Single Neuron Pid Control Based On RBF Neural Network Sequential Learning Algorithm

Single neuron PID control based on RBF neural network identification is a typical structure of neural network PID control. It obtains the Jacobian information of object by RBF neural network identification for the guidance of single neuron weights learning, then the on-line tuning of PID parameters are achieved. In this method, the approximation performance of RBF neural network is determined to a great extent by its structure and parameters. Due to the fixed network structure, the conventional RBF neural network has a poor adaptability for time-varying and complex nonlinear system. Therefore, RBF neural network is structured online in this thesis by sequential learning algorithm, to optimize the structure and parameters, and then the structure of single neuron PID control based on RBF neural network sequential learning algorithm is constituted. The simulation result indicates that sequential learning algorithm improves the accuracy and convergence rate of RBF neural network to a certain extent, then the adaptive performance of the controller is improved, and the control precision and response speed is improved, for nonlinear system the controller is able to obtain an ideal effect.The main constent of this thesis are listed as follows:(1) Neural network PID control method is studied, and then the principle and structure of single neuron PID control based on RBF neural network identification is introduced.(2) Several commonly used sequential learning algorithm of RBF neural network is studied, and an improved algorithm is proposed based on minimal resource allocating network algorithm, which exist some problems in the strategy of removing hidden layer neuron and adjusting threshold. According to the simulation of nonlinear system identification, the advantages and disadvantages of the improved algorithm is analyzed.(3) Dynamic orthogonal structure adaptive algorithm is studied, due to the algorithm exist some problems in the strategy of removing hidden layer neuron and network parameters adjustment, an improved dynamic orthogonal structure adaptive algorithm is proposed. According to the simulation of nonlinear system identification, compared with several commonly used sequential learning algorithms, the effectiveness and superiority of the improved algorithm is proved. (4) RBF neural network which adopt three different sequential learning algorithms respectively are used as identifier, combines with single neuron PID control, the structure of single neuron PID control based on RBF neural network sequential learning algorithm was constituted. A typical nonlinear system is used as control object, the simulation results show that the single neuron PID control based on RBF neural network sequential learning algorithm, compared with the single neuron PID control based on RBF neural network identification, have certain advantages, and it obtains a better control effect when the improved dynamic orthogonal structure adaptive algorithm are used in identification network of the control structure, than the nearest neighbor clustering algorithm and minimal resource allocating network algorithm.