|
|
Parallel Computing and Neural Networks in Behavioral Modeling
Vágnerová, Jitka ; Škvor,, Zbyněk (oponent) ; Dědková, Jarmila (oponent) ; Lukeš, Zbyněk (vedoucí práce)
This thesis is focused on methods for the aircraft equipment modeling. The first part provides a brief overview of classical system modeling approaches used for system description, identification, and modeling. Then adaptive, fuzzy and hybrid methods used mainly for black-box system modeling are introduced. Aim of the thesis is to develop an algorithm for identification and modeling of a general system, which can be nonlinear, dynamic and complex. Multiple inputs and multiple outputs of model are assumed. The main part of the thesis introduces a new method which falls into the hybrid systems. It combines fuzzy approach with parametrically defined rules and general regression neural network. Firstly, the fundamentals of simple general regression neural network and its smoothness parameter determination are presented. Secondly, the general regression neural network with the fuzzy rules is introduced. Third part of the thesis is focused on the parallel computing, one of the main objectives. The final algorithm is designed for the parallel machine, because the computational time can be significantly high and for the larger datasets, the model is not achievable when evaluated in single thread. Block diagram for parallel computing in Matlab and CUDA is provided, as well as the basic structure of CUDA source code. Finally, the method is verified on data obtained from the measurement of a miniaturized aircraft model using the antenna outside the aircraft and the probe inside the fuselage of the aircraft model. The validation of the method is done using mean squared error and compared to mean squared error of corresponding model performed using the multilayer neural network with backpropagation learning and Levenberg-Marquardt algorithm.
|
|
|
Parallel Computing and Neural Networks in Behavioral Modeling
Vágnerová, Jitka ; Škvor,, Zbyněk (oponent) ; Dědková, Jarmila (oponent) ; Lukeš, Zbyněk (vedoucí práce)
This thesis is focused on methods for the aircraft equipment modeling. The first part provides a brief overview of classical system modeling approaches used for system description, identification, and modeling. Then adaptive, fuzzy and hybrid methods used mainly for black-box system modeling are introduced. Aim of the thesis is to develop an algorithm for identification and modeling of a general system, which can be nonlinear, dynamic and complex. Multiple inputs and multiple outputs of model are assumed. The main part of the thesis introduces a new method which falls into the hybrid systems. It combines fuzzy approach with parametrically defined rules and general regression neural network. Firstly, the fundamentals of simple general regression neural network and its smoothness parameter determination are presented. Secondly, the general regression neural network with the fuzzy rules is introduced. Third part of the thesis is focused on the parallel computing, one of the main objectives. The final algorithm is designed for the parallel machine, because the computational time can be significantly high and for the larger datasets, the model is not achievable when evaluated in single thread. Block diagram for parallel computing in Matlab and CUDA is provided, as well as the basic structure of CUDA source code. Finally, the method is verified on data obtained from the measurement of a miniaturized aircraft model using the antenna outside the aircraft and the probe inside the fuselage of the aircraft model. The validation of the method is done using mean squared error and compared to mean squared error of corresponding model performed using the multilayer neural network with backpropagation learning and Levenberg-Marquardt algorithm.
|
host ::
RSS.