Proxy of Ti-Ni Shape memory alloy Actuators Based on Recurrent Neural Networks

The conventional experimental procedure involving Titanium–Nickel (Ti-Ni) shape memory alloys requires conducting dozens or even hundreds of heating and cooling cycles performed by the actuator to generate thermal hysteresis curves. This study proposes the development of a proxy model based on machine learning techniques, using experimental results, with the goal of replicating the actuator's function in this experiment. The proxy model should be capable of accurately predicting the actuator’s thermomechanical response based on time series data of heating and cooling cycles over time. It is important to highlight that this is not the traditional time series forecasting problem aimed at predicting future values, but rather a problem of predicting the dynamic responses of the actuator associated with new input profiles (temperature, mechanical stress, and strain). The proposed strategy is based on the use of deep neural network algorithms, aiming to capture the actuator’s dynamics from experimental data. The main architecture used for modeling temporal dependencies is the recurrent neural network (RNN), specifically the Long Short-Term Memory (LSTM) type, known for its ability to extract complex and nonlinear temporal patterns in time series data. To evaluate the performance of the proxy model, an experimental dataset was generated using a helical spring-shaped actuator under load. The model's predictions were compared with the experimentally obtained hysteresis curve in order to validate its generalization capability. The results demonstrate that the proposed technique is highly promising, achieving a mean squared error on the order of 1.2%.