引用本文:于洋,邓瑞,余刚,庞新富.奇异值分解下在线鲁棒正则化随机网络[J].控制理论与应用,2024,41(3):407~415.[点击复制]
YU Yang,DENG Rui,YU Gang,PANG Xin-fu.Online robust regularized random networks under singular value decomposition[J].Control Theory and Technology,2024,41(3):407~415.[点击复制]
奇异值分解下在线鲁棒正则化随机网络
Online robust regularized random networks under singular value decomposition
摘要点击 2358  全文点击 76  投稿时间：2022-09-29  修订日期：2024-04-09
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2023.20857
  2024,41(3):407-415
中文关键词  随机权神经网络  正则化  奇异值分解  磨矿过程  磨矿粒度
英文关键词  random vector functional link network  regularization  singular value decomposition  grinding process  grinding particle size
基金项目  矿冶过程自动控制技术国家重点实验室、矿冶过程自动控制技术北京市重点实验室项目(BGRIMM–KZSKL–2021–03), 国家自然科学基金项目 (61773269), 辽宁省自然科学基金项目(2021–BS–189)资助.
作者单位E-mail
于洋* 沈阳航空航天大学 ergejiayu@126.com 
邓瑞 沈阳航空航天大学  
余刚 矿冶过程自动控制技术国家重点实验室  
庞新富 沈阳工程学院自动化学院  
中文摘要
      在线鲁棒随机权神经网络(OR-RVFLN)具有较好的逼近性、较快的收敛速度、较高的鲁棒性能以及较小的 存储空间. 但是, OR-RVFLN 算法计算过程中会产生矩阵的不适定问题, 使得隐含层输出矩阵的精度较低. 针对这 个问题, 本文提出了奇异值分解下在线鲁棒正则化随机网络(SVD-OR-RRVFLN). 该算法在OR-RVFLN 算法的基础 上, 将正则化项引入到权值的估计中, 并且对隐含层输出矩阵进行奇异值分解; 同时采用核密度估计(KDE) 法, 对 整个SVD-OR-RRVFLN网络的权值矩阵进行更新, 并分析了所提算法的必要性和收敛性. 最后, 将所提的方法应用 于Benchmark数据集和磨矿粒度的指标预测中, 实验结果证实了该算法不仅可以有效地提高模型的预测精度和鲁 棒性能, 而且具有更快的训练速度.
英文摘要
      Online robust random vector functional link network (OR-RVFLN) has better approximation, faster convergence speed, higher robustness and smaller storage space. However, the OR-RVFLN algorithm can cause the ill-posed problem of the matrix in the calculation process, which makes the low precision of the hidden layer output matrix. To solve this problem, based on the singular value decomposition approach, this paper proposes the online robust regularized random vector functional link network (SVD-OR-RRVFLN). Firstly, the SVD-OR-RRVFLN introduces the regularization term into the OR-RVFLN algorithm, and the singular value decomposition approach is used for the hidden layer output matrix. Further, the kernel density estimation (KDE) method is used to update the matrix weight. Secondly, the necessity and convergence of the proposed algorithm are analyzed. Finally, the proposed method is applied to Benchmark data set and the index prediction of grinding particle size. The experimental results show that the proposed algorithm can not only effectively improve the prediction accuracy and robustness of the model, but also have faster training speed.