Modeling and sensorless tension control of SMA actuator using GRU

Abstract

This paper proposes a method for sensorless tension control of shape memory alloy(SMA) actuators. Sensorless control has advantages in cost and size. Gated recurrent unit(GRU) was adopted to model complex characteristics of the SMA actuators. GRU is one of the machine learning technologies, which is evolved from the recurrent neural network. Modeling of SMA actuators are not easy due to their characteristics such as nonlinearity, hysteretic behavior, effects of temperature and stress. They are often responsible for control difficulties such as inaccuracy and instability. In this study, GRU is used to estimate the relation among the applied voltage(duty ratio), electrical-resistance, tension, displacement and room temperature. Two SMA models were created by GRU: the sensing model and the inverse model. Both models were trained with the original dataset obtained experimentally. A control framework for the sensorless tension control including the two SMA models was proposed. The sensing model which estimates the tension and displacement from the applied voltage, electrical-resistance and room temperature substituted for sensors. The inverse model which estimates the applied voltage canceled or reduced the effects of the hysteresis and nonlinearity as the feedforward controller. The results of experiments with the proposed control framework show that the actual tensions closely followed the target tensions without sensors.