留言板

尊敬的读者、作者、审稿人, 关于本刊的投稿、审稿、编辑和出版的任何问题, 您可以本页添加留言。我们将尽快给您答复。谢谢您的支持!

姓名
邮箱
手机号码
标题
留言内容
验证码

基于机器学习的偏钛酸粒度软测量建模研究

路瑞芳 刘婵 孙伟 吴健春 孙蔷

路瑞芳, 刘婵, 孙伟, 吴健春, 孙蔷. 基于机器学习的偏钛酸粒度软测量建模研究[J]. 钢铁钒钛, 2021, 42(2): 36-42. doi: 10.7513/j.issn.1004-7638.2021.02.007
引用本文: 路瑞芳, 刘婵, 孙伟, 吴健春, 孙蔷. 基于机器学习的偏钛酸粒度软测量建模研究[J]. 钢铁钒钛, 2021, 42(2): 36-42. doi: 10.7513/j.issn.1004-7638.2021.02.007
Lu Ruifang, Liu Chan, Sun Wei, Wu Jianchun, Sun Qiang. Soft sensing modeling of metatitanic acid particle size based on machine learning[J]. IRON STEEL VANADIUM TITANIUM, 2021, 42(2): 36-42. doi: 10.7513/j.issn.1004-7638.2021.02.007
Citation: Lu Ruifang, Liu Chan, Sun Wei, Wu Jianchun, Sun Qiang. Soft sensing modeling of metatitanic acid particle size based on machine learning[J]. IRON STEEL VANADIUM TITANIUM, 2021, 42(2): 36-42. doi: 10.7513/j.issn.1004-7638.2021.02.007

基于机器学习的偏钛酸粒度软测量建模研究

doi: 10.7513/j.issn.1004-7638.2021.02.007
详细信息
    作者简介:

    路瑞芳(1986—),女,高级工程师,主要从事硫酸法钛白绿色、智能制造、工艺优化与开发及钛资源综合利用相关研究,E-mail:lulu195658@163.com

    通讯作者:

    孙蔷,女,副教授,主要从事功能材料的可控合成及应用基础研究、纳米材料的结构和粒径调控机制研究,E-mail:sunq@smm.neu.edu.cn

  • 中图分类号: TF823

Soft sensing modeling of metatitanic acid particle size based on machine learning

  • 摘要: 以某硫酸法钛白生产线的3608条数据为样本,采用皮尔逊系数和统计P值考察了工业钛液的五个属性变量与偏钛酸粒度D50的相关性,并采用LOF算法对数据进行离异值处理以提高数据质量。在此基础上,采用python语言基于Ridge (岭回归)、Lasso (套索回归)、KNN (K-近邻)、ANN (人工神经网络)、Random Forest (随机森林)、SVR (支持向量机)六种模型编写了偏钛酸粒度控制的回归模型算法,六种算法分别应用在整套数据上的回归预测效果差别不大,离异值处理后数据的RMSE都是在0.276上下波动,MAE则是在0.197上下波动,模型效果均优于离异值处理前模型效果。进一步的,通过对ANN、Random Forest、SVR三个模型进行集成学习模型搭建,回归预测效果得到显著提升,RMSE和MAE值分别降至0.245和0.192。
  • 图  1  机器学习回归模型构建过程示意

    Figure  1.  Schematic diagram of construction process of machine learning regression model

    图  2  ANN模型某次测试数据集预测过程

    Figure  2.  Prediction process of ANN model test data set

    图  3  集成学习(Ensemble learning)回归模型构建过程示意

    Figure  3.  Schematic diagram of the construction process of ensemble learning regression model

    表  1  五个属性变量和目标变量的分析指标

    Table  1.   Analysis index of five attribute variables and target variables

    属性\分析指标皮尔逊相关系数统计变量P值
    残余物0.01490.3697
    Ti3+含量−0.01920.2501
    TiO2含量−0.07891.602×10−6
    Fe/TiO20.11859.334×10−13
    F值−0.07921.934×10−6
    下载: 导出CSV

    表  2  离异值处理前后五个属性变量和目标变量的分析指标对比

    Table  2.   Comparison of attribute variables and target variables before and after outlier data processing

    属性\分析指标残余物Ti3+含量TiO2含量Fe/TiO2F值
    皮尔逊相关系数离异值处理前0.0149−0.0192−0.07890.1185−0.0792
    统计变量P值0.36970.25011.602×10−69.334×10−131.934×10−6
    皮尔逊相关系数离异值处理后0.0548−0.0344−0.02310.0560−0.1545
    统计变量P值0.00320.06430.21510.00267.147×10−17
    下载: 导出CSV

    表  3  离异值处理后数据的六种经典机器学习回归模型评估指标

    Table  3.   Evaluation indexes of six classical machine learning regression models after outlier data processing

