Abstract
Friction-induced stick–slip motion is a complex phenomenon in tribological systems. Accurate prediction and analysis of this behavior are crucial for understanding and controlling frictional dynamics. In this study, we integrated long short-term memory (LSTM) networks with rate- and state-dependent friction models to predict stick–slip behavior. A simulated dataset from a one-degree-of-freedom (1-DOF) friction system was used to train the proposed model. LSTM outperformed recurrent neural networks (RNNs) and gated recurrent units (GRUs) in univariate predictions, accurately capturing nonlinear dynamics. To address the unobservable variables in the experiments, we employed multivariate prediction using the spring force to estimate the velocity, acceleration, friction coefficient, and other state variables. Gaussian noise was added to the training data to enhance the robustness of the model, enabling accurate predictions even with noisy experimental data. Additionally, the inverse analysis capability of LSTM was used to estimate past stationary contact times based on the maximum friction coefficient observed before sliding. Validation against experimental data showed a strong consistency between the predicted and actual trends, demonstrating the potential of the model for analyzing stick–slip motion.
