Training sample selection is a key step to establish soft sensor models. According to traditional selection methods,the information of dependent variables cannot be used well. In addition, it is difficult to evaluate the impact of training samples on soft sensor models. To handle these issues,in the present paper,a new training sample selection strategy based on intelligent optimization algorithms was proposed. The objective function of our proposed method was combined with a loss function and a compression-ratio operator of training samples which can tune the direction of searching. The advantage is that it can make full use of dependent variables and the effectiveness of training samples in a certain soft sensor model. As a result,it can optimize the structure of selected training samples. The performance of our proposed methods was demonstrated by its practical applications on research octane number(RON)- near infrared(NIR)spectrum data set,which were selected from gasoline blending process. Besides,a diesel NIR spectrum benchmark data set were also provided. Based on these data sets,we analysis and discuss the impact of training samples on soft sensor model,some useful results were gained. Compared with traditional partial least squares method(PLS),locally weighted PLS,and several other modeling strategies,the proposed method was found to achieve good accuracy and robustness,it is very suitable for industrial application.%根据目标工况合理选择训练样本,是建立软测量模型的关键.传统的训练集样本选择方法难以充分利用因变量信息,而且难以综合考虑样本对模型的影响.为了解决上述问题,本文提出一种基于智能优化算法的训练集样本选择模型,定义了损失函数和样本压缩率,通过权重因子将二者融合为多目标适应度函数,可调整优化算法的寻优方向,使算法能够同时对建模样本组合结构与样本数量寻优,因此极大提高了所选建模样本的质量.为了验证方法的有效性,以汽油调和过程中采集的汽油近红外光谱-研究法辛烷值数据以及柴油近红外基准数据为例,与偏最小二乘、局部权重偏最小二乘等多种方法进行了比较研究,并分析了建模样本对软测量模型的影响.结果表明,本文方法在大规模降低训练集样本规模的同时能够保证软测量模型的精度和泛化性,非常适合工业应用.
展开▼
