A Neural Network‐Based Self‐Sensing Embedded Position Control System for Shape Memory Alloy Wire Actuators

Abstract

Shape memory alloy (SMA) wire transducers undergo a phase change in their crystal lattice when heated above a certain transformation temperature, which results in an actuation strain on the order of 4–6% as well as a change in the material electrical properties. The latter opens up the possibility of performing self-sensing, namely estimation of the wire mechanical deformation during actuation based on electrical measurements only. However, accurate SMA self sensing is challenging because the electromechanical and thermal characteristics are generally hysteretic and depend on various factors like actuation frequency, biasing mechanism, external load, and electrical heating strategy. To practically address this issue, in this work, an artificial-intelligence-based approach is proposed for SMA self-sensing. The proposed neural network consists of a combination of recurrent and simple neuron types to reconstruct the position of an SMA actuator in real-time, using applied electrical power and SMA wire’s electrical resistance as inputs. After training and validation, the neural network is implemented on a microcontroller-based embedded system to develop a self sensing closed-loop position control system. The accuracy of the self-sensing control scheme is validated against a sensor-based control solution, showing root mean squared errors of 0.028 and 0.021 mm respectively over a displacement range of 2mm.