What are learning algorithms

Machine learning algorithms

Machine learning algorithms in mobile applications to predict unknown states

One focus of automotive research is the development and functional expansion of vehicle dynamics control systems (such as ESP, ABS, ASR, ...). These stabilize the vehicle in critical driving maneuvers, among other things through suitable control of actuators (e.g. braking of individual wheels). Precise detection of the current driving status is essential for this. This is guaranteed by a large number of different measurement signals (sensor, bus signals). The measurement of the required parameters is sometimes a challenge, both from a feasibility point of view and from an economic point of view. A strategy that is in the focus of science to counter this problem is the estimation of the required measurands. Signals that are already available are used to estimate the variables, from which the missing variables are predicted. The physical relationships or models of the extent to which missing measured variables can be estimated using known measured variables are sometimes not known.

Classical control systems are often based on a model-based approach or follow a defined logic. However, these rule approaches are only able to learn or appreciate unknown relationships to a limited extent. Rather, the estimation of the missing variables represents an independent model that predicts unknown target variables from existing data. This learning and recognition of patterns from existing data is called machine learning and offers the possibility of integrating a “virtual sensor” into the real system. This simulated sensor can replace a “hardware” sensor if the real integration in the vehicle is associated with great effort or high costs.

For the integration of the virtual sensor, the underlying algorithm only needs to be trained using a prototype so that it can recognize regularities. The training data used for this is obtained from real sensor data. After training and validation of the model, the physical sensor can be dispensed with and the machine learning algorithm takes on the role of the sensor.

Project supervisor at the Chair of Mechatronics:

Participating chairs / departments at the UDE:

Affiliation to the main profile areas of the Faculty of Engineering at the UDE:

Link to the faculty (main profile areas): www.uni-due.de/iw/de/forschung/psp