Does your System inherently contain Feedback? We can Model it from Data!

Thursday 6 Jul 2023, 13:00 → 14:00 UTC

In common engineering problems, system identification is seen as an enabling tool to control design. In this presentation, we invert this paradigm by proposing a data-driven modelling approach that takes benefit of a preprocessing of the data performed using an off-line model-based controller. Motivated by mechanical applications, but with a broader scope, our work focuses on the modelling of nonlinear feedback systems. In essence, we identify the unknown nonlinear function in the feedback path by first inferring the nonlinear latent state trajectories of the system. For that purpose, we design a model predictive controller that tracks the output of the system based on a linearised model of the input-output dynamics. Next, the functional form of the nonlinearity can be learned. We perform this task by means of neural nets, although other machine learning mappings are eligible. Simulation and experimental data from a diverse set of nonlinear systems are employed to demonstrate our findings.

Register: https://andreapizzoferrato.notion.site/ML-meets-Engineering-fa48aefc1a7f40e4b98cb6f861f766cd