Current workshop diagnostics are based on manually-generated decision trees. This approach is increasingly reaching its limits due to a growing variant diversity, and the increasing complexity of vehicle systems. This session will describe BMW’s new Apache Spark-enabled approach: Use the available data from cars and workshops to train models that are able to predict the right part to switch, or the action to take.
You’ll get an overview and presentation of BMW’s complete pipeline including ETL, model training based on Spark 2.1, serializing results along with metadata and serving the gained insights as Web-App. You’ll also hear how Spark helped BMW leverage the information from millions of observations and thousands of features, and learn what pitfalls they experienced (e.g. setting up a working dev-toolchain, working with 50K features, parallelizing well), and how you can avoid them.
Session hashtag: #SFent2
– 2008-2014: Mechanical engineering at the Technical University of Munich (TUM), majors in automotive and computer science
– 2015-present: PhD. candidate for computer science at BMW
– 2016: Publication: “Design and optimization of an autonomous feature selection pipeline for high dimensional, heterogeneous feature spaces”, IEEE SSCI
– Skills: Extensive C/C++/Java Skills
– Personal: Passion for writing efficient code (C/Scala), brewing beer (projekt: bierbot.de), learning new things and doing sports.