异粒子约束收敛方向, 设置ε (y_i)提高学习效率; (5) 比较每个粒子的适应度, 计算每个粒子的局部和全局最优值; (6) 根据(4)和(5)的结果调整V_ijX_ij; (7) 若不满足结束条件重新并学习至符合要求。

	Figure Option View Download New Window
	图2 改进型PSO-SVM算法流程图Fig.2 Flowchart of the improved PSO-SVM algorithm

3 实验部分

采用Nicolet iS5型红外光谱仪, 扫描范围: 350~7 800 cm^-1, 分辨率为0.8 cm^-1。被测气体在井下实际采集, 由于井下气体中包含大量煤尘、粉尘和水蒸气, 所以首先将待测气体引入过滤池进行去尘去湿处理, 然后再导入气室进行检测。测试主要针对CH₄(0~10%), C₂H₆(0~1%), C₃H₈(0~1%), SO₂(0~1%), CO₂(0~1%)进行。混合气体共计800组样本数据, 取400组为样本数据, 其他为检验样本。对样本进行基线校正、降噪、归一化从而完成光谱数据的预处理。

为了使模型具有快速、高效的预测能力, 对于不同组分气体进行了独立建模处理, 对于主要检测气体CH₄而言, 粒子数群维数设置为2, 每维粒子数为20, 优化迭代200代, 惯性权重为0.9, 终止值为0.4, 完成逐代计算。根据上述模型优化后得到其粒子群优化误差曲线。如图3所示, 改进型PSO算法仅需43步可达全局最优解, 具有很高的收敛性, 200代的优化时间约为4 012 s, c=58.46, σ =4.67。将该优化结果带入SVM, 得到测试样本的预测均方差是4.8× 10^-8。同时, 将同组光谱数据采用BP网络进行光谱分离并反演各组分浓度(仅选测试范围内典型的三个值)。对井下五种主要待测气体的测试结果见表1。

	Figure Option View Download New Window
	图3 优化后粒子群优化误差曲线Fig.3 Error curve of particle swarm optimization

表1 改进型PSO-SVM和传统BP网络的测试结果对比 Table 1 Comparison of test results between improved PSO-SVM and traditional BP network

由于遍历算法过程范围大, 当步长过小时优化时间会大幅增加, 而步长过大时, 容易造成局部最优。将(c, σ )从(0, 0.01)开始以(100, 1)为步长遍历至(10^-4, 20), 由传感模型反馈完成参数的优化。经遍历优化构建五种气体模型的部分测试数据仿真结果如表2所示, 实测数据采用测量精度更高的气体传感器实现。

表2 基于改进型PSO-SVM的测试结果 Table 2 Test results based on improved PSO-SVM

通过表1、表2的数据比较可知, 采用传统BP网络的方法总计用时72 072 s, 而改进型PSO-SVM算法总用时为8 706 s, 提高了仅一个数量级, 相比本算法效率更高。而对整个测试范围内几个典型值进行比较可知, 基于传统BP网络的测试误差的总平均值为7.00%, 而基于改进型PSO-SVM算法的测试误差的总平均值为1.269%, 由此可见, 该算法的误差也相比传统处理方法要好。该多组分混合气体的定量分析模型具有较好的预测能力, 改进型PSO-SVM算法可以快速、准确地完成多组分气体的定量分析, 通过信息共享和变异粒子约束大大降低了寻优时间, 提升了收敛速度。

4 结论

研究了一种用于多组分气体定量分析的光谱数据处理算法, 该改进型PSO-SVM算法能够通过变异粒子约束、设置动态不敏感区间等方法提升预测精度和缩短收敛周期。与传统BP网络优化算法相比, 该算法优化速度更快、模型预测精度更高, 建模效率得到了大幅提升。验证了该模型符合实际测试要求, 在井下多组分气体定量分析具有一定的实际应用价值。

