An Improved Recurrent Neural Network for Industrial Control System Identification

Abstract:
The rapid development of industrial automation makes the control process more and more demanding for the accuracy of system identification. Finding an effective identification model for regulating and controlling the time-critical and nonlinear process remains a difficult problem. Therefore, this paper proposes an improved recurrent neural network for industrial control system identification, called SI-RNNs, to solve the problem of insufficient prediction accuracy of the traditional RNNs in the initial time steps. SI-RNNs adds the state initialization process to determine the initial operational state of the control systems and selects the optimal state initialization steps length by perturbating feature values approach. In practical cases, it is verified that our method can effectively improve the solution accuracy of the model for industrial control process, especially in the gated recurrent neural networks capable of long-term memory.