Sequence-to-sequence LSTM-based Dynamic System Identification of Piezo-electric Actuators

Abstract

During the past few year, recurrent neural network (RNN) has been proposed to model the nonlinear dynamics of various dynamic systems, such as nano positioning systems (e.g, piezo electric actuators (PEAs)). Although high modeling accuracy has been demonstrated using RNNs, it has been found that the conventional RNNs (such as vanilla RNN) are susceptible to gradient vanishing or exploding issue and hence difficult to train. Deep RNNs, such as Long short-term memory (LSTM), have been proposed to address these issues. However, due to the conventional training data construction, the training is susceptible to overfitting and the computation is extensive. In this paper, we propose a new type of LSTM in the application of PEA system identification: a sequence-to-sequence learning approach (namely, LSTMseq2seq). The structure of LSTMseq2seq and its training data construction are presented in detail. The efficacy of LSTMseq2seq in terms of modeling accuracy and computation speed is demonstrated by applying it for PEA system identification and comparing its performance with that of vanilla RNN.