参考文献

文献列表

[1]	Wu Hongpeng, Dong Lei, Zheng Huadan, et al. Nature Communication, 2017, 8: 15331. [本文引用:1]
[2]	LI Jing-yu, ZHANG Ying-wei, LI Wen-zhe, et al(李晶宇, 张影微, 李文哲, 等). Agricultural Mechanization(农机化研究), 2015, 6: 255. [本文引用:1]
[3]	SUN Ming-guo, MA Hong-liang, LIU Qiang, et al(孙明国, 马宏亮, 刘强, 等). Acta Optica Sinica(光学学报), 2018, 38(5): 0530003-1. [本文引用:1]
[4]	PENG Peng, DI Chang-an, ZHANG Yong-jian(彭澎, 狄长安, 张永建, 等). Acta Armamentarii(兵工学报). 2018, 39(7): 1397. [本文引用:1]
[5]	Cao Y, Sanchez N, Griffin R J, et al. Optics Express, 2015, 23: 2121. [本文引用:1]
[6]	LI Nan, ZHU Xiu-fang, PAN Yao-zhong, et al(李楠, 朱秀芳, 潘耀忠, 等). Journal of Remote Sensing(遥感学报), 2018, 22(4): 560. [本文引用:1]
[7]	ZHA Shen-long, LIU Kun, TAN Tu, et al(查申龙, 刘锟, 谈图, 等). Acta Photonica Sinica(光子学报). 2017, 46(6): 0612002-1. [本文引用:1]
[8]	PANG Tao, WANG Yu, XIA Hua, et al(庞涛, 王煜, 夏滑, 等). Acta Photonica Sinica(光子学报), 2016, 45(9): 0912003. [本文引用:1]
[9]	Yin Xukun, Dong lei, Wu Hongpeng, et al. Sensors and Actuators B, 2017, 247: 329. [本文引用:1]
[10]	LIU Quan-jin, ZHAO Zhi-min, LI Ying-xin, et al(刘全金, 赵志敏, 李颖新). Acta Electronica Sinica(电子学报), 2016, 44(4): 995. [本文引用:1]
[11]	WANG Zhi-wen, ZHANG Ji-long, WANG Zhi-bin, et al(王智文, 张记龙, 王志斌, 等). Transducer and Microsystem Technologies(传感器与微系统), 2011, 30(3): 67. [本文引用:1]
[12]	QU Jian, CHEN Hong-yan, LIU Wen-zhen, et al(曲健, 陈红岩, 刘文贞, 等). Chinese Journal of Sensors and Actuators(传感技术学报), 2015, 28(5): 774. [本文引用:1]
[13]	ZENG Yi-shuai, YANG You-liang, MA Cui-hong(曾怡帅, 杨友良, 马翠红). Chinese Journal of Luminescence(发光学报), 2016, 37(7): 859. [本文引用:1]
[14]	GUO Tian-tai, HONG Bo, PAN Zeng-rong, et al(郭天太, 洪博, 潘增荣, 等). Infrared and Laser Engineering(红外与激光工程), 2016, 45(6): 0617011-1. [本文引用:1]
[15]	Li Duan, Zhang Hongxin, Muhammad, et al. The Journal of China Universities of Posts and Telecommunications, 2017, 24(3): 83. [本文引用:1]

2017

0.0

... 引言煤炭是我国能源的主要组成部分, 而国内矿井类型复杂, 以地下矿、高浓度瓦斯矿居多, 也是我国矿难多发的主要因素之一^[1] ...

2015

0.0

... 多组分混合气体的定量浓度检测主要方法有电化学传感器法、气相色谱法以及光学检测法^[2,3] ...

2018

0.0

... 多组分混合气体的定量浓度检测主要方法有电化学传感器法、气相色谱法以及光学检测法^[2,3] ...

2018

0.0

... 为了提高光学检测法的精度, 通常采用人工神经网络^[4,5](artificial neural network, ANN)、支持向量机^[6,7,8,9](support vector machine, SVM)、蚁群算法等 ...

2015

0.0

... 为了提高光学检测法的精度, 通常采用人工神经网络^[4,5](artificial neural network, ANN)、支持向量机^[6,7,8,9](support vector machine, SVM)、蚁群算法等 ...

2018

0.0

... 为了提高光学检测法的精度, 通常采用人工神经网络^[4,5](artificial neural network, ANN)、支持向量机^[6,7,8,9](support vector machine, SVM)、蚁群算法等 ...

2017

0.0

... 为了提高光学检测法的精度, 通常采用人工神经网络^[4,5](artificial neural network, ANN)、支持向量机^[6,7,8,9](support vector machine, SVM)、蚁群算法等 ...

2016

0.0

PANG

Tao

, WANG

, XIA

Hua

, et al(庞涛, 王煜, 夏滑, 等). Acta Photonica Sinica(光子学报), 2016, 45(9): 0912003.

