Speaker
Description
Many physical systems, such as wind farms, aircraft, and particle
accelerators, are nonlinear, meaning their behaviour can be unpredictable and
difficult to model with simple linear differential equations. However, we need
linear representations of these nonlinear systems within their operating range
because control design and fault diagnosis for linear systems are
well-established and computationally efficient. With today's access to large
amounts of data, machine learning techniques offer promising tools for system
identification.
Yet, relying solely on machine learning can lead to complex, hard-to-interpret
digital twins of nonlinear systems.
In this poster, I present an alternative approach that combines physical
insights with data to identify a linear representation of nonlinear systems.
By integrating Koopman theory and Autoencoders, we can achieve a powerful
data-driven method for system identification, retaining the benefits of linear
analysis while capturing the underlying nonlinear dynamics.
I invite you to explore how this integration works and how it can be applied
to model the complex systems.
