Differential resistance based self-sensing recurrent neural network model for position estimation and control of antagonistic Shape Memory Alloy actuator

Abstract

The paper presents the development and experimental investigation of recurrent neural network (RNN) based self-sensing position estimation (SSPE) model for shape memory alloy actuator (SMA). RNN is used as an estimator in the position feedback control loop to replace the external additional position sensor. The model was inspired by the physics-based analogy of the Mass-Spring-Damper (MSD) system for the antagonistic SMA wire actuator. Actuator displacement presents the hysteresis and non-linear dynamic relationship with observed differential resistance (sensing signal) during phase transformation. The resistance variation causes due to dissimilar resistivity between primary phases of SMA material. The RNN based estimation model was considered because it consists of memory elements for storing the processed information and feedback connections for dynamic modeling of the system. RNN was trained with input sensing signal and target displacement datasets. The estimation accuracy of the model was real-time evaluated during trajectory tracking of reference signals in the feedback control loop. A quantitative performance analysis is assessed in terms of correlation (ℛ), mean absolute error (MEA), and root mean square error (RSME) of the learned model along with the developed actuator system. The tracking results confirm the close agreement between estimated and measured displacement at a reasonable accuracy.