A combined B-spline-neural-network and ARX model for online identification of nonlinear dynamic actuation systems

Michele Folgheraiter

Research output: Contribution to journalArticle

8 Citations (Scopus)

Abstract

This paper presents a block oriented nonlinear dynamic model suitable for online identification. The model has the well known Hammerstein architecture where as a novelty the nonlinear static part is represented by a B-spline neural network (BSNN), and the linear static one is formalized by an autoregressive exogenous model (ARX). The model is suitable as a feed-forward control module in combination with a classical feedback controller to regulate velocity and position of pneumatic and hydraulic actuation systems which present nonstationary nonlinear dynamics. The adaptation of both the linear and nonlinear parts is taking place simultaneously on a patter-by-patter basis by applying a combination of error-driven learning rules and the recursive least squares method. This allows to decrease the amount of computation needed to identify the model's parameters and therefore makes the technique suitable for real time applications. The model was tested with a silver box benchmark and results show that the parameters are converging to a stable value after 1500 samples, equivalent to 7.5. s of running time. The comparison with a pure ARX and BSNN model indicates a substantial improvement in terms of the RMS error, while the comparison with alternative nonlinear dynamic models like the NNOE and NNARX, having the same number of parameters but greater computational complexity, shows comparable performances.

Original languageEnglish
JournalNeurocomputing
DOIs
Publication statusAccepted/In press - Jul 8 2015

Fingerprint

Neural Networks (Computer)
Nonlinear Dynamics
Splines
Identification (control systems)
Neural networks
Dynamic models
Benchmarking
Least-Squares Analysis
Feedforward control
Silver
Pneumatics
Learning
Computational complexity
Hydraulics
Feedback
Controllers

Keywords

  • B-spline neural network
  • Hammerstein model
  • Humanoid robotics
  • Hydraulic actuation systems
  • Nonlinear dynamic model
  • Real-time model adaptation

ASJC Scopus subject areas

  • Computer Science Applications
  • Cognitive Neuroscience
  • Artificial Intelligence

Cite this

A combined B-spline-neural-network and ARX model for online identification of nonlinear dynamic actuation systems. / Folgheraiter, Michele.

In: Neurocomputing, 08.07.2015.

Research output: Contribution to journalArticle

@article{5e43a3f96b974c57a146561092631e94,
title = "A combined B-spline-neural-network and ARX model for online identification of nonlinear dynamic actuation systems",
abstract = "This paper presents a block oriented nonlinear dynamic model suitable for online identification. The model has the well known Hammerstein architecture where as a novelty the nonlinear static part is represented by a B-spline neural network (BSNN), and the linear static one is formalized by an autoregressive exogenous model (ARX). The model is suitable as a feed-forward control module in combination with a classical feedback controller to regulate velocity and position of pneumatic and hydraulic actuation systems which present nonstationary nonlinear dynamics. The adaptation of both the linear and nonlinear parts is taking place simultaneously on a patter-by-patter basis by applying a combination of error-driven learning rules and the recursive least squares method. This allows to decrease the amount of computation needed to identify the model's parameters and therefore makes the technique suitable for real time applications. The model was tested with a silver box benchmark and results show that the parameters are converging to a stable value after 1500 samples, equivalent to 7.5. s of running time. The comparison with a pure ARX and BSNN model indicates a substantial improvement in terms of the RMS error, while the comparison with alternative nonlinear dynamic models like the NNOE and NNARX, having the same number of parameters but greater computational complexity, shows comparable performances.",
keywords = "B-spline neural network, Hammerstein model, Humanoid robotics, Hydraulic actuation systems, Nonlinear dynamic model, Real-time model adaptation",
author = "Michele Folgheraiter",
year = "2015",
month = "7",
day = "8",
doi = "10.1016/j.neucom.2015.10.077",
language = "English",
journal = "Neurocomputing",
issn = "0925-2312",
publisher = "Elsevier",

}

TY - JOUR

T1 - A combined B-spline-neural-network and ARX model for online identification of nonlinear dynamic actuation systems

AU - Folgheraiter, Michele

PY - 2015/7/8

Y1 - 2015/7/8

N2 - This paper presents a block oriented nonlinear dynamic model suitable for online identification. The model has the well known Hammerstein architecture where as a novelty the nonlinear static part is represented by a B-spline neural network (BSNN), and the linear static one is formalized by an autoregressive exogenous model (ARX). The model is suitable as a feed-forward control module in combination with a classical feedback controller to regulate velocity and position of pneumatic and hydraulic actuation systems which present nonstationary nonlinear dynamics. The adaptation of both the linear and nonlinear parts is taking place simultaneously on a patter-by-patter basis by applying a combination of error-driven learning rules and the recursive least squares method. This allows to decrease the amount of computation needed to identify the model's parameters and therefore makes the technique suitable for real time applications. The model was tested with a silver box benchmark and results show that the parameters are converging to a stable value after 1500 samples, equivalent to 7.5. s of running time. The comparison with a pure ARX and BSNN model indicates a substantial improvement in terms of the RMS error, while the comparison with alternative nonlinear dynamic models like the NNOE and NNARX, having the same number of parameters but greater computational complexity, shows comparable performances.

AB - This paper presents a block oriented nonlinear dynamic model suitable for online identification. The model has the well known Hammerstein architecture where as a novelty the nonlinear static part is represented by a B-spline neural network (BSNN), and the linear static one is formalized by an autoregressive exogenous model (ARX). The model is suitable as a feed-forward control module in combination with a classical feedback controller to regulate velocity and position of pneumatic and hydraulic actuation systems which present nonstationary nonlinear dynamics. The adaptation of both the linear and nonlinear parts is taking place simultaneously on a patter-by-patter basis by applying a combination of error-driven learning rules and the recursive least squares method. This allows to decrease the amount of computation needed to identify the model's parameters and therefore makes the technique suitable for real time applications. The model was tested with a silver box benchmark and results show that the parameters are converging to a stable value after 1500 samples, equivalent to 7.5. s of running time. The comparison with a pure ARX and BSNN model indicates a substantial improvement in terms of the RMS error, while the comparison with alternative nonlinear dynamic models like the NNOE and NNARX, having the same number of parameters but greater computational complexity, shows comparable performances.

KW - B-spline neural network

KW - Hammerstein model

KW - Humanoid robotics

KW - Hydraulic actuation systems

KW - Nonlinear dynamic model

KW - Real-time model adaptation

UR - http://www.scopus.com/inward/record.url?scp=84964937330&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84964937330&partnerID=8YFLogxK

U2 - 10.1016/j.neucom.2015.10.077

DO - 10.1016/j.neucom.2015.10.077

M3 - Article

AN - SCOPUS:84964937330

JO - Neurocomputing

JF - Neurocomputing

SN - 0925-2312

ER -