基于改进特征选择RF算法的红外光谱建模方法  

Infrared spectrum modeling method based on RF algorithm of improved feature selection

作　　者：王凯[1] 王菊香[2] 邢志娜[2] 韩晓 Wang Kai;Wang Juxiang;Xing Zhina;Han Xiao(Graduate Students’Brigade,Naval Aeronautical&Astronautical University,Yantai Shandong 264001,China;Dept.of Airborne Engineering,Naval Aeronautical&Astronautical University,Yantai Shandong 264001,China;;8357 Military Representative Office of Navy in Tianjin Institute,Tianjin 300308,China)

机构地区：[1]海军航空工程学院,研究生管理大队,山东烟台264001 [2]海军航空工程学院,飞行器工程系,山东烟台264001 [3]海军驻天津8357所军事代表室,天津300308

出　　处：《计算机应用研究》2018年第10期3000-3002,共3页Application Research of Computers

摘　　要：针对线性红外光谱建模方法会导致模型的泛化能力受限,而非线性方法随着光谱特征数目增多会导致模型预测准确度下降的问题,对随机森林(RF)标准算法的特征选择方法进行改进。根据红外光谱与待测组分的相关性对光谱特征重要性进行度量,采用K-均值聚类算法划分光谱特征区,按特定比例从各特征区采样并建立决策树,最终构造随机森林。实验结果表明,改进算法建立较少的决策树就可以达到较高的准确度,将其与PLS、SVM和标准RF算法进行比较,证明改进RF算法能够提高红外光谱模型的准确度,同时降低模型的复杂度。Linear infrared spectrum modeling method can lead to limited generalization ability.Nonlinear method can cause a decline in prediction accuracy with the increased feature number.So this paper studied an improved random forests algorithm based on the feature selection.Firstly,it measured the correlation as importance of the spectral characteristics and classified the weights by the method of K-means clustering.Then,it selected the different characteristics from the different classes according to the specific proportion.Finally,it created the decision trees and the random forests.Experimental results show that the improved algorithm can build less decision trees and achieve higher accuracy.The improved algorithm improves the model’s accuracy and reduces the model’s complexity compared with PLS、SVM and standard RF algorithm.

关 键 词：特征选择 随机森林 比例采样 红外光谱 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]