基于KELM的连续搅拌反应釜模型辨识  

Model Identification of Continuous Stirred Tank Reactor Based on Extreme Learning Machine with Kernels

作　　者：李军[1] 石青[1] LI Jun SHI Qing(School of Electrical Engineering and Automation, Lanzhou Jiaotong University, Lanzhou 730070, Chin)

机构地区：[1]兰州交通大学自动化与电气工程学院,兰州730070

出　　处：《控制工程》2017年第10期2137-2143,共7页Control Engineering of China

基　　金：国家自然科学基金项目(51467008);光电技术与智能控制教育部重点实验室(兰州交通大学)开放课题(KFKT2016-3)

摘　　要：极限学习机(Extreme learning machine,ELM)是一种单隐层前馈神经网络(SLFNs),它随机选择网络的隐含层节点及其参数,训练时仅需调节输出层权值,因此ELM以极快的学习速度获得良好的推广性。考虑到ELM的特征映射函数未知时,可以将核矩阵引入到ELM中。针对模型未知的强非线性连续搅拌反应釜(Continuous Stirred Tank Reactor,CSTR),提出一种基于核极限学习机(Extreme Learning Machine with Kernels,KELM)的NARX模型辨识方法。以仿真的CSTR过程实例进行辨识实验,建立基于NARX-KELM的辨识模型。实验结果表明,在相同条件下,与带动量因子的BP神经网络、模糊神经网络(FNN)、GAP-RBF、MGAP-RBF神经网络、回声状态网络(ESN)、ELM等方法相比,KELM能够有效地改进辨识精度,而且性能更好,这表明了所提方法的有效性和应用潜力。Extreme learning machine（ELM） is a single hidden layer feedforward networks（SLFNs）, which randomly chooses the input weights and the parameters of hidden nodes and only needs to adjust the weights of output layer during the training process, hence ELM provides better generalization performance at a much faster learning speed. Considering that when the feature mapping function is unknown to users, one can apply the kernel matrix for ELM. A system identification method based on extreme learning machine with kernels（KELM） using the nonlinear autoregressive model with exogenous inputs（NARX） is applied to a strongly nonlinear unknown model of continuous stirred tank reactor（CSTR）. Identification experiments are carried out with the simulation of the CSTR process, and then the identification model based on NARX-KELM is established. Experimental results show that, compared to BP neural network with momentum factor, fuzzy neural network（FNN）, GAP-RBF neural network, MGAP-RBF neural network, echo state network（ESN） and ELM, the employed method can effectively improve the identification accuracy and has considerably better performance under the same condition, which shows the effectiveness and applicability.

关 键 词：核极限学习机 单隐层前馈神经网络 连续搅拌反应釜 NARX模型 辨识 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]