The Algorithm Research On The Online Decoupling Of The Neural Network

ABSTRACTThe controlled system is usually multivariable, coupling and time varying in the field of the industry control. Aiming at the situation that the coupling is affected by the change of the time or load, and taking the advantage of the Neural Network (NN), it is very important to study the online learning algorithm of the NN decoupler.The following aspects are studied principally in this thesis:A kind of online learning algorithm of the NN decoupler, which based on MIMO cross-correlation function, is proposed on the frame of distributed decoupling system .The algorithm makes a set of cross-correlation function of the MIMO act as the target function of the NN decoupler, and adopts genetic algorithm to train NN decoupler online. It is proved theoretically that the NN will approach the ideal decoupling segment as soon as the cross-correlation of the input and output becomes small in a linear system. It is feasible to select cross-correlation function as the target function.Combined with the coal-pulverized control system of the power plant, the algorithm is verified in simulation. When the decoupling is completed, the traditional PID is adopted to control the two 5150 systems. The control effect is satisfactory.This thesis discusses preliminarily the topic about decoupling of nonlinear MIMO system using distributed decoupling structure and online learning algorithm. Because of the complexity of the random signal while transiting in the nonlinear system, this thesis verifies the feasibility that the cross-correlation function is selected as the target function in numerical simulations. After the success in decoupling, single neural cell self-adapting PID is adopted to control nonlinear object. The simulation results show that the control strategy gets better effects.