Introducing Compressed Mixture Models for Predicting Long-Lasting Brake Events

Paper in proceeding, 2018

With tougher restrictions on emissions the automotive industry is in dire need of additional functionality to reduce emissions. We conduct a case study trying to predict long-lasting brake events, to support the decision-making process when the engine can beneficially be put to idle or shut down to achieve emission reduction. We introduce Compressed Mixture Models, a multivariate and mixed variate kernel density model featuring online training and complexity reduction, and use it for prediction purposes. The results show that the proposed method produces comparable prediction results as a Random Forest Classifier and outperform a Support Vector Classifier. On an urban road a prediction accuracy of 87.4 % is obtained, while a prediction accuracy of 76.4 % on a highway segment using the proposed method. Furthermore, it is possible to use a trained Compressed Mixture Model as a tool for statistical inference to study the properties of the observed realization of the underlying random variables.