Hybrid physics machine learning for ultrasonic field guided 3D generation and reconstruction of rail defects

Abstract

Ultrasonic technology is widely used in the field of rail defects detection. 3D reconstruction of rail defects can intuitively restore the 3D size and spatial position of the defects inside the rail. Currently, ultrasound-based 3D reconstruction requires a multi-probe or mechanical scanning platform in a laboratory setting, which is not suitable for the railway environment. In addition, 3D reconstruction requires a large amount of data, making it difficult to collect sufficient ultrasonic 3D defect data for long-distance rail inspections. This paper proposes an ultrasonic field-guided 3D reconstruction method combined with machine learning hybrid physics for rail defects. It combines both sound field GAN model to reconstruct the defect 3D model from the 2D B-scan data. The proposed method can generate a defect cross-sectional image using a deep learning algorithm guided by the acoustic field in the B-scan space, and extract the 3D size information of the defect from the 2D B-scan information by establish a defect echo model. By stablishing a spatial mapping relationship between the B-scan and the rail coordinate system, the position of the defect in the rail coordinate system is obtained. The defect data of standard damage rails are tested. Experiment results indicate that the defects in different parts of the rail can be reconstructed by the proposed method. The average size error rate is 9.56%–21.14 %, and the average height error is 3.458mm–6.353 mm.