Assessment of railway ballast fouling using GPR and AI-Based learning from LDCP and geoendoscopy data

Ballast is a key components of ballasted railway tracks. Its main function is to guarantee the vertical, lateral and longitudinal stability of the track for the passage of trains. These functions are compromised when ballast begins to deteriorate or becomes fouled, so it is imperative to monitor the rate of fouling index to determine the necessary maintenance or renovation actions. The objective of this study is to characterize the fouling index of the ballast using Ground Penetrating Radar (GPR) measurements with 400 MHz antennas and employing machine learning techniques. The proposed methodology focuses on the parametric development of GPR signals, incorporating both time and frequency domain analyses, along with specific analytical parameters. This comprehensive approach enables a more precise characterization of GPR signals, enhancing their interpretation and analysis in various geotechnical contexts. This analysis will be carried out using a historical database of French railways, consisting of 4700 km of GPR measurements and 12,000 soundings with the light dynamic cone penetration (LDCP)/geoendoscopy test principle. The determination of the target variable, which is the fouling state of the ballast layer, will be performed through the soundings. The results obtained show that the most appropriate model for estimating the fouling index is Random Forest, demonstrating an accuracy of 96% in the training phase. On the other hand, in the model evaluation phase with cases external to the database, the XGBoost model obtained the best result, with a maximum accuracy of 86%.