Locking Wire Fracture Classification Algorithm Based on Convolutional Neural Networks

Locking wire is generally used to prevent bolts from loosening. Any locking wire fracture may cause bolt loosening or loss, which will affect the normal operation of key components and threaten the operation safety of EMUs (Electric Multiple Units). However, due to little characteristic information and small feature area of broken locking wire, it is difficult to extract wire fracture features effectively based on traditional convolutional neural networks (CNNs) classification methods, which results in low classification accuracy. In this paper, using the deep learning method, a siamese network based on convolutional neural networks was presented to classify locking wire images through distance metric learning. In addition, a double margin contrastive loss function was proposed to further improve the classification performance of this model. Experiments demonstrate that the model based on the double margin contrastive loss function outperforms traditional classification methods based on cross entropy loss function. Through robustness testing, the proposed method can well overcome false anomalies caused by illumination, oil stains, water stains and component movement and has stronger robustness than other methods. © 2022, Department of Journal of the China Railway Society. All right reserved.