Synthetic Fuel Detection for Vehicles in Field using Machine Learning

In order to achieve the EU-wide climate targets and to realize the full CO2 savings potential, a reliable in-vehicle fuel type detection is inevitable. We developed an appealing solution for fuel type detection, which is inexpensive and robust. Our software approach uses existing hardware of diesel vehicles and can be integrated into standard engine control units, thus allowing retrofitting of vehicles in field. As a novelty, our approach uses the fuel temperature as the most important distinguishing feature and is based on several machine learning methods. For our research we used a large endurance run data set of Hydrogenated Vegetable Oil (HVO) and diesel fueled vehicles. We modeled numerous sensor models using time series regression analysis and calculated the differences between model outputs and the actual sensor values. These differences are the main input for our fuel type classifier, which is a feed forward neural network. Moreover, we developed a filtering method using the neural network output probabilities and a one-class-SVM to enhance the robustness of the fuel type detection. Our approach achieves an overall accuracy of 97%. To the best of our knowledge, our approach allows for the first time distinguishing HVO and diesel without additional hardware.