Remaining useful life prediction using sparse physics-informed neural network driven by degradation dynamics under incomplete data

In practical industrial scenarios, critical components are often replaced preventively before functional failure occurs due to safety and economic considerations. This results in acquired degradation data exhibiting significant truncation and incompleteness. Such incomplete degradation data severely limits the accuracy and generalization capability of data-driven remaining useful life (RUL) prediction. To address this challenge, this paper proposes a degradation-dynamics-driven sparse physics-informed neural network (PINN) prediction framework. First, a defect size evolution model consistent with real degradation mechanisms is constructed. Through dynamic simulation, physically coherent full-lifecycle degradation trajectories are generated, enabling structural completion of degradation information and effectively addressing data incompleteness. Second, guided by the dynamic simulation knowledge, a partial differential equation constrained model linking health indicators to RUL evolution is established. The Lasso sparse regularization mechanism is introduced to adaptively screen and constrain candidate equation terms and key parameters, thereby constructing a prediction framework integrating sparse regression with a PINN to achieve RUL under incomplete data conditions. Finally, the effectiveness and superiority of the proposed method are validated through experiments simulating various data missing scenarios and missing rates on two full-cycle bearing lifetime datasets: SMU-SY and XJTU-SY. Experimental results demonstrate that under high missing rate conditions, the proposed method still achieves high-accuracy, robust, and physically interpretable RUL prediction performance. This paper presents a novel and effective solution for rotating machinery life prediction under incomplete and degraded data conditions in complex industrial environments.