This conference paper was submitted for presentation at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

The method described below is part of the research project KI-MeZIS ('œAI Methods in Condition Monitoring and Demand-Adapted Maintenance of Rail Vehicle Structures' (funding code 19|21024D)). The aim of this project is to develop and apply the potential of artificial intelligence (AI) methods for monitoring rail traffic. With the help of sensors placed on the front, on the supporting structure and on the bogie of the train, AI methods should be able to evaluate and interpret the data. Nowadays, aerodynamics and lightweight constructions are the most important requirements for rail vehicles. However, standards are currently used for the design of rail components and most of them are not actual anymore and the origin has been lost during the years. Moreover, it often leads to an oversizing of the components and thus to higher costs in production and operation due to a higher mass. In order to respect our lightweight requirements, new materials can be used or the components can be designed according to the loads applied during operation. This has the advantage that the components can be designed according to the real load and thus lead to a mass reduction overall. In our case, sensors, such as accelerometers and strain gauges are used to determine the acting loads. The main goal is to use then these loads for Finite-Element-Method(FEM) simulations or fatigue analysis. Forces are the easiest loads to define on a FEM model. But the force cannot be so easily determined from accelerometers or strain gauge because of non-linear problems. That is why, Machine learning method is used to solve the inverse problem. Sensors data, named acceleration and strain are given to the neuronal network (NN) and as output, the force is given. In order to train the NN, a large amount of couple acceleration/force or strain/force are needed. To that end, a FEM model was created to generate these training data. A code has been developed to create and run automatically FEM simulations, changing the input force. Finally, the output force results from the NN model can be used for FEM simulations, design optimization, like topology optimization for example or/and fatigue analysis.