Deep learning methods for multibody modeling: application to railway dynamics

Details

Supervisors

Fisette, Paul ; Docquier, Nicolas

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électromécanicien, à finalité spécialisée: mécatronique

Abstract

In recent years deep neural networks have emerged as an effective approach to tackling various problems. In parallel, multibody simulation allows the generation of large amounts of data. This data can be used to train neural networks. For multibody modeling, a deep learning approach has two main advantages. It allows us to make accurate models which are time-efficient. They also do not require explicit mathematical equations between input and output data and are thus very suited for surrogate models. The main objective of this report is to demonstrate how deep-learning approaches can be used for multibody modeling. This master’s thesis focuses on two deep neural networks, applied to railway applications. During the first experiment a deep neural network predicts the vertical acceleration of a railway vehicle’s car body. The predictions are made based on the track profile. This network is composed of two independent sub-networks: one for handling low frequencies and another for high frequencies. Both sub-networks are mainly made of long short-term memory layers, treating the accelerations as time series. The second experiment performs parameter identification on the secondary suspension of a railway vehicle. The network is a multi-layer perceptron. This master’s thesis presents both experiments; how the networks were constructed, as well as their obtained accuracies.