Data-Driven Linear Parameter-Varying Model Identification Using Transfer Learning

Abstract

This letter proposes transfer learning methods to address a challenge in state-space linear parameter-varying (LPV-SS) model identification/learning using kernelized machine learning, when the distributions of the training and testing sets are different. Kernel mean matching is first employed to correct sample bias by resampling the data in the training set before the states in state-space model are estimated. Moreover, transfer component analysis is adopted to find a state-space basis transformation such that the transformed states follow similar distributions. The proposed methods are validated by testing on an ideal continuous stirred tank reactor (CSTR) model. Simulation results show that the proposed learning methods can enhance the accuracy of model identification and reduce the efforts involved in hyperparameters tuning.