Fig. 2: Numerical simulation and experimental data used in the transfer learning analysis.

The top row shows the deep learning model input signal as the kinetic energy and acoustic emissions, respectively, and the bottom row shows the target friction coefficient. a Finite-discrete element method (FDEM) time series are split into training/validation/testing segments (60/20/20%) shown in green, blue, and pink shades, respectively. The convolutional encoder-decoder is fully trained and tested using these data. b The experimental data (p4677) are split into training/validation/testing segments (20/20/60%) to include six cycles of stick-slip events for training the model latent space.