Prediction of bearing remaining useful life based on physical models and convolutional neural network

To overcome the limited physical interpretability of conventional data-driven approaches in predicting the remaining useful life (RUL) of rolling bearings, a dual-stream collaborative prediction framework that integrates physical modeling with deep learning is proposed. A multi-dimensional degradation feature system is constructed based on the Paris-Erdogan crack growth theory, incorporating time-frequency features of vibration signals and crack growth rate parameters to characterize the fatigue damage evolution process of bearings. The framework adopts a CNN-LSTM dual-stream hybrid architecture, in which the data-driven branch extracts deep features from vibration signals using convolutional neural networks combined with LSTM for temporal dependency modeling, while the physics-informed branch encodes mechanistic knowledge through fully connected networks. A feature fusion strategy is employed to achieve complementary enhancement between data-driven and physics-based features, and monotonicity constraints are integrated into the loss function optimization to preserve physical consistency in degradation trends. Experimental evaluation on the PHM2012 benchmark dataset demonstrates substantial improvements in RMSE, MAE, and R² metrics compared with mainstream single-model methods, thereby verifying the effectiveness of the physics-informed enhancement strategy in improving the reliability of RUL prediction.