研究業績

Fast and accurate prediction of the temperature field in wet-clutch friction components is critical for thermal failure warning in transmission systems. However, the heat partition characteristics of finite-thickness components remain unclear, and the classical heat partition coefficient proposed by Charron is not fully applicable to clutch systems. In addition, standard physics-informed neural networks require retraining for each operating condition and therefore lack cross-condition prediction capability. To address these limitations, this study proposes an extended physics-informed neural network (E-PINN) integrated with a thermodynamic model. In the thermodynamic model, the coupled effects of asperity contact and oil-film lubrication are considered, and the dynamic heat partition coefficient is determined through particle swarm optimization by enforcing interfacial temperature continuity. Pressure and rotational speed are introduced as additional network inputs, while adaptive sampling and dynamic weighting strategies are employed to improve training performance across different operating conditions. The results show that the dynamic heat partition coefficient evolves in three stages. In the initial sliding stage, it remains close to 0.30, in agreement with Charron's analytical solution. In the subsequent stage, the coefficient changes rapidly as the heat flux reaches the boundaries and interfacial heat exchange becomes significant. In the final stage, it approaches a steady value determined by material heat capacity and boundary conditions. Because the dynamic coefficient accounts for finite-geometry effects and interfacial heat transfer, it is physically more reasonable than the fixed coefficient and yields lower, more stable training losses. Specifically, for the separator disc, the PDE loss and the boundary loss decrease by 1.03 × 10−3 and 5.38 × 10−3, respectively, whereas for the friction disc, they decrease by 1.00 × 10−2 and 1.13 × 10−1, respectively. The proposed method achieves a mean relative error of 2.03%, a root-mean-square error of 1.72 °C, and a coefficient of determination of 0.997 for temperature prediction, outperforming the finite difference method and multiple data-driven models. Compared with the PINN, it provides comparable accuracy while maintaining a stable online time of 6.16 ms and avoiding retraining for new operating conditions. This work offers a physical basis for heat partition in finite-thickness disc-to-disc friction pairs and supports online thermal health monitoring of clutches.