Damage Detection of wire rope Broken wire Damage Based on deep Semi- supervised Learning

Abstract

Aiming at the problems such as the existence of a large amount of background information for wire rope image, the lack of efficient and accurate damage area extraction methods. A novel wire rope image preprocessing method is proposed, which can automatically adjust the pose of wire rope, remove the background information of wire rope image, and extract the damage area of wire rope. In response to the problem that deep learning based wire rope damage detection methods require a large amount of labeled data, a convolutional neural network model WR-CNN adapted to the steel wire rope dataset is designed, And combine WR-CNN with the semi-supervised learning example provided by Temporal Ensembling to form WR-TE-CNN. Aiming at the problem that the existing detection methods of broken wire damage do not consider the difference information between different types of damage and the similar information between the same type of damage, the distance loss function is introduced into the loss function of WR-TE-CNN and the characteristic quantity of calculating distance loss function is improved ( Before the improvement it was called WR-TE-DLCNN, after the improvement it was called WR-TE-IDLCNN ). The experimental results show that When the proportion of labeled data is 0.5, the wire breakage recognition rate of WR-TE-IDLCNN exceeds WR-CNN and is 99.04%. WR-TE-IDLCNN can further utilize the information between different types of damage and between the same types on the basis of WR-TE-DLCNN.