Convolutions on graphs for learning vehicle crash behaviour

Abstract

Convolutional Neural Networks (CNN) have shown successful results in the recent years, especially within the area of image analysis. The idea of learning to predict the result of a crash simulation using machine learning rose from the analogy between images and Finite Element models (FE-models) used in crash simulations. However, the data used when training a machine learning model using CNN needs to be structured in a consistent way, as images are. FE-models however are represented as graphs and do not have the grid-like structure that images have and can therefore not be directly processed using CNN. The purpose of this project was to investigate the possibility to transform FE-models into image-like embeddings and to use CNN to explore these embeddings. Two graph convolutional methods were investigated for the creation of the embedding. The first one was the Neural Graph Fingerprint (NGF) method suggested in the literature for the original purpose of parsing molecular graphs. The second one was developed during this project, called the FEMBEDDING method, and was to parts inspired by NGF and the Graph Neural Network model that also has been suggested in literature. Three datasets of crash simulations with varying geometrical complexity were developed during the project. It is shown here that embeddings created by using both methods can successfully be used to train a CNN and predict the outcome of the test sets with a good level of accuracy already with only randomly initialized embedding weights. The FEMBEDDING method made the embeddings richer in information and performed consistently better than the NGF method. For the more geometrical complex dataset it is shown that the value of the FEMBEDDING embeddings increases with an increased neighbourhood depth taken into account while parsing the the FE-graphs.