Context Discovery for Personalised Automotive Functions

Context: With the growing prevalence of machine learning applications within the automotive domain, we observe an increase of built-in sensors and generated data. This data includes complex data such as point clouds for the environmental model but also less sophisticated data like temperature and road type. When developing data-driven functions there is often no need for additional sensors since a huge amount of data is already sensed, but often hidden. Aim: We aim to discover vehicle signals that describe a system's or user's behaviour when interacting with vehicle functions. These signals could directly serve as input for data-driven vehicle functions. Method: Based on supervised feature selection algorithms we propose an approach to discover these vehicle signals. This approach can either be deployed in the backend on test fleet data or onboard the vehicle. Results: Based on seven test cases, we evaluated the approach with 17 feature selection algorithms on eight customer vehicle data sets. To evaluate the resulting signal subsets, we trained machine learning models that in turn were able to predict the behaviour of the user. Conclusion: These trained models achieved high accuracy in the prediction, which shows that current vehicles already collect enough data to predict the user's behaviour and the proposed approach was able to discover the appropriate vehicle signals. Considering the huge amount of data and vehicles as well as the highly diverse behaviour of every user, a scalable discovery approach considering every user is inevitable.