Hybrid Modeling of Mechanical Systems Based on First Principle and Data-Driven Approaches

Lightweight robots are becoming increasingly interesting due to their complementary advantages - low energy and resource consumption, low mass and high working speed. However, due to their lightweight construction and the associated considerable structural flexibility, they are more susceptible to vibrations. At the same time, a highly dynamic mode of operation is required for optimum productivity of the robot systems. As a result, physical effects such as elasticities, inhomogeneous friction effects etc., which could be neglected in conventional, slow-moving systems, are becoming increasingly influential.

In this context, modeling and simulation are crucial for the entire product life cycle, from engineering to commissioning and model-based control through to maintenance. Past studies have shown that in the purely physical modeling of complex systems, strong discrepancies sometimes occur between the simulated system behavior and the real behavior, e.g. due to insufficiently identified effects. In the following, these discrepancies are referred to as the simulation-to-reality gap (Sim2Reality Gap). The composition of a hybrid overall model Σ from a physics-based and a data-based model corresponding to Σ = ΣPhysics + ΣData can be very profitable here.

Data-based modeling is particularly interesting for changing system components. Permanent model adaptation of ΣData can take place through monitoring. For this reason, when setting up the model architecture and deriving the models, care is taken to ensure that the models can adapt permanently.

The more complex the real system to be mapped, the more complex the process of hybrid model creation. Expert-driven, purely manual configuration of hybrid models of robots for new complex processes does not make economic sense. In addition, the current shortage of engineers limits the availability of experts. This results in the challenge of developing a workflow with which engineers with limited expert knowledge can also use hybrid models in an industrial context.