Research Of Modeling Methods And Control Strategies Based On Neural Networks

As a new intelligent control method, artificial neural networks (ANN) with nonlinearmapping and high parallel information processing capabilities pave a new way to solve theproblem of identification and control of nonlinear systems. Taking the level and heaterexchanger as the object in this thesis, some modeling approaches and real-time controlstrategies based on ANN are studied. The main contributions are summarized as follows: ⑴An open hardware architecture of the process control system is presented.Especially, the technique of developing the open software systems is introduced usingVisual C++, Matlab and LabWindows/CVI. The real-time control results based on aboveplatform are also given. ⑵The property for modeling of the level and heater exchanger is analyzed, and twomodeling approaches via static back propagation (BP) network and dynamic diagonalrecurrent neural network (DRNN) are adopted to identify these nonlinear systems. Thecomparison between the two experimental results indicates that the DRNN model processesnonlinear dynamic mapping characteristics and suits identification of nonlinear systems. ⑶An extended linearized neural state space (ELNSS) model is a novel neural networkand processes recurrent architecture. Thus the nonlinear property of the model is expounded,its convergence of training process is also proved. When ELNSS model is adopted toidentify the level and heater exchanger process, the serial-parallel structure is adopted andthe extended Kalman filter (EKF) method is used to train the parameters. Thegeneralization ability of the model is tested by training samples. The results demonstratethe effectiveness of the ELNSS model for the identification of nonlinear systems. ⑷The convergence of genetic algorithm (GA) is discussed. Because of simple geneticalgorithm (SGA) tendency local optimization, a modified adaptive genetic algorithm (MGA)is put forward to get over the insufficiency. A new formula to computer crossoverprobability and mutation probability is developed. The MGA is used as optimizing methodof identification and trained the weight of ANN. The identification results to heaterexchanger show that the GA can promote the accuracy of model identification. ⑸Some adaptive control methods based on ANN are used to control a level process, IIIthe methods include a single neuron adaptive PID control, BP neural network adaptivecontrol, adaptive PID control algorithm based on BP neural network and DRNN. Theexperimental results indicate that the DRNN possesses satisfactory stability to deal with adynamic system and can realize optimum PID adaptive control. Meanwhile, in order toovercome the influences of the initial parameters on performance, the MGA is used to trainneural network to get initial weight and PID parameters before adaptive control. Thereal-time control results indicate that the above approach can accelerate the convergence ofthe system and improve its performance. ⑹For nonlinear multivariable systems, the static decoupling method and dynamicdecoupling method based on ANN are presented especially on their limitation. Two hybriddecoupling methods on the basis of decoupling and control are presented, one method isneuron decoupling control, the other approach is DRNN-PID decoupling control based onGA. The principle of hybrid decoupling is that the decoupler and the controller is designedas a combined system. According to the requirement of the system object function, theoutput is mapping to the input, and tracing the set point, the decoupling problem is resolvedby this means. Simulations are presented to a nonlinear coupling system using the proposedmethods, also, it lays a foundation for the decoupling application. ⑺In order to make up the weakness that disturbance affects the dynamic performancewhen heat exchange occurred, a forward-feedback dynamic matrix control (DMC) isadopted. The approach of feedback compensation to...