基于选择性更新的在线核极限学习机建模  

On-line Kernel Extreme Learning Machine Modelling based on Selective Update

作　　者：孙朝江[1] 汤健[1] 魏忠军[2] 赵立杰[3] 

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

出　　处：《控制工程》2013年第4期659-662,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)

摘　　要：针对每样本递推更新的在线建模方法计算消耗大、常用的人工智能建模方法学习速度慢的缺点,为能够对软测量模型进行有效更新和提高在线模型的学习速度,提出了一种基于选择性更新的在线核极限学习机(KELM)建模方法。该方法首先采用近似线性依靠(ALD)条件判别新样本与建模样本间的线性独立依靠程度,选择满足设定条件、含有足够新信息的样本对软测量模型进行更新,降低了模型在线学习次数;然后选择学习速度快、泛化性强的KELM方法建立软测量模型,有效地避免了极限学习机(ELM)模型固有的随机性和支持向量机(SVM)模型求解的复杂性;最后将ALD条件和KELM算法有效结合,提高了在线软测量模型的学习速度和预测性能。通过合成数据的仿真实验结果验证了所提方法的有效性。On-line modeling method based on every new sample recursive update needs big compute consumption. The generally used intelligent modeling approaches have low learning speed. In order to effective update soft sensor model and improve on-line model＇s learning speed, a new on-line kernel extreme learning machine （KELM） modeling approach based on selective update is proposed in this paper. The approximate linear dependence （ALD） condition is used to judge the linear independent degree between the new sample and modeling Samples. Only these samples that satisfy the ALD condition and contain interesting new information are used to update the soft sensor model. Thus, the on-line model＇s updating times is reduced. The KELM algorithm with the character of fast learning speed and strong generalization is used to construct soft sensor model, which overcome the randomization of ELM and the complexity solution proeess of SVM. Therefore, with the integration of ALD condition and KELM algorithm, the learning speed and the prediction performance of the on-line model are proposed. The simulation results base on synthetic data verifie the validity of the proposed method.

关 键 词：选择性更新 近似线性依靠(ALD) 核极限学习机(KELM) 在线建模 

分 类 号：TP27[自动化与计算机技术—检测技术与自动化装置]