A data-driven analysis approach for rate- and state-dependent frictional sliding behavior using LSTM network

抄録

Friction-induced stick-slip motion is widely observed in various industries, particularly in mechanical systems where tribological interactions play a crucial role. Understanding and predicting this complex phenomenon is essential for improving manufacturing processes and system performance. However, the behavior of stick-slip motion, especially under cyclic loading conditions, is intricate due to its nonlinear nature and dependence on multiple factors, unlike more straightforward mechanical responses. Moreover, there are limited verifications of constitutive equations suitable for predicting stick-slip behavior in numerical simulations and finite element method analyses. The aim of this study is to establish a prediction method for stick-slip motion using deep learning techniques, specifically Long Short-Term Memory (LSTM) networks, integrated with a rate- and state- dependent friction model. Here, since the model needs to accurately describe cyclic behavior, the LSTM architecture is adopted for its ability to capture long-term dependencies in time series data. The LSTM model contains multiple parameters related to various phenomena, such as the evolution of friction coefficient, velocity dependence, and state variable dynamics. The values of these parameters were identified from the results of comprehensive simulated datasets generated from a one-degree-of-freedom (1-DOF) friction model. As an initial approach, the model was trained and validated using two types of simulated tests: a constant load test and a constant velocity test. To enhance the model's robustness, Gaussian noise was added to the training data, simulating real-world measurement uncertainties. The model's performance was evaluated using load cells and displacement sensors in experimental setups. Numerical simulations of cyclic stick-slip tests using the trained LSTM model showed good agreement with experimental results. Additionally, it was found that incorporating the model's inverse analysis capability was necessary to accurately predict essential frictional parameters and stationary contact times from experimental observations. In conclusion, the stick-slip motion in tribological systems can be represented accurately by the LSTM-based deep learning model, offering a promising approach for predicting complex frictional behaviors in engineering applications.