Class Of Model Structures Research Articles

Machine learning with nonlinear state space models

Abstract In this dissertation, a novel class of model structures and associated training algorithms for building data-driven nonlinear state space models is developed. The new identification procedure with the resulting model is called local model state space network (LMSSN). Furthermore, recurrent neural networks (RNNs) and their similarities to nonlinear state space models are elaborated on. The overall outstanding performance of the LMSSN is demonstrated on various applications.

Identification of reheat furnace temperature models from closed‐loop data—an industrial case study

AbstractThis work deals with the application of prediction error (PE) method to identify the furnace temperature models from a set of closed‐loop data. In this study, we gather a set of measurement data of the set points of the furnace zone temperatures as inputs and the furnace sidewall temperatures as outputs under a dynamic change of the slab pace rate. Owing to the complexity of its process dynamics, we assume no knowledge about the nature of the feedback mechanism. By treating the slab pace rate as an additive external signal, we show that the closed‐loop data is informative for applying a direct approach to the closed‐loop identification using the PE method, but only for a particular class of model structures. Model validation results support this analysis, in which the identified ARX, Box–Jenkins, and state‐space models are reasonably better than the identified FIR models, according to the Akaike's index and the one‐step‐ahead prediction criteria. The residual analysis reveals that the identified ARX, Box–Jenkins, and state‐space models do satisfy a 99% confidence region of its auto‐ and cross‐correlation functions. Moreover, we find out that, for the collected data, there is no significant difference in the model predictive quality when applying the MISO and MIMO PE methods using the state‐space model structure. © 2006 Curtin University of Technology and John Wiley & Sons, Ltd.

The parsimony principle for a class of model structures

A new form of the parsimony principle is presented for a class of model structures containing a common subset of parameters. The principle is applied to the problem of estimating the autoregressive (AR) parameters of an AR process corrupted by white measurement noise.