A Lightweight Physics-Informed Neural Network Model Based on Causal Discovery for Remaining Useful Life Prediction

Predicting the Remaining Useful Life (RUL) holds a pivotal role in intelligent prognostics and health management strategies. Lightweight neural networks and micro neural networks are gradually becoming research hotspots due to their unique advantages in recent years. In this paper, we propose a lightweight physics-informed neural network model based on causal discovery (Cau-AttnPINN) for remaining useful life prediction. Specifically, causal discovery algorithms are employed to explore complex causal chains and feedback mechanisms among variables, selecting significant features for dimensionality reduction. Self-attention mechanisms are utilized to learn differences and interactions among important features for further dimensionality reduction, and the predictive interaction neural network (PINN) is utilized to map features to RUL. Its advantage lies in model lightweighting, reducing computational complexity and memory requirements. The proposed framework demonstrates commendable predictive accuracy and interpretability while utilizing fewer trainable parameters when tested on a real C-MAPSS benchmark dataset.