Authors:
Laurent Bourgois
;
Gilles Roussel
and
Mohammed Benjelloun
Affiliation:
Université du Littoral - Coôte d’Opale, France
Keyword(s):
Semi-physical modeling, gray-box, inverse dynamic model, neural network, model fusion.
Related
Ontology
Subjects/Areas/Topics:
Informatics in Control, Automation and Robotics
;
Intelligent Control Systems and Optimization
;
Machine Learning in Control Applications
;
Neural Networks Based Control Systems
;
Signal Processing, Sensors, Systems Modeling and Control
;
Signal Reconstruction
Abstract:
This study proposes to examine the design methodology and the performances of an inverse dynamic model by fusion of statistical training and deterministic modeling. We carry out an inverse semi-physical model using a recurrent neural network and illustrate it on a didactic example. This technique leads to the realization of a neural network inverse problem solver (NNIPS). In the first step, the network is designed by a discrete reverse-time state form of the direct model. The performances in terms of generalization, regularization and training effort are highlighted in comparison with the number of weights needed to estimate the neural network. Finally, some tests are carried out on a simple second order model, but we suggest the form of a dynamic system characterized by an ordinary differential equation (ODE) of an unspecified r order.