A self-adaptive PINN for hidden-state degradation feature learning and remaining useful life prediction of wet friction components

Wet friction components serve as critical elements in clutch systems for powertrain and torque regulation, and their remaining useful life (RUL) prediction is of great significance for ensuring the reliable operation of transmission systems. Traditional data-driven methods fall short in fully characterizing the underlying physical degradation mechanisms, while purely physics-based models are often limited by inaccessible parameters and modelling complexity. To address these challenges, the present work proposes a Transformer-based self-adaptive physics-informed neural network (Transformer-SAPINN) for RUL prediction of wet friction components. The proposed approach first employs a Transformer encoder structure to capture temporal dependencies and inter-feature interactions in degradation signals. Moreover, a physics-guided regulator (PGR) is developed to self-adaptively learn the unknown nonlinear partial differential equation (PDE) that governs the degradation evolution, and this physics prior is embedded into the PINN framework. Furthermore, the method introduces a task uncertainty-based adaptive weighting mechanism, which dynamically adjusts the relative contributions of the data and physics loss terms throughout training. Results indicate that the proposed method achieves superior performance compared with other models, with the highest coefficient of determination ( R 2 ) reaching 0.9978, and reductions in mean absolute error ( MAE ) and root mean square error ( RMSE ) up to 89.33% and 85.23%, respectively. Ablation studies reveal that, compared with the model without Transformer, MAE and RMSE are reduced by 65% and 68%, respectively. Furthermore, interpretability analysis identifies the Thickness, maximum circumferential temperature difference ( T c ) and maximum radial temperature difference ( T r ) as the three most critical indicators influencing lifetime prediction.