Fatigue-aware force estimation for NiTi SMA actuators using LSTM recurrent networks

Abstract This study focuses on the force estimation of NiTi shape memory alloy (SMA) actuators, particularly addressing the challenges posed by material fatigue and force degradation over repeated actuation cycles. Traditional modeling approaches for SMAs, which typically include constitutive, phase transformation, and heat transfer sub-models, often fail to account for these degradation phenomena, leading to inaccuracies in long-term predictions. To overcome these limitations, we developed and implemented multivariate long short-term memory (LSTM) recurrent neural networks for accurate force estimation using measured parameters such as electrical power (current, voltage), strain, and temperature. An extensive experimental setup was designed to cyclically and electrothermally (Joule heating) activate the SMA actuators at varying strain levels, and the degradation in force production was quantified, with a reduction of up to 20.62% observed after 1000 cycles. The study introduced both single-stage and two-stage LSTM architectures, with the SLF model, which incorporates historical force data along with other measurements, achieving the highest accuracy (RMSE of 0.003 N). More sensor-limited and deployment-oriented variants also performed well: SLWF and SLWTF achieved RMSEs of 0.089 N and 0.110 N, respectively, and the two-stage MTSL (with learned temperature as a hidden state) achieved 0.096 N. The proposed LSTM-based models demonstrated their efficacy in providing reliable force estimations under conditions of actuator degradation, making them suitable for real-time applications where traditional analytical models fall short. These findings offer valuable insights for applications requiring stabilized forces and consistent actuation performance in the presence of material fatigue.