Machine Learning for Predicting the Damaged Parts of a Low Speed Vehicle Crash

Using time series of on-board car data, this research focuses on predicting the damaged parts of a vehicle in a low speed crash by machine learning techniques. Based on a relatively small and class-imbalanced dataset, we present our automatic and for small datasets optimized method to use time series for machine learning. Based on 3982 extracted features, we are using feature selection algorithms to find the most significant ones for each component. We train random forest models per part with its most relevant set of features and optimize the hyper-parameters by different techniques. This so-called part-wise approach provides good insights into the model performance for each part and offers opportunities for optimizing the models. The final F1 prediction scores (reaching up to 94%) show the large potential of predicting damaged parts with on-board data only. Furthermore, for the worse performing parts of this small and imbalanced dataset, it indicates the potential for reaching good prediction scores when adding more training data. The utilization of such method offers great possibilities, e.g., in vehicle insurance processing for automatized settling of low speed crash damages.