Data-driven control and transfer learning using neural canonical control structures*

TL;DR

This paper introduces a neural network-based data-driven control method that uses auto-encoders for system identification and transfer learning, enabling adaptable control across systems with different parameters.

Contribution

It presents a novel neural canonical control structure for indirect control and transfer learning, combining feedback linearization with neural networks for system identification and adaptation.

Findings

Effective transfer of control strategies between systems with different parameters.

Successful demonstration on academic and industrial examples.

Neural network auto-encoders accurately identify system transformations.

Abstract

An indirect data-driven control and transfer learning approach based on a data-driven feedback linearization with neural canonical control structures is proposed. An artificial neural network auto-encoder structure trained on recorded sensor data is used to approximate state and input transformations for the identification of the sampled-data system in Brunovsky canonical form. The identified transformations, together with a designed trajectory controller, can be transferred to a system with varied parameters, where the neural network weights are adapted using newly collected recordings. The proposed approach is demonstrated using an academic and an industrially motivated example.