This paper introduces a novel approach for generating synthetic data to enhance model generalization and robustness in system identification, particularly when training data is scarce. The method leverages knowledge transfer from similar systems using a pre-trained meta-model (a Transformer) to predict system behavior and generate synthetic output sequences. A hyperparameter balances the influence of training and synthetic data in model estimation, tuned using a validation dataset. The efficacy is demonstrated through a numerical example involving Wiener-Hammerstein systems.
Publisher
Published On
Authors
Dario Piga, Matteo Rufolo, Gabriele Maroni, Manas Mejari, Marco Forgione
Tags
synthetic data
model generalization
system identification
knowledge transfer
Transformer
hyperparameter tuning
Wiener-Hammerstein systems
Related Publications
Explore these studies to deepen your understanding of the subject.
