A recurrent neural network–based model for predicting bending behavior of ionic polymer–metal composite actuators

Abstract

The high application potential of ionic polymer–metal composites has made the behavior identification of this group of smart materials an attractive area. So far, several models have been proposed to predict the bending of an ionic polymer–metal composite actuator, but these models have some weaknesses, the most important of them are the use of output data (in autoregressive models), high complexity to achieve a proper precision (in non-autoregressive models), and lack of compatibility with the behavioral nature of the material. In this article, we present a hybrid model of parallel non-autoregressive recurrent networks with internal memory cells to overcome existing weaknesses. The validation results on experimental data show that the proposed model has acceptable accuracy and flexibility. Moreover, simplicity and compatibility with the behavioral nature of the material promote using the proposed model in practical applications.

Application‐Oriented Modeling of Soft Actuator Ionic Polymer–Metal Composites: A Review

“…[43,44,46,47,49,51] The prediction of the entire deformation shape of the IPMC [44,46,53] and the position prediction without external data can be achieved. [53][54][55][56] Some models have been developed for higher precision controllers. [47]…”

“…[ 54 ] Zamyad et al [ 55 ] proposed a hybrid model of parallel nonautoregressive recurrent networks with internal memory units, and the accuracy of the proposed recurrent network is comparable to that of the autoregressive neural network model (Figure 3c) [ 54 ] and the ANFIS‐NARX model. [ 46 ] However, compared with the two nonautoregressive MLP models based on Volterra [ 53 ] and Laguerre, [ 54 ] the model in another study [ 55 ] has significant advantages over the Volterra model, and there is no significant difference in performance between the model and the Laguerre MLP model. [ 54 ] Annabestani et al [ 56 ] proposed a hybrid model combining resistor–capacitor (RC) distributed model and a deep MLP (D‐MLP) neural network model (Figure 3b).…”

“…[58] Some studies use digital cameras to record the bending deformation of IPMCs to overcome the shortcomings of laser displacement sensors. [53][54][55][56] The digital camera is usually placed on one side of the IPMC actuator (Figure 4b), and the tip displacement of IPMCs can be measured using a scale to obtain a grid paper. The charge-coupled device (CCD) video camera is also important for displacement measurement in IPMC.…”

“…The intelligent identification model takes the input voltage and output displacement of the IPMC actuator as the input and output of the neural network. Such models can predict the displacement of IPMC actuators with high accuracy, [43,44,46,47,[49][50][51][53][54][55][56] and can characterize the hysteresis or creep of IPMC. [43,44,46,47,49,51] The prediction of the entire deformation shape of the IPMC [44,46,53] and the position prediction without external data can be achieved.…”

A Deep Neural Network-Based Model for Quantitative Evaluation of the Effects of Swimming Training

“…e training optimization problem of deep neural networks is a highdimensional nonconvex optimization problem, which is very different from the traditional machine learning optimization and the general mathematically studied nonconvex optimization problems [7]. Deep neural network training optimization is characterized by large scale, extensive data, and high-dimensional nonconvexity.…”