99 / 2025-05-14 23:47:05

Physics-Informed Neural Network For Angular Dynamic Model Parameters Estimation Of Faulty Bearing Oriented To Optical Encoder Signal

rolling element bearing,instantaneous angular speed,angular dynamic model,physics-informed neural network,parameter estimation

Special Sessions > SSL 01 Advanced signal processing enabled intelligent fault diagnosis and measurement

Dynamic models of bearings have been widely recognized as an effective approach to synthesizing diverse bearing fault samples for data-driven intelligent diagnosis. To tackle the key challenge of dynamic parameter estimation in instantaneous angular speed (IAS)-oriented bearing fault models —specifically, mitigating the substantial discrepancy between simulated signals and real-world operating conditions induced by empirical parameter dependence. A physics-informed neural network (PINN) framework for dynamic parameter estimation is proposed. This framework embeds a bearing dynamic model into the neural architecture to inversely identify model parameters through PINN-based optimization. First, a three-degree-of-freedom (3-DOF) angular-domain dynamic model is established specifically tailored for bearing raceway faults. Subsequently, a physics-constrained loss function is formulated by rigorously embedding angular-domain differential equations into the network architecture, where torsional damping and stiffness parameters are implicitly encoded within the neural network parameters. Finally, under the synergistic drive of physical mechanisms in dynamic model and measured signals, model parameters estimation is achieved through neural network parameter updating mechanisms. Experimental results demonstrate the effectiveness of proposed methodology.