Abstract: In order to eliminate the influence of unavoidable outliers in soft sensor sample data on model's performance,a novel adaptive weighted least square support vector machine regression method(AWLS-SVM) was proposed.Firstly,in AWLS-SVM,least square support vector machine regression was employed for the sample data to develop model and obtain the sample datum fitting error.Then,according to the fitting error,the initial error weight was calculated.Secondly,the lever weight was defined according to the space distribution of the sample data.Finally,combining the error weights and lever weights,the adaptive sample weights were obtained via the proposed adaptive iterative method.Thus,the model was developed by AWLS-SVM with the adaptive sample weights.The simulation experiment results show that the outliers influence on the model's performance is eliminated in AWLS-SVM,and that the prediction performance is better than those of LS-SVM and radial basis function network method.Further,AWLS-SVM was applied to develop the soft sensor model for 4-carboxybenzaldehyde concentration in terephthalic acid,which is the most important intermediate product of p-xylene oxidation to terephthalic acid reaction.The satisfied result is obtained.onclick="SelectDisplayDiv('EnDivSummaryMore','EnDivSummaryHuanYuan');HideSpanDiv('EnDivSummaryMoreSummary')">

Key words: Outliers , Weighted , Least square support vector machine , Soft sensor

摘要： 针对软测量建模样本中数据难以避免存在粗差、以及粗差数据对模型性能的影响,提出了一种自适应加权最小二乘支持向量机(AWLS-SVM)回归建模方法。AWLS-SVM基于建模样本数据,根据最小二乘支持向量机回归模型的拟合残差确定各样本的残差权值,根据样本的空间分布确定杠杆权值,进而通过迭代运算,自适应确定各建模样本的权值,在有效减小粗差点对模型性能影响的同时,保留了其所提供的有效信息。仿真实验表明,AWLS-SVM能有效克服粗差样本数据的影响,其模型的预测性能明显优于LS-SVM和径向基函数网络。最后,应用AWLS-SVM建立粗对苯二甲酸中4-CBA含量软测量模型,获得满意结果。

关键词: 粗差 , 加权 ,  最小二乘支持向量机 , 软测量