Online Decoupling And Control Based On Neural Network For A Class Of Nonlinear Multi-variable System

Decoupling is an issue with a long history during the development of control theory. There are many mature approaches in theory and engineering about decoupling control of linear multi-variable system at present.But there are more nonlinear multi-variable systems in the actual industrial control.By means of the neural network(NN),a nonlinear system can be approximated with arbitrary accuracy in the close compact set, and the NN decoupling for nonlinear multi-variable system can be achieved.The NN is often trained in offline mode.Compared to the offline mode,it is suitable for the changing work conditions and real-time online learning.When the work condition is changed,the NN can modify the weights to eliminate the couple of the variables in the system on time.Due to the complexity of the nonlinear system,it is classified as five kinds of discrete nonlinear multi-variable system according to nonlinear operator.Based on the class of nonlinear multi-variable system with the nonlinear operator and couple appearing in the input of system,the online decoupling control with NN is studied in this thesis.The proposition and proof of the target function of the online decoupling,the measure for target function difficulty,the realization of online decoupling based on the static NN and dynamic NN and the analysis of algorithmic convergence are targets for research.The main content can be summarized as follows.1.The signal to train the NN is analyzed.It will not only satisfy the demand of decoupling but also not disturb the system normal work.Aiming at the train signal which is white noise and uniform random step sequence,the generalized cross-correlation function between the input sequence and the output sequence of coupling channel is proposed as the target function of online decoupling.The target function is proved to characterize the relation of coupling channel.A frame of online decoupling with NN is proposed.The trained NN will be switched to work by a switching system.2.Based on the frame of online decoupling and control with NN,a class of nonlinear multi-variable system is proposed which can be decoupled online by means of distributed decoupling and feedforward decoupling.Aiming at the target function of online decoupling,the fitness-distance correlation,the correlation function of the random walk sequence,NK-model and the best first order function are used to comprehensive analyze and test the difficult of target function.3.The algorithm of online decoupling with the static NN is designed.For the two sub-classes of nonlinear multi-variable system,the genetic algorithm,the genetic algorithm plus simulated annealing algorithm and the genetic algorithm plus pattern search algorithm are used to train the static NN respectively.The decoupling effect of different frames and algorithms is analyzed and the single loop system is controlled by the signal neuron or the PID controller after decoupling.4.The algorithm of online decoupling with the dynamic NN is designed.Based on the feedforward decoupling frame,the time delayed NN,Elman net and internal time delayed NN are respectively used to decouple the multi-variable system.The difficulty of the optimization target is analyzed by the means of the correlation function of the random walk sequence and the best first order function.The three dynamic NNs are respectively trained using the genetic algorithm plus pattern search algorithm.The decoupling effect is compared among the different dynamic NNs and the single loop system is controlled after decoupling.5.Based on the online decoupling algorithm with the static NN and dynamic NN, the quantitativeanalysis of convergence is achieved.The indicator of convergence rate is defined and analysis results are given to illustrate the feasibility of the online decoupling algorithm.