Abstract: In order to obtain the very generalized and accurater mathematic model, a regression neural network algorithm with frequency decomposition power function and two grade learning objectives was proposed. The network structure is divided into frequency decomposition, input layer, hidden layer and output layer. The input signal is decomposed into several frequency range and sent to the hidden layer. The transfer function of hidden layer is power function. The hidden layer and output layer have learning objectives respectively, and the neural network has local and globe feedback. The hidden layer adopts the local gradient algorithm based on vector angle and the output layer uses the global linear regression algorithm. The neural network model was used to adjust the PID parameters of control system; the controlled variable was achieved by modified iterative learning algorithm, then the PID parameters were turned by constrained linear least squares algorithm. Simulation shows that the neural network model is generalized and accurate; the quality of control system is excellent than traditional turned methods.

Key words: generalization, frequency decomposition, power activation function, vecter angle, iterative learning, PID tuning