The exact prediction of the dynamic behavior of technical systems by means of physical modeling approaches poses a challenge where the acting mechanisms can only be represented with great effort. This is the case, for example, with friction phenomena, which add nonlinear, dissipative elements to the system behavior. Detailed modeling approaches are usually accompanied by high complexity and application specification. Where physical models are not available or can only be implemented with great effort, models based on machine learning methods can provide a solution. The combination of both approaches to a hybrid model allows an efficient modeling of dynamic systems.
Starting point is the modeling and validation of the dynamic behavior of a double pendulum with adjustable friction in the joints. For this purpose, the mass and stiffness distributions of the double pendulum are already identified and stored in a physical model. This model is extended by a sub-model for the representation of the dissipative components by means of machine learning methods.
In the DFG-funded project "Hybrid modeling for data-driven multi-objective optimization of multi-body systems", the methodology for hybrid modeling is extended to general multi-body systems and used in multi-objective optimization. Again, the double pendulum serves as an initial example. Subsequently, the developed methodology is applied to a (multi-link) rear axle. For this purpose, the detailed axle model already available at the chair serves as a reference.
Further information on the project "Hybrid modeling for data-driven multi-objective optimization of multi-body systems" can be found here.