Modeling And Control Of Nonlinear Systems Using Self-Organizing Map Multiple Models

Last decades have witnessed the expansion of the systems with more complex structures and broader working conditions, such as unmanned aerial vehicles (UAV). When they are in the status of large angle maneuval, the UAV are aften viewed as the systems with strong nonlinearity. For these systems needed to be controlled, the traditional control method with single controller designed for only one working condition may not achieve the satisfied performance. The multiple models control method thus appeared and filled the gap between the control theory and its application. This method itself is also one of the most important parts of the current research in control theory.This thesis focuses on the modeling and control of nonlinear system using multiple models method based on the self-organizing map neural network. The main contributions of this thesis are summarized as follows.1. The use of self-organizing map neural network in multiple models approximation of nonlinear system is analyzed. Based on the analysis of the way in which the neural network partitions the space into local regions, the conclusion is made that the multiple model approximation using self-organizing map neural network can not obtain the optimal approximating result. The essential reason is that the goal of neural network partitioning the space does not coincide with that needed in the multiple models approximation.2. An active learning method is proposed for the self-organizing map neural network used in multiple models approximation. The method feeds back the approximation errors of local models into the selection of the learning samples and, as a result, achieves the different distribution of the neural nodes. The resulted neural network has more appropriate distribution of neural nodes and greatly improves the modeling performance.3. For the local linear multi-input-multi-output models, the robust inverse Nyquist array method is used to analyze the diagonal dominance and stability of the models with parameter uncertainty. An improved method is proposed to approximately estimate the width of the robust Gershgorin bands of the models when there are identification errors of parameters in the models. The method reduces the conservation of the traditional estimations and facilitates the controller design for the uncertain models.4. A new multiple models direct inverse control structure using self-organizing map neural network is proposed. In the scheme, a self-organizing map is used to partition the working space of nonlinear system into several local regions in which the inverse models are derived. The models are also used as the inverse controllers and one of them is activated in a local region respectively. The switching between the controllers is also realized by the neural networks. The boundness of the control error in this method is proved. Further more, a multiple models adaptive inverse controller is derived by adding an adaptively tuned inverse model. The new control method can asymptotically track the stationary signal.5. The application of modeling and control of large angle maneuver of UAV is investigated. The problems of training samples selection and model structures selection are discussed for the modeling method using self-organizing map multiple moedels. The steps of modeling and controller design are also given.6. The use of multiple models based on the self-organizing map neural network in the nonlinear system filtering problem is explored. The multiple models provide the Jacobian matrix in the extended Kalman filter which will otherwise be computed pointwise. The method reduces the computation burden in the extended Kalman filter and at the same time improves the filtering performance.