Two New Neural Networks And Their Applications In Nonlinear System Control

Reactor, as an essential tool of production, takes the very important role in the field of oil production, petrochemical industry and bio pharmacy etc. Its operation state can affect the production efficiency and quality index directly. Continuous stirred tank (CSTR) is one of the most important reactor, those parameters of reactants in is which such as temperature, pressure and density etc. will impact the product quality of the chemical production process directly. So, it is of great importance in guarding the quality of chemical production process by measuring and adjusting these reactant characteristic parameters. In this paper, a continuous stirred tank will be taken as a study object and a control system will be designed for it.The control system of CSTR is always a nonlinear, time-varying system with lag behind, noise disturbance and the other impact at the same time. So classical PID control theory would hard to gain a good effect. The essential reasons of poor precision of control system have been found by analyzing systemically the control properties of CSTR and the control system design schemes have been proposed about CSTR based on the neural network.Firstly, this paper proposed and proofed a non-simplify recurrent Newton algorithm, which has improved the calculation speed of traditional BP network and avoided of the information loss during model being simplification. The common BP neural network have been developed. One of the key problems in BP neural network is study algorithm, which is always slow convergence speed or even divergence sometimes. Karayiannis proposed a second order algorithm, which has been simplified during dealing with the recursion formula by adjusted some rights of Hessian matrix in BP neural network. While, these treatment would omitted some useful information about the second derivative of the objective function, which would lead to the uncompleted derivation and even more amount of calculation. Aimed at such problems, a non-simplify recurrent Newton algorithm is proposed and proofed. It is the same order with Newton algorithm, the second order convergence speed. New Algorithm has an inertia item, which can avoid vibration of the algorithm, so that the convergence speed can improved. Meanwhile, new algorithm can embody all the right error information in first derivation and second derivation. It is a real second order algorithm. Seen form the amount of calculation, new algorithm is a little more than the common recurrent lease square method and less than Karayiannis algorithm.Secondly, it proposed a new mixed algorithm to analysis the contracture, parameter and rights of the radial basis function neural network (RBFNN). RBFNN is another kind of neural network, which is also this research objective. In RBFNN, due to the parameters of active function would be adjusted in the hidden layer, using nonlinear parameter optimization method, the study speed is slow. Meanwhile, the study process in RBFNN is always in two layers, which is a mixed study process. Based on the basic principle of RBFNN, a new optimal selection cluster algorithm is proposed. Combined with orthogonal least squares method and gradient method usedin the technology of optimization, a new mixed algorithm, which can be used as identify the structure of neural network, parameter and rights. This new mixed algorithm can greatly improve the identify precision of RBFNN and develop the generalization ability.Then, a PID Self-tuning control scheme of CSTR has been proposed, based on the improved BP neural network. BP neural network, with its properties of approximation and adaptive, can greatly improve the control effect. Compare the properties from the simulations between the BP based self-tuning control PID scheme and classical PID scheme. The result shows that given the same condition of transient response time, the frontier has the smaller overshoot amount and the smaller output in controller.At last, the direct control scheme and adaptive control scheme have been designed separately based on RBFNN. In these two scheme, the design of control of the closed-loop system and stability analysis are applied individually. Analyzed from the Lyapunov function, these two scheme are all asymptotic stability, and the control schemes are all effective by the results of simulation.The control system of CSTR based on neural networks belongs to the category of intelligent control, which kind of control schemes are no necessary of knowing the accuracy mathematic model from the objection, strong robustness, and stronger inhibition to the noise interference. It is very suitable used in such kind of unknown model or various complex control system, is of great industrial practical value.