基于进化算法优化的混合核极限学习机建模  

Kernel Extreme Learning Machine Modelling Approach with Mixed- Kernel Function Based on Evolutionary Algorithm

作　　者：张德全[1] 魏忠军[2] 汤健[3] 赵立杰[4] 

机构地区：[1]中国人民解放军92493部队98分队 [2]中国人民解放军92721部队 [3]中国人民解放军92941部队 [4]沈阳化工大学信息工程学院

出　　处：《控制工程》2013年第6期1127-1130,共4页Control Engineering of China

基　　金：国家自然科学基金项目(61203102);国家自然科学基金重点项目(60534010);中国博士后自然科学基金(2013M532118)

摘　　要：基于核方法的软测量模型的核类型、核参数及学习参数影响模型泛化性能,而且核类型和核参数还与建模数据相关,难以有效选择;常用的基于支持向量机(SVM)的建模算法虽然泛化性好,除了模型的学习参数难以选择外,其学习速度较慢。为解决这些问题,提出了基于进化算法优化的混合核极限学习机建模方法。该方法选用具有较快学习速度和较好泛化性能的核极限学习机(KELM)算法建立软测量模型,其核函数则采用具有局部和全局特性的径向基(RBF)核函数和多项式核函数加权得到的混合核函数。软测量模型的相关参数,即混合核的权系数、核参数及和KLELM模型的惩罚参数通过进化算法同时优化选择。最后采用基于近红外谱(NIR)数据建立的软测量模型验证了所提方法的有效性。The generalization of soft sensor model based on kernel method correlates with the model＇ s kernel type, kernel parameter and learning parameter. Moreover, with different modelling data, the kernel type and the kernel parameters are different. Thus, It is very difficult to effective select the model＇ s learning and kernel parameters. Although the normally used modelling algorithm based on support vector machines （SVM） have good generalization, except kernel type and parameters problem, its learning speed is very low. Aim to these problems, kernel extreme learning machine （KELM） modelling approach with mixed - kernel function based on evolution- ary algorithm is proposed in this paper. The soft sensor model is built based on mix - kernel KELM algorithm with fast learning speed and high generalization. The mix - kernel function is constructed based on radical base function （RBF） with local characteristic and po- ly function with global characteristic. The kernel parameters of RBF and poly function, weighting coefficients, costing parameters of KELNM are optimized selected based on evolutionary algorithm simultaneously. Soft sensor model based on the NIR spectral data is con- structed, which verifies the validity of the proposed method.

关 键 词：软测量 混合核函数 核极限学习机(KELM) 进化算法 

分 类 号：TP29[自动化与计算机技术—检测技术与自动化装置]