LSTM-Based Self-Sensing Application of Shape Memory Alloy Wire Actuators Under Practical Loading Conditions

The paper discusses the self-sensing feature in Shape Memory Alloy (SMA) wire actuators, where the change in electrical resistance of the wire can be utilized to estimate the degree of actuation, eluding the requirement of any feedback sensor. SMA wire actuators are generally driven by resistive heating and are very sensitive to varying ambiences. The existing techniques to model such behavior involve complex mathematical models that are computationally costly or require empirical parameters based on experimental data. To obviate this, a neural network-based methodology has been proposed in this study to simulate the SMA behavior under such practical loading conditions. Two neural networks comprising Long Short-Term Memory (LSTM) layers are developed to model the behavior of the SMA wire-actuated linear system and a non-linear rotary manipulator. The developed network comprising 3 LSTM layers, estimates the actuation generated in each of the systems from the change in the electrical parameters during the actuation, for a set of forced cooling cases of varied magnitudes, intensities, and time durations. The results obtained from the trained network are presented and compared against the experimentally obtained actuation, and was found to have an RMS error of around 2.5–3% across all test cases, thus revealing the potential of the proposed networks.