Interpretable fusion physical information neural network with multiple physical constraints for remaining useful life prediction

Abstract Remaining useful life prediction for industrial equipment faces critical challenges regarding insufficient model credibility and weak cross-device generalization. This paper proposes an Interpretable Fusion Physical Information Neural Network (MPIE-PINN) that synergistically integrates multi-physical constraint adaptive loss (MPC-ALPINN) with task-aggregation-based meta-learning (TA-RMLO). By embedding laws of monotonicity, acceleration, and Wiener processes as ex-ante constraints, the framework ensures physical consistency while achieving rapid adaptation to new devices with limited data. The proposed framework provides a physics-consistent and empirically verifiable form of interpretability by linking ex-ante physical constraints with ex-post representation analysis. Experimental results on NASA commercial modular aviation propulsion system simulation (C-MAPSS) and laboratory gearbox datasets demonstrate that MPIE-PINN achieves superior accuracy, with an average root mean square error of 11.84 on C-MAPSS subsets. A multi-dimensional visual interpretability module further validates that the model’s decision-making aligns with actual mechanical degradation mechanisms, providing a robust and transparent solution for equipment health management.