Deep Neural Network-Based Physics-Inspired Model of Self-Sensing Displacement Estimation for Antagonistic Shape Memory Alloy Actuator

Abstract

The paper introduces a self-sensing feature that utilizes differential resistance measurement of an antagonistic Shape Memory Alloy (SMA) actuator to estimate linear displacement. The external position sensor used for control feedback becomes extraneous while utilizing an electrical resistance as feedback. The self-sensing capability provides additional advantages such as the overall reduction in size, weight, and interface complexity of an actuator. SMA actuator wires were used in the antagonistic configuration for the bi-directional actuation of the targeted application. A Deep Neural Network (DNN) having Long Short Term Memory (LSTM) layers were used to estimate the dynamic relation of resistance and displacement. A novel DNN model was developed for estimation purposes inspired by the physics-based description of the SMA actuation phenomenon. Accordingly, DNN was developed using the first and last layers as LSTM and a middle layer of feedforward neural network. Estimation results are presented with performance evaluation in terms of the R-squared index and mean absolute error (MAE) of the proposed model; also, the model has been compared with the generic LSTM 2-layered model.