Speaker
Description
Using artificial intelligence to correct weather related vehicle data
Hellweg, M.(1) ; Stiller, C.(1)
(1) Institute for Measurement and Control Systems, Karlsruhe Institute of Technology (KIT)
The project “Fleet-Weather-Map” aims to use vehicles as mobile weather stations. With the vehicular data the improvement of spatial and temporal resolution of weather forecasts and nowcasting is targeted. With this safe autonomous driving functions can be guaranteed. In the project, the AUDI AG cooperates with the “Deutscher Wetterdienst” (national weather service of Germany). The close collaboration between the car manufacturer and the national weather service ensures the development in a realistic vehicle environment and the test for usability of the meteorological data for model based applications.
The fleet vehicles collect meteorological data, such as air temperature, relative humidity, air pressure and rain intensity, from mass-produced sensors. The technology carrier built for the project is equipped with additional sensors (see Figure a)) to validate the quality of the measurement results of the mass-produced sensors. In this project, every measurement is considered individually to account for potential interfering influences and systematic measurement deviations it may be exposed to.
Prior to deployment of the vehicular data as input data for weather models, a correction of the sensor data is necessary. In this project we implement and compare two different approaches. First, we have implemented a classical model-approach that filters individual falsifying effects. Secondly, a data driven approach using artificial intelligence has been implemented. Especially for measured quantities with a variety of influencing factors and a complex linkage between these factors, using an artificial intelligence appears to be the more beneficial approach.
The artificial intelligence approach is based on supervised learning. A reference sensor is fixed on the roof of the technical carrier. There it is mostly unaffected by effects induced by the dynamics of the vehicle and thereby provides data useable as ground truth. In addition to the raw signal of the relative humidity, quantities such as air pressure, vehicle speed and air temperature are used as features for the model. Increasing the complexity of the model by using many features does not necessarily improve the performance of the model.
Currently a database of a single test drive with 12.936 data points is used. A preliminary result is presented in Figure b). Further test drives to extend the data basis are scheduled for January 2020 and will be reported at the workshop. The artificial intelligence model is expected to produce more precise results with a more extensive data basis.
