Real-Time Neural Network Control of Temperature-Dependent Shape Memory Alloy Actuators Using Self-Sensing Feedback

Shape memory alloy (SMA) actuators are smart materials that generate motion through a thermally induced phase transformation between martensite and austenite states. Compared to traditional actuation technologies such as electric or pneumatic motors, SMAs provide a unique combination of high force output, compact form factor, and silent operation. These advantages make them attractive for applications in robotics, aerospace mechanisms, and biomedical devices where space and weight are critical. However, SMA actuators present significant challenges in achieving precise and repeatable control due to their nonlinear, hysteretic, and temperature-dependent behavior.

In this work, we address these challenges using a compact neural network model designed for real-time control on a microcontroller platform. The network was trained on experimentally collected data using resistance, power, and temperature as input features and displacement, measured by a laser sensor, as the output. During operation, the laser measurement is used only for validation; the network relies solely on self-sensing signals for prediction and control. To capture the actuator’s dynamic characteristics, the neural network processes sequences of ten prior input states, each spaced 100 ms apart, effectively covering one second of temporal dependencies in resistance and power behavior. This approach enables accurate, repeatable, and sensorless prediction of SMA wire displacement, demonstrating the feasibility of real-time, learning-based control for compact smart actuation systems.