Active learning algorithm in the application of hydraulic AGC system

Abstract

In view of the strip thickness control precision problem in hydraulic automatic gauge control system, put forward a kind of active learning algorithm based on dynamic neural network. In order to improve the generalization ability of the network, the algorithm, based on a kind of modified uncertainty sampling strategy in the active learning, selecting samples for training the dynamic neural network from a large number of unmarked samples. In order to improve learning rate, the network reducing neurons which with small sensitivity value through sensitivity analysis, and inserting a new neurons bases on the winning mechanism principle when the ability dealing with problems of the network is not good enough. The algorithm is applied to the hydraulic AGC system. To control the thickness deviation caused by the change of various parameters during the system working process, it adjusts PID controller parameters online by combining improved structure dynamic neural network with traditional PID controller. Shows the active learning algorithm in dynamic network training and dynamic network structure adjustment step stimulation experiments show that the active learning algorithm based on dynamic neural network can effectively implement strip thickness control, and it is with strong dynamic performance.