模式识别在烟草化学中的过拟合现象研究  

Research on overfitting phenomenon in pattern recognition in tobacco chemistry

作　　者：居雷[1] 朱仲良[1] 陈磊[2] 余苓[2] 

机构地区：[1]同济大学化学系,上海200092 [2]上海烟草集团有限责任公司,上海200082

出　　处：《计算机与应用化学》2013年第5期561-564,共4页Computers and Applied Chemistry

摘　　要：烟草化学建模过程中,化学成分间的多重共线性是常见的问题,使用偏最小二乘判别分析可以有效克服这一问题,但是模型容易出现过拟合的现象,即模型的构建效果好,但是预报能力差。本文选取湖南烟区3个种植大区,即湘南、湘中、湘西北种植的53种不同烟叶,使用偏最小二乘判别分析建立了烟叶主要化学指标与地区大类之间的模型,但是由于变量过多以及噪声的干扰,模型的预报精度差,偏最小二乘判别分析方法出现了过拟合现象,模型的稳健性受到破坏。本文采用了多种模式识别的方法,逐步筛选变量,准确提取出特征变量9个,对产地变量有更好的解释能力,并能够有效地在模型预测的过程中避免变量间的多重共线性以及仪器检测的噪声干扰,建立了有效的烟叶一产地识别模型。模型预报的准确率由未筛选变量之前的75%提高到87.5%,模型的稳健性得到很大提高,改善了模型的过拟合现象。In the tobacco chemical modeling process, the chemical composition of multicollinearity was a common problem. PLS-DA can effectively overcome this problem, but the model was prone to overfitting phenomenon, i.e. the model-building effects well, but the prediction ability is poor. The paper selected 53 kinds of different tobacco leaves, which was planting in three large areas, Xiang Nan, XiangZhong and XiangXibei, using the PLS-DA method to establish the model between the tobacco major chemical indicators and regions. However, with too many variables and noise interferences, as well as the poor prediction accuracy of the model, the PLS-DA method appeared overfitting, and the robustness of the model was damaged. In this paper, a variety of pattern recognition methods were adopted to stepwise select variables, in order to extract 9 characteristic variables, which had a better explanation of the origin variables, and was able to effectively avoid multicollinearity between the variables in the model to predict the process as well as the detection of the instrument noise interference, and establish effective tobacco-origin recognition model. The accuracy of the model was raised to 87.5 % from 75 % that unfiltered variables, the robustness of the model had been greatly enhanced and the model overfitting phenomenon improved as well.

关 键 词：多重共线性 偏最小二乘判别分析 过拟合 

分 类 号：TQ015.91[化学工程] O6-39[理学—化学]