Computer Engineering and Applications ›› 2019, Vol. 55 ›› Issue (10): 193-198.DOI: 10.3778/j.issn.1002-8331.1802-0156

Previous Articles     Next Articles

Catenary Pillar Image Anomaly Detection Combined with SVDD and CNN

WU Jingfeng, JIN Weidong, TANG Peng   

  1. Department of Electrical Engineering, Southwest Jiaotong University, Chengdu 610036, China
  • Online:2019-05-15 Published:2019-05-13


吴镜锋,金炜东,唐  鹏   

  1. 西南交通大学 电气工程学院,成都 610036

Abstract: In the train running system, the accurate access to the catenary pillar number is the prerequisite for the intelligent monitoring of the catenary pillar state, to determine whether the abnormal state of the pillar has become an urgent problem to be solved. Because of the difficulty in obtaining and marking the abnormal data in the actual situation, the traditional pattern recognition method has the limitation of only a large number of normal classes non-equilibrium data processing. Through the semi-supervised learning idea, the improved Lenet-5 neural network and support vector data description are used to detect anomalies with only a few markers, and the depth image features are learned and extracted by the neural network. The support vector data description algorithm is used to train these image features, and obtain the normal domain model of the column number image feature, finally, determine whether the new catenary pillar number is abnormal.

Key words: Convolution Neural Network(CNN), Support Vector Data Description(SVDD), anomaly detection

摘要: 在列车运行系统中,准确获取接触网支柱编号是接触网状态智能监测的前提,判断支柱编号的状态是否异常成为亟待解决的问题。由于实际情况中较难获取和标记异常类数据,针对传统模式识别方法对只存在大量正常类的非平衡数据处理的局限性,基于半监督学习的思想,利用改进的Lenet-5神经网络和支持向量数据描述(Support Vector Data Description,SVDD)相结合的方法,在仅有少量标记异常数据时进行异常检测;由神经网络学习和提取深度图像特征,利用SVDD算法训练这些图像特征,得到了支柱编号图像特征的正常域模型;最后判别新的接触网支柱编号是否异常。

关键词: 卷积神经网络(CNN), 支持向量数据描述(SVDD), 异常检测