Colloquia

Summary

In recent years, machine learning (ML) and artificial intelligence (AI) have become more prominent in the field of model engineering. A physics-informed neural network (PINN) combines the highly nonlinear capabilities of machine learning with the fundamental principles and physical constraints governed by real-world physics. In particular, the Deep Lagrangian Networks (DeLaN) have been researched before for their ability to generate feedforward control of compliant actuated systems.

This thesis presents the integration of a dedicated damping network into the DeLaN, extending it to DeLaN+D. The objective is to determine if the DeLaN+D PINN framework is a viable, automated option for modelling system dynamics. By constructing a neural network using Lagrangian mechanics, the data-driven DeLaN+D framework can learn complex dynamics with a physics-supported, generic structure. The framework is used to identify and replicate the dynamic behaviour of two parallel kinematic manipulators (PKMs) with compliant joints. 

The results indicate that DeLaN + D can replicate the dynamics of complex systems with a purely data-driven approach. When used as a feedforward control system, neural network models can compete with traditional modelling approaches, boasting similar performance-enhancing results. However, the results also highlight risks regarding the usage of neural networks, where undesirable controller behaviour can be interpreted as system dynamics. Although caution is necessary, the findings suggest that the usage of DeLaN+D can be a viable, simple, and automated approach to modelling the system dynamics of the considered PKMs.