Towards Improving the Performance of the RNN-based Inversion Model in Output Tracking Control

Abstract

With the advantages of high modeling accuracy and large bandwidth, recurrent neural network (RNN) based inversion model control has been proposed for output tracking. However, some issues still need to be addressed when using the RNN-based inversion model. First, with limited number of parameters in RNN, it cannot model the low-frequency dynamics accurately, thus an extra linear model has been used, which can become an interference for tracking control at high frequencies. Moreover, the control speed and the RNN modeling accuracy cannot be improved simultaneously as the control sampling speed is restricted by the length of the RNN training set. Therefore, this article focuses on addressing these limitations of RNN-based inversion model control. Specifically, a novel modeling method is proposed to incorporate the linear model in a way that it does not affect the existing high-frequency control performance achieved by RNN. Additionally, an interpolation method is proposed to double the sampling frequency (compared to the RNN training sampling frequency). Analysis on the stability issues which may arise when the proposed new model is used for predictive control is presented along with the instructions on determining the parameters for ensuring the closed-loop stability. Finally, the proposed approach is demonstrated on a commercial piezo actuator, and the experiment results show that the tracking performances can be significantly improved.