According to the needs of full range and high precision measurements of methane concentration in coal mine safety production monitoring system, a full-scale integrated laser methane sensor was designed based on a Tunable Diode Laser Absorption Spectroscopy (TDLAS). A 1 653.72 nm Distributed Feedback Laser(DFB) semiconductor laser was employed as the system light source. Laser driving, temperature controlling, signal modulation and demodulation, and concentration calculation were realized by a single board circuit. To meet a high measurement precision and a large dynamic measurement range simultaneously, the wavelength modulation technique was used to detect the methane at a low concentration. When the gas concentration is higher than the threshold set in the system, the direct absorption detection technique was switched to automatically. Experimental results show that, the measurement error is less than ±0.06% in the range of 0~5% and less than ±6% of the actual value in the range of 5~100%, and the response time is less than 15 s, which meets the requirements of the mine measurement in practice.

1. Anhui Provincial Key Laboratory of Photonic Devices and Materials, Anhui Institute of Optics & Fine Mechanics, Chinese Academy of Sciences, Hefei 230031, China; 2. University of Science and Technology of China, Hefei 230026, China

根据煤矿安全生产监控系统对测量甲烷浓度全量程高准确度的需要，基于可调谐半导体激光吸收光谱技术，设计了一种全量程一体化激光甲烷传感器.采用1 653.72 nm 分布式反馈半导体激光器作为系统光源，单板电路实现激光器驱动、温度控制、信号调制与解调、浓度反演.为兼顾高测量准确度和大动态测量范围，系统在低浓度时利用波长调制技术进行甲烷浓度在线检测；当气体浓度大于阈值时，自动切换到直接吸收检测技术.实验结果表明，该传感器在浓度范围为0～5%内误差小于±0.06%，在浓度范围为5～100%内误差小于真值的±6%，响应时间约为15 s，满足矿井实际测量需要.

... 为了提高光学检测法的精度, 通常采用人工神经网络^[4,5](artificial neural network, ANN)、支持向量机^[6,7,8,9](support vector machine, SVM)、蚁群算法等 ...

2017

0.0

... 为了提高光学检测法的精度, 通常采用人工神经网络^[4,5](artificial neural network, ANN)、支持向量机^[6,7,8,9](support vector machine, SVM)、蚁群算法等 ...

2016

0.0

LIU

Quan-jin

, ZHAO

Zhi-min

, LI

Ying-xin

, et al(刘全金, 赵志敏, 李颖新). Acta Electronica Sinica(电子学报), 2016, 44(4): 995.

A new PSO algorithm is proposed in this paper for feature selection.Distances within the same class and between different classes are used as the index for distinguishing different classes,and thus can be used to construct the fitness function of particles in PSO.The direction of particles for searching optimal features which can result in close intra-class distance and far inter-class distance is determined by the current best solution of the particle and the optimal individual in particle neighborhood,weighted by the fitness function.Meanwhile,the PSO algorithm is aggregated by the weighted voting method to improve its stability and robustness.The experiment results on 5 high dimensional datasets show that the ensemble PSO algorithm is effective and feasible.

1. College of Science, Nanjing University of Aeronautics and Astronautics, Nanjing, Jiangsu 210016, China; 2. Department of Physics, Anqing Normal College, Anqing, Anhui 246011, China; 3. Beijing Jingwei Textile Machinery New Technology Co Ltd, Beijing 100176, China; 4. Beijing Light Industry and Textile Machinery Engineering Research Center for Machine Vision, Beijing 100176, China; 5. Jiangsu Institute of Standardization, Nanjing, Jiangsu 210029, China

提出了一种新的PSO特征选取方法.以粒子对应特征组合的同类近邻样本和异类近邻样本间的距离关系作为类别可分性和粒子适应度函数.以适应度函数加权的群体历史最佳、粒子历史最佳和粒子邻域内最佳个体信息共同指导粒子运动方向,搜索类内紧密、类间分离的最佳特征组合;同时,利用加权集成方法对PSO特征选取方法进行集成,以提高特征选取方法的稳定性和鲁棒性.在5个高维数据集上的特征选取实验结果表明集成PSO特征选取方法的有效性和可行性.

... 刘全金^[10]等利用RBF-ANN实现了传感器阵列采集的方式, 大幅降低了运算数据量, 经实验验证了其方法的可行性 ...

2011

0.0

... 王智文^[11]等曾对多组分气体建模, 利用主成分筛选特征标志, 从而优化ANN网络, 实现了检测时间从4 000~8 s的提高 ...

2015

0.0

... 曲健^[12]等通过优化SVM参数, 对CO₂进行定量分析, 其分类收敛精度提升了1个数量级 ...

2016

0.0

... 本文在改进了粒子群优化(particle swarm optimization, PSO)^[13,14,15]算法的基础上, 完成了支持向量机参数的确定, 实现了对多组分气体混合光谱数据的有效降维 ...

2016

0.0

2017

0.0