A Novel Single-Domain Generalization Method for Remaining Useful Life Prediction under Missing Data

As equipment becomes increasingly complex, it often operates under a variety of conditions, rendering traditional predictive models built on historical data less effective. Additionally, data loss during online operation presents significant challenges to the accuracy of model predictions. Therefore, this paper proposes a single-domain remaining useful life (RUL) prediction method based on tri-path contrastive learning. This method employs convolutional neural networks and a newly proposed skip-attention mechanism to extract features from three pathways: complete data, missing data, and simulated missing data. The RUL is then predicted using a predictor constructed with a multi-head attention mechanism. To enhance the model's generalization performance, a loss function integrating feature alignment and label alignment is designed. Finally, the effectiveness of our method is validated using the N-CMAPSS dataset.