A New Feedforward Neural Network Training Algorithm And Its Applications In Control

Neural network control is a very important branch of intelligent control and one of research hotspots of control field. The research work in this paper is about the training algorithm of Feedforward Neural Network (FNN) and its application in Automation Control. It is well known that the Global Extended Kalman Filter (GEKF) algorithm has much better performance than popular gradient descent with error backpropagation in terms of convergence and quality of solution. But the GEKF algorithm achieves better performance at the expense of much greater computational and storage requirements, which make the'GEKF algorithm prohibitive for training FNN with a relatively large number of nodes. Some researchers have developed some new algorithms such as the Decoupled Extended Kalman Filter (DEKF) algorithm and the Multiple Extended Kalman Algorithm (MEKA) based the idea of dimensional reduction and partitioning of the global problem. In this paper, a new training algorithm, called Local Linearized Least Squares (LLLS) algorithm, is presented, which is based on viewing the local system identification at neuron level as recursive linearized least squares problems. The new algorithm is shown by simulations to give better convergence results in comparison to the DEKF algorithm and the MEKA Algorithm for highly coupled applications and to approach the performance of GEKF Algorithm better.Then, a new adaptive control method for nonlinear systems is presented by integrating robust modeless learning adaptive control theory for nonlinear systems and FNN. A FNN that is trained online by the presented LLLS algorithm is used to be the system identifier to quickly model the nonlinear time-varying system. A adaptive control law in order to obtain the control input is computed through the FNN identifier. The control simulations show that the presented adaptive control method is very computationally simple and quickly tuning and has a satisfactory robustness to process dynamics variations and disturbances. At last, the reasons of the simulation results are given.