Integrating Machine Learning in Geotechnical Engineering: A Novel Approach for Railway Track Layer Design Based on Cone Penetration Test Data

Abstract

The cone penetration test (CPT) has emerged as a cost-effective and time-efficient method for assessing soil conditions relevant to railway track infrastructure. The geotechnical data obtained from the CPT serve as crucial input for asset managers in designing optimal sublayers and form layers for track renewal works. To properly assess the condition of soil layers, various soil behavior type charts and machine learning models based on CPT data have been published to help engineers classify soils into groups with similar properties. By understanding the properties of the soils, an optimal substructure can be designed to minimize extensive maintenance and reduce the risk of derailment. However, when analyzing multiple CPTs, the diversity and non-uniformity of subsoil characteristics pose challenges in designing a new optimal trackbed. This study presents an automated approach for recommending thicknesses of sublayers and form layers in railway tracks based on CPT data, employing machine learning algorithms. The proposed approach was tested using CPT data from the Belgian railway network and showed very good agreement with results from traditional soil investigation interpretations and layer design. A Random Forest classifier, fine-tuned through Bayesian optimization with a cross-validation technique and trained on 80% of the datasets, achieved an overall accuracy of 83% on the remaining 20%. Based on these results, we can conclude that the proposed model is highly effective at accurately designing sub-ballast layers using CPT data.