ABOUT THIS POSITION

The fatigue life of riveted connections in steel bridges is of a growing concern since the age of these bridges and the traffic volume increases. TNO has gathered 629 fatigue tests from literature, and evaluated these with the conventional net section stress resulting in a largely scattered S-N plot. If this is applied, the fatigue assessment method will have to be very conservative for a large number of riveted connections. Therefore, TNO has proposed a new fatigue driving force for riveted connections, based on a more local stress parameter, and developed an analytical model that estimates this stress parameter. The analytical model is deterministic, mostly based on physics, and therefore uses the joint geometry as input. This model reduces the scatter significantly, indicating the dependency of the fatigue life on the geometry. The goal of this proposal is to evaluate the fatigue tests with a machine learning (ML) algorithm based on symbolic regression instead of based on physics. Symbolic regression is a method to determine equations that presents a good fit of a dataset. It combines variables (input), with mathematical operators, constants, and analytical functions. Because the output are the equations, the ML algorithm provides insight in the relations between the different variables, and is therefore interpretable.

This internship will include the following steps:

• Understanding and implementation of an open-source symbolic regression ML algorithm.
• Training of the ML algorithm on a part of the fatigue test dataset (training set). This includes determining the coefficient of determination, and Pearson correlation coefficients. Multiple combinations of analytical functions, variables and mathematical operators will be tested.
• Application of the predicted equations from the ML algorithm to the rest of the fatigue test dataset (test set), to determine the performance of the algorithm.
• A sensitivity analysis will be performed to determine how the predicted equations from the ML algorithm hold for values of the variables outside of the tested ranges.
• The results of 3. will be compared to the developed analytical model.