Automatic analysis of railway ground penetrating radar: Using signal processing and machine learning approaches to assess railroad track substructure

Abstract

Prior to track rehabilitation works on the French railway network, train-mounted ground penetrating radar (GPR) allows for fast and nondestructive data acquisition. Substructure condition evaluation is of paramount importance as it has an impact on the quality of the planned renewal works. The interpretation of GPR data remains a challenging and time-consuming task which requires expert intervention. This research seeks to enhance and automate the analysis of the GPR data using two different approaches. A signal processing approach based on entropy analysis determines the layer thicknesses, evaluates ballast fouling, and locates areas with water retention. Field comparisons showed concordances with the proposed method. The second approach relies on deep learning to detect mud pumping defects. It showed promising results for the detection of high intensity mud pumping.