Complas 2023

COMPLAS 2023

Use of Neural Networks to Classify the Conformity of Wheel-Rail Contact Based on Families of Profiles

Moreira Lopes, Modesto Valci (Polytechnic School, University of São Paulo)
Alves de Lima, Vinicius (Polytechnic School, University of São Paulo)
Machado, Izabel Fernanda (Polytechnic School, University of São Paulo)
Martins Souza, Roberto (Polytechnic School, University of São Paulo)
Kiyoshi Fukumasu, Newton (Polytechnic School, University of São Paulo)
Profito, Francisco José (Polytechnic School, University of São Paulo)
Driemeier, Larissa (Polytechnic School, University of São Paulo)

In session: THa-600 - - CT600 - Contact Mechanics

Please login to view abstract download link

The analysis of wheel-rail contact has been highly encouraged in recent years, due to its significant influence on transportation safety and costs. Traditionally, one of the topics under investigation is the formation of defects on the wheel-rail rolling surfaces, corroborating the need for inspections and maintenance. In order to minimize time and costs and improve these analyzes, machine learning techniques have been increasingly used for this purpose. One way to predict whether the failure at the wheel-rail contact will be predominantly associated with rolling contact fatigue (RCF) or wear involves assessing the conformity of the contact. RCF will be predominant in non-conformal contacts, while wear will prevail in conformal and closely conformal contacts. For the determination of contact type, the referenced articles proposes a metric called maximum separation (or "s" parameter), which comprises the distance between the rail and wheel surface along the rail centroid after contact. The objective of the present work is to propose an Artificial Intelligence model, through Convolutional Neural Networks (CNNs), to classify the conformity of the wheel-rail contact. The inputs are the rail and wheel profiles, obtained directly from field measurements, and the output is a classification according to the level of the contact (s < 0.1 - Closely Conformal, 0.1 < s < 0.4 - Conformal, s > 0.4 - Non-Conformal). The database was generated from a validated 2D quasi-static finite element (FE) analysis, performed in Abaqus software. The analysis also required the development of a Python script to automate the simulation process of various combinations of different wheel and rail profiles and calculate the "s" parameter. Finally, after training and validation, the CNN presented satisfactory results, allowing the direct evaluation of the contact between the profiles, without complex and expensive analyzes by FE software.