A siamese network-based non-contact measurement method for railway catenary uplift trained in a free vibration test

Abstract

The vibration of the catenary that is initiated by the passing pantograph has a direct influence on the pantograph–catenary contact performance. Monitoring the dynamic uplift of the catenary can help inspectors to evaluate the railway operation conditions and investigate the mechanism of pantograph–catenary interaction further. In this paper, a non-contact measurement method based on the deep leaning method is proposed to monitor the real-time vibration of the catenary. The field test for the catenary free vibration is designed to validate the method’s performance. The measurement method is developed based on the fully convolutional Siamese neural network, and the contact wire is taken as the tracking target. To reduce the recognition errors caused by the changes in the shape and grayscale of the moving contact wire in images, the class-agnostic binary segmentation mask is adopted. A developed down-sampling block is used in the neural network to reduce the image feature loss, which effectively enhances the recognition effect for the catenary vibration under variable lighting conditions. To validate the performance of the proposed measurement method, a series of field tests of catenary free vibration were conducted under various lighting conditions and different excitations, and the recognition results were compared with traditional target tracking methods. The results show that the proposed method performs well for catenary vibration identification in the field test. Additionally, the uplift data extracted from the identified images agree with the numerical results, and also help to further investigate the wave propagation and damping characteristics in the catenary structure.