When analysing vehicle data from field tests in real world situations, it can often be helpful to be able to show differences between vehicles. A direct comparison of the data is usually not possible in a meaningful way, as the differences between the vehicles are often small. Even if two vehicles were driven on identical routes, the influences of driver, weather and traffic will overshadow the vehicle differences to be analysed. Similarly, the analysis of ageing effects in real road traffic is challenging.
Therefore, a new AI-based approach was developed at FKFS. An AI-based Digital Twin is generated for each real vehicle. Afterwards, the digital twins can be compared on virtual test drives with exactly identical ambient and driver conditions, allowing analyses that are not possible in real test drives.
To create a digital twins an AI model based on LSTM cells is trained on vehicle measurements. The LSTM cells are able to learn inertias, e.g. in thermal behaviour or in the power train, as well as complex temporal dependencies in the dynamic behaviour. In the demonstrator project, the powertrain of a power-split hybrid and two battery electric vehicles are modelled with the subsystems component temperatures and cooling, battery & electrical path and ECU. Only ECU signals and the series sensors mounted in the vehicle, whose number and signal quality are often very low, were used as input values. The low quality of the input data can be partially compensated by the huge amount of data from a very large number of test drives from field and fleet trials. The poster shows the model quality that can be achieved with about 5,000 km of training data from real driving recordings.