Using the SEKF to Transfer NN Models of Dynamical Systems with Limited Data

TL;DR

This paper introduces the use of the Subset Extended Kalman Filter (SEKF) to efficiently adapt pre-trained neural network models of dynamical systems to new systems with minimal data, reducing data requirements and computational costs.

Contribution

It presents a novel application of SEKF for neural network model transfer in dynamical systems with limited data, demonstrating effective adaptation with as little as 1% of original data.

Findings

SEKF-based adaptation captures system dynamics with minimal data.

Finetuning with SEKF reduces computational cost.

Method improves generalization with limited data.

Abstract

Data-driven models of dynamical systems require extensive amounts of training data. For many practical applications, gathering sufficient data is not feasible due to cost or safety concerns. This work uses the Subset Extended Kalman Filter (SEKF) to adapt pre-trained neural network models to new, similar systems with limited data available. Experimental validation across damped spring and continuous stirred-tank reactor systems demonstrates that small parameter perturbations to the initial model capture target system dynamics while requiring as little as 1% of original training data. In addition, finetuning requires less computational cost and reduces generalization error.