Identification of Nonlinear Local Model State Space Networks

点击翻译按钮获取中文摘要

This project shall investigate a new approach for nonlinear system identification. It fuses ideas from the machine learning and system identification communities. The goal of this project is to develop a novel class of model structures and associated training algorithms for building data-driven nonlinear state space models. They are based on local linear model networks and will offer many benefits compared to existing approaches. The methods contribute to the demand for trustworthy artificial intelligence in industry. A training algorithm for constructing local model state space networks (LMSSNs) from data will be developed. Thereby, the focus will be on the transfer of incremental tree-construction strategies to state space modeling. Various challenges arise from the recurrency of the state vector and are addressed in this project. In order to carry out the training of LMSSNs in an effective and efficient way, new optimization schemes inspired by machine learning approaches are investigated. A suitable parameterization of the local models needs to be found by considering canonical forms, block-oriented structures, and/or regularization techniques. Hereby special care is given to the crucial issues of robustness and extrapolation behavior. Finally, the developed methods are compared to state-of-the-art alternatives on various benchmark problems. Furthermore, their practicability is tested on a number of real-world processes.