期刊文献+

具有奖罚机制STDP的Spike-CNN模型的机械臂故障分类

Classification of Robot Execution Failures Using a Spike-CNN Model Underlying STDP Learning by Reward and Punishment Mechanisms
下载PDF
导出
摘要 在计算机视觉领域中,卷积神经网络取得了举世瞩目的成就,但其能耗问题一直未能得到很好解决.基于此问题,本文主要研究无监督学习范式下的Spike-CNN分类性能以及计算力.首先,本文设计了一种基于CNN和SNN的混合结构,在层级结构上实现脉冲机制;其次,为减少模型训练时间,本文提出了ReLU-ROC编码方案;最后,为使兴奋性神经元快速做出决策,本文提出了具有决策能力的RP-STDP学习方案:计算每对突触前与突触后兴奋性神经元的相对时间差.实验结果表明:以工业机器人采集到多元时间序列数据解决机械臂不同工作状态的3分类、4分类、5分类问题,在没有引入其他分类器的情况下,本文提出的具有奖罚机制的STDP的Spike-CNN方法平均准确率为LP1(91.07%)、LP2(96.66%)、LP4(93.95%). Although convolutional neural networks(CNNs)has made a great achievement in the computer vision area,its energy consumption problem has never been well solved.Thus,this paper will focus on the classification performance and computational power of the deep spiking convolutional neural networks(DSCNNs).Firstly,based on the CNNs and spiking neural networks(SNNs),a hybrid architecture is proposed in this paper to realize the pulse mechanism in the hierarchical structure.Secondly,the ReLU-ROC encoding scheme is proposed for reducing the training time of the model.Finally,an RP-STDP learning scheme with decision-making capabilities is proposed to get excitatory neurons to make quick decisions,which can figure out the relative time difference between each pair of pre-and post-synaptic excitatory neurons.The result of the experiment shows that without introducing other classifiers,and using multivariate time-series data collected from the industrial robot manipulator to solve classification problems of the three-to five-categories in the manipulator,the average accuracy of the proposed model reaches LP1(91.07%),LP2(96.66%),and LP4(93.95%)by employing the proposed STDP method with reward and punishment mechanism.
作者 刘颖 周恩辉 张薇 王秀青 吕锋 LIU Ying;ZHOU En-hui;ZHANG Wei;WANG Xiu-qing;LV Feng(Hebei Provincial Engineering Research Center for Supply Chain Big Data Analytics&Data Security,Hebei Provincial Key Laboratory of Network&Information Security,College of Computer and Cyber Security,Hebei Normal University,Shijiazhuang 050024,China;School of Mathematical Sciences,Hebei Normal University,Shijiazhuang 050024,China;School of Electronic Information Engineering,Hebei University of Technology,Tianjin 300387,China)
出处 《小型微型计算机系统》 CSCD 北大核心 2022年第6期1285-1292,共8页 Journal of Chinese Computer Systems
基金 国家自然科学基金面上项目(61673160)资助 河北省自然科学基金项目(F2018205102)资助 河北省高等学校科学技术研究项目(ZD2021063)资助 河北师范大学重点基金项目(L2019Z11)资助 河北师范大学2019年研究生创新项目(CXZZSS2019060)资助.
关键词 脉冲神经网络 STDP学习规则 卷积神经网络 机械臂故障诊断 分类 spiking neural networks STDP learn rule convolutional neural networks manipulator failure diagnosis classification
  • 相关文献

参考文献5

二级参考文献30

  • 1王守觉,曹文明.半导体神经计算机的硬件实现及其在连续语音识别中的应用[J].电子学报,2006,34(2):267-271. 被引量:3
  • 2Sasiadek J Z. Sensor fusion[J]. Annual Reviews in Control,2002,26(2) :213-228.
  • 3Luo R, Yih C, Su K. Multisensor fusion and integra- tion: Approaches, applications and future research directions[J]. IEEE Sensors Journal, 2002, 2(2): 107-117.
  • 4王秀青,侯增广,谭民.多传感器数据融合技术在移动机器人中应用的新进展[J].哈尔滨工业大学学报,2006,38(S):1030-1034.
  • 5Berghbfer E, Schulze D, Rauch C. ART-based fusion of multi-modal perception for robots[J]. Neuroeom- paring, 2013, 107: 11-22.
  • 6Liu S H, Huang T S, Yen J Y. Comparison of sen- sor fusion methods for an SMA-based hexapod biomi- metic robot [J]. Robotics and Autonomous Systems, 2010, 58: 737-744.
  • 7Pateraki M, Baltzakis H, Trahanias P. Visual esti- mation of pointed targets for robot guidance via fu- sion of face pose and hand orientation[J]. Computer Vision and Image Understanding, 2014, 120: 1-13.
  • 8Vapnik V. The nature of statistical learning theory[M]. New York: Spring-Verlag, 1995.
  • 9Boser B E, Guyon 1, Vapnik V. A training algorithm for optimal margin classifiers[C] // Proceedings of the Fifth Annual Workshop on Computational Learning Theory. Pittsburgh: ACM Press, 1992: 144-152.
  • 10Kohonen T. The self-organizing map [J]. Proceed- ings of the IEEE, 1990, 78(9) : 1464-1480.

共引文献413

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部