PhD - Hybrid Physics-Based Neural Network Models for Predicting Nonlinear Dynamics in Mechatronic Applications

When
27-10-2021 from 17:00 to 19:00
Where
leslokaal Rudolf E. Richter, gebouw 131 Volta, gelijkvloers, Technologiepark Zwijnaarde 131, 9052 Zwijnaarde
Language
English
Website
https://tinyurl.com/PhD-Defense-WannesDeGroote
Add to my calendar

Public defense of the doctoral thesis - Hybrid Physics-Based Neural Network Models for Predicting Nonlinear Dynamics in Mechatronic Applications

Mechatronic applications such as robots, production machines and electric vehicles are paramount in modern society. More than ever, the development of these applications is substantiated by accurate system models. The original development strategy of these system models was primarily based on physical insights. Fundamental laws were used to make simplified representations of the considered application.

However, in practice, there is often insufficient expert knowledge to capture all interactions at play by physical laws, leading to discrepancies between the model and measurement data. During recent years the use of machine learning methods have gained increasing interest to learn the system behavior. These models do not require prior knowledge about the system because they can learn patterns directly from the data by fitting the (non-physically interpretable) model parameters to the measurements. Nevertheless, these blackbox approaches are often insufficient interpretable and their reliability cannot always be assured.

Hence, both physics based and data-driven approaches have their own strengths and weaknesses. In this PhD research, we searched for modeling methods that exhibit the benefits of both physics-inspired and data-driven approaches. More specifically, we combined physics-inspired models with neural networks, to obtain interpretable, accurate and robust models that can predict the nonlinear behavior of mechatronic applications.