.png){kind=icon}
.png){kind=icon}
Return
A highly-accurate identification method for linear systems using transferred knowledge
.png){kind=icon}
PRE
AI
DOI：10.1016/j.automatica.2024.112016.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The identification accuracy of systems relies on sufficiently-rich measurement information. Collecting large numbers of informative data is costly and burdensome due to various factors. This paper proposes a novel knowledge transfer identification (KTI) method, which utilizes the extra knowledge from a source system to improve the identification accuracy of the target system. Specifically, we transfer the source knowledge in the form of probability distribution and use it as an additional constraint on the target posterior distribution. Then, the identification problem is solved by inferring an optimal distribution minimizing the Kullback-Leibler divergence between the transfer posterior conditioned on both the transferred knowledge and the target data while the posterior only conditioned on the target data. To further highlight the advantages of the proposed KTI method, a data transfer identification (DTI) strategy is presented to directly transfer the raw source data. Moreover, we explore the essence of the proposed KTI and derive the condition of avoiding negative transfer. The results are validated by a numerical example and a quadruple water tank example, which confirm that transferring probability distribution outperforms transferring raw measurement data. (c) 2024 Published by Elsevier Ltd.
Keywords:
System identification
Transfer identification
Data transfer identification
Negative transfer