    RidgeLassoKNNANNRandom ForestSVR
    RMSEMAERMSEMAERMSEMAERMSEMAERMSEMAERMSEMAE
    1 0.2648 0.2005 0.2684 0.2088 0.2671 0.2043 0.2688 0.2062 0.2710 0.2074 0.2752 0.2136
    2 0.3435 0.2010 0.3515 0.2042 0.3542 0.2118 0.3482 0.2081 0.3519 0.2117 0.3515 0.2094
    3 0.2466 0.1889 0.2456 0.1923 0.2565 0.1971 0.2460 0.1904 0.2577 0.1951 0.2504 0.1950
    4 0.3767 0.2279 0.3778 0.2277 0.3814 0.2268 0.3766 0.2270 0.3744 0.2231 0.3823 0.2261
    5 0.4396 0.2409 0.4418 0.2418 0.4439 0.2481 0.4447 0.2447 0.4464 0.2488 0.4498 0.2451
    6 0.2279 0.1755 0.2248 0.1750 0.2335 0.1791 0.2272 0.1757 0.2432 0.1851 0.2365 0.1816
    7 0.3370 0.2087 0.3328 0.2081 0.3354 0.2112 0.3337 0.2068 0.3302 0.2063 0.3424 0.2123
    8 0.3396 0.2230 0.3381 0.2296 0.3412 0.2224 0.3371 0.2239 0.3342 0.2225 0.3477 0.2292
    9 0.3152 0.2199 0.3171 0.2184 0.3080 0.2167 0.3145 0.2190 0.3169 0.2236 0.3243 0.2245
    10 0.2646 0.1968 0.2688 0.1983 0.2571 0.1869 0.2656 0.1970 0.2596 0.1911 0.2626 0.1936
    11 0.2636 0.2008 0.2739 0.2077 0.2552 0.1999 0.2658 0.2017 0.2586 0.2050 0.2597 0.1971
    12 0.2599 0.1863 0.2658 0.1935 0.2579 0.1850 0.2606 0.1868 0.2586 0.1849 0.2587 0.1815
    13 0.2927 0.2158 0.3000 0.2224 0.2917 0.2169 0.2946 0.2173 0.2921 0.2183 0.2945 0.2134
    14 0.2661 0.1960 0.2706 0.1964 0.2704 0.1952 0.2647 0.1933 0.2640 0.1927 0.2660 0.1963
    15 0.2197 0.1757 0.2286 0.1818 0.2239 0.1801 0.2206 0.1754 0.2184 0.1750 0.2144 0.1696
    16 0.2215 0.1836 0.2285 0.1879 0.2226 0.1841 0.2204 0.1826 0.2229 0.1822 0.2138 0.1716
    17 0.2433 0.1902 0.2491 0.1935 0.2423 0.1929 0.2384 0.1889 0.2375 0.1902 0.2450 0.1913
    18 0.2780 0.1866 0.2826 0.1886 0.2810 0.1900 0.2793 0.1861 0.2845 0.1948 0.2811 0.1835
    19 0.2286 0.1741 0.2245 0.1715 0.2230 0.1730 0.2262 0.1731 0.2246 0.1713 0.2219 0.1676
    20 0.2705 0.2086 0.2779 0.2133 0.2770 0.2132 0.2710 0.2087 0.2695 0.2100 0.2726 0.2077
    21 0.2473 0.1803 0.2489 0.1805 0.2484 0.1789 0.2474 0.1796 0.2459 0.1785 0.2447 0.1735
    22 0.2867 0.2148 0.2866 0.2129 0.2879 0.2137 0.2836 0.2114 0.2817 0.2095 0.2809 0.2043
    23 0.3342 0.2224 0.3322 0.2221 0.3262 0.2131 0.3315 0.2213 0.3313 0.2179 0.3358 0.2183
    24 0.2052 0.1733 0.1950 0.1644 0.1933 0.1632 0.2009 0.1697 0.1967 0.1661 0.1905 0.1606
    25 0.2135 0.1704 0.2014 0.1558 0.2120 0.1721 0.2254 0.1827 0.2330 0.1887 0.2051 0.1626
    26 0.3247 0.2135 0.3338 0.2303 0.3255 0.2174 0.3294 0.2217 0.3402 0.2339 0.3316 0.2231
    27 0.1965 0.1527 0.2050 0.1591 0.1960 0.1516 0.1973 0.1520 0.2047 0.1568 0.1947 0.1486
    28 0.1872 0.155 0.1795 0.1467 0.1923 0.1614 0.1928 0.1614 0.1992 0.1646 0.2024 0.1691
    29 0.2395 0.1984 0.2335 0.1964 0.2489 0.2059 0.2363 0.1970 0.2432 0.2026 0.2353 0.1916
    30 0.3144 0.2004 0.3148 0.2024 0.3195 0.2080 0.3149 0.2023 0.3186 0.2055 0.3153 0.1981
    下载: 导出CSV

