Multiview deep learning based on tensor decomposition and its application in fault detection of overhead contact systems

• DOI: 10.1007/s00371-021-02080-y
• Publication Date: 2021-02-19T20:38:33Z
• Authors: Xuewu Zhang; Yansheng Gong; Chen Qiao; Wenfeng Jing
• Keywords: Feature (linguistics); Tucker Decomposition

Abstract This article mainly focuses on the most common types of high-speed railways malfunctions in overhead contact systems, namely, unstressed droppers, foreign-body invasions, and pole number-plate malfunctions, to establish a deep-network detection model. By fusing feature maps shallow deep layers pretraining network, global local features malfunction area are combined enhance network's ability identifying small objects. Further, order share fully connected network reduce complexity model, Tucker tensor decomposition is used extract from fused-feature map. The operation greatly reduces training time. Through images collected Lanxin railway line, experiments result show that proposed multiview Faster R-CNN based had lower miss probability higher accuracy for three faults. Compared with object-detection methods YOLOv3, SSD, original R-CNN, average improved model this paper decreased by 37.83%, 51.27%, 43.79%, respectively, increased 3.6%, 9.75%, 5.9%, respectively.