شناسايي و كنترل سيستم هاي غير خطي با استفاده از شبكه هاي عصبي موجك بازگشتي

چكيده انگليسي :

Identification and Control of nonlinear dynamic systems using recurrent wavelet neural networks Afsaneh Ghadirian a ghadirian@ec iut ac ir Date of Submission 2011 03 15 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language Farsi Supervisor Maryam Zekri mzekri@cc iut ac ir Abstract Feed forward neural networks NNs have been shown to obtain successful results in system identification and control NNs are static input output mapping schemes that can approximate a continuous function to an arbitrary degree of accuracy A recurrent neural network RNN is a dynamic mapping network which is more suitable for describing dynamic systems than the NN According to RNN structure it can deal with time varying input or output through its own natural temporal operation For this ability to temporarily store information the structure of the network is simplified Therefore fewer nodes are required for system identification The combination of wavelet theory and neural networks has lead to the development of WNNs WNNs are feed forward neural networks using wavelets as activation function In WNNs both the position and the dilation of the wavelets are optimized besides the weights The WNN keeps generalization approximating property of NNs and the capability of wavelet decomposition WNNs combine the capability of artificial NNs in learning processes and wavelet localization property and have high accuracy and fast learning ability Accordingly the concept of the wavelet neural network WNN has become increasingly important and can be used for identification and control of the complex nonlinear systems On the other hand the recurrent wavelet neural network RWNN combines the properties of attractor dynamics of the RNN and good performance of the WNN The RWNN can deal with time varying input or output through its own natural temporal operation In RWNN structure the mother wavelet layer is composed of internal feedback neurons to capture the dynamic response of a system Since the proposed RWNN is a modified model of the wavelet neural network WNN the RWNN includes the basic ability of the WNN such as fast convergence and localization property Besides the RWNN has a property unlike the WNN that the RWNN can store the past information of the network The objective of this thesis is to introduce an on line recurrent wavelet neural network controller RWNNC for single input single output nonlinear dynamic systems The purpose of control is to determine control signal u such that the closed loop system output can track a desirable output The RTRL algorithm is applied to adjust the shape of wavelet functions feedback weights and the connection weights Since the architecture of the RWNNC model is able to preserve past states of the networks the RWNNC model has the capability to deal with temporal problems and also the controller has the other advantages such as simple structure and good generalization performance to nonlinear systems We train the proposed controller as on line in the closed loop system by the tracking error of nonlinear system Controller inputs are the tracking error and the first order derivative of the tracking error Then the RWNNC model is applied to two control problems In the first example the RWNNC is applied to control a nonlinear servomechanism and in the second example the proposed controller is used for a single link robot The obtained results of simulation examples demonstrate that a desirable tracking response can be achieved by the RWNN controller even under the change of parameters of system and the proposed controller is quite effective in control nonlinear dynamic systems Keywords Recurrent wavelet neural network Nonlinear systems Real time recurrent learning Control

قديريان بهارانچي، افسانه

شناسايي و كنترل سيستم هاي غير خطي با استفاده از شبكه هاي عصبي موجك بازگشتي

الگوريتم يادگيري بلادرنگ