    表  4  离异值处理前后样本的六种回归模型评估指标(30次预测的平均值)对比

    Table  4.   Comparison of six regression model evaluation indexes of samples before and after outlier data processing

    指标RidgeLassoKNNANNRandom ForestSVR
    离异值处理后RMSE0.27500.27660.27580.27580.27660.2762
    MAE0.19610.19770.19730.19710.19830.1953
    离异值处理前RMSE0.30200.30370.30020.29800.30230.2995
    MAE0.20920.21190.20830.20820.21130.2055
    下载: 导出CSV

    表  5  离异值处理后样本进行集成学习(Ensemble learning)回归模型评估指标

    Table  5.   Ensemble learning regression model evaluation indexes of samples after outlier data processing

    指标12345678910
    RMSE0.22870.19730.23340.25410.28220.23120.32340.25600.35220.2376
    MAE0.18980.16240.18430.17420.20560.18480.24500.17660.28350.1805
    指标11121314151617181920
    RMSE0.27750.20610.27490.20510.19860.18570.24010.26700.25430.2147
    MAE0.22000.15730.20840.16340.16930.15530.18660.19060.19270.1779
    指标21222324252627282930
    RMSE0.22300.24680.35600.20350.24040.27120.17950.21050.23570.3124
    MAE0.17800.19870.26880.16280.18480.21690.14640.18400.19140.2488
    下载: 导出CSV

    表  6  单个回归模型和集成模型学习评估指标(30次预测的平均值)对比

    Table  6.   Comparison of single regression model and ensemble model learning evaluation index

    指标RidgeLassoKNNANNRandom forestSVREnsemble learning (Random forest+SVR+ANN)
    RMSE0.27500.27660.27580.27580.27660.27620.2454
    MAE0.19610.19770.19730.19710.19830.19530.1919
    下载: 导出CSV
  • [1] Santacesatia E. Kinetics of titanium dioxide precipitation by thermal hydrolysis[J]. Journal of Colloid and Interface Science, 1986, 111(1): 45-53.
    [2] Duncan J F, Richards R G. Solution equilibriums, kinetics and mechanism[J]. N. Z J. Sci., 1976,19(2):179−183.
    [3] Chen Hongyun, Jin Bin, Dong Yingjie, et al. Study on determination of the average particle size of TiO2 by use of laser particle sizer[J]. Paint & Coating Industry, 2002,32(6):40−42. (陈洪云, 金斌, 董英杰, 等. 激光粒度仪测定钛白粉平均粒径的研究[J]. 涂料工业, 2002,32(6):40−42. doi: 10.3969/j.issn.0253-4312.2002.06.016
    [4] (张立德, 牟季美. 纳米材料与纳米结构[M]. 北京: 科学出版社, 2002.)

    Zhang Lide, Mou Jimei. Nanometer materials & nanometer structure[M]. Beijing: Science Press, 2002.
    [5] (GB/T 19627—2005. 粒度分析光子相关光谱法[S].)

    GB/T 19627—2005. Particle size analysis- photon correlation spectroscopy[S].
    [6] Tian Congxue, Hu Hongfei, Du Jianqiao, et al. Determination of the particle size distribution of metatitanic acid by photon correlation spectroscopy[J]. Iron Steel Vanadium Titanium, 2010,31(2):15−19. (田从学, 胡鸿飞, 杜剑桥, 等. 用光子相关光谱法测定偏钛酸粒度分布[J]. 钢铁钒钛, 2010,31(2):15−19. doi: 10.7513/j.issn.1004-7638.2010.02.004
    [7] (何桢. 六西格玛管理[M]. 北京: 中国人民大学出版社, 2014.)

    He Zhen. Six Sigma management[M]. Beijing: China Renmin University Press, 2014.
    [8] Kadlec P, Grbic R, Gabrys B. Review of adaptation mechanisms for data-driven soft sensors[J]. Computers & Chemical Engineering, 2011,35(1):1−24.
    [9] Kano M, Ogawa M. The state of the art in chemical process control in Japan: Good practice and questionnaire survey[J]. Journal of Process Control, 2010,20(9):969−982. doi: 10.1016/j.jprocont.2010.06.013
    [10] Zhao F, Lu N, Lu J. Quality control of batch processes using natural gradient based model-free optimization[J]. Industrial & Engineering Chemistry Research, 2014,53(44):17419−17428.
  • 加载中
图(3) / 表(6)
计量
  • 文章访问数:  210
  • HTML全文浏览量:  22
  • PDF下载量:  40
  • 被引次数: 0
出版历程
  • 收稿日期:  2020-07-21
  • 刊出日期:  2021-04-10

目录

    /

    返回文章
    返回