| Artificial neural network has been widely applicated in many fields, including prediction of time series, modeling of nonlinear system and intelligent control, because of the property for nonlinear mapping, self-learning, self-adaption and large-scale parallel processing. Universal Learning Network (ULN) has some advantages in structure than feedforward neural network, such as simple, compact and easy-to-use. Nodes have multiple, bidirectional connections. The branches among nodes and time delay in each branch are set arbitrarily.This thesis studied universal learning network in many fields, including prediction of chaotic time series, multivariable system identification and internal model control:1,A hybrid particle swarm optimization is proposed by conventional particle swarm optimization algorithm combined with gradient descent method. The new algorithm is used in the training of universal learning network. It is successful to be applied for prediction of typical logistic chaotic time series. Simulation results proved the effectivity of the proposed algorithm.2,A kind of system identification based on state equation with universal learning network is proposed in the fermentation of Saccharomyces cerevisiae. The advantages of nodes and branches in ULN are utilized well. Simulations have shown that the proposed network structure is much better than conventional one, and the purpose and physical meaning are more clearly and the accurate is much more exactly.3,Through the training of time delays on the braches of ULN to the optimal ones, the modeling precision for continuous stirred tank reactor (CSTR) which is complicated nonlinear system with large lag was greatly improved.4,An improved ULN internal model control method (improved ULN-IMC) based fuzzy control theory was proposed. The tracking and constant value control with external disturbance in simulation by this new method for CSTR have good performances. |
PDF Full Text Request